PHM Society European Conference

A Comprehensive Literature Review of State of Safety (SoS) for Maritime Battery Management Systems (BMSs)

Eric Aaslund — 2026-07-03

Battery-powered vessels can help reduce greenhouse gas emissions in the maritime industry, which is crucial to achieving the International Maritime Organization’s (IMO) ambition of net-zero emissions by 2050. However, batteries also introduce unique safety risks, since failures can lead to catastrophic outcomes such as thermal runaway and onboard fires. Effective Prognostics and Health Management (PHM) is essential for early detection of hazardous states to prevent critical failures or unsafe conditions. This paper investigates the emerging concept of state of safety (SoS), a metric for real-time quantification of battery safety, which is essential for ensuring safe operation of battery-powered vessels. A PRISMA-based literature review is conducted to address three research questions: (1) What are the current definitions of SoS in the maritime and other industries? (2) What are the existing methods to estimate SoS? (3) What gaps remain in defining and implementing SoS in the maritime industry?
The review reveals recent advancements in SoS research from the automotive and energy storage domains, while identifying a notable lack of studies addressing SoS in maritime battery systems. Existing estimation approaches and key battery parameters relevant for SoS assessment are reviewed and critically discussed. To advance the practical implementation of SoS, the paper provides a structured overview of the battery parameters required for comprehensive SoS implementation, including measurement methods and associated challenges. Furthermore, safety hazards identified in the European Maritime Safety Agency (EMSA) battery guidance are systematically allocated across the battery management system, SoS, and design levels, thereby positioning SoS as a complementary layer to the safety functions ultimately governed by the Battery Management System (BMS). Finally, future research directions are outlined for advancing SoS estimation in marine applications, focusing on interpretability, granularity, ultrasound-based strain monitoring, and integration in the maritime environment.

Reinforcement Learning Control for Natural Circulation in a Marine Pressurised Water Reactor Cooling System

Felipe Montana — 2026-07-03

In safety-critical systems, a system fault response can lead to a system shutdown. While safe at a component level, this poses safety challenges for the system as a whole, requiring an additional system to manage this process. In pressurised water reactor (PWR) submarines a loss of coolant pump can force a shutdown by dropping the control rods, referred to as SCRAM. To avoid this, a possible response is to use natural circulation, a degraded operating mode characterised by strong non-linearities in system dynamics, to provide a limited level of functionality. Under these conditions, conventional model-based control approaches become difficult to apply, as the assumptions underlying nominal system models no longer hold. This paper investigates the feasibility of using reinforcement learning (RL) as a fault-response control strategy for systems operating under degraded and poorly modelled conditions. RL provides a data-driven framework capable of learning control policies directly from a black-box model or simulator, without requiring an explicit analytical model. However, when applied in a safety-critical fault management context, understanding and validating the learnt control policy is essential. We analyse the policy learnt through RL by approximating it with a transparent surrogate model and through visualisation of the policy actions. We further assess the robustness of the policy to modelling errors, providing insight into its sensitivity to discrepancies between the simulated environment and the real system. The proposed
approach is evaluated using a simplified submarine reactor cooling loop model that captures key features of fault-induced operation, including changes in system dynamics due to platform pitch and cascading faults. The results demonstrate the potential of reinforcement learning for interpretable control
under faulted conditions.

Shape of complexity

Manuel Guillermo Reta Palacios — 2026-07-03

Failure in complex engineered systems does not occur, as a simple or linear accumulation of independent component faults. Their failure modes are often relational: degradation propagates through feedback loops, coupling pathways, and many-body interactions among sensors, controllers, and ac tuators. This creates a gap for prognostics and health man agement, where many established approaches still interpret system health primarily through single-channel indicators or pairwise summaries. This paper argues for a broader PHM perspective in which system health is read from the structure of interactions rather than from isolated signals alone. Simplicial complexes pro vide a natural representation for this purpose because they encode both pairwise and higher-order relations, while topo logical descriptors such as Betti numbers compress that re lational structure into interpretable, threshold-robust signa tures. Within this perspective, we use an interconnectivity pipeline based on mutual information, temporal lag, coupling modal ity, and O-information as one concrete example of how mul tichannel data can be converted into a simplicial complex and analysed topologically. We validate and tune our method using an analytic toy model in which connectivity between components can be controlled. As part of this, we discuss what different interaction measures can and can not recover. A double-loop controller motor experiment then illustrates the PHM value of the approach: edge density, mean edge strength, and persistent loop structure vary systematically across fault conditions even when no single signal provides an equally clear separation. Together these results provide evidence that relational and topological descriptions can extend PHM be yond the single-signal view of system health.

A Context-Aware Edge-Cloud Multi‑Agent PHM Framework for Multi‑Tool CNC Turning Machines Using a Single Vibration Sensor

BK Ramesh — 2026-07-03

CNC turning machines operate under highly variable and context-dependent conditions, where multiple cutting tools and machine subsystems share a common structural vibration path. This creates a practical monitoring challenge: a single sensor captures a superposed vibration response, while tool wear and machine degradation evolve differently across tools, operations, and cutting conditions. This paper presents a context-aware edge-cloud multi-agent Prognostics and Health Management (PHM) framework for multi-tool CNC turning machines using a single spindle-mounted vibration sensor. The framework assigns a logical PHM agent to each tool, while a context router uses controller-side process information, such as tool identity, spindle speed, feed rate, and depth of cut, to transform the shared vibration stream into tool-specific feature streams. Each agent performs in situ baseline calibration, health indicator (HI) estimation, monitoring mode selection, online degradation tracking, and, where applicable, Remaining Useful Life (RUL) estimation. To improve deployment robustness, the library includes adaptive threshold estimation, online threshold refinement, and explicit handling of anomaly-dominant and degradation-dominant operating regimes. The framework is implemented as an edge-cloud architecture in which the edge performs feature extraction, context-aware routing, HI computation, and alert generation, while the cloud ingests telemetry and curated feature streams for visualization, storage, and lifecycle management. The study includes an industrial CNC case study with three tool types and an auxiliary evaluation on the UC Berkeley milling dataset. Results show that the framework can isolate tool-specific degradation behavior using a shared sensor, provide early warnings for rapidly degrading tools, and support low-latency edge deployment.

Cost-Sensitive Deep Learning for Scania Component X: Minimising Operational Cost via Asymmetric Threshold Optimisation

Abdelhakim Mraihi — 2026-07-03

Maintenance decisions in industrial fleets must balance the cost of unnecessary interventions against the higher cost of missed failures. This paper presents a cost-sensitive deep learning approach for the Scania Component X benchmark. Rather than predicting the original five degradation classes directly, the problem is reformulated as a binary task that identifies whether a vehicle is healthy or at risk based on recent operational data. The predicted risk score is then converted into a maintenance decision through asymmetric threshold optimisation under the official Scania 5×5 cost matrix. Three temporal deep learning models—CNN, Transformer, and Temporal Convolutional Networks—are evaluated under the same cost-aware training and decision setting. Results on real data from thousands of heavy-duty trucks show how cost-sensitive learning affects the trade-off between failure avoidance, maintenance workload, and total operational cost.

Designing a Prognostics Framework for Pharmaceutical Development: Applying PHM Principles to Computational Drug Discovery for Novel HIV-1 C(SA) Protease Inhibitors

Charmaine Kahiya — 2026-07-03

Prognostics and Health Management (PHM) has transformed the engineering industry through accurate pre-initialisation failure prediction. Here, those principles are applied to pharmaceutical development by treating a physics-based computational chemistry model as a pre-synthesis failure predictor for candidate drug molecules. The target is the HIV-1 subtype C (South African, C(SA)) protease, the dominant strain across sub-Saharan Africa, against which the existing subtype-B-optimised protease inhibitors lose potency. Twenty pentacycloundecane (PCU) cage peptoid candidates in the C→N backbone orientation were evaluated against the HIV-1 C(SA) protease (PDB 3U71) by molecular dynamics and MM-GBSA binding free energy calculations (AMBER 20, ff19SB, n = 4 replicates), with nine FDA-approved protease inhibitors evaluated under an identical protocol as positive controls. Glu-PCU-Glu ranked first (ΔGbind = −89.24 ± 7.42 kcal mol⁻¹); the conformationally constrained Pro-PCU-Pro ranked last (−22.22 ± 2.20 kcal mol⁻¹). The MM-GBSA screen functions as a Stage 1 health check that rules out the primary failure mode, insufficient target affinity, before any synthesis cost is incurred. Three compounds, Glu-PCU-Glu, Cys-PCU-Cys and Tyr-PCU-Tyr, pass the check and are carried forward.

Keywords: prognostics and health management; pre-initialisation failure prediction; computational drug discovery; MM-GBSA; HIV-1 C(SA) protease

Hybrid Physics-Informed UKF–Transformer Framework for Lithium-Ion Battery SOH Estimation

Abdel Rahman EL KHATIB — 2026-07-03

The accurate estimation of the State of Health (SOH) of lithium-ion batteries is essential for ensuring reliable operation and enabling prognosis in energy storage systems. Model-based approaches such as the Unscented Kalman Filter (UKF) provide physically interpretable estimates and stability properties that can be analyzed under standard modeling and noise assumptions. However, their performance is constrained by the need for an explicit and accurate battery model, as well as careful tuning of process and measurement noise covariances. As a result, standalone UKF implementations may strugglw in the presence of nonlinear aging effects, parameter drift, and real-world operating variability. On the other hand, data-driven approaches—particularly transformer-based architectures—excel at modeling nonlinear systems and capturing long-range temporal dependencies. However, they typically require large and diverse datasets, are sensitive to the scenario distribution used during training, and may lack stability or physical interpretability. Despite these challenges, transformers can reduce the overall data requirement by efficiently learning both temporal relationships and cross-feature interactions within the input signals. In this context, we propose a hybrid physics-informed framework that combines a UKF with a transformer model. The UKF provides a physically grounded intermediate SOH estimate, while the transformer compensates for unmodeled dynamics and nonlinear degradation patterns. The transformer further improves the estimator’s efficiency by capturing temporal dependencies and cross-feature relationships, which reduces the amount of training data required while maintaining robustness and generalization. Using the strengths of both techniques, the proposed hybrid approach improves estimation accuracy, robustness to noise, and generalization capability compared to either method used alone.

Methodology for Estimating the Corrosion-Induced Degradation Trajectory of 304L Stainless Steel in Nitric Acid

Abdoulaye Affadine Haoua — 2026-07-03

Corrosion-induced degradation is a major challenge for the integrity management of industrial equipment operating in aggressive nitric acid environments, particularly in the nuclear industry. Austenitic stainless steel 304L is widely used in such conditions due to its corrosion resistance. However, predicting its long-term degradation remains difficult because corrosion kinetics depend on multiple operating parameters (e.g., temperature, nitric acid concentration, exposure history, and oxidizing species), while available experimental data remain sparse, heterogeneous, and mostly limited to laboratory studies.

This work proposes a methodology for estimating the corrosion-induced degradation trajectory of 304L stainless steel using data extracted exclusively from the scientific literature. First, a structured database of gravimetric corrosion tests was built from published studies and standardized by converting mass losses into equivalent thickness losses under the assumption of uniform corrosion. The collected data were categorized according to exposure conditions, and only renewed nitric acid environments were retained as representative of industrial operating conditions.

Based on these data, a power-law degradation model was identified to describe thickness loss as a function of time. The parameters of this model were then estimated using machine learning approaches based on decision-tree regression, allowing the prediction of degradation parameters as functions of operating conditions such as temperature, nitric acid concentration, and environmental descriptors. Model performance was evaluated using a leave-one-out cross-validation strategy adapted to the limited dataset size.

Finally, the predicted degradation parameters were combined with operating-condition sequences in order to reconstruct cumulative degradation trajectories under variable conditions. The proposed approach provides a reproducible and physics-guided framework for estimating degradation trajectories despite limited data availability and constitutes a promising basis for prognostics and remaining useful life (RUL) assessment of equipment exposed to nitric acid environments.

Physics-Aided RUL Prediction of Pick-and-Place Robotic Arm Using Dynamic Simulation and GMM-Based Health Indicator

Mujin Kim — 2026-07-03

Pick-and-place robotic arms driven by closed-loop servo motors operate under repetitive trapezoidal motion profiles (acceleration–constant speed–deceleration–stop), which yield non-stationary and multi-modal signal distributions and can undermine conventional distance-based anomaly detection and health indicators. This paper proposes a physics-aided prognostics and health management (PHM) framework for a one-degree-of-freedom servo-driven robotic joint. Based on failure mode and effects analysis (FMEA), bearing degradation (represented as increased friction torque) and shaft misalignment (represented as an eccentric disturbance torque) are selected as critical degradation mechanisms. To address the scarcity of industrial run-to-failure data, a high-fidelity Modelica dynamic model of a motor–shaft/bearing–compliance–load drivetrain is developed to synthesize progressive degradation scenarios. We first show that principal component analysis (PCA) visualization and a Mahalanobis-distance-based indicator can be unreliable because healthy data form multiple clusters induced by the servo operating phases. To explicitly capture healthy multi-modality, we train a Gaussian mixture model (GMM) using healthy modes and define a probabilistic health indicator as the negative log-likelihood (NLL) under the learned healthy distribution; the indicator shows an increasing trend across the designed degradation scenarios with degradation progression. The health indicator is then normalized between healthy and failure thresholds to construct a normalized remaining useful life (Normalized-RUL) index that provides a relative life measure without requiring ground-truth failure-time labels. Finally, continuous Normalized-RUL trajectories are estimated using a continuous wavelet transform–long short-term memory (CWT–LSTM) model that learns time–frequency degradation patterns and their temporal evolution. In the most severe simulated condition, the predicted remaining life decreases to approximately 40% of the nominal life, demonstrating the potential of the proposed framework for predictive maintenance of servo-driven robot joints.

Prediction and simulation of battery pack usage for intelligent service robot deployment at a train station

Alexander Löwen — 2026-07-03

In order to make rail transportation more attractive compared to motorized private transport, intelligent solutions are required across the entire mobility chain. Train stations, as places of connection and transfer, offer the greatest potential in this context. Thus, the deployment of autonomous service robots at train stations offers a wide range of possibilities for supporting passengers on their journeys by rail, for example, by providing information, accompanying them to the next link in the mobility chain, or transporting their luggage. Since such robots are battery-powered, one challenge is to carefully plan activities based on the remaining battery capacity.
Therefore, the aim of this work is to predict battery usage and, in turn, battery state, in order to inform passengers, and to enable intelligent planning of its usage.
In this work, realistic operational conditions of a service robot are systematically assessed through measurements on the service robot itself and through passenger surveys at a train station. The estimated operational conditions are experimentally replicated to simulate the heterogeneous use of the battery pack, while acquiring condition monitoring data throughout its use. The battery pack is modeled based on the single particle model, which can robustly predict the battery state by simulating its usage considering past operational conditions. The approach is validated using experimental data obtained from simulating realistic usage. On the one hand, additional battery packs are employed, and on the other hand, the load is varied. The advantage of this approach lies in the ability to account for future changes, such as higher loads than anticipated, to provide robust predictions.
This enables the intelligent use of service robots, offering passengers an improved service experience at train stations.

Representation Learning–Based Fault Diagnosis of Angular Contact Ball Bearings for Machine Tool Spindles

Jooho Hwang — 2026-07-03

As a precision mechanical component designed to reduce friction between moving parts, the rolling bearing is widely employed across various industries. Predicting the remaining useful life (RUL) of bearings is critical for preventing unexpected failures and ensuring safe and reliable equipment operation. Equally important is the role of lubrication, which not only supports proper bearing performance but also ensures the accurate operation of CNC spindles. The primary objective of predicting lubricant degradation is therefore to estimate the time at which the lubricant no longer fulfills its intended function

In this study, bearing defect frequencies and lubricant anomalies are investigated through vibration-signal analysis under different operating conditions using a representation learning approach. To be more specific, a representation framework is employed to reconstruct input signals. A target frequency band, extending from three times the rotational frequency to 10,000 Hz, is first defined. Spectral root mean square (Spectral RMS) features are then extracted from the low-to-middle frequency portion of this band, where defect-induced impulsive excitations are effectively captured. The framework is trained using labeled healthy-bearing datasets to establish an anomaly-detection threshold, which is subsequently applied during the testing phase to evaluate reconstruction loss. The threshold is defined as the 95^th percentile of the reconstruction means square error obtained during training. This criterion enables the identification of anomalies in labeled fault-condition data. The effectiveness of the proposed method is demonstrated by employing two practical datasets. The results indicate that the proposed method effectively detects anomalies in unseen data and achieves robust diagnostic performance

Self-Powered Multi-Parameter Wireless Sensing for Condition Monitoring of Marine Propulsion Shafts

Van Ai Hoang — 2026-07-03

Self-powered multi-parameter wireless sensing enables autonomous condition monitoring of rotating marine machinery, where wired power delivery and frequent maintenance are impractical. This paper presents a self-powered wireless sensor system (SP-WSS) that integrates a compact electromagnetic energy harvester (EH) with sensors for shaft speed, torsional strain/vibration, temperature, and power monitoring. The system was installed on a 300 mm training-ship propulsion shaft and evaluated for 7.2 h under real operating conditions. The harvester delivered an average power of 487.87 mW, exceeding the system demand of 374 mW by 30.4%, and maintained wireless data acquisition during the investigated period. The measured torsional responses captured operational shaft behavior and provided fatigue-relevant loading histories. These results confirm the feasibility of the proposed SP-WSS as a practical sensing platform for prognostics and health management (PHM) applications in marine propulsion systems.

A Similarity-Based Ensemble Framework for Remaining Useful Life Prediction of a Subway Door System

Kai-Lin Yang — 2026-07-03

Remaining useful life prediction is challenging when only a small number of run-to-failure trajectories are available and the evaluation emphasizes early prognostic accuracy. This paper addresses the PHME 2026 Data Challenge on remaining useful life prediction for a subway door system. We propose a similarity ensemble that characterizes each operating cycle with statistical features, captures position-feedback degradation behavior, and estimates the failure cycle by comparing partial trajectories with historical run-to-failure cases. The resulting estimates are fused and constrained using model-disagreement and operating-condition information, and the final remaining-life sequence is generated as a monotonic countdown from the estimated failure cycle. The method is evaluated using stratified cross-validation on the official training set and further assessed through the challenge leaderboard submission. Results show that the proposed framework provides accurate and stable predictions under limited-data conditions, achieving a validation score of 0.9467 on observable-degradation scenarios and an official leaderboard score of 0.9963. The study demonstrates that constrained similarity-based failure-cycle estimation is an effective and interpretable strategy for small-sample prognostics.

Agentic AI for PHM Pipeline Development: A Human-Supervised Case Study in the PHME 2026 Data Challenge

Takanobu Minami — 2026-07-03

Developing an effective prognostics and health management (PHM) prediction pipeline requires iterative decisions on data structure, degradation behavior, feature representation, evaluation metrics, and failure cases. This paper reports a human-supervised Agentic AI workflow for PHM pipeline development through a case study in the PHME 2026 Data Challenge, which focused on remaining useful life (RUL) prediction for a subway door servomotor system.

In the proposed workflow, a human researcher supervised the overall exploration, while AI-assisted roles supported planning, experiment specification, implementation, review, case-wise error analysis, and knowledge accumulation. Through this process, we developed a case-aware RUL prediction pipeline combining health-indicator-based estimation, operating-condition-aware correction, training-data-based calibration, temporal consistency checks, and submission-format verification. The final submission achieved an official challenge score of 0.9983.

This case study suggests that Agentic AI can help structure, accelerate, and document iterative machine learning research, while human supervision remains essential for selecting directions, managing risks, and interpreting results.

Gated Residual End-of-Life Prediction for Variable-Prefix RUL in the PHME2026 Data Challenge

Giuseppe Mannone — 2026-07-03

This paper presents a gated residual end-of-life (EOL) method for variable-prefix remaining useful life (RUL) prediction in the PHME 2026 Data Challenge. The task is to estimate the lifetime of partially observed subway-door actuation runs under hidden operating conditions. We frame the problem as EOL estimation: the model first predicts the cycle at which failure occurs and then derives RUL by subtracting the current cycle index. This gives prefixes of different lengths a common lifetime scale. The approach uses engineered prefix features that summarize electrical load, position behavior, shock events, trends, and source-model predictions. Several source models provide an initial EOL prior, while disagreement among them serves as an uncertainty signal. Rather than replacing this prior with one unconstrained predictor, the method refines it through bounded residual corrections. To keep the prediction stable, the method first forms a conservative anchor and calibrates it to the inference setting using released test measurement features. Two bounded specialists then propose residual corrections: one from similar training prefixes and one from shock, position, and current signal evidence. When these specialists disagree, a clipped gate controls how much each correction affects the final EOL estimate. Local diagnostics indicate where the source prior is sufficient, where corrections add value, and where difficult prefixes remain uncertain.

A Collective Learning Workflow for Remaining Useful Life Estimation of Building Assets Under Sparse Degradation Data

Edgar Segovia — 2026-07-03

Building maintenance in the commercial sector is still dominated by reactive and schedule-based strategies, yet the shift to predictive maintenance is held back by a chronic shortage of degradation data: individual assets such as fan coil units, air handling units and pumps rarely accumulate enough failure or degradation history to train standalone prognostic models, and the records that do exist are short and heavily skewed towards normal operation. Existing approaches do not resolve this. Deep sequence models require continuous multi-year recordings that newly instrumented buildings cannot provide and overfit on small, imbalanced datasets; semantic ontologies organise building data but carry no prognostic capability; and baseline forecasting methods generate operating features without estimating degradation. This paper presents a collective-learning workflow that groups functionally identical assets by their Brick semantic class, pools their operational records into a shared dataset, and trains a single Random Forest model to estimate each asset's daily degradation increment, which is then accumulated into a remaining-useful-life projection and updated as maintenance is recorded. Applied to operational data from a pilot building, the workflow produced per-unit daily degradation estimates and remaining-useful-life projections with quantified uncertainty for fan coil units that individually lacked sufficient history to be modelled in isolation. Under leave-one-out cross-validation, operational features alone did not recover the per-unit degradation level (cross-validated R² below zero), whereas fusing a static condition indicator with the operational dynamics was required to do so (R² = 0.67, CVRMSE 20 percent), reported as a preliminary finding rather than a validated prognostic capability. The results show that pooling sparse degradation labels across semantically aligned assets makes data-driven prognostics feasible for building portfolios under data scarcity, providing a transferable route to predictive maintenance that does not depend on multi-year failure records.

A Conceptual Framework to Integrate Prognostics and Health Management with Maintenance, Repair and Overhaul for a Hydrogen-Electric Aircraft Propulsion System

Lisandro A. Jimenez-Roa — 2026-07-03

This paper proposes a conceptual integration framework that maps Prognostics and Health Management (PHM) stages onto the Maintenance, Repair and Overhaul (MRO) process chain, motivated by a use case on Hydrogen-Electric Aircraft Propulsion System (HEAPS). Drawing on scientific literature, standards, and domain expertise, we identify failure modes, mechanisms, and sensor measurands for HEAPS, and examine maintenance considerations through PHM and MRO lenses. The framework reveals a fundamental asymmetry. From PHM to MRO, diagnostics is the only stage with operational links, feeding condition-based updates to the Aircraft Maintenance Programme (AMP) without altering its structure; prognostic capabilities show potential for planning the unscheduled at the Part 145 level, but beyond diagnostics PHM outputs remain underutilised due to missing certification pathways and because the schedule-driven AMP revision cycle can be slower than the dynamic decision-making advanced PHM promises. From MRO to PHM, Part 145 provides the ground-truth feedback essential to train and validate PHM models, yet structured end-to-end feedback pipelines remain a challenge in scale and architecture.

A Convolutional Autoencoder for Fast Compressive Sensing Reconstruction of Vibration Signals

Imen Tounsi — 2026-07-03

n many health monitoring applications, large volumes of high-frequency measurement data must be acquired and processed to extract reliable health indicators for fault detection and identification. Compressive sensing (CS) provides an effective framework to reduce data dimensionality at the acquisition stage by exploiting signal sparsity, enabling sub-Nyquist sampling and lowering storage and transmission requirements. However, practical CS deployment is often limited by the reconstruction step, which typically relies on iterative optimization algorithms that are computationally expensive and difficult to implement in real-time monitoring systems. This work proposes a learned reconstruction strategy that replaces conventional CS solvers with a convolutional autoencoder based approach. The sensing process follows the standard CS formulation, where the original signal is projected onto a lower-dimensional measurement space using a fixed random sensing matrix. During training, the autoencoder is constrained so that its encoder reproduces the measurement operation, while the decoder learns a data-driven inverse mapping to reconstruct the original signal from compressed measurements. At inference time, compressed measurements are directly fed into the decoder, eliminating iterative reconstruction. Experimental results obtained on simulated gearbox signals and real vibration measurements demonstrate that the proposed method significantly reduces reconstruction time compared with classical CS algorithms while preserving diagnostically relevant information for fault detection.

A Framework for the Integration of Hybrid Models in Digital Twin Architectures for PHM

Jill Mercedes Linneweber — 2026-07-03

Hybrid model structures combine physics-based and machinelearning
(ML) models to leverage complementary strengths
in prognostics and health management (PHM). While both
hybrid modeling and digital twin (DT) architectures are widely
studied, their structural interaction is rarely addressed systematically.
In practice, hybrid models are often developed
application-specifically, and their structural integration into
DT architectures remains weakly formalized.
This paper analyzes hybrid model structures focusing on their
architectural composition and establishes a requirement-driven
configuration framework based on core PHM constraints. Four
hybrid coupling strategies are classified according to their
structural integration principles and positioned within a design
space defined by structural dominance and integration
depth. Based on this configuration framework, architectural
integration requirements for DT environments are derived and
operationalized in a project-specific DT implementation.
The study contributes (1) a structured classification of hybrid
coupling strategies, (2) a requirement-driven configuration
framework for hybrid model structures in PHM contexts,
and (3) a structured integration workflow for embedding
hybrid models into DT architectures. The results provide
a consistent foundation for managing hybrid model structures
within scalable DT environments.

A Graph Auto-encoder Framework for Spatio-temporal Anomaly Detection of Corrosion across a Fleet of Offshore Wind Turbines Using ICCP Data

Orkun Temel — 2026-07-03

Offshore wind farms are exposed to severe marine conditions, which can lead to long-term structural integrity concerns due to corrosion-induced degradation processes. Here, we propose a spatio-temporal anomaly detection methodology using the Impressed Current Cathodic Protection (ICCP) data from an offshore wind farm. First, we employ a graph autoencoder (GAE) to infer the spatial variations in the measurements. We construct a graph based on the spatial proximity between wind turbines, where nodes and edges correspond to wind turbines and distance between turbines. Then, the latent representation of the measurements obtained by the GAE, are passed to a long-short term memory (LSTM) model, which infers the temporal evolution of measured signal and predict the next state. Finally, we perform anomaly detection using a combined scoring that includes graph reconstruction errors, latent prediction errors and observation-space prediction errors. Our results highlight the potential of integrating graph‑based and sequence‑based approaches for industry‑relevant anomaly detection and demonstrate that the proposed methodology can identify turbines and corresponding time periods exhibiting deviations from fleet‑level behavior.

An MBSE Driven Framework for Automated Generation of RAM Risk Models from Maritime Vessel System Architectures

Yanlai Wang — 2026-07-03

Reliability, availability, and maintainability (RAM) analysis in maritime systems depends on risk models that are typically developed manually from design documentation. This process is time-consuming, error-prone, and weakly connected to system architecture models. This work presents an MBSE-driven framework that transforms SysML system models into RAAML-based risk models, establishing a direct link between system design and RAM analysis.

The framework defines a multi-layer mapping approach covering concept, semantic, and meta-model levels. Structural, behavioural, and interface elements in SysML are systematically converted into risk modelling constructs, enabling the automatic generation of reliability block diagrams, mission profiles, and failure propagation models. Quantitative parameters, including failure rates and maintenance data, are incorporated to support system-level reliability evaluation. The derived propagation paths and critical dependencies also provide a basis for sensor placement and condition monitoring design, supporting PHM applications.

A maritime case study based on a remotely operated vehicle (ROV) demonstrates the applicability of the approach. The results show a reduction in RAM modelling effort of 34% to 55% compared to conventional methods, with greater benefits observed for more complex systems. The framework improves modelling consistency, traceability, and scalability by deriving risk models directly from system architecture.

A Methodological Framework for Prognosis Using Control-Oriented Models: Application to an Aeronautical Power Converter

Mohamed SKAIK — 2026-07-03

Developing Prognostics and Health Management (PHM) for safety-critical systems faces a major challenge. Obtaining degradation and failure data is both expensive and time-consuming, especially during the design and development phases. Models built at this stage for control specification and verification, such as those in MATLAB/Simulink using the Specialized Power Systems toolbox, were not designed to capture component faults or track degradation over multi-year aging horizons. In addition, their single-domain electrical focus neglects important multi-physics interactions like electro-thermal feedback. To address this gap, this work applies a parametric four-stage workflow that leverages the computational efficiency of electrical models, making long-duration degradation simulations practical for data generation. We apply this approach to an aeronautical power converter, with the Silicon Carbide (SiC) Metal-Oxide-Semiconductor Field-Effect Transistor (MOSFET) selected as the most reliability-critical component. The increase in on-state resistance is used as the principal degradation indicator. Without altering the model's structure, degradation is introduced through controlled parametric fault injection to generate structured degradation datasets, followed by systematic feature engineering of electrical signatures to identify degradation-sensitive patterns. Top-ranked features are then used to train regression models that provide a diagnostic estimate of fault severity. This end-to-end workflow validates the PHM pipeline using available design-stage tools, establishing a performance baseline and reducing technical risk before transitioning to higher-fidelity multi-physics co-simulation.

A Methodology for Progressive Physics Integration in Data-Driven Anomaly Detection: Application to Circuit Breaker Monitoring

Melvin FERNANDES NOVO — 2026-07-03

Physics-informed machine learning has emerged as a promising paradigm for industrial health monitoring, yet practical guidance on when and how to integrate domain knowledge into detection pipelines remains limited. This paper proposes a structured methodology for progressive physics integration in unsupervised anomaly detection, organised into three levels of increasing depth. Level 0 refers to purely data-driven models operating on raw signals. Level 1 injects operational covariates such as temperature or equipment subtype through stratification or conditioning. Level 2 integrates physical knowledge about the internal structure of the signal—its segmentation into electromechanical phases of differing diagnostic relevance—through phase-aware representations, scoring, or end-to-end architectural integration. The methodology is applied systematically to three method families—statistical envelopes, isolation forests and variational autoencoders—for monitoring medium-voltage circuit breaker coil currents, where the breaker’s protection-switching function makes condition monitoring critical. At the deepest level, a physics-informed conditional VAE (PicVAE) injects domain knowledge through phase-segmented inputs, FiLM-conditioned architecture, and a phase-weighted reconstruction loss. Validated on real operational data with expert-labelled anomalies, the results reveal two findings: (i) operational conditioning at Level 1 consistently improves detection across all three method families; (ii) structural physics injection at Level 2 has a method-dependent impact, yielding clear benefits for phase-aware representation learning while introducing trade-offs for simpler models. The PicVAE achieves the best overall performance (AUC-ROC =0.951, Youden J = 0.854). The proposed methodology provides a reproducible template for integrating domain knowledge into anomaly detection pipelines.

A Metric-Driven Framework for Evaluating Prognostic Based Failure Estimation on Spare Part Inventory Management

Tanmay Daga — 2026-07-03

In recent years, the aviation sector has been at the forefront of adopting Industry 4.0 technologies, including artificial intelligence, additive manufacturing, cyber-physical systems, big data analytics, and the Internet of Things (IOT). These technologies have accelerated the development of advanced main-
tenance strategies, such as Predictive and Prescriptive Maintenance, especially in the field of spare part inventory management. By leveraging insights from Prognostic Health Management (PHM) technologies, logistics and maintenance serviceproviders can optimize inventory levels to reduce costs while maintaining service levels, thereby minimizing aircraft downtimes. Despite these potential advantages, the widespread adoption of PHM strategies in the Maintenance, Repair and Overhaul (MRO) industry remains a challenge, primarily due to their modelling complexity, high cost of adoption, regulatory challenges, data availability, and impact assessment. However, a more targeted allocation of development resources, to address these barriers, can be achieved if economic benefits can clearly be demonstrated for individual stakeholders (such as logistics) and different PHM technology maturity levels. Therefore, the aim of this study is to quantify the benefits of prognostic-based inventory policies in comparison to traditional reliability-based approaches across different demand patterns. Specifically, this study investigates the influence of different prognostic accuracy and prognostic horizon levels on key performance indicators, such as total cost and service level. It also evaluates the robustness of the proposed methodology against noise factors like prediction biases and false alarms. Based on these comparisons, minimum performance requirements for prognostics based policies can be established to ensure tangible benefits. Consequently, this study not only provides the readers with a methodology to quantify the impact of prognostics-based failure prediction on spare part inventory management, but it also proposes a lightweight framework which could act as surrogate for prognostic models to assist in the development of future prescriptive maintenance strategies

A Novel High-Level Reasoning Architecture for Aircraft Prescriptive Maintenance

Haroun El Mir — 2026-07-03

State-of-the-art Integrated Vehicle Health Management (IVHM) systems and digital twins (DTs) integrate physics-based and data-driven methodologies for predictive maintenance. Such systems commonly incorporate multiple DT instances to estimate substantive outputs. Nonetheless, they exhibit key limitations: lack of multi-DT orchestration mechanisms, limited uncertainty quantification, and insufficient prescriptive decision-support capability. To this end, this paper introduces an innovative High-Level Reasoner (HLR) decision support architecture for aircraft systems. The proposed HLR architecture comprises a multi-layer data-transfer structure, with the principal HLR layer consisting of adaptable specialist modules that facilitate a robust decision support implementation for query-driven prescriptive maintenance. The developed architecture is illustrated on an aircraft landing gear system (ATA 32), orchestrating multiple federated subsystems; represented by the Brake Temperature DT and Tyre Pressure DT. The contribution is a modular architecture that provides a reusable framework for extension across aircraft systems and wider IVHM applications. It therefore serves as an enabling technology that advances beyond existing diagnostics and prognostics solutions for asset utilisation.

A Proposal for Application of Physics-Informed Digital Twin and Particle Filtering for the Detection and Prognosis in Harmonic Drives

Roberto Guida — 2026-07-03

Electro-mechanical actuators (EMAs) in aerospace and robotics increasingly rely on harmonic drives, whose compliant architecture introduces nonlinear dynamics and specific degradation mechanisms. Conventional Prognostics and Health Management (PHM) approaches remain limited: data-driven methods require extensive fault datasets and lack interpretability, while physics-based models are often too computationally demanding for embedded real-time use. This work proposes a physics-informed digital-twin-based framework for fault detection and prognosis in harmonic drives. The digital twin is implemented through a Physics-Informed Neural Network (PINN), so that governing mechanical relations are embedded directly into the training process. Wear evolution is described through a physics-based degradation model, while Remaining Useful Life (RUL) is estimated via particle filtering to provide probabilistic prognostic predictions. The study focuses on the digital twin and prognostic modules within a broader PHM architecture. Preliminary results show accurate reconstruction of nonlinear dynamics, physically coherent degradation tracking, and consistent probabilistic RUL prediction.

A Probabilistic Baseline Learning Framework for SCADA-Based Wind Turbine Aging Diagnosis and Multi-Scale Performance Degradation Analysis

Qianling Wang — 2026-07-03

Accurate diagnosis of wind turbine aging from SCADA data is essential for reliable long-term operation, yet conventional
wind-speed-binning methods often fail to capture the nonlinear and condition-dependent nature of power degradation. To
address this issue, this study proposes a probabilistic baseline learning framework for SCADA-based wind turbine aging
diagnosis and multi-scale degradation analysis. A two-stage mean–variance XGBoost model with uncertainty calibration
is developed to estimate both the healthy-reference power output and its predictive uncertainty under varying operating
conditions. The deviation between measured power and the probabilistic healthy baseline is then used to quantify degrada-
tion across temporal, wind-speed, directional, and joint windspeed–direction dimensions. The results show that turbine
aging is cumulative but strongly condition-dependent, with the most severe degradation concentrated in specific operating sec-
tors rather than uniformly distributed across all inflow states. Furthermore, some directional degradation hotspots appear
to correspond to potential upstream turbine positions, suggesting a possible wake-related contribution, although further
verification is still required. More importantly, the proposed framework not only improves aging diagnosis accuracy, but
also provides wind farm operators with interpretable information on the conditions under which degradation is most severe,
thereby supporting targeted inspection, maintenance planning, and future evaluation of wake-aware mitigation strategies such
as active wake steering.

A Prognostic Framework for Railway Track Geometry: Tamping Detection, Settling-Aware Estimation, and Spatially Resolved RUL Prediction

Abdelhamid Ghoul — 2026-07-03

Accurate estimation of railway track geometry degradation and reliable prediction of remaining useful life (RUL) are essential for cost-effective infrastructure management. This paper presents a self-contained Kalman filtering framework that integrates unsupervised tamping detection and Settling-aware post-tamping state management, jointly addressing state estimation, short-term forecasting, and RUL prediction for longitudinal level and twist parameters using only historical onboard monitoring (OBM) and measurement train (MT) data, without requiring traffic load, environmental, or maintenance metadata. The framework comprises four integrated stages: (i) source-specific outlier detection exploiting the distinct noise characteristics of OBM and MT instruments; (ii) unsupervised tamping detection through adaptive jump analysis on monthly representative values; (iii) a damped Kalman filter with adaptive noise modelling and spatial fusion of neighboring positions, augmented by a Settling-aware state management mechanism that detects the rapid post-tamping consolidation phase and injects physically informed velocity priors to prevent filter lag; and (iv) a RUL prediction module that propagates the final Kalman state forward under damped dynamics until the predicted geometry violates the Alert Limit (AL) or Immediate Action Limit (IAL) defined by EN 13848-5. The complete pipeline is evaluated on 50 consecutive track positions spanning a 50-metre segment of Sudostbahn line 870 (Switzerland), using 6 MT and 33 OBM observations per position collected over five years (2016-2021). Results demonstrate accurate estimation through both degradation and maintenance phases, with six-month forecast confidence bands. The multi-position RUL analysis classifies 98% of positions as safe beyond a 2-year horizon and identifies the remaining positions with finite RUL values, enabling spatially targeted maintenance prioritization.

A Reliability Digital Twin Architecture for Real-Time Fleet Monitoring and Predictive Maintenance of Hydrostatic Transmission Components

Bruno Dandine — 2026-07-03

As Digital Twins (DTs) are increasingly developed for Prognostics and Health Management (PHM), many published works remain conceptual, or rely solely on Data-Driven (DD) algorithms, that are difficult to deploy under fleet or industrial constraints. This paper proposes a reliability-centered DT architecture, specifically designed for fleet-level deployment with limited sensing, rare failures, and strong legacy knowledge.

The contribution lies in structuring and integrating heterogeneous estimation mechanisms, relying on physics-based lifetime models, Bayesian techniques for risk control of model updates, soft sensing, and with the adjunction of unsupervised anomaly indicators. The goal here is to have a coherent, explainable, and industrially deployable DT pipeline. Central to the framework is a two-level fusion strategy: (i) a centralized reliability fusion operating on cumulative damage, accelerated life testing (ALT) models and sparse failures, and (ii) decentralized embedded diagnostics, providing complementary health indicators under operational variability.

The architecture is demonstrated on hydrostatic transmission components using test-to-failure databases (DBs), accelerated life models, temperature soft sensing through Extended Kalman Filter (EKF) and neural networks (NN), and vibration-based anomaly detection/degradation quantifier via SOM–MQE. The paper explicitly addresses scalability, explainability, and statistical risk control, which remain open challenges in DT deployments for PHM. The proposed framework targets practitioners seeking DT implementations compatible with ISO‑13374 logic, uncertainty guarantees, and industrial or fleet asset management constraints.

A Study on the Impact of Categorical Alarm Data in Power Estimation and Anomaly Detection of Photovoltaic Inverters

Jorge Ruiz Amantegui — 2026-07-03

This study investigates the impact of incorporating categorical inverter data into power-forecasting and anomaly-detection frameworks. Three forecasting models are evaluated on their ability to estimate power output on a large dataset coming from a fleet of multiple photovoltaic plants, over one hundred inverters and an approximate total of 33.2 installed MW. The forecasting models employed are Multi-Layer Perceptron, Long Short-Term Memory, and Extreme Gradient Boosting. Two encoding strategies for categorical alarm codes are compared: one-hot encoding and entity embeddings. Anomaly detection is performed by analysing residuals between predicted and measured power output. By systematically evaluating the integration of categorical inverter data into PV monitoring models,this work addresses an important gap in the literature and provides a foundation for future research exploring advanced methods for exploiting categorical operational data in photovoltaic systems.

A Sustainable Anomaly Detection Framework for Autonomous Surface Ship: Adaptive Subsystem-Level Anomaly Detection Algorithm via MLOps

Minji KIM — 2026-07-03

For the stable operation of Maritime Autonomous Surface Ships (MASS), this study proposes a sustainable anomaly detection framework that integrates a subsystem-level Condition-Based Maintenance (CBM) model with an adaptive MLOps pipeline. The main engine is decomposed into 14 functional units, each monitored by a hybrid algorithm that combines an Attention-LSTM-AutoEncoder and an Isolation Forest to detect subtle anomalies. To address model performance degradation caused by gradual data drift in maritime environments, an Autonomous Maintenance Mechanism is developed. This mechanism utilizes state severity (Z-Score) and drift velocity (ΔZ) indicators to algorithmically distinguish between sudden physical faults and gradual sensor drift. Based on this distinction, the MLOps pipeline accumulates confirmed drift in a buffer and selectively retrains and redeploys models using local onboard data once sufficient evidence has been gathered, while bypassing suspected fault conditions to avoid learning anomalous patterns. Experiments on an engine testbed indicate that the proposed system can suppress the Anomaly Rate (AR_t) during data drift and help restore diagnostic reliability, suggesting a practical basis toward self-sustaining condition monitoring for MASS.

A Two-Stage Machine Learning Approach for Quantitative Gear Crack Detection Using Vibration Signal Analysis

Paarth Singh Rathore — 2026-07-03

Early detection of gear tooth cracks is essential for preventing catastrophic failures in rotating machinery, yet existing approaches struggle to accurately detect small incipient cracks due to their distinct vibration characteristics compared to larger cracks. Current machine learning methods optimize for overall performance, sacrificing sensitivity to early-stage damage where preventive maintenance is most effective. This study presents a regime-aware feature-driven two-stage modeling framework employing separate polynomial Ridge regression models for small cracks and large cracks, selected through systematic correlation analysis and exhaustive grid search optimization using vibration features extracted from residual signals. The small crack model utilizes wavelet
detail coefficients, while the large crack model employs clearance factor, envelope peak, and wavelet d1 std features, both validated using Leave-One-Out cross-validation with limited training data. Simulation results demonstrate that the proposed approach achieves R2 = 0.9982 under simulated, data-scarce conditions, with performance evaluated using leave-one-out cross-validation on 19 samples.

ACE – Automating Causal Extraction: Leveraging Large Language Models for Bowtie Diagram Generation in Failure Analysis

Priyank Venkatesh — 2026-07-03

This paper investigates whether open-source, instruction-tuned large language models (LLMs) can automate the generation of Bowtie diagrams from Failure Mode and Effects Analysis (FMEA) documentation. Three pipelines are developed: Retrieval-Augmented Generation (RAG), Optical Character Recognition (OCR) based extraction, and a vision-enabled dual-LLM approach. Each is designed to handle both structured FMEA tables and unstructured narrative text. Three models (Mistral, Qwen-2.5, and LLaMA-3) are evaluated using Sobol sensitivity analysis, stochasticity experiments, and expert Likert scoring on narrative outputs. With strict schema-constrained prompts, models frequently achieve Node and Edge F1 scores above 0.8 on tabular data. Outputs were identical across repeated runs under fixed settings. Sobol analysis shows that prompt strictness and prompt type are the dominant drivers of Bowtie quality, whereas decoding parameters have a negligible effect. On unstructured narrative text, all models struggled, producing hallucinated nodes, incorrect role assignments, and diagrams that deviated from expert references. The results establish a working approach for automating Bowtie generation from FMEA tables and identify the specific obstacles to extending this to narrative sources.

AI for Sustainable Building Operations: Data-Driven Anomaly Detection in Ventilation Systems

Anahid Wachsenegger — 2026-07-03

Detecting deviations in building time series data is essential for robust heating, ventilation, and air conditioning (HVAC) operation and energy-efficient facility management. In practice, however, building management system (BMS) data are often incomplete, heterogeneous, and lack reliable fault labels.
This paper presents a benchmarking and feasibility study of data-driven anomaly detection on multivariate air-handling unit (AHU) time series data under realistic deployment constraints. We construct a unified dataset and define a domain-informed rule-based baseline as an interpretable operational reference and source of weak labels. We further evaluate classical unsupervised methods and representation-learning approaches using Temporal Convolutional Network (TCN) and Time Series Mixer (TSMixer) autoencoders, considering both a joint multivariate representation of all selected sensors and subsystem-based representations in which sensors are grouped by AHU function. Additionally, SHapley Additive exPlanations-based (SHAP) attribution is used to improve interpretability by identifying the sensor-level contributions to detected deviations.
The results show that rule-based methods capture explicitly defined conditions, while data-driven approaches identify additional statistically unusual and temporally structured deviations, with representation-learning models flagging 1.1–1.4% of windows in the global setting and up to 4.7% in subsystem-based analyses. High-consensus events (~0.8%) occur during temporally localized episodes with agreement across multiple models, indicating robust, structured deviations. These detections represent candidate anomalies that require further validation.
Our results show that combining rule-based, classical, and representation-learning methods provides complementary insights into AHU behavior and helps screen for relevant deviations in performance and energy use.

AI-Driven PHM for Floating Offshore Wind Turbines: Review and Main Challenges

Son Hai Nguyen — 2026-07-03

Floating Offshore Wind Turbines (FOWTs) offer a transformative solution for capturing wind energy in deep waters, where fixed-bottom installations become economically unfeasible. However, Operations and Maintenance (O&M) costs, which represent up to 30% of total energy costs, remain a major barrier to widespread deployment. The harsh marine environment and limited accessibility demand intelligent and autonomous monitoring systems, making prognostics and health management (PHM) essential for cost-effective FOWTs operations. This paper presents a review of AI-based PHM studies specifically for FOWTs, addressing a significant gap in the existing literature. Particularly, most of existing reviews predominantly focus on offshore operations, digital twin concepts, structural dynamics, or control strategies, none have comprehensively analyzed AI applications tailored to the unique PHM challenges of FOWTs systems. Through a literature review of AI-based PHM studies using the Web of Science and Google Scholar databases, we identify a FOWT-specific monitoring emphasis on structural and station-keeping assets. In addition, we propose a comprehensive end-to-end PHM lifecycle for FOWTs, integrating a hierarchical taxonomy of critical components with a systematic mapping of AI methods to key PHM tasks. By synthesizing the state of the art and identifying critical technological gaps, this work outlines priority research directions essential for enabling reliable, scalable, and autonomous offshore operations.

Addressing the Cold-Start Challenge in Building Predictive Maintenance: Translating Facility Manager Expertise into Criticality Indices

Edgar Segovia — 2026-07-03

Modern smart buildings rely on complex mechanical and electrical systems such as HVAC units, pumps, and lifts to function effectively. However, determining the criticality of these assets for Predictive Maintenance (PdM) prioritisation is often constrained by the ’cold start’ problem, where newly commissioned buildings lack the historical failure data required for data-driven ranking. While industry standards provide generic equipment lists, they often fail to capture the context-specific operational risks recognised by Facility Managers. This knowledge is typically tacit, subjective, and poorly documented, limiting its integration into interoperable digital strategies.

This paper presents a method to translate qualitative human expertise into a quantitative engineering metric. The approach begins with qualitative interviews to elicit latent decision-making criteria specifically safety, business continuity, and occupant comfort. These insights inform a psychometric survey mapped to standardised asset classes using the Brick Schema ontology, enabling consistent asset categorisation across buildings. To process these inputs, the study employs a Mamdani fuzzy inference system with centroid defuzzification, which handles the linguistic uncertainty of human responses and produces a continuous criticality index for each asset class. The research shows how qualitative expert judgement can be structured into a reproducible Criticality Index (CI) and Asset Health Index (AHI). These indices allow asset owners to prioritise PdM resources based on a transparent, expert-informed assessment of operational risk rather than generic heuristics, providing a semantically grounded foundation for deploying predictive algorithms in data-scarce environments.

Anomaly Detection in Multivariate Industrial Signals: LLMs, TSFMs, or Classical Deep Learning

Allen Baranov — 2026-07-03

Large language models (LLMs) offer several distinctive advantages over other machine learning models. First, they are trained as general-purpose models and are readily available, which eliminates the need for task-specific training and allows them to improve rapidly over time. Second, they can be applied directly, without constructing domain-specific or signal-specific models. Third, they are easy to integrate into existing systems and can be deployed without requiring an additional training step. Finally, they are inherently interactive because users can direct them with natural language. In this paper, we investigate whether LLMs can achieve multivariate anomaly detection. To fully exploit the aforementioned benefits, we define a set of guiding principles (such as avoiding pre-learning or representation learning on the signals) to ensure the LLMs remain general-purpose models. Based on these principles, we then propose several algorithmic approaches for building multivariate anomaly detection pipelines. We compare our approaches with two alternatives: (i) classical deep learning pipelines trained specifically for anomaly detection, and (ii) a foundation-model-based approach, in which domain-specific or general purpose time-series foundation models are trained without explicit supervision for anomaly detection but are then used for this purpose. The comparison highlights trade-offs along three key dimensions: anomaly detection accuracy, computational cost, and the amount of domain knowledge required to develop the pipeline. We evaluate our methods through two case studies. The first uses a benchmarking testbed designed for anomaly detection, while the second examines real-world data from wind turbines with known anomalous events.

API-Based Integration Framework for Dual-LLM Prescriptive Maintenance Report Generation in PHM-Enabled Digital Twin Applications

Atuahene Barimah — 2026-07-03

Digital Twins (DTs) have emerged as key enablers of Prognostics and Health Management (PHM) for predicting asset failures and optimizing maintenance strategies. However, translating predictive insights into actionable prescriptive maintenance plans remains a significant implementation challenge. This paper proposes an API-based framework that extends PHM-enabled DTs by incorporating a prescriptive maintenance layer aligned with enterprise operational constraints, including workforce scheduling, inventory availability, cost considerations, and compliance requirements. The framework integrates a generator model to produce maintenance recommendations and a checker model to evaluate report quality against operational criteria using a sequential model loading approach. Prescriptive maintenance reports using enterprise data as context are generated for a hydraulic system undergoing MCD scenarios. The framework proposed in this paper provides a low-cost implementation for integrating LLMs for prescriptive maintenance reporting for DT applications. This study contributes to LLM implementation use cases for DT applications for fleet asset management.

Active Sim-to-Real Gap Reduction for Industrial Inspection via Digital Twin and Embedding Analysis

Huimin Zhuge — 2026-07-03

Simulation-based training is increasingly used in automated industrial inspection, where collecting and annotating real-world inspection data is costly and often impractical. While synthetic data generated from digital twins enables scalable training, models trained solely in simulation suffer from a significant sim-to-real gap under real inspection conditions such as varying lighting, surface properties, and sensor noise. In this work, we propose a data-efficient sim-to-real adaptation framework that combines representative sample selection via k-determinantal point processes (k-DPP) with embedding-level alignment using Kullback–Leibler (KL) divergence. The key idea is to actively identify a small set of representative synthetic samples, acquire the corresponding real images, and align their latent feature representations while retaining the coverage provided by the larger synthetic dataset. We first train an RF-DETR(Detection Transformer) detector on 550 synthetic inspection images, achieving near-perfect performance in simulation but only 0.2516 mean Average Precision (mAP) on real-world images. Using only 50 paired real images (approximately 10% of the synthetic training set) together with 500 unpaired synthetic images, the proposed method increases real-world mAP from 0.2516 to 0.8853. The k-DPP sampling strategy maximizes the diversity of selected samples, reducing the risk of bias introduced by limited real-world data, while KL-based embedding alignment further reduces domain discrepancy between synthetic and real images. The proposed framework provides a lightweight and practical approach for reducing sim-to-real gaps in a representative industrial inspection setting where real data collection is limited.

Aging States Estimation and Monitoring Strategies of Li-Ion Batteries Using Incremental Capacity Analysis and Gaussian Process Regression

Moritz Landwehr — 2026-07-03

Existing approaches for battery health forecasting often rely on extensive cycling histories and continuously monitored cells. In contrast, many real-world scenarios provide only sparse information, e.g. a single diagnostic cycle. In our study, we investigate state of health (SoH)- and remaining useful life (RUL) estimation of previously unseen lithium-ion cells, relying on cycling data from begin of life (BOL) to end of life (EOL) of multiple similar cells by using the publicly available Oxford battery aging dataset. The estimator applies incremental capacity analysis (ICA)-based feature extraction in combination with data-efficient regression methods. Particular emphasis is placed on a multi-model Gaussian process regression ensemble approach (GPRn), which also provides uncertainty quantification. Due to a rather cell invariant behaviour, the mapping of ICA features to SoH estimation is highly precise and points out a normalized mean absolute error (NMAE) of 1.3%. The more cell variant mapping to RUL estimation is challenging, reflecting in a NMAE of 5.3%. Using the estimation results, a RUL monitoring strategy is derived. The objective is to safely operate a battery cell from BOL to EOL by only taking sparse diagnostic measurements. On average, only four diagnostic measurements are required during a cell’s lifetime of 3300 to 5000 cycles.

Analysis of fault-induced electromechanical disturbance effect in a closed-loop system

Urszula Jachymczyk — 2026-07-03

This paper examines whether motor current can provide a useful proxy for bearing faults in a closed-loop drive system. In such systems, the controller continuously adjusts current to maintain the operating point, which can hide small fault-related increases in losses and make overall current consumption a weak health indicator. To address this problem, current signatures from the Paderborn University bearing dataset were analyzed using 23 indicators covering global current level, band-limited residual ripple, envelope-based features, and classical sideband measures. These indicators were ranked according to class separability, stability in healthy measurements, and robustness to residual operating-condition variation. Their physical relevance was further explored by predicting selected current indicators from vibration features. The analysis shows that better discrimination is obtained when the analysis is shifted from total current level to residual ripple content, where fault-induced effects are significant enough to be visible and distinguishable from noise. Among the tested features, the band 105-2000 Hz features consistently provided the best separation between healthy and faulty bearings across 4 operating conditions. These results suggest that, for closed-loop drives, band-limited current features are promising monitoring inputs for sensor-light diagnostics.

Bayesian Online Changepoint Detection Algorithm for Degradation Monitoring and Prognostics of Electrical Devices

Roman Mukin — 2026-07-03

Electrical devices that undergo frequent switching operations can suffer mechanical wear over numerous cycles. Identifying early signs of degradation through routine measurements is important for condition monitoring and predictive maintenance. This work applies Bayesian Online Changepoint Detection (BOCD) to multivariate signal data from an electrical device to detect the onset of degradation and estimate remaining useful life (RUL) under uncertainty.

Classical approaches such as rolling-window trend fitting react slowly to regime changes and require careful tuning. BOCD instead maintains an online posterior over the run length and recursively updates a predictive model as new observations arrive. We extend the classical formulation to a multivariate regression setting with operation-dependent trends, conjugate matrix-normal inverse-Wishart priors with the regression-coefficient precision calibrated from prescribed alarm bounds to confine initial predictions to the operational envelope, and a Weibull hazard function that encodes wear-out behavior. The resulting algorithm performs online inference without storing the full measurement history while providing calibrated predictive uncertainty.

The framework is first evaluated on synthetic degradation scenarios covering abrupt, gradual, stepwise, and accelerating fault modes, and then applied to real mechanical endurance-test data from a run-to-failure experiment. On the synthetic scenarios, abrupt changepoints are detected without delay and gradual degradation is flagged once the cumulative drift exceeds the noise level, with the lower, median, and upper RUL bounds converging consistently at end of life. On real endurance-test data, the algorithm maintains a stable RUL estimate throughout the healthy device life despite noise-induced changepoints, and provides a timely end-of-life warning once sufficient post-changepoint evidence confirms the degradation trend.

Bayesian Post-Repair Prognostics for Reliable RUL Prediction

Mariana Salinas-Camus — 2026-07-03

Prognostics aims to predict the Remaining Useful Life (RUL) of engineering assets and is essential for effective Predictive Maintenance (PdM). Unlike preventive maintenance, PdM offers substantial cost benefits by scheduling maintenance only when needed. However, most existing prognostic models assume that repairs return an asset to an “as-good-as-new” condition. In practice, repairs are often imperfect, as they only partly restore the asset and may change its subsequent degradation behavior. This mismatch represents a major limitation of current prognostic approaches, as poor prognostic performance can lead to unnecessary maintenance actions or unexpected failure.

This paper proposes a fully Bayesian prognostic model, named the Sequential Bayesian Semi-Markov Framework (SBSM), that explicitly accounts for imperfect repair while being trained exclusively on data from non-repaired assets. The framework combines a Hidden Semi-Markov Model (HSMM) to represent the degradation model with a particle filter for the predictive step. Repair actions are incorporated as prior distributions that represent repair effectiveness, enabling repair uncertainty and prognostic uncertainty to be treated in a unified manner. This formulation allows post-repair degradation trajectories to differ from pre-repair behavior without retraining the model.

The approach is experimentally validated using an in-house dataset of aluminum open-hole specimens subjected to constant-amplitude fatigue loading, repaired via cold spray deposition, and tested to failure. Baseline (non-repaired) specimens are used for training, while repaired specimens are used exclusively for testing. The proposed method is compared against a Convolutional Neural Network (CNN) baseline. Results show that the SBSM achieves lower prediction error and improved probabilistic calibration, particularly under significant distributional shifts induced by more effective repairs. The framework demonstrates robust post-repair RUL prediction and well-calibrated uncertainty estimates, highlighting its potential for real-world predictive maintenance applications involving imperfect repair.

Benchmarking Time-Series Anomaly Detection Algorithms for Photovoltaic Plants

Hannes Bradl — 2026-07-03

Photovoltaic (PV) systems are increasingly important for renewable energy generation, but faults can reduce yield if they remain undetected. This paper presents a benchmark of multivariate time-series anomaly detection (TSAD) methods for PV monitoring using two complementary datasets: a newly collected real-world twin-plant PV dataset (FHC), in which controlled faults were physically introduced during operation and additional anomalies were injected after data collection, and the publicly available PV fault dataset (UTFPR). Using the TimeEval framework, we evaluate a broad range of unsupervised and semi-supervised TSAD algorithms on the FHC dataset, and only unsupervised algorithms on the UTFPR dataset, as the latter does not contain fault-free days required for semi-supervised training. On the FHC dataset, the highest performance is achieved by semi-supervised reconstruction-based methods. On the UTFPR dataset, fully unsupervised distance- and density-based methods perform best. The ablation study further shows that environmental features do not necessarily improve detection performance and may introduce confounding operating-regime structure for some algorithms. These findings highlight the importance of dataset characteristics and feature selection when applying TSAD methods to PV monitoring.

Bifurcated Remaining Useful Life Prediction: A Hybrid Approach for Realistic Uncertainty Characterization

Xabier Belaunzaran — 2026-07-03

This study presents a novel hybrid prognostic framework for uncertainty-aware Remaining Useful Life (RUL) estimation in turbofan engines using the NASA C-MAPSS dataset. The framework employs a state-aware strategy that bifurcates the engine's operational lifespan into `healthy' and `degraded' regimes. An LSTM-based autoencoder, trained strictly on nominal data (RUL > 150 cycles), monitors reconstruction error to act as a robust state classifier. For the healthy regime, a Conditional Weibull Survival Analysis is used for Mean Residual Life estimation. For the degraded regime, a Probabilistic Neural Network with Monte Carlo Dropout captures both aleatoric and epistemic uncertainties. Rather than using rigid binary labels, a calibrated sigmoid function converts the autoencoder’s output into continuous state probabilities, dynamically weighting the final ensemble prediction. The primary strength of this framework is its generation of physically consistent uncertainty bands, yielding high-confidence predictions near end-of-life while accurately reflecting the inherent variance of early operation, providing a robust tool for risk-informed maintenance.

Bridging Methods and Data in System-Level Prognostics: A Comprehensive Review

A. Madjid CHERGUI — 2026-07-03

System-level prognostics (SLP) is a core problem in prognostics and health management (PHM) that seeks to predict future health states or remaining useful life (RUL) in complex multi-component systems where degradations interact through functional and operational interdependencies.
Despite significant progress, the SLP literature remains fragmented, and existing reviews largely offer taxonomies without systematically linking modeling assumptions, data availability, and validation choices to reported performance and reproducibility.
This paper offers a data‑centric synthesis of the SLP landscape by integrating insights from influential reviews, technical contributions, and publicly available prognostics datasets.
We conduct an analysis to characterize publication trends and research domains, identify and quantify the coverage of recurring SLP challenges, and assess how current methodologies address (or overlook) these issues.
We also curate a dataset catalog to quantify gaps between methodological ambitions and benchmarking resources.
The study concludes by outlining priority directions for advancing reproducibility, data diversity, and deployment‑oriented SLP research.

Classification and Assessment of Propeller Faults in Electric Unmanned Aerial Vehicle Drive Trains

Immo Schmidt — 2026-07-03

Propellers are critical to the safe operation of multicopter unmanned aerial vehicles (UAVs), as faults can decrease the efficiency of the propulsion system and affect flight performance. Depending on the type and extent of the fault, the effects can range from a slight reduction in performance to a significant loss of thrust that could compromise safety. Because of the limited amount of sensor data available on board a UAV, propeller damage cannot be measured directly. Therefore, a data-based prediction using available sensors is required. This paper focuses on establishing and predicting a health index for damaged propellers.
A test bench is used to investigate the effects of two different types of damage: broken propeller tips and notches at the leading edge. Each type of damage is examined at three levels of severity. Based on sensor data collected from the test bench, a health index is defined to characterize the remaining performance of the damaged propellers. A two-stage approach for the data-based health prediction is implemented by first classifying the type of the propeller faults, and then employing a random forest regressor to estimate the remaining health.

Comparative Evaluation of Prognostic Models for Medium-Term Pressure Prediction in Gas Transmission Units

Houda SARIH — 2026-07-03

Gas pressure reduction and delivery stations are critical components of gas transmission networks, where pressure drifts may signal early degradation or control malfunction. Despite the availability of measurements supporting data-driven Prognostics and Health Management (PHM), medium-term forecasting remains difficult due to non-stationarities, operator adjustments, and heterogeneous setpoints. This study investigates 5-day-ahead pressure forecasting for maintenance planning and early detection of evolutions preceding threshold exceedances. A robust preprocessing pipeline addresses irregular sampling, missing data, and level shifts under real operating conditions. Two interpretable models are evaluated: SARIMAX, for seasonal and exogenous effects, and LightGBM, for nonlinear dynamics with feature-importance analysis. Performance is assessed on a homogeneous subset of gas pressure measurement recorders using standard regression metrics. The results emphasize not only predictive capability but also interpretability, positioning transparent forecasting models as scalable and auditable building blocks for PHM in gas transmission networks.

Contrastive, Autoencoding, and Variational Representations for Telemetry-Driven RUL Prediction

Mahmoud Rahat — 2026-07-03

Predictive Maintenance in safety-critical industries relies on accurate Remaining Useful Life (RUL) estimation from multivariate telemetry. Still real-world datasets are often dominated by censored observations and frequently lack explicit failure annotations. These constraints limit the effectiveness of purely supervised learning and motivate the need for approaches that can leverage unlabeled data. This paper presents a Pseudo RUL Guided Semi Supervised Learning framework that combines unsupervised representation learning with physics and statistics based soft failure indicators to enable robust RUL prediction under scarce failure labels. Compact latent representations are learned from censored telemetry using three encoder families i.e. autoencoders, variational autoencoders, and contrastive learning. The learned representations are subsequently used as inputs to a lightweight regression model trained on the available labeled samples. In scenarios where no failures are recorded, soft-failure transitions are used to construct pseudo-RUL targets, allowing training to proceed even in fully censored settings. Experiments on three diverse multivariate time-series datasets demonstrate that the learned representations consistently reduce prediction error relative to raw features while also reducing model size.

Cost-effective Inspection and Maintenance Rule for Train Control Beacons

Dersin Pierre — 2026-07-03

A substantial part of the French railway network is equipped with ground-based beacons that control train speed. This
“KVB” system (French initials for beacon-based speed control) plays an important role in ensuring train operation safety.
When a beacon is failed, maximum speed compliance over the section where failure occurred is potentially at risk. Therefore,
detecting failed beacons and replacing them in a timely manner but not too often is fundamental to guarantee safe operation of the network while keeping maintenance costs under control. This is accomplished by means of regular commercial trains which detect failed beacons (albeit imperfectly). Upon a number p of successive detections of the same beacon as failed by different trains over a certain period (typically one day), this beacon is signaled as failed to the maintenance control center and is replaced as soon as possible. The question therefore arises of optimizing the signaling rule, i.e., determining the best number p of apparent detections that should trigger a maintenance intervention. Three main contributions are reported in this paper: 1) Determination of steady-state operational availability as a function of the failure rate, the headway, the mean time to restore and the fault detection probability; 2) A method for the optimal choice of the number p in the signaling rule; 3) An algorithm for diagnosing whether a beacon or a train is defective, thereby reducing detection time and false positives. Several sensitivity analyses are also conducted, both of the availability and the total cost,
with respect to the various relevant parameters. Generalization to other train control systems, such as the European-wide
Pierre Dersin et al. This is an open-access article distributed under the terms of the Creative Commons Attribution 3.0 United States License, which permits unrestricted use, distribution, and reproduction in any medium, provided
the original author and source are credited. ETCS (European Train Control System) should be straightforward. To the authors’ knowledge, it is the first time that such algorithms for decision optimization under uncertainty are applied in the context of train control system maintenance.

Cloud-Enabled Autoencoder-Based Anomaly Detection for Gas Turbine Faults

nima ameri — 2026-07-03

Engine Health Monitoring (EHM) is critical to reducing service disruption and operational costs in the aerospace industry.
While traditional physics-based algorithms have been effective, the increasing volume and complexity of sensor data
necessitate more scalable, automated, and accurate detection methods. This paper discusses the transition of Rolls Royce
EHMcapability- for civil and commercial aircraft engines toward a modern Machine Learning (ML) paradigm within
the Rolls-Royce Data Science Environment (DSE). Leveraging a cloud-based architecture and its tech stack, we address
the challenges of manual recalibration and high false-positive rates. We present a case study of an anomaly detection frame
work utilizing a two-stage approach: a deep neural network for input-output residual generation followed by an autoencoder with a custom loss function for latent representation. Furthermore, we outline the integration of MLflow to ensure
robust experiment tracking, as well as the use of a cloud-based unified data governance framework. This work demonstrates
how end-to-end ML lifecycle management maintains model performance and operational trust in a highly regulated envi
ronment.

Cost-Efficient Prognostics Framework for Heliostat Drive Units

Dominik Steinberg — 2026-07-03

In concentrating solar power tower plants, heliostat drive units are critical components, as they control the precise two-axis alignment of thousands of mirrors, so-called heliostats, that focus incoming solar radiation onto a central receiver. Due to the large field sizes and the corresponding long heliostat–tower distances, even small angular deviations in the milliradian range (1 mrad ≈ 0.057 degree) result in significant focal point displacements at the receiver. Consequently, the reliable operation of heliostat drive units is essential for the stable and safe operation of solar tower plants.

However, existing research on heliostat operation and maintenance (O&M) predominantly focuses on optical aspects such as mirror soiling (i.e., the accumulation of dust and sand on reflective surfaces), mirror calibration and tracking algorithms, and the influence of wind loads on heliostat performance and structural behaviour. In contrast, the operational health of heliostat drive units remains largely unexplored. To close this research gap, this study presents a cost-efficient prognostics framework for the recording and the subsequent maintenanceoriented analysis of operational data of the heliostat drive units. For this purpose, an extensive measurement campaign is conducted in the heliostat field of the DLR solar tower research facility in Juelich, Germany. In addition to the existing industrial-grade reference sensors and data loggers at the solar tower research facility in Juelich, this study develops a low-cost Arduino-based data acquisition system and performs a comparison between those conventional and cost-efficient monitoring architectures.

The results demonstrate that the recorded measurement data provide a robust foundation for monitoring the heliostat drive units. The proposed prognostics framework is experimentally validated and successfully applied to selected heliostats in the field: first, this shows that sufficiently precise measurements, adequate sampling rates, and straightforward installation and handling can be achieved for field deployment. Second, it demonstrates the capability to identify and analyse real-world operational anomalies. And third, it enables reliable and costefficient monitoring significantly reducing the barriers to scalable prognostics and health management (PHM) deployment. Although developed for heliostat drive units, the diagnostics and prognostics methodology presented in this work may be transferable to a wide range of electromechanical systems in industrial PHM applications, as it integrates sensors and instrumentation with anomaly detection, and supports conditionbased and predictive maintenance strategies.

Decision at First Sight: An Attention Network for Direct Maintenance Optimization from Sensor Data

mbadfar — 2026-07-03

Maintenance planning is a crucial strategy in industrial systems, where maintenance costs can consume up to 40% of total production expenses, and downtime costs can reach hundreds of thousands of dollars per hour. Despite its importance, the implementation of advanced maintenance approaches remains limited due to challenges such as insufficient resources, lack of expertise, inadequate funding, and difficulty converting vast operational data into actionable decisions. This paper introduces a novel attention-based deep learning model for maintenance scheduling that bypasses traditional degradation modeling and optimization techniques. The proposed model operates directly on sensor data, leveraging a multi-head attention mechanism within an encoder-decoder architecture to generate maintenance schedules. The cost function of the model is flexible and can be customized to accommodate different maintenance scenarios, making it adaptable to various operational requirements. The model's performance is validated through comparisons with the state-of-the-art predict-then-optimize benchmark, demonstrating its ability to generate cost-effective maintenance schedules.
For commercial lithium-ion battery fleets, ATOM achieves a 22–35% reduction in maintenance expenses relative to predict-then-optimize approach.
This approach provides a scalable, data-driven solution for dynamic and complex maintenance environments, eliminating the need for explicit remaining useful life (RUL) estimates or predefined degradation models.

Development of Bearing Fault Detection Models using Multibody Simulation Training Data

Luca Giraudo — 2026-07-03

This study evaluates the performance of simulation-trained fault detection models on large spherical roller bearings vibration data. A high-fidelity multibody (MB) model of a SKF 22240 CCK/W33 bearing is developed through Simscape Multibody to reproduce the coupled dynamics of inner and outer rings, cage, and 38 rolling elements. Localized defects on raceways are represented through a pointcloud contact formulation, where selected nodes are radially displaced to emulate faults. The model outputs triaxial accelerations at the outer ring under realistic loading and speed conditions that mirror an experimental test campaign.
Simulation signals are processed through bandpass filtering, envelope analysis, and segmentation. A set of 23 time and frequency domain features is extracted from each segment, then each feature vector is normalized. The same processing chain is applied to experimental data acquired on a medium-to-large bearing test rig at Politecnico di Torino, mounting SKF 22240 CCK/W33 bearings with machined inner race, outer race, and rolling element defects.
A supervised Artificial Neural Network (ANN) classifier is trained only on the simulated feature dataset and then directly evaluated on the independent experimental dataset, in a process free of any data transfer. The network addresses a two-class problem (healthy and damaged), and its performance is assessed through standard classification metrics computed over multiple bootstraps of both training and test sets.
Despite the intrinsic differences between simulated and experimental signals, the ANN trained purely on simulations provides reliable and selective fault detection on real measurements. Most residual classification errors are concentrated in low-speed inner race damage conditions, where fault signatures are weak and partially overlap with healthy observations, while high-speed and outer race damage conditions are recognized more robustly.
These results show that MB simulation can generate sufficiently realistic vibration data to train ANN-based fault detection models that generalize experimental measurements for large spherical roller bearings. The proposed framework introduces an alternative to costly fault campaigns and offers a flexible way to expand training datasets across loads, speeds, and defect sizes in industrial condition monitoring applications.

Development of a hybrid PHM system for flight control actuators fusing data analytics with physical knowledge

Andrea De Martin — 2026-07-03

It is well known that the development of PHM systems can be approached with a large variety of techniques. In general, the different prognostics techniques are classified as "data driven", or "physics model based", depending on whether the assessment of a system health status and its evolution with time is performed with an intelligent statistical analysis of present and historical data (data driven approach), or the system actual operating status derived from different sensors is compared with the expected system status in the same operating conditions (model based approach). When a large database of historical data is available, data driven techniques can be accurate and do not require a knowledge of the underlying physics as for the case of model-based techniques. On the other hand, a data driven approach detects only anticipated faults, while a physics-based model approach can also detect unanticipated faults, that never occurred in the past. Flight control actuators of aircraft in revenue service are a typical application in which a large historical database can be available from maintenance, repair and overhaul departments, still the prediction of their health status may fall short of the necessary accuracy without a model description of their physics. For this reason, Safran Actuation, a leading manufacturer of flight control actuators, is conducting an extensive R&D work, together with Politecnico di Torino and Forvis Mazars, aimed at developing an effective PHM system with specific reference to the spoiler actuators of the Airbus A320. This use case is of a particular interest due to the very large number of this type of aircraft in service, to their expected continued use in the years to come and to the number of actuators per aircraft. With reference to this use case, this paper shows how an intelligent fusion of data analytics with physical knowledge can be a multiplying factor in improving accuracy and reliability of the PHM system, with the objective of developing a technological demonstrator for its validation in the lab prior to the implementation of a prototype.

Developing Generalized Health Index of Electric Vehicle Drivetrain

Kyung-Woo Lee — 2026-07-03

Motors and reducers are core components of an electric vehicle’s drivetrain. If either the motor or reducer fails, the vehicle cannot operate, and at high speeds, this poses a significant safety risk. Therefore, preventing failures in these components is critical for customer safety. However, most existing fault diagnosis models for electric vehicle drivetrains show limited performance under real driving conditions because load and speed vary continuously.

In this study, we propose a novel vibration signal generalization method that combines order tracking with physics-based amplitude adjustment techniques to improve diagnostic accuracy under variable operating conditions. Furthermore, we developed an AI model incorporating a health index that enhances generalization performance, enabling scalability across 5 different vehicle types

To achieve this, we designed a data processing technique that standardizes measurement data from various vehicle types by integrating domain knowledge, such as order analysis based on CAN bus speed information. The resulting health index successfully distinguishes deteriorated vehicles from normal ones regardless of vehicle type or driving conditions.

The findings of this study are expected to play a key role in applications under variable speed conditions.

Diagnostics for Mechanical Systems with Unknown Fault Modes: A Novel Open Set Recognition Approach

Jiaxuan Song — 2026-07-03

A common challenge in condition-based maintenance is that not all fault modes of the system are known from historical data, particularly in systems with evolving operating conditions, or are newly developed. Conventional data-driven diagnostic methods typically rely on a closed-set assumption, where all possible fault modes are represented during training. As a result, previously unseen fault modes are often incorrectly assigned to known ones with high confidence, potentially leading to ineffective or even risky maintenance decisions. To address this limitation, this paper proposes an open-set diagnostic approach that integrates supervised contrastive learning with a simplified Hopfield energy score. An encoder is trained using a supervised contrastive loss function to obtain well-separated embeddings of known system states. During inference, the alignment between a test observation and the learned state prototypes is quantified using the simplified Hopfield energy score. Observations with low similarity to known states are identified as unknown through thresholding. Experimental results on a benchmark dataset demonstrate that the proposed method effectively distinguishes unknown states while maintaining an accurate classification of known states, achieving competitive performance compared to established baselines. By explicitly identifying unknown states, the proposed approach enables more reliable and risk-aware maintenance decisions, particularly in safety-critical applications.

Direct and indirect Structural Health Monitoring of steel railway bridges: A state-of-the-art review and future challenges

Christos Sakaris — 2026-07-03

Steel railway bridges are a vital part of the transportation infrastructure, and their normal operation is crucial to a functioning society. However, aging bridges are subjected to traffic loads and harsh environmental conditions, which can lead to deterioration mechanisms. When damages caused by such mechanisms reach a critical level, they can lead to catastrophic bridge failures, high maintenance costs, and loss of human lives. Thus, early damage detection, localization, quantification, and the estimation of the remaining useful life of a bridge are crucial. Structural Health Monitoring (SHM) systems based on vibration measurements have been developed for bridge monitoring. SHM is characterized as direct or indirect (drive-by) depending on how the sensors are used. In direct SHM, vibration sensors are mounted on the bridge to measure the response of the bridge as the trains pass, while in indirect SHM, vibration sensors are installed on passing trains to measure the response of the bridge. The high cost for the deployment and maintenance of direct SHM instrumentation across the large number of bridges in a typical railway network limits its scalability, making network-wide deployment economically impractical. As a result, indirect SHM has been explored as a less costly alternative for network-level monitoring, while direct SHM remains highly valuable for critical assets, high-risk structures, and validating indirect monitoring methods. Despite growing interest, the main research gap is the existence of only two review papers on SHM in steel railway bridges, with the studies referred to in the review papers covering only direct SHM and mainly damage detection, localization, and quantification. The goal of the current review article is to address this research gap by reviewing the state-of-the-art in SHM methods applied to steel railway bridges between 2010 and 2025. The state-of-the-art encompasses direct SHM studies with numerical, experimental, and field validation on full-scale bridges, and indirect SHM studies with numerical and field validation on full-scale bridges and experimental validation on laboratory-scaled bridges.} {Within indirect SHM, frequency identification, which recovers bridge natural frequencies from train-mounted sensors, is treated as an enabling monitoring task that supports SHM workflow. In addition, the review article provides recommendations for future challenges.

Domain Adaptation of Automatic Speech Recognition Models for Diagnostic Applications

Aman Kumar — 2026-07-03

Automatic speech recognition (ASR), or speech-to-text (STT), is becoming an important interface for AI systems in diagnostic workflows, but general-purpose ASR models often degrade in specialized technical domains. In diagnostic applications such as fault identification, root cause analysis, and repair recommendation, general-purpose ASR systems struggle with domain-specific terminology, abbreviations, part identifiers, and measurement expressions, leading to elevated transcription errors. This work presents a domain adaptation pipeline that unifies three components: a synthetic benchmarking framework in which domain-specific technical text is converted to speech via text-to-speech~(TTS) synthesis and transcribed by open-source ASR models to establish baseline performance; Low-Rank Adaptation~(LoRA)-based fine-tuning of Whisper Large-v3 using those synthetic audio-text pairs; and transfer validation on curated real-world automotive YouTube recordings to assess generalization beyond synthetic conditions. Using automotive technical language as a representative diagnostic domain, a data-scaling study employing progressively larger subsets of in-domain training data evaluates performance on a held-out test set via word error rate~(WER), character error rate~(CER), normalized error metrics, alphanumeric error rate, semantic similarity, and Bidirectional Encoder Representations from Transformers Score~(BERTScore). Results show consistent gains from lightweight domain adaptation on both held-out synthetic data and real-world recordings, confirming that synthetic data generation combined with LoRA-based fine-tuning is an effective and computationally practical strategy for improving ASR accuracy in specialized technical domains where labeled speech is scarce.

Dual-Threshold Maintenance Optimisation for Hydraulic Floodgates under Runaway Stochastic Degradation

Jack Lally — 2026-07-03

A stochastic predictive degradation modelling framework for hydraulic floodgates operating with combined environmental and operational influences is presented. A condition-based maintenance approach is optimised for a multi-gate system to identify the most cost effective negative opportunistic preventative replacement policy, in order to minimise concurrent maintenance interventions. Our research focuses on the hydropower context, where evolving operating practices and regulatory constraints have increased exposure to the highly degrading fluid-induced vibration mechanism. A coupled simulation of the system is developed integrating (i) a synthetic hydrological forcing process, (ii) a reservoir-operation control model, and (iii) a stochastic degradation model based on a gamma process with state-dependent parameters. The degradation dynamics capture the interacting feedback loops that connect seal wear, leakage, vibration and erosion, with environmental conditions and operating decisions, leading to a run-away degradation effect beyond critical thresholds. A dual-threshold preventive maintenance policy is implemented using vibration as an observable proxy for a hidden degradation state, considering for inspection intervals, degradation uncertainty, and operational constraints on gate availability. The policy performance is evaluated via long-term Monte Carlo simulation, optimising expected annual cost under variations in conditions. The expected cost is highly sensitive to the lower preventive threshold, which effectively mitigates runaway degradation, while the upper threshold shows limited influence due to increasing end-of-life uncertainty. This undermines the effectiveness of negative opportunistic maintenance strategies aimed at avoiding simultaneous interventions for systems with runaway degradation behaviour. The findings emphasise the critical role of degradation non-linearity and information limitations in maintenance decision-making, and indicate that improved monitoring may be necessary to effectively make use of condition-based maintenance policies for similar applications.

Early Fault Detection in Rotating Machinery via Multivariate Autoencoder-Based Indicator Fusion

Faras Jamil — 2026-07-03

This research proposes a normal behaviour multivariate autoencoder, which fuses multiple condition indicators into a single high-level health indicator to provide a comprehensive overview of the mechanical component's health. The model is trained exclusively on healthy data to learn normal behaviour and detect faults by observing deviations from this learned normal behaviour. The proposed method is validated in real time by monitoring run-to-failure bearing experiments on an FE8 bearing test rig. It is employed to detect blind faults in real time during ongoing experiments, resulting in the termination of the experiment to analyse the cause of fault initiation. Furthermore, historical data is utilised to quantify the lead time between the proposed method's detection and the final termination triggered by traditional condition monitoring methods. A comparative analysis of the physical bearing damage after the completion of tests demonstrates the capability of the proposed method to detect blind faults at an early stage. The results suggest that this approach identifies fault at an earlier damage stage as compared to traditional methods. It improves the conditions for studying bearing fatigue initiation by avoiding the interference of secondary damage or extensive spalling. In addition, the real-time blind fault detection capability demonstrates the practical application of the proposed framework in preventing catastrophic failures within high-value industrial assets.

Edge AI - Enabled Smart Sensors for Predictive Maintenance of Marine Vessels

Igor Makienko — 2026-07-03

Traditional Predictive Maintenance deployment on marine vessels requires continuous acquisition and transmission of large volumes of sensor data, resulting in high complexity, cost, and cybersecurity exposure that limit large-scale adoption.

This work presents RSL Smart Sensors, an Edge AI - enabled sensing platform that performs distributed on-sensor intelligence to significantly reduce data transmission, operational costs, and cyber risks. Sensor data are processed locally, with only health indicators and diagnostic insights transmitted to the cloud for decision support.

A novel machine learning algorithm is introduced to estimate machine operating conditions directly from vibration data under strict power constraints, without access to external operating parameters. Validation on marine auxiliary generator turbochargers demonstrates high accuracy and confirms the feasibility of high-performance predictive maintenance at the edge under limited power and information conditions.

Edge-Deployed Generative Language-Based Retrieval for Aerospace Asset Health Management

Lukmon Rasaq — 2026-07-03

Aerospace asset health management increasingly relies on access to large volumes of maintenance documentation; however, operational environments are often constrained by limited connectivity and computational resources, restricting the use of cloud-based intelligence systems. This paper presents a fully offline, edge-deployable retrieval-augmented generation (RAG) framework for aerospace maintenance and prognostics using technical documentation. The framework integrates a locally hosted lightweight large language model with vector retrieval and cross-encoder reranking to support natural language querying of Airworthiness Directives (ADs) and maintenance records. Deployed on an NVIDIA Jetson Orin Nano 8GB device, the system performs document ingestion, indexing, retrieval, reranking, and response generation entirely on-device without cloud connectivity. Experimental evaluation on real aerospace maintenance documents demonstrates the ability to identify failure mechanisms, failure modes, root causes, affected components, and maintenance procedures described in FAA documentation. The cross-encoder reranking stage improves retrieval precision by refining semantically overlapping maintenance evidence prior to generation. The framework achieved an average fidelity score of 0.83, indicating that most generated responses remained grounded in retrieved FAA evidence. Across representative AD queries, the system achieved practical edge inference latency of approximately 4–10 seconds on the Jetson platform. The results demonstrate the feasibility of privacy-preserving, low-latency generative artificial intelligence for aerospace maintenance decision support on resource-constrained edge devices.

Edge-Server Collaborative System for Real-Time and In-Depth Damage Detection of Wind Turbine Blades using Acoustic Signals

Zhi Zhu — 2026-07-03

Efficient health monitoring is indispensable for the reliable operation of wind turbines. Damage to wind turbine blades, such as cracks and holes, typically generates whistle-like sounds during rotation. This study proposes a two-stage edge-server collaborative system for detecting blade damage using acoustic signals captured by arrays built from commodity microphones. The first stage employs a lightweight attention-based convolutional neural network to run on edge devices for the real-time binary classification to determine whether anomalous sounds are present. Suspicious time segments are stored for further analysis. The second stage uses a time-frequency sound event detection model that employs a detection transformer with an audio spectrogram transformer backbone to identify the time and frequency ranges of sound events via bounding boxes in the spectrograms. Owing to its high computational demand, this in-depth analysis is performed on a server. To validate the proposed system, acoustic signals were recorded intermittently for more than a year using micro-electromechanical system (MEMS) microphones externally attached to wind turbine towers. The models were trained and evaluated on a manually annotated dataset comprising 4,210 audio clips (15 s each) containing 14,420 sound events. The experimental results demonstrated that the binary classification model achieved an area under the receiver operating characteristic curve (AUC) of 0.920, whereas the sound event detection model attained an average precision at a 50% intersection-over-union threshold (AP50) of 0.510. Furthermore, evaluations on test data under unseen conditions, comprising 496 clips with 135 sound events recorded by handheld recorders at different locations, yielded an AUC of 0.867 and an AP50 of 0.440. The results highlight the robustness of the proposed system to variations in microphone types, recording locations, and environmental noise, demonstrating its strong potential for practical continuous automatic damage detection in wind power infrastructure.

Engine Health State Index (EHSI)

Rohit Deo — 2026-07-03

Assessing the health of diesel engines is challenging due to multiple coexisting failure modes, overlapping fault signatures, and highly imbalanced operational data. This paper proposes an Engine Health State Index (EHSI), a probabilistic health metric that aggregates risk estimates from multiple failure-code–specific models.

The framework employs a collection of binary classifiers, each trained to estimate the likelihood of a specific failure code from historical telemetry and diagnostic data. At each time step, the resulting failure risk vector provides a distributed representation of latent fault exposure rather than a single dominant failure mode. EHSI maps this risk distribution to a scalar health index using normalized uncertainty measures, enabling continuous tracking of health degradation without relying on explicit fault triggers.

Experiments on real-world diesel engine datasets show that EHSI produces smooth and interpretable health trajectories that correlate with impending failures while remaining sensitive to early-stage degradation. The proposed approach is model-agnostic, extensible to additional failure modes, and suitable for large-scale fleet monitoring applications.

Estimating the Battery State of Health with Quantified Aleatoric and Epistemic Uncertainty

Joshua Bogaert — 2026-07-03

Batteries are crucial in the transition towards a sustainable society. There is therefore an increased interest in battery health management. In battery health management, one of the key quantities is the State of Health (SoH), i.e., the maximum capacity of a battery, which decreases over time as the battery degrades. Accurate SoH estimations are needed to plan operations and battery replacements. A crucial challenge in SoH estimation is to quantify the uncertainty of the estimates. Two types of uncertainty must be considered. First, aleatoric uncertainty is irreducible and caused by inherent noise in the data. Quantifying this uncertainty gives a lower bound on the SoH. Second, epistemic uncertainty is reducible and is caused by, among other factors, a lack of training data. Epistemic uncertainty can be used to identify if a test sample differs from the training samples, i.e., if it is Out-Of-Distribution (OOD). In this paper, we estimate the SoH during discharge based on the current and voltage measurements obtained during charge. For this, we employ a Bidirectional Gated Recurrent Unit (Bi-GRU) neural network with attention. We estimate the aleatoric uncertainty using Simultaneous Quantile Regression (SQR), while we estimate the epistemic uncertainty by applying Orthonormal Certificates (OC). We test our approach on the fast charging dataset of Toyota. We achieve good results with a high accuracy, with a RMSE of only 0.00343 Ampere hours, and a good calibration. The model estimations become less accurate near the End of Life (EoL) of the batteries, but the corresponding data samples are correctly identified as OOD due to the high epistemic uncertainty.

Estimating Spall Severity in Rolling Element Bearings: A Supervised Learning Approach With Naturally Progressing Spalls

Stephan Baggerohr — 2026-07-03

Estimating the severity of localized defects in rolling element bearings is critical for accurate Remaining Useful Life (RUL) estimation, yet it remains challenging under non-stationary operating conditions with fluctuating speeds. Existing data-driven methods struggle to generalise due to the lack of high-fidelity, damage-progression data, and susceptibility to machine-specific structural transfer functions. A Siamese Transformer based neural network is utilized to predict continuous spall size directly from concurrent vibration measurements across three selected operating speeds, reducing the need for long-term trending. Using the amplitudes at the ball-pass frequency and its harmonics from spectra, and a data augmentation strategy, the proposed approach aims to decouple the fault signature from the system transfer function. Trained on a single run-to-failure dataset of one N209 ECP bearing with automated ground truth sizing for labels, the network acts on a regression target to learn the mapping between spectral features and the defect size. Preliminary results suggest that this proof-of-concept framework shows promise for generalization to unseen speed combinations and synthetic transfer function profiles within the scope of the studied experiment.

Evaluation of Input Presentation in Transfer Learning for Bearing Fault Detection

Amirhossein Berenji — 2026-07-03

Transfer learning is a promising technique to overcome data insufficiency, a noticeable barrier to real-world application of intelligent maintenance solutions. Although the importance of data preparation and preprocessing is widely acknowledged, no study in particular has investigated the effect of vibration data input presentations on transferability capabilities. This study aims to fill in this gap by conducting experiments across three benchmark datasets to evaluate direct transfer, catastrophic forgetting and data efficiency for bearing fault classification. Moreover, we explore the opportunity to employ source model pseudo-labeling to reduce the need for data labeled by human experts. Our findings show that not only does the choice of preprocessing pipeline significantly affect target-set performance, but also that the vulnerability to catastrophic forgetting varies accordingly. Thus, we conclude that finding the right data processing routine is also a key component in achieving supreme transfer learning performance and, indeed, it deserves more attention. The code base of this study is open-sourced and made publicly available to support reproducibility, transparency, and further research.

Evaluation of Remaining Useful Life Prediction Algorithms in the Absence of Run-to-Failure Ground Truth Data

Indranil Roychoudhury — 2026-07-03

Accurate evaluation of Remaining Useful Life (RUL) prediction algorithms is fundamental to the deployment of Prognostics and Health Management solutions. However, for critical industrial assets with extended operational lifespans, run-to-failure ground truth data is typically not available. Preventive maintenance intentionally precludes failure events, creating a fundamental challenge of assessing prognostic accuracy without observing actual end-of-life. This paper presents an algorithm-agnostic framework for the continuous online evaluation of RUL predictions in the absence of run-to-failure data. The innovation is a retrospective methodology that treats the asset’s current sensor state as a pseudo ground truth, enabling the evaluation of whether past predictions correctly anticipated the trajectory leading to the present condition. The framework includes two evaluation modes: (1) Measurement-based evaluation that assesses past sensor forecast accuracy against current observations, and (2) RUL-based evaluation that treats the current sensor value as a virtual degradation threshold and evaluates whether past RUL estimates correctly predicted the time to reach the present condition. The RUL-based evaluation adapts the well-established α–λ accuracy framework (Saxena, Celaya, et al., 2008) by replacing the unknown end-of-life with the current time as a pseudo ground truth reference, enabling continuous online assessment without failure observations. Individual prediction verdicts are aggregated using configurable weighting schemes into a single Service-Level Indicator suitable for performance monitoring. Experimental results across several industrial systems demonstrate the framework’s generalizability across diverse degradation mechanisms, sensor modalities, and prediction algorithms. The framework requires only historical sensor measurements and RUL predictions at different times.

Evaluating the Impact of Data Partitioning and Client Selection on Federated Remaining Useful Life Prediction in Aviation

Faruk Ozdemir — 2026-07-03

Federated Learning (FL) is increasingly explored for Remaining Useful Life (RUL) prediction in aviation, motivated by the distributed nature of operational data across operators and platforms, the need to learn from heterogeneous fleet conditions, and the requirement to preserve data ownership and intellectual property by avoiding raw data sharing. While existing studies report promising results, they rely on subjectively defined benchmarking setups, where non-Independent and Identically Distributed (non-IID) data partitioning, client selection, and comparison criteria are selected without systematic examination of the bias they may introduce. Consequently, it remains unclear whether reported performance differences arise from the learning method itself or from unexamined configuration choices. This paper investigates the bias induced by data partitioning and client selection configurations in FL for aviation RUL prediction. Representative heterogeneity and client selection scenarios, including operating-condition shift, are evaluated under systematic learning settings to isolate their effect on model outcomes. The results show that both partitioning and selection choices can materially influence reported performance independent of the underlying model, demonstrating that selection bias alone can alter fleet-level RUL estimates. These findings highlight the need for harmonised and well-grounded benchmarking practices to support objective comparison and credible evaluation of FL approaches in aviation applications.

Failure-Mode-Informed Development of Remaining Useful Life Prognostics

Kiavash Fathi — 2026-07-03

Various environmental and operating conditions affect the degradation behavior of physical assets, leading to various degradation trajectories and ultimately to distinct failure modes. To obtain accurate Remaining Useful Life (RUL) prediction, it is important to distinguish between such degradation trajectories and their associated failure modes. In this paper, we develop a framework where we analyze the latent space of autoencoders using spectral clustering to evaluate the similarity in degradation trajectories and failure modes in training datasets. This failure-mode-informed training sets are then used to develop failure-specific regressors for RUL prediction. On one hand, this reduces the amount of data needed to effectively train prognostics models. In addition, the accuracy of the RUL predictions is further improved. We demonstrate this using the C-MAPSS dataset, which provides fleet-based run-to-failure sequences under varying operating conditions and failure modes. We argue that latent information about different degradation mechanisms can be inferred from sensor readings, enabling the construction of failure-mode-specific RUL regressors. Our results show that this failure-mode-informed data separation reduces the amount of training data needed to generate RUL prognostics by up to 55%, while simultaneously improving prognostics accuracy - the Root Mean Square Error (RMSE) is reduced by 3%.

Fault Diagnosis of Harmonic Reducers Used in Industrial Collaborative Robots via DWT, MB-FFT, EMD and Machine Learning Classifiers (RF, XGBoost, SVM and KNN)

Samuel Ayankoso — 2026-07-03

Harmonic reducers are critical components in industrial collaborative robot joints but are prone to faults because they operate under cyclic motion and fluctuating load conditions. This work focuses on three representative failure modes in harmonic reducers: gear tooth of the flexspline breakage, flexible bearing outer race defects, and wear at the flexspline–circular spline interface. To enhance weak fault signatures, three signal preprocessing schemes are evaluated: Discrete Wavelet Transform (DWT), Multiband Fast Fourier Transform (MB-FFT), and Empirical Mode Decomposition (EMD), followed by the extraction of 11 time-frequency domain features from each processed signal set. The resulting features are used to train four classical Machine Learning (ML) classifiers, Random Forest, Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM) and K-Nearest Neighbors (K-NN). Experimental results show that the proposed MB-FFT technique provides the lowest computational cost while delivering superior classification performance, achieving 100% accuracy when combined with random forest and XGBoost for both vibration and current signals. Compared with deep learning models, the results demonstrate that signal enhancement can significantly improve classification performance despite weak fault characteristics, and that current signals can serve as effective indicators for harmonic drive fault diagnosis in cobots.

Frequency-Domain Feature Analysis for Early Gear Damage Detection in Planetary Gearboxes

Lisa Binanzer — 2026-07-03

The earliest possible detection of pitting damage in gearboxes is a central objective of vibration-based condition monitoring. Machine learning enables the automated analysis of vibration signals, but reliable detection of very early pitting damage requires a detailed understanding of which frequency ranges and frequency resolutions contain damage-relevant information. This work applies machine learning as a data-driven analysis tool to systematically quantify the relevance of frequency-based vibration features for pitting damage detection and pitting size classification. The investigations are based on experiments with three identical single-stage planetary gearboxes and four defined pitting sizes ranging from 0.5 % to 4 %. The measured time signals are transformed into the frequency domain using the Fast Fourier Transform, and the amplitudes of individual frequency bins are used as features. Since the bin width depends on the selected segment length, the influence of frequency resolution on the identification of damage-relevant features is also analyzed. A tree-based gradient boosting algorithm is used for classification, and the importance of individual frequency features is quantified by permutation analysis. The evaluation follows a two-stage approach. First, healthy and damaged states are compared to identify generally relevant frequency features. Second, the healthy state is contrasted separately with each pitting size to determine when specific features become relevant and how their importance changes with increasing damage size. In addition, feature consistency across operating conditions, sensor positions, and the three identical gearboxes is examined. The results support the targeted selection of frequency-based features for subsequent machine-learning-based damage detection and damage size classification and provide guidance for sensor placement, frequency resolution, and measurement system design in application-oriented condition monitoring.

From Machine Health to Elderly Health: A Sustainability-oriented Elderly-centric Social Robot System enabled by PHM

Deepti Mishra — 2026-07-03

The purpose of designing machines is to serve humans and assist them in completing their work on time. A machine, such as a social robot, can be particularly supportive for older individuals who may have reduced strength and be more susceptible to various health issues. The present paper proposes a scenario in which elderly health is monitored by a social robot. As an observer, a social robot, through perceiving its environment, can capture several moments (including daily activities or falls) of an older adult and can provide timely reminders and alerts to the elderly and their caregivers, ensuring the elderly's well-being. Therefore, validation of the machine's health (functioning) is necessary. To accomplish this, the current paper suggests utilization of Prognosis and Health Management (PHM) for the elderly-centric robot system. Furthermore, a PHM-enabled elderly-centric robot system has two main entities: an older adult and a robot, hence it is important to analyze the sustainability dimensions from various perspectives. There are two main objectives of this paper: (i) to develop a comprehensive list of sustainability topics under various dimensions achieved from the examination of three sustainable frameworks: the Triple Bottom Line, Responsible Research and Innovation, Sustainable Assessment and Sustainability Awareness Framework (SuSAF). (ii) to apply the comprehensive list of sustainability dimensions to the proposed case of PHM-enabled elderly-centric social robot system. The results suggest that SuSAF is the most comprehensive and suitable framework for the sustainability assessment of the proposed system. Furthermore, the use of sustainable dimensions can ensure improved robot health and, hence, the health of the elderly.

From MCU to Neuromorphic Chip: A Zero-Gradient Spiking-Compatible Engine for Cross-Domain Predictive Maintenance

Jianwei Lou — 2026-07-03

Real-time predictive maintenance on resource-constrained edge hardware demands sub-millisecond inference, online adaptation without retraining, and reliable generalization across heterogeneous industrial domains. We present a zero-gradient neural dynamics engine: a population of computational units governed entirely by local plasticity rules, discrete population-level gating, and resource-constrained structural adaptation. The frozen model occupies 50–100 KB with no GPU dependency. We evaluate nine task configurations across five physical datasets (bearing fault diagnosis on CWRU and Paderborn, satellite telemetry anomaly detection on SMAP, bridge structural health monitoring on Z24, audio machine monitoring on DCASE, human activity recognition on UCI HAR, and cross-domain transfer), plus one real-world tunnel construction monitoring deployment, using a single unmodified engine configuration. The engine achieves a mean improvement of +22.6 percentage points over raw-feature baselines across all nine configurations; on the four bearing fault tasks, multi-seed consensus voting further pushes accuracy to ≥99.9%. End-to-end inference latency is 53 µs on x86, 103 µs on ARM (Raspberry Pi 5), and∼350 µs on a Cortex-M7 MCU at 600 MHz. All engine primitives operate without surrogate gradients or backpropagation, and exact SNN mapping to a neuromorphic chip has been verified. Engine routing is fully unsupervised: online k-means discovers K regimes from the input stream without consuming labels. A one-time calibration step of 50–200 labeled samples, supplied either by a domain expert or by a cloud-based LLM oracle, then maps these discovered regimes to domain-specific fault categories; thereafter the engine runs autonomously.

From Sensor Data to Maintenance Actions: An Industrial PHM Application for UltrasonicWelding Assembly Machines

Francesco Cancelliere — 2026-07-03

Ultrasonic welding machines are widely used in high precision manufacturing processes, where progressive component degradation can lead to quality losses, increased reject rates, and unplanned downtime. Although modern machines generate large volumes of high-frequency process data, such as welding time, amplitude, and pressure, deploying effective predictive maintenance solutions remains challenging due to strong process variability, the lack of explicit failure labels, and the absence of historical run-to-failure datasets. This paper presents an industrial Prognostics and Health Management (PHM) framework developed for a multi-station ultrasonic welding machine used in pharmaceutical assembly, with the objective of enabling early detection of performance degradation and supporting predictive maintenance decisions. The proposed approach focuses on the construction of an interpretable, data-driven healthindicator at component level, derived from sensor data and explicitly designed to be understandable by machine experts. Domain knowledge provided by the machine manufacturer is integrated to interpret the health indicator evolution and to translate detected degradation patterns into concrete maintenance recommendations. A healthy reference behavior is established using data from machines operating under stable conditions, enabling relative deviation analysis and trend-based monitoring across heterogeneous stations. The framework was deployed in an industrial pilot and demonstrated the ability to identify abnormal behaviors associated with component wear and process sensitivity, including cases where conventional maintenance actions showed limited effectiveness. The results indicate that the proposed indicators can reveal degradation patterns earlier than traditional reject-rate monitoring, thereby supporting maintenance prioritization at component level. This work illustrates how interpretable, data-driven PHM methodologies, co-designed with machine experts, can be successfully integrated into real manufacturing environments, bridging the gap between raw process data and actionable maintenance insights.

Hybrid As-Operated Digital Twin of an Aircraft Brake

Vinayak Chandran — 2026-07-03

This paper proposes a Digital Twin-based approach to predict friction wear of an aircraft braking system between required brake overhauls, based on individual aircraft operating conditions. The classification of Condition-Based Maintenance (CBM) as a Type III approach in SAE ARP6887 introduces significant challenges for purely statistical or data-driven algorithms. Models must be component-specific and sensitive to varying operational conditions, leading to domain shift, data sparsity, and limited generalization across fleets. Ensuring robustness, interpretability, and certifiability under these constraints are open research challenge. Conventional data analytics approaches are of limited use in scenarios characterized by a lack of run-to-failure data. On the other hand, the usage of model-based approaches can be limited due to complex physics modeling. Hybrid technologies are a more recent research area in the field of CBM, integrating prior physical knowledge with Machine Learning (ML) solutions. In this context, the As-Operated Digital Twin paradigm has emerged as a promising framework, representing the virtual counterpart of physical assets to reflect actual in-service condition and usage history. It can provide several benefits in the field of CBM, as it provides real-time insights into the health and degradation status of the monitored component and reduces the common challenges of a lack of run-to-failure data and a lack of collocated sensors with respect to the source of degradation. This article proposes a Hybrid As-Operated Digital Twin architecture, relying on the Archard equation to model the degradation phenomenon. The unknown degradation parameters of Archard’s law are usually modeled with empirical equations based on laboratory data, limiting the validity of the model to the laboratory domain. To improve model generalization and reduce the number of experiments to fit the degradation parameters, a physics-informed Recurrent Neural Network (RNN) model is proposed. The proposed Hybrid As-Operated Digital Twin includes the physics-informed RNN model and the heat-sink thermal model of the brake system. The robustness of the model is studied, employing Monte Carlo methods for uncertainty quantification. The discussed methodology is demonstrated on a sample aircraft brake model with laboratory-domain data.

Hybrid Detection for Heat Pump Contamination Using Physics-Informed Machine Learning

Ahmed Qarqour — 2026-07-03

The deployment of residential heat pump systems is a key enabler of the decarbonization of the heating sector. However, their long-term reliability remains a barrier to sustained performance and user acceptance. A major degradation driver is water contamination within the hydraulic circuit, which leads to fouling, scaling, and corrosion of components such as plate heat exchangers – ultimately reducing efficiency and shortening system lifetime. Although installation procedures and operational filtration measures, including magnetite filtration, aim to reduce particle accumulation, continuous condition-based monitoring of component degradation remains limited. To address the scarcity of real-world failure data for training predictive models, this paper proposes a physics-informed, data-prior approach that combines physical knowledge with machine learning. Instead of embedding physics into the model architecture or loss functions, the approach incorporates it at the data level by generating labeled healthy and faulty scenarios through a physics-based laboratory setup. This enables the model to learn degradation patterns grounded in physical behavior, supporting early fault detection and producing outputs that remain interpretable and plausible for domain experts. The approach is demonstrated on a plate heat exchanger contamination use case. A design-of-experiments campaign in a climate chamber generated labeled data representing healthy, moderately contaminated, and severely contaminated states. A Random Forest classifier achieved consistent cross-validation performance (AUC ≈ 0.98) with low variance across folds. Precision–recall analysis revealed robust early fault detection, with average precision values of approximately 0.96 for moderate contamination and 0.97 for severe contamination. Cumulative gain and lift analyses indicated that inspecting the top 20–40 % of systems ranked by model risk can identify 80–100 % of the contaminated cases, supporting efficient maintenance prioritization. Model-derived feature importance was assessed using Gini importance and subsequently validated through expert review, enabling interpretable failure logic for condition-based maintenance strategies. The results demonstrate that combining physically grounded data, supervised machine learning, and explainable diagnostics provides a transferable hybrid approach for interpretable reliability assessment and condition-based monitoring beyond the specific case.

Impact of Reactive Power Assumptions on Physics-Informed Assessment of Transformer Ageing in Distribution Networks

Junyi Lu — 2026-07-03

Power distribution networks are undergoing a fundamental shift in their utilisation driven by the rapid increase in the prevalence of residential photovoltaic (PV) systems, coupled with the increasing penetration of inverter-based household loads (such as heat pumps), changing expected reactive power flows. However, many Distribution System Operators (DSOs) have limited reactive power monitoring and instead rely on historical constant power factor assumptions, masking the true transformer current and thermal stress, resulting in biased lifetime estimates. This paper proposes an assessment methodology that integrates a low-voltage distribution network model with the physics-based IEEE C57.91 transformer thermal-ageing model to quantify transformer loss of life. The methodology evaluates different LCT penetration levels using representative load and reactive-power profiles. The methodology is demonstrated using a spatially coherent real-world distribution network model, derived from geographical network data, with PV generation and residential load profiles. We conduct a comparative analysis between the standard DSO assumption (constant power factor) and the actual reactive power profile. The results indicate that the assumption of a constant power factor of 0.95 significantly over/underestimates total current and thermal stress, resulting in biased transformer health assessments. The study demonstrates that a revised reactive power assumption is an important consideration for transformer health assessment and asset-management decisions.

Improved State of Health Assessment for Lithium-Ion Batteries Based on Electrochemical Impedance Spectroscopy Measurements

Gabriele Patrizi — 2026-07-03

Conventionally, battery State of Health (SOH) is defined through the measurement of discharge capacity. However, such approaches are poorly suited for online and in-vehicle applications, as they require full charge/discharge cycles and accurate current integration over long periods. For this reason, indirect health indicators are widely adopted, especially in automotive PHM frameworks. In this context, Electrochemical Impedance Spectroscopy (EIS) has proven to be an effective tool for investigating battery degradation, as it provides detailed insight into internal electrochemical processes. Nevertheless, EIS measurements are strongly influenced by the operating conditions of the tested device, which reduces their practical value under high or variable stress levels. To address these limitations, this work proposes a robust procedure to extract a one-dimensional Health Indicator (HI) from EIS measurements performed after the electric vehicle (EV) charge phase. Instead of relying on full-spectrum fitting or equivalent circuit modeling which are often computationally intensive and difficult to implement online, the proposed method extracts multiple physically meaningful geometrical features directly from Nyquist plots. The features are normalized and evaluated through an adaptive selection process that identifies the most informative ones for degradation tracking and prognostics, ensuring robustness under varying stress conditions. The selected features are then combined through an innovative algorithm to generate a single HI that accurately reflects the battery degradation trend, enabling integration into automotive Battery Management Systems (BMS). To ensure generalizability and repeatability, the approach is validated at different processing stages using a custom dataset of lithium-ion cells tested under highly stressful charge and discharge conditions for a total of 400 cycles.

Integrating Deep Autoencoders and Bayesian Inference for Diagnostics and Health Management of Industrial Inkjet Systems

Gianluca Nicchiotti — 2026-07-03

Modern printing industry requires extreme reliability to achieve zero-defect production. This study contributes to the paradigm shift from reactive to predictive maintenance in industrial inkjet systems, reducing downtime, material waste, and operational costs. We present an integrated Diagnostics and Health Monitoring strategy leveraging piezoelectric self-sensing with deep autoencoders and Bayesian inference for real-time fault diagnostics. The system captures residual pressure waves in the ink chamber—exploiting the dual actuator-sensor function of piezoelectric crystals—as "acoustic signatures" to detect subtle nozzle deviations from clogging, air bubbles, or mechanical wear. Nozzle pressure signals feed into a multimodal autoencoder (AE) architecture, where each AE specializes in a distinct fault class (jetting, non-jetting, deviated, intermittent). AE outputs combine with Gaussian Mixture Models (GMM) and Bayesian inference to provide high-confidence classification, even with imbalanced industrial datasets. Tests on industrial printheads (Ricoh MH5420) demonstrate 99.4% accuracy in detecting critical failures and enabling preventive maintenance. However, while the system excels at detecting fluidic obstructions, challenges remain in classifying deviated and intermittent faults. The prognostic layer reuses the jetting autoencoder's reconstruction error (RE) as a continuous Health Indicator, correlating pressure-induced degradation with nozzle health. Controlled experiments varying ink chamber pressure reveal a parabolic RE-pressure relationship, with minimum RE at nominal operating range. This enables early degradation detection, as RE increases progressively before functional failure, supporting condition-based maintenance strategies.

Interpretable Operational Regime Classification for a Wind Turbine PHM Digital Twin Architecture

Gard Indrekvam — 2026-07-03

Wind turbines operate under varying environmental and control conditions that make reliable interpretation of SCADA data challenging. Without contextual information about the current operating state, normal variability may be mistaken for abnormal behaviour, reducing the reliability of diagnostic and prognostic analysis. This study develops a Digital Twin framework for wind turbines using SCADA data, where periodically updated models represent turbine structure, behaviour, and operating context. To address this, physics-informed operational zones are defined based on power curve characteristics and control logic, providing structured labels. These are used to train interpretable rule-based algorithms, Repeated Incremental Pruning to Produce Error Reduction (RIPPER) and First-Order Inductive Learner (FOIL), which generate explicit IF-THEN rules linking measured variables to operational states. Evaluation using overall accuracy, macro-F1 score, and per-class precision and recall shows that both methods achieve classification while producing compact, physically interpretable rule sets aligned with known turbine behaviour. The study demonstrates that rule-based learning enables transparent and effective operational regime classification, forming a critical contextual layer for prognostics and health management (PHM) oriented Digital Twins, with applicability beyond wind turbines to other industrial assets.

Integrating Survival-Based Aging Models with Data-Driven RUL Prognostics

Abishek Srinivasan — 2026-07-03

Predictive maintenance requires reliable remaining useful life (RUL) estimation. Existing methods mainly follow two paradigms: wear-based aging models that capture cumulative degradation and sensor-driven data models that reflect instantaneous health conditions, each providing only partial information. In this work, we propose a probabilistic fusion framework that integrates wear-based and sensor-based prognostic components through failure probability distributions. Based on explicit structural assumptions linking wear, latent health, sensor observations, and failure, we derive a principled combination rule that enables uncertainty-aware integration with adaptive weighting of the components. Experimentally, we assess this combination rule by learning the wear-based component using a parametric survival model and the sensor-based component using a 1D convolutional neural network (1D-CNN) with a post-hoc uncertainty model. Evaluation on multiple N-CMAPSS datasets demonstrates that the fused model improves point accuracy, preserves the C-index, and produces narrower yet well-calibrated prediction intervals compared to either component alone. The results highlight the complementary roles of wear-based survival model and sensor-based deep learning model, and show that their probabilistic integration provides a structured pathway toward more robust and consistent prognostics over the life-time.

Learning the Language of Vibration: A Self-Supervised Transformer Foundation Model for PHM

Giuseppe Mannone — 2026-07-03

Industrial prognostics and health management (PHM) increasingly relies on vibration-based deep learning, yet field deployment remains limited by two practical constraints: (i) fault labels are scarce and expensive, and (ii) models trained on one machine or dataset often degrade under distribution (domain) shift (different sensors, sampling rates, loads, and signal conventions). These constraints motivate vibration foundation models: reusable encoders trained once on large collections of unlabeled raw vibration recordings and adapted to new assets with minimal supervision. This paper presents VibFM, a Transformer encoder trained via self-supervised masked spectrogram modeling in the spirit of masked language modeling and masked autoencoders. Raw waveforms from 16 open datasets totaling ≈ 400 hours are standardized into 128 × 128 log-magnitude short-time Fourier transform (STFT) spectrograms and paired with a compact conditioning vector that encodes sampling rate and time/frequency resolution. Pre-training reconstructs masked time–frequency patches, encouraging the encoder to capture transferable vibration primitives such as persistent narrowband ridges, modulation sidebands, and impulsive transients. Transfer is evaluated on the held-out Paderborn University and KAt DataCenter bearing benchmark (excluded from pre-training) using leakage-resistant bearing-level splits. On three-class fault diagnosis, frozen VibFM features substantially improve over training from scratch, while end-to-end fine-tuning provides the strongest performance. For reconstruction-based anomaly detection, adapting a decoder on healthy target data yields reconstruction-error scores that separate healthy from damaged states across operating conditions. Masked reconstructions and pooling-attention visualizations provide qualitative audits of learned time–frequency structure, and the limits of these interpretability probes are discussed.

Leveraging Time Series Foundation Models Embeddings for Remaining Useful Life Prediction

Ilias ABDOUNI — 2026-07-03

Recent advances in Remaining Useful Life (RUL) prediction rely heavily on task-specific deep learning architectures, such as CNNs, LSTMs, and Transformers. While effective, these data-intensive models frequently struggle to generalize across varying operating conditions. Time Series Foundation Models (TSFMs) offer a promising zero-shot alternative to this training-from-scratch approach, yet directly applying their forecasting-optimized representations to prognostics often fails to capture the physical constraints of equipment degradation. To resolve this task-objective mismatch, we propose a domain-agnostic adapter architecture that applies the Wide & Deep learning paradigm to repurpose the frozen Chronos-2 foundation model for continuous RUL regression. Our methodology explicitly bridges the domain gap by extracting and flattening the model’s abstract multivariate embeddings (Deep), and fusing them with raw, normalized physical measurements (Wide). Experiments on the full C-MAPSS benchmark demonstrate that this approach achieves state-of-the-art performance, reaching an average RMSE of 10.32. By outperforming recent specialized architectures without requiring backbone fine-tuning, this work proves that lightweight adaptation of generalized temporal representations offers a scalable, robust alternative to traditional prognostic modeling.

Leakage-Safe, Reproducible Benchmarking for Vibration-Based Fault Diagnosis

Pawel Knap — 2026-07-03

Vibration-based bearing fault diagnosis is a widely studied predictive maintenance problem, but reported results are often difficult to compare. Performance depends not only on the model itself, but also on the evaluation protocol, the train--test split, and the type of domain shift considered. In particular, leakage-prone window-level splitting and loosely defined source--target settings can lead to overly optimistic conclusions that do not reflect real transfer performance across changing operating conditions, acquisition regimes, or bearing identities. To address this issue, this paper introduces a leakage-safe and reproducible benchmark for cross-domain bearing fault diagnosis on the Case Western Reserve University and Paderborn University datasets. The benchmark defines six fixed source--target scenarios, enforces recording-level train--test separation, and evaluates both machine-learning and deep-learning baselines under a common protocol. Final reporting is based on a consistent evaluation setup, with repeated-seed follow-up used where necessary to support reliable conclusions for deep-learning models. The results show that scenario difficulty is highly heterogeneous. Some transfer settings are effectively saturated, while others remain substantially more challenging. Deep-learning models often achieve stronger performance, but their conclusions can be sensitive to initialization and require repeated-seed validation. Overall, the benchmark provides a reproducible basis for scenario-level evaluation and more reliable comparison of cross-domain bearing diagnosis methods. The code for this study is publicly available at https://github.com/1Sensor/pdm-bench.

Load-Aware Stochastic Degradation Modeling and Lifetime Characterization for PEM Fuel Cells

Mouhamad Houjayrie — 2026-07-03

The degradation of proton exchange membrane fuel cells (PEMFCs) is strongly dependent on load history and presents significant variability across nominally identical stacks. This paper proposes a physics-based, load-aware stochastic degradation framework for PEMFCs, with the objective of characterizing lifetime under static and dynamic operating profiles and providing a modeling basis for future Prognostics and Health Management (PHM) applications. The cell voltage is described through a polarization model parameterized by degradation-sensitive quantities, namely the normalized electrochemically active surface area (ECSA), the membrane ohmic resistance, and the hydrogen crossover current. Catalyst degradation is represented by a simplified ECSA state driven by platinum dissolution--oxidation kinetics, while membrane ageing is described through a stochastic cumulative damage state modeled as a non-homogeneous Gamma process whose mean evolution is matched to a semi-empirical membrane degradation law. Membrane thickness and conductivity are reconstructed consistently from this damage state, and the resulting ohmic resistance and crossover current are fed back into the voltage model. Load dependence is enforced through the coupling between mission demand, operating-point computation, and voltage-driven degradation dynamics. The resulting framework captures both intra-stack stochasticity, through the membrane damage process, and inter-stack variability, through dispersion in selected model parameters. A Health Index (HI) is defined as the normalized virtual rated-point voltage. Simulation studies under static and dynamic load profiles illustrate the influence of load level, load cycling, and parameter variability on degradation trajectories and lifetime distributions. Although state estimation and remaining useful life prediction are not addressed here, the proposed framework is intended to serve as a compact modeling basis for such future PHM developments.

Lubrication Condition Assessment of Planetary Roller Screw under Varying Operating Parameters

Zhichao Dong — 2026-07-03

The planetary roller screw mechanism (PRSM) is a critical transmission component in electromechanical actuators. Under prolonged operation in harsh service conditions, PRSMs are prone to lubrication deficiencies, which significantly increase the risk of mechanical failure. Consequently, it is essential to develop evaluation techniques for the lubrication condition of PRSMs to enable efficient health management. This study proposes an assessment method for PRSM lubrication status during operation based on a vibration mechanism model. The model integrates the contact characteristics of the threads and operating parameters to characterize the influence of internal lubrication conditions on the system's vibration response. Based on this model, a dimensionless evaluation metric was developed that depends on the amount of lubricant used. Using a dedicated experimental test rig, vibration signals were collected under varying screw rotational speeds, axial loads, and lubricant quantities to validate the method. The experimental results demonstrate the effectiveness of the proposed method in evaluating the lubrication state under varying operating conditions using a small amount of vibration data.

Maintenance Planning for Prognostic Health Management of Multi-Component Systems: A Case Study of Multi-Objective Optimisation for Railway Vehicles

Bernd Wagner — 2026-07-03

Condition-based maintenance (CBM) offers an opportunity to improve classical preventive maintenance strategies by transitioning to a dedicated prognostics & health management (PHM) strategy for individual components. This means that time-based fixed maintenance intervals are being replaced by condition-dependent and component-specific interventions.

For multi-component, complex systems, such a component-condition-based maintenance strategy yields multiple remaining useful life (RUL) prognostics. Maintainers must combine these into optimal maintenance plans that account for reliability, availability, maintainability, safety and costs (RAMS-C). Combining multiple RUL prognostics with the system's real-world dynamics, including architecture, anticipated future operating conditions, economic interdependence, stochastic interdependence and structural interdependence, yields complex maintenance planning decisions. To address this problem, several studies propose solutions based on multi-objective optimisation algorithms. However, these are often based on a priori knowledge and do not account for the day-to-day dynamics of operational variations. This paper presents a methodology for managing the complexity of maintenance scheduling under these conditions. The model is applied to the maintenance history of a Voith Maxima 30CC cargo locomotive. A priori maintenance models are compared with real-life implementation and optimised models using evolutionary multi-objective optimisation (EMO) to assess the rigour of the a priori plans under varying operating conditions. The study shows that applying such algorithms can significantly reduce costs while increasing overall system reliability and availability.

Manufacturing Quality–Informed Prognostics: A Novel Approach to Past Uncertainty Management

Benjamin Brito Schiele — 2026-07-03

The future behavior of a system is largely determined by its manufacturing process, as variations in production quality can lead to different performance outcomes over time. In the context of prognostics, all sources of uncertainty that exist prior to the system’s commencement of operation are collectively referred to as past uncertainty, yet its role is rarely recognised in the prognostics and health management (PHM) community. Most existing approaches to uncertainty and its management focus only on model parameters, leaving the management of past uncertainty largely unexplored. This work introduces a framework to explicitly manage this source by incorporating manufacturing quality control (MQC) data into hidden semi‑Markov model (HSMM) prognostics. The method creates quality‑specific HSMMs, each tailored to a particular manufacturing quality (MQ) type, and combines them during inference using MQC‑informed Bayesian model averaging.
The framework is validated on composite specimens with pristine, oil‑induced, and Teflon‑induced defects. Ultrasonic scans provide MQC inputs, while strain data describes degradation. Results show that accounting for MQ reduces uncertainty in RUL predictions and highlights the importance of correctly identifying the MQ type for effective uncertainty management in prognostics.

Meets Expectations: System Health Analysis and Prognosis for Embedded and Cyber-physical AI

Michael Borth — 2026-07-03

The health management of AI-based systems differs strongly from that of traditional systems, forcing a rethinking and partial reinvention of established techniques: The AI inside provides the capacity to adapt to an operational context – and that leads to a variability in system behaviors that renders conventional health and performance indicators possibly obsolete. Also, even core AI functionality, like perception, can hardly be assessed without complicated reasoning about whether a lack of performance is circumstantial or an actual system health issue. For these reasons, we introduce a system health monitoring methodology that checks health and key performance indicators against expectations while factoring in mitigating circumstances, like environmental effects.

This methodology, which is based on probabilistic reasoning, allows the detection of system health degradation and root-cause analytics and is used by us to ensure the operational fitness of safety-critical systems, i.e., Automated Vehicles. As this domain is subject to temporal changes like seasons that impact a system’s performance more than many developing health issues, we combine health monitors with domain monitors and drift detection. Overall, this provides probabilistic health management that looks at expectations, sets of observations, their distributions and their dynamics to determine whether an embedded AI is still fit for its purpose, whether the cyber-physical system embodying it continues to meet the AI’s operational requirements, and whether observations indicate a health or fitness trend that will result in a lack of safety. A key aspect of this novel approach to reasoning about system health is that it addresses unique properties of AI-based systems: It works with the hit-or-miss behavior of AI that occasionally fails even on seemingly comparable inputs, and it can investigate adaptive processes by looking at the health of information flows that define the decision-making of AI-based systems.

Monitoring Model Drift in Feedback-Controlled Systems via Efficacy and Efficiency Metrics

Leonardo Baldo — 2026-07-03

This work explores the possibility of monitoring model drift inside a control loop by leveraging efficacy and efficiency metrics. The methodology is based on a fundamental energy–error relationship that links tracking performance to the control effort required to sustain it.
The proposed strategy is explicitly designed to deal with feedback-controlled systems and relies only on basic loop signals and simple indicators on time windows: a control-energy metric and an error-performance metric. Based on the similarity between model drift and fault detection problems, the method enables early detection and tracking of progressive faults or drifts that would otherwise remain hidden. The monitoring relies on residuals obtained by comparing the joint evolution of the two metrics against a nominal-condition baseline model. The approach is demonstrated through a MATLAB/ Simulink simulations across various degradation scenarios, showing consistent sensitivity to drifts and enabling their timely detection well before loss of performance becomes apparent at the output level. These results support a lightweight, explainable pathway for model drift monitoring in feedback-controlled systems without requiring additional sensors, the development of complex machine learning/high-fidelity physics models, or structural modifications to the control architecture.

Ontology-Based Graph Transformer Network for Robust Bearing Fault Diagnosis under Unseen Operating Conditions

Yong Hun Park — 2026-07-03

Conventional data-driven diagnostic models for rotating machinery often exhibit limited generalization under varying operating conditions. Signal-based approaches typically rely on condition-dependent training data and statistical feature learning, making them vulnerable to domain shifts and complex fault scenarios. In particular, frequency-domain features are commonly treated as unstructured statistical descriptors, without capturing the physical relationships among characteristic frequencies such as harmonic and sideband structures intrinsically linked to bearing fault mechanisms. Consequently, the learned representations tend to overfit to the training distribution, leading to pronounced performance degradation under unseen operating conditions. These limitations highlight the need for diagnostic frameworks that incorporate physically grounded relational structures to achieve robust generalization.

To address these challenges, this study proposes OFG-GTN (Ontology-based Frequency Graph–GTN), an ontology-based bearing fault diagnosis framework that integrates structured frequency representation and relational graph learning. Vibration signals are transformed into ontology-based frequency graphs by encoding physically defined characteristic frequencies, harmonic relations, and sideband dependencies derived from bearing fault mechanics. A GTN is employed to perform graph-level fault classification by capturing higher-order relational dependencies among frequency components. Experimental results demonstrate that OFG-GTN achieves robust and generalizable bearing fault diagnosis across diverse operating conditions, including those not encountered during training.

Ontology-Grounded Large Language Models for Reliable Querying of Wind Turbine Inspection Knowledge

Louis Verstraeten — 2026-07-03

Inspection reports of industrial assets contain valuable diagnostic knowledge, but their unstructured nature makes automated reasoning difficult. This paper presents an ontology-grounded question answering framework for querying wind turbine gearbox inspection reports using natural language. Inspection data are automatically parsed into structured representations consisting of a domain ontology and a knowledge graph. On top of this representation, a large language model translates user questions into SPARQL queries. To improve robustness, we employ example-based query generation combined with an Ontology-Based Query Checker (OBQC) that validates generated queries against ontology constraints and iteratively repairs violations before execution. The approach is evaluated on real-world inspection reports using 50 diagnostic prompts of varying complexity, achieving a 96\% successful execution rate. Results demonstrate that combining ontology grounding with constrained LLM-based query generation enables reliable and flexible diagnostic reasoning over inspection documentation.

Particle Filter-based Degradation Modeling and SOH Prediction for Lithium-ion Batteries of Autonomous Systems

Donghoon Seo — 2026-07-03

Physics Based and Data Driven Anomaly Detection Methods Using Vibration Data for Early Gear Damage Detection in Planetary Gearboxes

Lukas Merkle — 2026-07-03

Various methods are known for using vibration sensors to distinguish anomalies such as pitting damage from normal operating conditions in gear drives. These methods are capable of detecting severe damage. However, to take full advantage of prognostics and health management (PHM) strategies, it is necessary to detect damage in a very early damage state. The objective of this work is therefore to analyze, improve and expand the known data evaluation methods in order to achieve the earliest possible detection of gear pitting damage during operation. The research question is: What is the smallest detectable pitting size in planetary gears using vibration data, and which methods are best suited for detection? A high-resolution vibration data set from a single-stage helical planetary gearbox is available for this study. The vibrations were recorded at different speeds, torque levels, and sensor positions. The evaluation methods include entirely physicsbased methods in the frequency and time domain, especially a selective analysis of characteristic frequencies and sidebands. These methods are supplemented with data driven approaches such as similarity analysis of frequency spectra. In contrast to physics-based methods, data- riven approaches aim to detect deviations of data sets regardless of their origin. To exclude false positives, these approaches inevitably require the use of multiple measurements without damage as a reference. The application of these methods results in the calculation of 130 condition indicators (CIs). This study provides statistical evidence for the detectability of small damage sizes, particularly using data-driven methods. The key findings of the study are that most of the information about the damage is contained in a comparison of full spectra. A comparison of the full spectral range provides a much clearer picture of damage compared to the analysis of individual characteristic frequencies.

Physics-Informed Domain Generalization for Bearing Prognostics Under Unseen Operating Conditions

Seyed Ali Hosseinli — 2026-07-03

Modern condition-based maintenance of rotating machinery increasingly relies on data-driven prognostic models to estimate bearing health and remaining useful life (RUL). While machine-learning approaches have demonstrated strong performance under known operating conditions, their reliability often degrades under unseen loads, speeds, and degradation regimes, limiting their industrial applicability. This work addresses bearing prognostics under unseen operating conditions through a physics-informed real-to-real transfer learning framework. To improve physical consistency and long-term prognostic stability, the proposed approach incorporates constraints inspired by fatigue crack growth theory into the learning process. In particular, monotonic degradation behavior consistent with Paris-law-type dynamics is enforced by learning bias, promoting physically plausible degradation evolution and RUL estimation. Building on this physics-guided foundation, the framework further addresses domain shifts through domain generalization and feature disentanglement. The method accounts for both marginal and conditional domain shifts across multiple source domains representing different operating conditions and degradation trajectories. By disentangling domain-invariant degradation features from domain-specific operational characteristics, the model enables zero-shot generalization to previously unseen target conditions without requiring target-domain data. The proposed method is validated using experimental bearing run-to-failure datasets and demonstrates robust prognostic performance under unseen operating conditions while maintaining physically consistent degradation behavior. The results highlight the potential of combining physics-informed learning with domain generalization for reliable industrial bearing prognostics.

Physics-Informed Machine Learning-Assisted State Estimation for Degrading System Considering Sensor Degradation Impacts

Trung Thanh Nguyen Thai — 2026-07-03

Degradation estimation is a fundamental component of prognostics, yet it is often compromised by the idealized assumption of perfect sensor fidelity. In harsh industrial environments, the concurrent degradation of both primary assets and monitoring sensors introduces severe observational ambiguity. Traditional state estimation methods, such as Kalman Filters (KF) or standard Sequential Monte Carlo (SMC), often fail to track states when sensor degradation causes the observation model to become non-stationary and complex. To address this, this paper proposes a hybrid degradation estimation framework that integrates Physics-Informed Machine Learning (PIML) into SMC inference. The joint evolution of asset and sensor degradation is modeled through stochastic Wiener processes, capturing both deterministic drift and diffusion. A learnable observation model is then implemented using a Multilayer Perceptron (MLP) to map the relationship between degradation states and measurements within a recursive Bayesian optimization framework. Numerical validation demonstrates that the proposed method achieves robust tracking accuracy under non-stationary conditions, performing competitively with model-based filtering approaches, presenting a promising approach in supporting prognostics frameworks considering sensor degradation.

Physics-Constrained Deep Learning for Interpretable Anomaly Detection in Large Battery Packs with Limited Monitoring Granularity

Antoni Plonczak — 2026-07-03

Condition monitoring of large-scale lithium-ion battery packs is often constrained by limited measurement granularity, as operational systems typically rely on aggregated pack-level signals rather than detailed cell-level data. While numerous methods for anomaly detection and health estimation have been developed for individual battery cells, pack-level modeling remains relatively underexplored. Moreover, the impact of reduced monitoring granularity on battery health assessment is still insufficiently understood.

In this work, we present PILSNet, a Physically Interpretable Latent State Network that enables the inference of physically interpretable aging indicators from current-voltage time series. The model combines a convolutional neural network with an equivalent circuit model (ECM), which acts as a constraint to enforce latent variables corresponding to internal resistance and capacity. We apply this hybrid framework to detect abnormal aging behavior in battery packs, achieving both high anomaly detection performance and interpretability of the underlying degradation processes.

Using a physics-based battery simulation framework, we conduct a systematic study of the effect of monitoring data granularity by comparing anomaly detection performance at the cell-level and under aggregated pack-level measurements. The results show that reduced monitoring granularity leads to a significant decrease in anomaly detectability, particularly under realistic scenarios with varying operating conditions and incomplete degradation trajectories. The proposed hybrid model mitigates this performance loss and consistently outperforms purely data-driven and feature engineering-based baselines, especially under constrained data conditions.

In addition to improved detection performance, the inferred latent variables provide direct insight into degradation mechanisms, enabling a pathway from anomaly detection toward fault diagnostics. Beyond battery systems, this work highlights the broader importance of systematically analyzing the relationship between monitoring design and achievable performance in complex systems.

PHM Framework for Discrete IGBTs in EV Applications: A PoF-Based Approach with a State-of-Damage Metric

JuHwan Kim — 2026-07-03

Power converters are increasingly required to deliver higher power density and longer service life, which increases the need for accurate remaining useful life (RuL) prediction of their switching devices. Although module-type switching devices dominate many applications, discrete insulated-gate bipolar transistors (IGBTs) remain commercially important and are deployed in traction inverters, making robust lifetime prediction models (LPMs) for discrete devices directly relevant to prognostics and health management (PHM).

This paper presents a Physics-of-Failure (PoF)-based LPM for discrete IGBTs that models die-attach solder degradation as the dominant wear-out mechanism. A novel State-of-Damage ( ) framework is further proposed to unify high- and low-cycle fatigue degradation within a single state variable, enabling damage accumulation under mixed-stress mission profiles. The proposed LPM and metric are experimentally validated through power cycling tests.

The proposed approach provides a practical alternative to data-driven methods, which typically require extensive failure datasets and additional sensing or data acquisition hardware that are costly to obtain and often unavailable in industrial environments. By grounding RuL estimation in the dominant degradation mechanism, the framework remains computationally efficient and data-lean while enabling reliable health assessment and RuL prediction for discrete IGBT-based power electronic systems.

Predictive Analysis for Safe and Optimal Operation of the CoBra High-Temperature Heat Pump

Jina Yoo — 2026-07-03

High-temperature heat pumps are a key technology enabling the decarbonization of conventional heating processes. The CoBra (Cottbus Brayton) high-temperature heat pump is a novel pilot plant, based on the reverse Brayton cycle, developed within DLR which achieves desired temperatures of more than 250 °C. However, during experiments, the pilot CoBra could no longer operate properly once the temperature of hydraulic oil, which is required to lubricate the compressor, exceeded the safety-relevant temperature limit. Emergency shutdowns were triggered to prevent compressor and components failure. Notably, the oil temperature rise occurs within a few minutes and shows no clear prior indication. As a consequence, the experimental objectives could not be fulfilled.
This study develops a predictive model for the compressor oil temperature based on historical experimental data such as operating pressure and compressor speed, using deep learning techniques. The model aims to provide operators with early warnings during experiments and to recommend appropriate countermeasures that can prevent system shutdowns. Furthermore, this work investigates long-term operational strategies for the CoBra under varying input conditions through dynamic simulation. The oil temperature prediction model and process simulation model are combined into a unified framework, and a dynamic optimization problem, with rolling horizon approach, is formulated to simultaneously address process performance and machinery safety constraints. IPOPT is selected as the solver for this optimization problem as a proof of concept. The proposed method enables safer and more efficient operation, improves cost efficiency and extends equipment lifetime.

Prescriptive Maintenance through Workload Allocation for Synchronizing Parallel Machines

Sacha Todoskoff — 2026-07-03

This paper addresses workload allocation for parallel machines subject to stochastic degradation in order to improve maintenance synchronization. In such systems, maintenance actions are often performed simultaneously across machines, which creates a trade off between premature maintenance and unexpected failures. Existing workload allocation strategies mainly focus on reducing failure probability or balancing degradation but do not explicitly aim at synchronizing maintenance conditions.

To address this issue, this paper proposes a framework combining stochastic degradation modeling and workload optimization. Machine degradation is modeled using a Gamma process in which the degradation rate depends on the assigned workload. A Sequential Quadratic Programming optimization is used to dynamically allocate workloads in order to maximize the probability that all machines reach the maintenance window defined by preventive and failure thresholds at the same time. Remaining Useful Life estimates are used to trigger workload reallocations and maintenance decisions.

Monte Carlo simulations compare the proposed strategy with a deterioration based workload allocation method from the literature. The results show that the proposed approach maintains a low proportion of corrective maintenance while achieving a high proportion of preventive maintenance with fewer workload reallocations.

Prognosis of Mission-Aware Remaining Useful Life for ROV Thrusters Using a Physics-Consistent Simulation Framework

George Fourlas — 2026-07-03

Remotely operated vehicles (ROVs) are now widely used to perform underwater missions that vary in duration, maneuvering intensity, and propulsive load. However, these differences, although they typically have a significant impact on how propulsion components degrade over time, are not considered by predictive models. As a result, the remaining useful life (RUL) of the thrusters is typically estimated assuming that future operating conditions will be like past ones, an assumption that is often unrealistic in mission-driven underwater operations.

In our work, mission-aware prediction of the remaining useful life of ROV thrusters is investigated, focusing on how planned mission characteristics shape the degradation progression and RUL estimation. We use a physics-consistent simulation environment representative of the BlueROV2, with applicability to related platforms such as the BlueBoat surface vehicle, in order to study this interaction in a controlled and repeatable manner where mission profiles are defined through a set of descriptors that capture thrust demand, load variability, duty cycle, and maneuvering aggressiveness, allowing for systematic comparison of different operational scenarios. Degradation is introduced by gradually modifying the electromechanical performance parameters of the thrusters, producing distinct degradation trajectories under the same initial conditions.

Health indicators derived from simulated measurements, such as motor current, rotational speed, temperature, and thrust, are used to monitor the progression of degradation, and predictive estimates are obtained by propagating the estimated health state forward across the candidate mission profiles rather than assuming a single RUL independent of the future mission. The results show that missions with similar time durations, but different propulsion command characteristics can lead to substantially different RUL predictions, even when the initial health state is the same.

The simulation framework, in addition to observing these differences, allows us to have a systematic investigation of the sensitivity of the prognosis results to the parameters of each mission. By varying the mission descriptors separately, we can determine which aspects of a mission, such as sustained high thrust versus intermittent peak loads, dominate the degradation behavior and leading to uncertainty in the RUL predictions. This type of analysis is difficult to perform only with field data, and it is quite important for understanding the limits of prognosis in operational environments. Also, these results suggest that the RUL should be considered as dependent on how a system is expected to be used, rather than as a single fixed value. Since different mission characteristics have been shown to affect degradation in different ways, considering the planned mission leads to predictive estimates that are easier to interpret and more directly linked to operational choices. In practice, the framework allows for the comparison of different mission options based on their expected impact on the thruster life and highlights missions that are likely to accelerate degradation while offering the ability to prioritize maintenance actions based on expected operational requirements.

Finally, our study through simulations offers a simple and repeatable way to investigate this behavior in underwater robotic systems, without requiring extensive real-world failure data, and can serve as a basis for future work on mission-aware PHM methods.

Physics-Informed Virtual Sensing for Isentropic Efficiency: Enabling Sensor Reduction in Heat Pumps

Savvas Eftychis — 2026-07-03

Virtual sensors are increasingly used in Industrial Internet of Things (IIoT) systems to estimate quantities that cannot be directly measured or when physical sensors are unavailable. In heat pump systems, compressor isentropic efficiency is a commonly used thermodynamic performance indicator typically computed from pressure and temperature measurements at both compressor inlet and outlet. This study presents an analysis in virtual sensing of isentropic efficiency under sensor reduction, studying the effect of physics-informed features. A comprehensive feature space was constructed from raw measurements and thermodynamic properties computed via CoolProp software. Feedforward neural networks were trained for all feasible combinations of two to four input features across multiple sensor removal scenarios. Model performance was assessed using structured data splits that allow for evaluation of generalizability from in-distribution training data to out-of-distribution unseen operating conditions. Results show that excluding the suction temperature sensor yields the most favorable trade-off between in-distribution accuracy and out-of-distribution robustness. Analysis across all sensor removal scenarios reveals that feature composition is the primary determinant of out-of-distribution performance, rather than model architecture or hyperparameter tuning. Robust feature sets consistently include discharge entropy together with suction pressure and saturation temperature, reducing out-of-distribution error by up to 70% compared with a raw-sensor baseline, at a modest cost in in-distribution accuracy.

Quantifying Reliability and Resilience in Interdependent Infrastructure Networks

Risat Rimi Chowdhury — 2026-07-03

Infrastructure networks such as power, water, and communication systems are interdependent, and disruptions in one network can propagate across others, causing performance loss and delayed recovery. Although prior studies have modeled recovery and evaluated resilience in such systems, few have examined their operational limits through reliability analysis. This study develops a simulation-based framework to quantify reliability and resilience in two partially interdependent networks while considering node failure and recovery. The critical percolation threshold is identified from the cascading failure process and used in a voting system model to compute system reliability. Results show that higher coupling strength shifts the percolation threshold to a larger value and reduces system reliability, whereas lower coupling strength allows the system to sustain connectivity under a greater degree of node failure. Resilience is evaluated through a cascading recovery process following complete disruption of one network. Recovery of a node enables restoration of its dependent counterpart in the other network, propagating sequentially across both networks. Multiple simulation scenarios are employed to represent uncertainty in the node recovery process. Results indicate that higher coupling strength leads to greater initial disruption in the dependent network but facilitates faster and more coordinated recovery in both networks. Thus, reliability and resilience exhibit opposing trends with coupling strength. An interdependency index is also introduced to support management decisions for interdependent infrastructure networks. This framework provides a structured basis for understanding how inter-network connectivity affects system reliability and recovery dynamics.

Quantum-Aided Bayesian Learning for the Prediction and Uncertainty Quantification of Remaining Useful Life

Giorgio Tosti Balducci — 2026-07-03

To make predictions on the future states of engineering systems, prognostics has been increasingly relying on data-driven models and machine learning. Recent work has also looked at the potential of quantum machine learning to provide insight on the health state of systems. Nevertheless, these approaches treated the quantum component only as a deterministic predictor, in fact disregarding the uncertainty information. In this work, we propose a different approach to exploiting quantum circuits in prognostics. We focus on Bayesian learning of neural networks for Remaining Useful Life (RUL) prediction and uncertainty quantification. Here, the quantum circuit is introduced not as a data-driven model, but as a generator of neural network weights. Using Variational Inference, the quantum circuit can be trained to approximate the true posterior of the classical machine learning predictor. The overall method retains the data-processing ability of state-of-the-art machine learning models, while exploiting the quantum circuit to introduce uncertainty by sampling the model
space from a distribution that is classically nontrivial. We validate our approach on the task of predicting the End of Discharge of Li-ion batteries, using data generated from a simulator with tunable process uncertainty, and we compare the predictions obtained through quantum sampling with those from Flipout Bayesian neural networks, heteroscedastic neural networks and Monte Carlo Dropout. The results show that our quantum circuits learn to approximate the weight posterior and that the resulting data-driven models demonstrate accuracy and uncertainty quantification that is comparable if not superior to the baselines. Overall, our work demonstrates the potential of quantum computing for uncertainty-aware prognostics, and sets the stage for further investigations in this area.

Reducing Negative Transfer in Domain Adaptation for Vibration Fault Diagnosis

Pawel Knap — 2026-07-03

Unsupervised domain adaptation (UDA) for vibration-based fault diagnosis can improve transfer across changing operating conditions, but its reliability remains a practical concern. In particular, large domain shifts can lead to negative transfer, where an adapted model underperforms a source-only baseline. This motivates evaluation beyond average gains, with emphasis on worst-case behavior and failure modes relevant to deployment. This study proposes a lightweight safeguard for discrepancy-based UDA that does not require labeled target data. The approach augments standard adaptation with an unlabeled monitoring rule based on target prediction entropy and alignment-loss trends. When adaptation appears unstable, training is paused and the model is rolled back to a safer checkpoint. The safeguard is designed as a small reliability layer on top of existing UDA pipelines rather than as a new adaptation method. We evaluate source-only training, standard UDA, source-pretrained UDA, and safeguarded UDA on the Case Western Reserve University (CWRU) and Paderborn University (PU) bearing datasets under multiple cross-condition transfer tasks. Experiments include raw time-domain, FFT-based, and STFT-based representations with MMD- and CORAL-based adaptation. Results show that negative transfer is a repeatable phenomenon, particularly on more challenging CWRU shifts, while source-pretrained UDA substantially affects reliability. The safeguard shows partial mitigation of harmful adaptation in selected PU cases but does not consistently prevent degradation across all scenarios. Overall, the results highlight that monitoring adaptation dynamics can improve reliability in some settings, but that safe deployment of UDA for fault diagnosis still requires explicit consideration of worst-case behavior and baseline comparisons.

Reliability Analysis of Rolling Bearings Using a Weighted Nonlinear Mixed-Effects Degradation Model

HUNG-TSE HSU — 2026-07-03

Reliability assessment of rolling element bearings is critical for the predictive maintenance of industrial rotary machinery.
This study proposes a Quadratic-Exponential Weighted Model (QEWM) based on Nonlinear Mixed-Effects (NLME) to characterize the degradation process of bearings. Utilizing the IMS Bearing Dataset (Set No. 2), we define the failure threshold based on the latest ISO 20816-3:2022 vibration severity standards, setting the critical RMS limit at 0.4 mm/s for Zone D. Unlike traditional models, the proposed QEWM incorporates a weight function to address heteroscedasticity, which typically intensifies during the rapid degradation phase. Model comparison based on the Akaike Information Criterion (AIC)
demonstrates that QEWM significantly outperforms linear and unweighted quadratic models. To quantify the uncertainty of
the estimation, a parametric bootstrap method with 5,000 replications was employed. The results identify a B10 life (t0.1) of
165.3 hours, supported by a precise 95% confidence interval of [162.7, 168.6] hours. This research provides a robust statistical framework for bearing life prediction that aligns with international industrial standards, ensuring high precision in
prognostic assessments.

Remaining Useful Life Prediction Using Constraint Guided Learning with Limited Physical Knowledge

Ilyas Lemmens — 2026-07-03

Unexpected failures in industrial assets can cause significant downtime and costs. Anticipating the Remaining Useful Life (RUL) enables proactive maintenance strategies that mitigate such risks. While deep learning models have proven themselves to be adept at RUL prediction, they do not inherently guarantee physically consistent predictions. This paper explores two alternative approaches for incorporating physical constraints into data-driven RUL prediction models. The first approach extends prior work on physics-guided loss functions by integrating domain knowledge directly into the training objective. Physical assumptions are encoded as additional penalty terms that act as regularizers, discouraging physically implausible behavior while allowing trade-offs with predictive accuracy. The second approach leverages Constraint-Guided Gradient Descent (CGGD), a recent optimization framework which enforces constraints at the optimization level rather than through the loss function. CGGD monitors constraint satisfaction during training and dynamically modifies gradient updates only when violations occur, steering the solution back into the feasible region. Both methods aim to improve model interpretability and robustness without requiring detailed or fully specified physical knowledge, making them applicable to a wide range of industrial settings. We evaluate
these strategies across multiple experimental setups, comparing standard predictive accuracy with additional physics-based evaluation metrics that assess adherence to physical assumptions. The findings provide useful insights into the advantages and limitations of constraint enforcement techniques and their
overall impact on predictions, contributing to the development of trustworthy Prognostics and Health Management (PHM) models.

Risk-Aware Optimization of Charging Time and Route Selection for Electric Vehicles Under Uncertainty

Jorge E. Garcia Bustos — 2026-07-03

Effective decision-making during fast-charging sessions is becoming increasingly critical as Electric Vehicles (EVs) are deployed at scale. Drivers must make operational decisions under uncertainty arising from charging-time requirements, travel constraints, and the risk of battery depletion before destination. Moreover, from the EV driver's point of view, the reality of popular EV charging stations is far from ideal: charging station billing schemes, charging power decreases over time, and the applicable charging protocol depends on both the charger and the vehicle. In this setting, a driver charging at a given station who intends to reach a predefined destination faces two key operational questions: how long to charge the vehicle and which route to take, given that multiple feasible routes may be available. Accordingly, the driver must determine the required charging time to reach the destination, accounting for uncertainty in the estimated State of Charge (SoC) and stochasticity in route-dependent energy demand, while ensuring that the route can be completed without violating a voltage-based feasibility condition associated with power cut-off. Against this backdrop, we formulate the charging decision as a probabilistic optimization problem that captures the trade-off between charging time, travel time, and the probability of energy shortfall under a user-defined level of risk tolerance. The problem is centered on estimating the likelihood of an End-of-Power-Availability (EPA) event, thereby enabling route-aware range prediction under uncertainty. We conduct experiments using real charging curves, time-based tariffs, and two alternative routes in Costa Rica. Results show a clear trade-off between charging duration and EPA probability and consistent improvements over simple heuristics such as charging to full capacity or targeting a fixed SoC threshold. These results position the proposed approach as a practical decision-support tool for EV users, enabling charging and routing decisions to be made explicitly under uncertainty and user-defined risk preferences.

Robust Multi-Modal Hamilton-Jacobi Reachability Prognostics: An Application to Battery Health Management

Boutrous Khoury — 2026-07-03

This paper presents a novel and robust prognostics frame-
work for battery health management based on Hamilton–
Jacobi reachability analysis. A two-dimensional degradation
state space is constructed from the state of health, obtained
from discharge-capacity measurements, and a normalized
impedance feature extracted from electrochemical impedance
spectroscopy. Within this joint state space, a failure region is
defined to capture both capacity fade and impedance growth
under a unified EOL criterion. The Hamilton–Jacobi partial
differential equation is then solved backward in a minimum-
time-to-reach setting to generate state-dependent remaining
useful life maps. To account for battery-to-battery variability,
uncertainty in the degradation dynamics is estimated empiri-
cally from experimental data and incorporated through nom-
inal, worst-case, and best-case drift scenarios, thereby yield-
ing corresponding remaining useful life predictions. The re-
sulting maps are computed offline and interpolated online,
making the framework computationally efficient in deploy-
ment. Validation on the NASA battery dataset shows that the
proposed approach delivers physically interpretable remain-
ing useful life estimates together with informative uncertainty
bounds that successfully encapsulate the true remaining use-
ful life trajectories.

Robust Anomaly Detection Under Contaminated Data: A Comprehensive Evaluation Across PHM Contexts

Stefano Donné — 2026-07-03

Robust anomaly detection under contaminated training data is an important challenge in Prognostics and Health Management (PHM). In semi-supervised anomaly detection, models are typically trained on data assumed to represent normal behavior. In practice, this ``normal'' set often contains an unknown fraction of abnormal or degraded samples, which can harm diagnostic performance. This work presents a comparative evaluation of several techniques designed to mitigate the effects of contaminated training data across four public datasets representative of diverse PHM contexts, spanning tabular and multivariate time-series data, as well as both discrete anomalies and gradual degradation processes. The results show that contamination-mitigating techniques can improve anomaly detection performance over classical baselines when constructing a training set consisting solely of normal instances is not feasible. However, the benefits offered by contamination-mitigating approaches vary according to dataset characteristics. The largest gains are observed on the time-series datasets considered here, suggesting that refinement techniques may offer a clearer advantage over contamination-sensitive baselines in these settings. These gains, however, come at a substantially higher computational cost. The experiments also suggest that the effect of contamination depends not only on its ratio, but also on the structure and distribution of anomalies.

Robust Real-Time Thrust Fault Diagnosis for UAVs: A Physics-Informed Framework DecouplingWind Disturbances

Taegyun Kim — 2026-07-03

Operational reliability of multi-rotor Unmanned Aerial Vehicles (UAVs) is frequently compromised by the ambiguity between external wind disturbances and internal thrust faults. This paper proposes a physics-informed fault diagnosis (PIFDI) framework that explicitly decouples wind-induced effects from total observed disturbances. By integrating an Extended Kalman Filter (EKF) for real-time wind estimation and a Disturbance Observer (DOB) for total torque monitoring, the framework isolates a clean fault residual through physical coefficient mapping. High-fidelity 6-DOF simulations involving Dryden turbulence and non-stationary discrete gusts demonstrate a rapid detection latency of 0.18 s for a 20% thrust loss, maintaining near-zero false alarms even during peak gust periods. Furthermore, a 300-trial Monte Carlo simulation confirmed high fault isolation accuracy, demonstrating
superior statistical robustness across varying wind intensities and randomized fault modes. The proposed physicsinformed
decoupling approach significantly enhances diagnostic resilience, providing a critical foundation for real-time fault-tolerant control in mission-critical UAV operations.

RUL-Aware RRT*: Degradation-Balanced Motion Planning for Robotic Manipulators

Haibo LI — 2026-07-03

Industrial robotic manipulators operating over long durations may suffer from uneven joint degradation, causing the weakest actuator to fail prematurely and limiting the lifetime of the entire system. Although Prognostics and Health Management (PHM) techniques can estimate component health and remaining useful life (RUL), such information is rarely incorporated into online motion planning. To address this gap, this paper proposes an RUL-aware RRT* method that integrates joint RUL information into the trajectory generation process through a health-aware cost formulation and an adaptive joint weighting mechanism. A deterministic cumulative joint-usage surrogate is used to represent planning-level degradation in simulation. The method is evaluated in a Local Degradation Scenario, where one joint starts from a severely weakened condition and acts as the dominant lifetime bottleneck. Results show that the proposed method reduces the motion assigned to the weak joint, delays the first system failure, and maintains more balanced degradation than the baseline RRT*. These findings demonstrate that integrating prognostic information into motion planning provides a practical pathway toward health-aware robotic decision-making.

Scalable Model-Based Discrete Mode Estimation for a Lunar Rover Power System

Annabel Gomez — 2026-07-03

Reliable autonomous operation of planetary rovers requires robust onboard monitoring and diagnosis of power systems under uncertain and dynamic conditions. This work presents a model-based mode estimation framework for a simplified rover power system that combines physics-based modeling with probabilistic reasoning to estimate system and environmental state, including anomalous behavior. The approach adapts the Miniature Mode Estimation (Mini-ME) architecture to perform real-time belief tracking over discrete component modes based on noisy measurement signals including voltage, current, temperature, and state of charge, and command inputs into the system. Mini-ME is a system health management tool that monitors components, diagnoses faults in real-time, and supports downstream system recovery. We simulate the physics of the simplified power system under varying loads using a lumped parameter equivalent circuit model that captures the electrical and thermal behavior of the battery. A compiled automaton is used to encode admissible probabilistic state transitions and observation likelihoods derived from sensor models. By compiling out the combinatorial search and integrating these representations within a Bayesian update process, the system efficiently detects, isolates, and diagnoses anomalies arising from faults or unexpected operating regimes. The simulation results demonstrate accurate recovery of hidden system modes during charge-discharge cycles, illustrating how model-based reasoning supports autonomous fault diagnosis and system health management in autonomous missions. This of the paper provides a thorough set of computational experiments to evaluate the performance of compiled mode estimation, versus uncompiled mode estimation, on space relevant benchmarks.

Sensor Fault Detection via Virtual Smart Heat Metering with Spatial-Temporal Graph Neural Networks

Keivan Faghih Niresi — 2026-07-03

Sensor faults and miscalibrated sensors represent an important challenge in district heating networks, where measurement errors, drift, calibration inaccuracies, or communication issues can compromise the reliability of thermal and hydraulic monitoring. Detecting such issues in a timely manner is essential for maintaining operational efficiency and ensuring the trustworthiness of system data. A promising approach for addressing these challenges is to compare physical measurements with estimates generated by virtual sensors. Virtual sensing enables the reconstruction of unmeasured or unreliable variables using data-driven models and existing measurements, thereby providing an estimate of the expected measurement value under the current operating and environmental conditions, which can serve as reference against which anomalous or inconsistent sensor behavior can be identified. In this work, we develop a virtual-sensor–based framework for sensor fault and miscalibration detection using a heterogeneous spatial–temporal graph neural network (HSTGNN). The proposed model learns both the spatial relationships among sensors and the temporal dynamics of their measurements to construct accurate virtual smart heat meter outputs. To evaluate the approach, we use a controlled laboratory dataset collected at the Aalborg Smart Water Infrastructure Laboratory, which provides synchronized high-resolution measurements of flow, temperature, and pressure representative of district heating operating conditions. Experimental results demonstrate that the proposed HSTGNN improves fault detection performance compared to several baseline methods.

Simulation-Based Inference for the Amortized Determination of Physical Model Parameters of a Mechanical System for Diagnostics and Prognostics

Cedric Schenkel — 2026-07-03

Many diagnostic and prognostic applications rely on complex measurements, including time series and high-dimensional
data. In a common approach, one extracts key features that capture the system degradation while ignoring nuisance effects. Physical system properties are of interest due to their (direct) relation and relevance to degradation and failure modes, often allowing for superior interpretability compared to purely data-driven approaches. However, determining them
from the observed data is difficult due to the inherent non-identifiability in already rather simple models. Recently developed simulation-based inference (SBI) approaches, based on neural posterior estimates (NPE) and conditional invertible neural networks (cINN), allow the incorporation of domain knowledge in the form of simulation capabilities to extract model parameters as physics-based features for diagnostics and prognostics. This is demonstrated in the case of a mechanical actuator that operates medium voltage breakers. Simulations of a simplified multi-body model are used as input for the training of a cINN that not only provides a set of physical parameter values but also their respective uncertainties. By using simulated data as synthetic measurements and
conducting a number of statistical checks, the performance of the trained cINN is confirmed. We demonstrate that accurate multi-body parameter estimation is possible for some parameters, whereas those that cannot be identified from either the
opening or closing motion remain distributed according to the prior distribution, without significantly affecting the others.
We further show that the opening and closing motions are sensitive to different parameters, complementing each other
in this respect. Data from a real mechanical endurance test are used to demonstrate the method’s effectiveness in a real-
world application. Its integration into a diagnostics or prognostics framework is discussed as an outlook.

Spatiotemporal Graph Neural Networks for Fault Detection and Structural Learning in Chemical Processes: Use Case on the Tennessee Eastman Process

Rayane AMMAR KHODJA — 2026-07-03

Classical Fault Detection and Diagnosis (FDD) methods, including many data-driven approaches, assume a static normal operating space and interpret deviations from a fixed reference as fault indicators. When Operating Conditions (OCs) vary over time, this assumption breaks down: legitimate transitions trigger false alarms while moderate faults go undetected. We propose a spatiotemporal Graph Neural Network (GNN) framework that decomposes the normal operating space into OC-specific subspaces linked by transition functions, with a dual learning objective combining reconstruction loss and a Deep Support Vector Data Description (DeepSVDD) one-class term. The framework learns adjacency matrices through Graph Attention Networks (GATs) and integrates spatial modelling with temporal encoding to represent process dynamics under evolving OCs.

This paper evaluates the foundational components of the framework — fault detection, spatial graph learning via GATv2, and temporal encoding — on the Tennessee Eastman Process (TEP) benchmark, with training performed exclusively on fault-free data. The spatiotemporal architecture achieves competitive detection performance from reconstruction error alone, with similarity-based feature selection improving both accuracy and graph structure diversity. We then evaluate the physical interpretability of the learned attention matrices against 22 ground-truth sensor pairs derived from the TEP control structure and process topology. The GATv2 attention does not recover all the necessary known physical pairs across multiple hyperparameter configurations, suggesting a structural limitation of reconstruction-driven attention rather than a tuning issue. This result challenges a common assumption in GNN-based FDD: that learned attention weights provide a basis for fault diagnosis and root-cause analysis. The architecture detects faults effectively, but the learned graph does not encode the physical topology needed for interpretable diagnosis, motivating physics-informed graph construction.

Structural Representation Learning for Thermal Turbulence Detection in Infrared Imagery using YOLO

Akash Deep — 2026-07-03

Thermal turbulence degrades imaging performance in long-range infrared systems by introducing spatially varying distortions that appear as irregular intensity fluctuations and curvilinear patterns. Detecting these regions is challenging due to the absence of well-defined boundaries and their diffuse nature. This work investigates how structural characteristics of thermal turbulence influence automated detection using deep learning–based object detectors. A systematic study is conducted to evaluate different structural representations derived from thermal imagery, including rolling guidance filter (RGF), variance-based fluctuation maps, curvature-based features from the Hessian matrix, and multi-scale vesselness responses using the Frangi filter. These descriptors are incorporated as multi-channel inputs within a YOLO-based detection framework and evaluated on annotated infrared turbulence data. Results show that while deep detectors can capture turbulence cues from raw thermal images, structural representations improve the visibility of distortions and enhance detection robustness. In addition, intensity-based enhancement strategies are analysed to examine whether simple contrast amplification alone can improve turbulence detection performance. A structural fusion of thermal intensity and complementary feature representations achieves the best overall performance, improving localisation accuracy and recall. The findings highlight the importance of representation design in detecting diffuse thermal patterns and provide a more reliable framework for turbulence-aware detection in infrared imagery.

Strain-based condition monitoring of inner raceways of deep-groove ball bearings using FBG sensors

Fernando de la Hucha Arce — 2026-07-03

Rolling element bearings are critical components of rotating machinery, whose failure is one of the main causes of downtime and maintenance. Traditional methods of condition monitoring of bearings, based on accelerometers or acoustic emission sensors and vibration analysis, are prone to signal attenuation and interference in the transfer path due to other machine components. As an alternative, fiber Bragg grating (FBG) sensors allow for quasi-distributed sensing of the local strain of the bearing, as they can be integrated in a single optical fiber bonded directly onto a bearing raceway. They offer several advantages, such as compactness, immunity to electromagnetic interference (EMI), and resistance to corrosion. Proximity and the quasi-distributed nature of FBG-based strain sensing are key properties to obtain significantly higher signal-to-noise ratio (SNR) and sensitivity, enabling enhanced fault diagnosis and localization.

In this work, we analyze the strain signals obtained during an accelerated lifetime test (ALT) of a deep-groove ball bearing. The FBGs are instrumented on the rotating inner raceway of the bearing, going beyond the implementation on the static outer raceway performed by several recent research works. We study the behaviour of the FBG signals and their features during the complete evolution of a surface-initiated fatigue fault on the inner ring, and evaluate their capabilities for simultaneous fault detection and localization. We observe that two features are reliable indicators for fault detection and localization, the RMS of the high-pass filtered FBG signals, and the sum of values of their squared envelope spectrum (SES) at the harmonics of the ball pass frequency on the inner race (BPFI), while the peak-to-peak (P2P) value is not.

System-Level CBM/PBM Aggregation: A Unified Framework for Proactive and Reactive Metrics in Redundant Architectures

Franck DESSERTENNE — 2026-07-03

Assessing the in-service performance of Condition-Based Maintenance (CBM) and Predictive-Based Maintenance (PBM) solutions remains a challenge, as traditional reliability formulas often fail to capture the true nature of proactive operations. First, this paper formally proves that applying classic reactive formulas to CBM/PBM KPIs (Key Performance Indicators) leads to a quantifiable underestimation of the actual Mean Time Between Failures (MTBF). While previous work established aggregation laws for "in-series" equipment configurations, this study extends the mathematical framework to "in-parallel" architectures. Indeed, such specific aggregation laws are essential for evaluating the global performance of CBM/PBM solutions applied to fault-tolerant systems involving redundancy.

Subsequently, this paper introduces a unified theory based on two distinct frames of reference: the "Operational Timeline," using the Mean Time Between Proactive Removals (MTBPR) to measure logistical workload, and the "Effectiveness Timeline," using the Mean Lifetime Reduction (MLR) correction term to account for the residual life lost due to early removal. Following a review of the laws governing MTBF and the reliability function, we demonstrate the laws applicable to key CBM/PBM indicators, including Recall and Precision-based metrics.

Finally, this framework elevates equipment-level performance into system-wide strategic indicators. Ultimately, this enables a holistic evaluation of the global performance of a CBM/PBM solution at the system or aircraft level by balancing the Proactive Aggregated Recall (PAR) for detection coverage, and the Proactive Aggregated Precision (PAP) for detection confidence.

Supervised and Unsupervised Methods for Detecting Anomalies in an Autonomous Long-Range Lunar Rover

Sofie Claridge — 2026-07-03

Autonomous space systems must navigate complex environments and accomplish concurrent tasks without continuous input and supervision from human operators. Interactions between subsystems and with the environment may lead to unintended behavior, resulting in downtime, delays, goal degradation, loss of functionality, or even catastrophic failures. The Endurance mission concept requires a lunar rover to traverse thousands of kilometers over a long period of time (multiple years) across the South Pole-Aitken Basin on the far side of the Moon. Onboard system health management is therefore required to identify anomalies and faults that pose a risk to mission objectives. To this end, we propose training both unsupervised and supervised machine learning models to detect and isolate anomalies onboard a rover over the course of a mission, supported by a pipeline for anomaly data generation that enables training and evaluation. We generate synthetic anomaly signatures using a low-fidelity mission simulator that outputs labeled datasets to enable supervised learning. We present results from a field experiment in which we deploy this supervised model as a ROS network node in a rover-ground network. In parallel, we train unsupervised models to detect anomalies in the mobility system by training on experimental field data and present results that verify the ability to detect anomalies observed by field operators as well as anomalies that were not detected by the operators. Our work demonstrates how machine learning models can detect anomalies onboard by leveraging multiple data sources, including pre-launch test data and operational data from earlier phases of the mission, and provides a pathway for improving anomaly detection as a rover's mission progresses.

Time‑Series Retrieval for Grounding Multimodal Language Models in Remaining Useful Life Prediction

Valeriu Dimidov — 2026-07-03

Large language models (LLMs) and agentic AI systems are increasingly being explored for domain-specific maintenance and prognostics tasks, raising the question of whether they can effectively support prognostics and health management (PHM). In this paper, we investigate remaining useful life (RUL) estimation with multimodal large language models (MLLMs) grounded through time-series retrieval. We propose a framework in which historically similar degradation segments are retrieved from the training set and, together with the test trajectory, transformed into a visual comparison artifact that is processed by the MLLM through a structured multimodal prompt. The approach is evaluated on the FD001 partition of the C-MAPSS benchmark under repeated experiments comparing retrieval-based inference against a non-retrieval baseline based on random reference selection. The results show that time-series retrieval consistently improves MLLM-based RUL prediction across the evaluated models, yielding lower error and more stable performance. At the same time, the magnitude of the benefit depends on model capacity, indicating that retrieval is most effective when the underlying MLLM is able to exploit the retrieved evidence. Overall, the study shows that time-series RAG is a promising mechanism for improving multimodal prognostic reasoning, while also highlighting the current limitations of MLLM-based RUL estimation in practical PHM settings.

Through-Life Monitoring of Resource-constrained Systems and Fleets

Felipe Montana — 2026-07-03

A Digital Twin (DT) is a representation of a physical system that provides information to make decisions that add economic, social or commercial value. DTs are widely used for prognostics and anomaly detection by continuously comparing measured system behaviour with the DT predictions to identify deviations and estimate degradation. The behaviour of a physical system changes over time; a DT must therefore be continually updated with data from the physical system to reflect its changing behaviour. In this paper, we consider a DT of a complex, non-linear and dynamic system subject to slow nominal degradation, disturbances and the risk of anomalies. The DT runs on a resource-constrained system, making up dating non-trivial due to limitations in computation, storage,
and data transfer bandwidth. Consequently, only a subset of the generated data can be retained or transmitted, making data prioritisation essential. Data must be evaluated online in order to select the most relevant subset with which to perform the update. DT updating must address the continual learning
challenge of adapting to new system behaviours, such as the response to previously unseen operating conditions, while retaining knowledge of previously observed behaviours. This paper presents a framework for updating a data-driven DT of a resource-constrained system. The proposed solution consists of: (1) an on-board, lightweight DT that enables the prioritisation and parsimonious transfer of data generated by the physical system; and (2) an off-board system for robust DT updating that enables the reliable detection of anomalous be haviours across the asset’s lifetime. The framework allows the DT to accurately replicate the behaviour of the system throughout its life, improve sensitivity to anomaly detection, and reduce the risk of forgetting previous system behaviours after updating the DT. An in-service gas turbine engine case study is used for demonstration.

Towards Green PHM: Adaptive Early Stopping for Sustainable Neural Architecture Search in Industrial Applications

David Solıs-Martın — 2026-07-03

Neural Architecture Search has revolutionized Prognostics and Health Management, yet adoption is often hindered by the massive carbon footprint generated during the evaluation of candidate architectures. To address this sustainability challenge, this work introduces a Green AI framework that significantly reduces the energy consumption of such searches through an intelligent and adaptive early stopping mechanism. The approach utilizes Prototypical Networks for regression to extrapolate learning curves from partial data, predicting final model performance early in the training process.

A distinct sustainability advantage of using Prototypical Networks lies in the inherent data efficiency and superior generalization capabilities of the architecture. By leveraging metric learning, the framework avoids the energy intensive process of training task specific predictors from scratch. Instead, it enables few shot transfer across diverse domains, minimizing the total computational overhead of the search process. Furthermore, a key contribution of this framework is the dynamic adaptation of the decision logic as the optimization process evolves. By utilizing a decision tree classifier that adjusts thresholds based on the progression of the search, the system becomes increasingly selective and prioritizes computational resources for the most promising candidates.

\textcolor{blue}{The proposed framework was validated across sixty one thousand learning curves from fifty diverse datasets. Experimental results demonstrate a drastic reduction in total computational hours, achieving 57\% of decrease in training time while maintaining high diagnostic fidelity. The system consistently identified top tier model configurations, reaching a mean selection rank of 0.9 across tested industrial scenarios. These results prove that high performance industrial intelligence can be achieved without the prohibitive environmental costs typically associated with large scale architecture optimization.

Transformer-Based Architectures for Machinery Prognostics: A Review

Maxime Pierfederici — 2026-07-03

Machinery prognostics requires robust modeling of multivariate degradation signals under noise, non-stationarity, variable operating conditions, and limited run-to-failure labels. Transformer-based deep learning architectures have recently attracted strong interest because self-attention can capture long-range temporal dependencies and inter-sensor interactions more directly than purely recurrent or convolutional models. This focused review presents Transformer-based approaches for machinery prognostics, with emphasis on remaining useful life (RUL) estimation and degradation representation learning.} \rboth{The literature is organized using a consistent taxonomy covering PHM task, Transformer backbone, hybridization strategy, and input representation.} \redit{We also analyze preprocessing choices that strongly influence performance, including windowing, health-indicator construction, tokenization, embedding, and positional encoding. Across benchmark datasets, studied studies frequently show gains from Transformers and hybrid attention models, especially when long temporal context and multivariate dependencies are central. However, improvements are not universal and remain sensitive to evaluation protocol, signal representation, and model complexity. Key open challenges include data efficiency, computational cost, cross-condition generalization, interpretability, and uncertainty quantification. The review concludes by identifying methodological gaps in the current literature and outlining research directions for robust, efficient, and deployable Transformer-based prognostics.

Trolley-Assisted Charging for Full-Shift Operation of Battery-Electric Load-Haul-Dump in Underground Mining

Lukas Gleisner J. — 2026-07-03

Underground mining operations are transitioning to battery-electric fleets to reduce diesel emissions and ventilation requirements. However, current battery technology cannot sustain a full production shift in battery-electric load--haul--dump (LHD) vehicles, forcing mid-shift charging stops or battery swaps that disrupt production. Trolley-assisted systems (TAS) address this limitation by supplying external power along selected route segments through two complementary mechanisms: dynamic charging, in which an overhead catenary delivers energy while the vehicle travels, and static connection, in which the LHD connects to the trolley rail during brief scheduled stops, such as dumping. Together, these mechanisms reduce net battery discharge and extend battery autonomy without dedicated charging downtime. This paper presents a simulation-based framework to quantify the operating capability of TAS-equipped LHDs and evaluate trolley-length allocation at the production level. A physics-informed power and energy model, built on the Sandvik LH518iB architecture, is integrated into a route-level simulator that reproduces a full draw-control schedule on a real Chilean copper mine layout. The trolley length is treated as a configurable design parameter and systematically varied to assess its effect on battery autonomy, idle time, and throughput. Results show that TAS can eliminate mid-shift battery swaps, improve productivity relative to battery-electric operation without TAS, and reduce energy cost per ton by up to 75 % compared with diesel operation. Furthermore, route-to-route energy variability motivates a two-stage stochastic optimization framework to determine routing and charging strategies for a given trolley configuration.

Trustworthy Abnormality Detection from Welding Images Through Class-Conditional Conformal Learning and Bayesian Cost Minimization

Zhenling Chen — 2026-07-03

Industrial fault/abnormality detection is often criticized for lacking explainability and robustness. Furthermore, practical industrial datasets are frequently highly imbalanced and operate under extreme risk asymmetry, i.e., false negatives carry penalties orders of magnitude higher than false alarms, which poses significant challenges to reliable detection. In this paper, we develop a trustworthy AI framework to improve confidence in welding defect detection. The proposed framework integrates two primary techniques: class-conditional conformal learning and Bayesian cost minimization. First, the conformal learning model quantifies the trustworthiness of model predictions. Instead of forcing a binary classification, the model outputs an "uncertain" state when confidence is low, facilitating informed human intervention. Second, a Bayesian cost minimization algorithm is used to avoid over-conservative predictions that yield too many "uncertain" predictions. Results on a real-world welding quality inspection dataset show that the developed method adapts robustly to dynamic intervention costs and mitigates worst-case cost spikes. The framework is deployment-oriented: it is not uniformly optimal in every setting, but it consistently avoids catastrophic failures and maintains a favorable cost–accuracy–intervention trade-off across heterogeneous base-model qualities.

Turbofan Sensor-FDI-Bench: A Synthetic Dataset for Sensor Fault Detection & Isolation under Degradation and Operating Variability

Aytunc Yildirim — 2026-07-03

Aircraft engine monitoring relies on sensor measurements to assess gas path condition and to distinguish gradual degradation from abrupt performance changes. In practice, however, sensor signals are influenced simultaneously by the engine state, changing operating conditions, and sensor side effects such as step, drift, random outliers, and measurement noise. This makes it difficult to determine whether an observed deviation originates from the engine, the environment, or the sensing system. For the development and fair comparison of sensor fault detection and isolation methods, a benchmark is required that represents these effects in a controlled and labelled manner. Most publicly available turbofan datasets, however, are primarily intended for remaining useful life prediction and do not provide standardised sensor fault cases for reproducible FDI evaluation. To address this gap, the Turbofan Sensor-FDI-Bench is introduced as a synthetic steady state benchmark dataset generated with a physics based turbofan performance model. The benchmark consists of cruise operating point snapshots and provides, for each flight, environmental conditions, an extended sensor package, and gradual multi component performance degradation. Structured sensor faults with controlled onset and severity are superimposed, including step and drift faults as well as stochastic measurement disturbances. The benchmark is organised as a progressive suite of subsets with increasing complexity, covering fixed and variable operating conditions as well as single fault and multi fault diagnosis settings. For each engine unit, clean reference sensor values are released alongside noisy or faulty measurements, enabling supervised denoising and controlled evaluation of sensor fault diagnosis methods. The resulting benchmark provides a reproducible basis for comparing sensor fault detection and isolation methods under degradation and operating variability.

Turbograd: An Open-source Differentiable Performance Model for Aeroengine Condition Monitoring

Daniel Cisneros Acevedo — 2026-07-03

Combining physics-based and data-driven models for aero-engine condition monitoring has attracted increasing research interest in Prognostics and Health Management. Hybrid methods that incorporate physics-based models into deep neural networks typically rely on pre-trained surrogate models of
the engine performance model, which serve as a differentiable proxy during training. However, constructing such surrogates requires extensive exploration of the parameter space to generate representative datasets, resulting in a rapidly increasing computational burden as the number of model parameters grow. To address this limitation, we present TurboGrad, an open-source differentiable aeroengine performance model that reformulates the Gas Turbine Simulation Program (GSPy) in PyTorch. Because the performance model is tracked as a computation graph, gradients with respect to any model parameter follow directly via backpropagation. We compared TurboGrad against GSPy for a single-spool turbojet, finding relative errors within 0.3%. Furthermore, we demonstrate gradient-based estimation of compressor and turbine efficiencies, converging to the ground truth after 30 epochs. TurboGrad is open-source and provides a differentiable foundation for integrating physics-based aeroengine models directly into deep learning pipelines.

Uncertainty-Aware Bearing Remaining Useful Life Prediction Based on Conformal Prediction

Shun Wang — 2026-07-03

Accurate prediction of the remaining useful life of rolling element bearings is a critical task in prognostics and health management. Although deep learning methods have shown strong predictive capability, purely data-driven approaches still face two important limitations: they may produce physically inconsistent predictions that contradict the irreversible nature of bearing degradation and often fail to provide reliable uncertainty estimates. To address these issues, this paper proposes a physics-informed probabilistic framework to predict the remaining useful life. First, a health index is constructed from logarithmic envelope spectrum features using a variational autoencoder, enabling the extraction of a monotonic degradation indicator without requiring labeled fault data. Second, a Transformer-based predictor is trained with a monotonicity constraint that explicitly enforces the predicted remaining useful life to be non-increasing over time. Third, Monte Carlo dropout is used to quantify epistemic uncertainty, and a post-hoc conformal calibration strategy is applied to construct finite-sample prediction intervals with guaranteed marginal coverage by leveraging historical degradation data. Experiments on the XJTU-SY full-lifecycle bearing dataset show that the proposed framework improves point prediction accuracy relative to controlled feature and model ablations. More importantly, the uncertainty results reveal a substantial mismatch between raw Monte Carlo dropout intervals and the observed prediction errors: the average prediction interval coverage probability increases from 0.4835 before calibration to 0.9445 after conformal calibration. The resulting wider intervals should not be interpreted only as a loss of sharpness, but as a correction of the severe overconfidence of the uncalibrated model under heterogeneous degradation trajectories. Bearings with more irregular or non-stationary degradation behavior require wider calibrated intervals to maintain reliable coverage, indicating that the proposed framework can expose trajectory-dependent prediction difficulty and support risk-aware maintenance decisions.

Uncertainty-Aware Surrogate for Fatigue Assessment of Moorings in Offshore Wind Turbines

Rohit Kumar — 2026-07-03

Floating Offshore Wind Turbines (FOWTs) deployed in deep waters extract wind energy via floating platforms that are station kept by mooring lines subjected to complex, highly dynamic loading conditions. The cyclic nature of these dynamic loads induces fatigue damage in the mooring lines, which can culminate in catastrophic failures with substantial operational, economic, and safety implications. The remote offshore location of FOWTs renders conventional, sensor-intensive structural health monitoring in deep water both costly and logistically challenging. Indirect sensing approaches offer a promising alternative; however, existing methods typically neglect inherent aleatoric material uncertainties arising from manufacturing variability, installation effects, and long-term corrosion, thereby limiting their reliability for informed decision making.
To overcome this limitation, the present study introduces an uncertainty-aware surrogate-based indirect sensing framework that quantifies mooring line fatigue damage in a fully probabilistic manner by constructing confidence regions of damage conditional on the prevailing environmental loading. This probabilistic characterization supports more reliable, risk informed inspection, maintenance planning, and life-extension strategies. The surrogate model is trained over a wave scatter table representative of the Gulf of Khambhat region, encompassing a wide range of sea states. Training and validation datsets are generated using high fidelity numerical simulations generated with OpenFAST, based on the NREL 5 MW OC4 semisubmersible wind turbine reference model.

Uncertainty-Aware and Risk-Controlled Identification of Abnormal Parametric Changes in Space Launcher Electrical Valve Actuators

Luis Basora — 2026-07-03

In the context of health monitoring for the next generation of reusable space launchers, this work presents an uncertainty-aware method for detecting and diagnosing off-nominal parameter variations in the electrical system that drives engine valves. The study relies on data generated from a physics-based model, where deviations of nine key parameters simulate realistic faults and degradations.

The proposed pipeline combines domain-driven segmentation to isolate valve-motion intervals, automated statistical feature extraction, and multiclass gradient-boosting-based classification, together with out-of-distribution detection using an isolation forest. To enable uncertainty-aware decisions with user-defined confidence levels, all predictive stages are calibrated through a learn--then--test risk-control framework, providing finite-sample guarantees for out-of-distribution rejection, fault detection, and diagnosis via prediction sets.

Numerical results on the available data demonstrate the effectiveness of the pipeline and an improvement over our previously published approach. However, a class-separability analysis reveals intrinsic limitations of the available signals for near-nominal fault classes, underscoring the need for improved observability or alternative modeling assumptions in future real-data deployments.

Understanding the Impact of Temporal Aggregation on Uncertainty in Quality Indicator Prediction for Industrial Processes

Thanos Kontogiannis — 2026-07-03

Industrial predictive datasets often rely on coarse synchronized targets obtained by aggregating serial measurements over predefined segments. Although such aggregation is usually imposed by storage constraints and synchronization requirements, the resulting labels are commonly treated as deterministic in downstream modeling. This can be misleading when the underlying process is serially dependent, because aggregation then injects uncertainty at the label level before any model is trained and directly affects the performance that can realistically be achieved. This work examines that effect using segment-level coiling temperature prediction in hot strip steel manufacturing as a real-world example, where meter-level coiling temperature measurements are synchronized to tracked material segments and averaged to form prediction targets. A serial, dependence-aware, and deployable formulation is introduced to quantify the uncertainty associated with these aggregated targets and propagate it to the downstream predictive task. Results show that serial dependence persists in the meter-level coiling temperature series even after downsampling, and that slower sampling increases the uncertainty associated with the aggregated labels. The estimated aggregation uncertainty is further shown to be of the same order as the error achieved by an extensively optimized downstream predictor, indicating that a non-negligible portion of the apparent prediction error is attributable to noise introduced during target construction rather than to deficiencies of the predictive model alone. The findings highlight that aggregation design should not be treated as an innocent preprocessing choice. Instead, it should be considered as a factor that directly shapes attainable predictive performance, with sampling frequency emerging as an actionable lever for improving the performance ceiling of industrial datasets constructed from serial measurements.

Vibration Analysis for Damage Detection and Classification for Condition Monitoring on Worm Gears

Philipp Häderle — 2026-07-03

Worm gears exhibit a higher proportion of sliding motion during tooth meshing than spur gears. Typically, this type of gear consists of a soft worm wheel, often made of brass, and a hard steel worm. These characteristics promote the occurrence of abrasive wear and further fatigue damage during operation. The objective of this study is the vibration data analysis for a sensor-based detection and classification of two types of damage in worm gears during operation. For this purpose, a data set is available containing the measurement data from an accelerometer under various operating conditions with regard to rotational speed and torque. The data set comprises measurements of the undamaged condition, artificial tooth thickness reduction on the worm wheel to reflect wear, and artificial breakouts reflecting pitting damage on the worm. For damage detection the accelerometer data is evaluated in the frequency domain. The amplitudes at different frequencies are analyzed for each type of damage. Both types of damage exhibit distinct characteristics in the analyzed frequency spectra. Based on these differences, vibration-based indicators are derived from the frequency spectra, enabling the detection and classification of breakout damage on the worm and wear damage on the worm wheel. Breakout damage on the worm is characterized by a discrepancy in the acceleration amplitudes at the harmonics of the gear mesh frequency when compared to measurements obtained under undamaged conditions. For the wear damage on the worm wheel, a systematic difference compared to the undamaged reference measurement can be detected in the frequency range between 7.5 kHz and 8.5 kHz. In addition to the effect of the damage, a distinct influence of the rotational speed and torque is observed on the extracted indicators for damage detection. Since different data evaluation methods are best suited for detecting each type of damage, damage classification is possible. In the context of PHM, this enables health management measures to be implemented according to the severity of the detected type of damage.

Video Motion Magnification for Vibration Measurement in Hydropower Applications

Florian Fritzsche — 2026-07-03

Ensuring the mechanical integrity of hydropower plants requires robust structural health monitoring to detect issues like rotor imbalance and cavitation. Video motion magnification offers a promising non-contact alternative for vibration measurement. This paper presents an experimental comparison of three state-of-the-art algorithms (phase-based, learning-based, and Swin Transformer-based) for quantitative vibration measurement. Rather than evaluating only the final output, a novel framework analyses motion signals across multiple stages of the algorithms' processing pipelines to identify optimal extraction points. The frequency detection capabilities of these algorithms are then evaluated using both industrial and consumer-grade cameras. The focus is on comparing their ability to accurately measure vibrations with different input data quality. The results demonstrate the importance of the quality of the input data on the performance of the algorithm, as the compressed videos from the consumer-grade camera performed significantly worse than the uncompressed videos from the industrial camera. The learning-based method demonstrated the best overall performance, particularly with high-quality video data. This enabled the oscillation frequency to be measured at amplitudes over 70 times smaller than a pixel.

Virtual Temperature Sensors in Power Transformers Using Neural Ordinary Differential Equations

Berk Hadzhamolla — 2026-07-03

Accurate modeling and forecasting of power transformer thermal behavior are critical for ensuring reliability, extending asset lifetime, and enabling optimized power system operation. Numerical approaches, such as finite element methods (FEM) and computational fluid dynamics (CFD), offer high fidelity but suffer from prohibitive computational costs, complex mesh generation, and limited feasibility in real-time or large scale applications, as well as often unknown geometries. Lumped-parameter thermal models provide a more practical alternative but depend on transformer-specific thermal constants and often fail to capture dynamic responses under varying operating and environmental conditions. Purely data-driven machine learning (ML) methods, including artificial neural networks (ANNs), convolutional neural networks (CNNs), and recurrent architectures such as long short-term memory (LSTM) networks, have shown success in forecasting transformer oil, winding, and hotspot temperatures; however, they typically require large volumes of high-quality training data and risk producing physically inconsistent or uninterpretable results. To overcome these limitations, hybrid frameworks such as physics-informed neural networks (PINNs) embed physical laws into the learning process, enabling physically consistent solutions while reducing data demands. This paper applies a physics-aware modeling of Neural Ordinary Differential Equations (Neural ODEs) adapted for forecasting transformer thermal behavior using real-world time-series data. Neural ODEs model system dynamics in continuous time, enabling smoother predictions, robustness to irregular sampling, and improved extrapolation capabilities compared to discrete-time models such as LSTMs. A key contribution of this work is the integration of simplified heat-transfer equations for power transformers directly into the Neural ODE, enabling a physics-aware formulation of the thermal dynamics. The model’s performance and generalization capabilities are evaluated across datasets from fifteen distinct transformers located in different regions of Norway, and characterized by varying designs and cooling mechanisms. The results demonstrate the success of the developed Neural ODEs framework to serve as a standardized, physics-aware, and robust forecasting tool for heterogeneous transformer units.

Causal Inference for Root Cause Analysis in Safety-Critical Engineering Systems

Evangelia Perivolaki — 2026-07-03

Root-cause analysis (RCA) in safety-critical engineering relies heavily on correlation-based methods and expert judgment. These approaches are useful but limited: they can identify which signals are associated with a fault, but not why that fault occurred or which factors genuinely caused it. This research explores how causal inference techniques, combined with domain engineering knowledge, can produce more reliable, explainable, and reusable diagnostic tools for Prognostics and Health Management (PHM). The work is grounded in a structured literature review and early prototyping and will be validated on a real industrial dataset from a large civil aerospace engine programme.

Health-Aware Load Allocation and Joint Energy--Maintenance Optimization for Multi-Stack PEM Fuel Cell Systems

Mouhamad Houjayrie — 2026-07-03

Multi-stack proton exchange membrane fuel cell (PEMFC) systems are promising for transportation applications because they are compact, provide high power density, operate at low temperature, and produce no direct CO₂ emissions during operation. However, high cost and insufficient durability still hinder large-scale deployment. Under real driving conditions, variable loads and frequent transients accelerate degradation and shorten system lifetime.

To address this challenge, this thesis develops a prognostics and health management (PHM) framework for improving the durability and lifecycle performance of multi-stack PEMFC systems. In a first stage, a load-dependent degradation and prognostics framework is developed to estimate the health state online and to predict end of life (EOL) and remaining useful life (RUL), together with associated uncertainty, under projected future load scenarios. In a second stage, these prognostic outputs are used for decision-making through health-aware energy management and maintenance planning. By coordinating load allocation and maintenance actions, the proposed framework aims to extend system lifetime, improve availability, and reduce lifecycle cost.

Multi-Modal 3D Neural Representations for Scene Modeling: Towards Building Health Management and Energy Evaluation

Chenghao Xu — 2026-07-03

Prognostics and Health Management (PHM) for buildings requires accurate digital representations that support energy-efficiency assessment and predictive analysis throughout the asset lifecycle. This is particularly important for building envelopes, whose condition directly affects thermal performance, energy efficiency, occupant comfort, and long-term structural health. However, most existing buildings lack simulation-ready digital models, while current inspection and modeling workflows remain labor-intensive, geometrically limited, and inadequate for integrating thermal and physical information. This thesis proposes a cost-efficient and automated framework for reconstructing simulation-ready, multi-modal three-dimensional building representations from visual inputs alone. The research is organized around three main contributions. First, it develops semantic scene reconstruction methods that combine neural implicit surface modeling with semantic prediction to recover detailed 3D building envelopes and estimate key characteristics such as window-to-wall ratio and footprint. Second, it investigates learning-based estimation of thermophysical parameters through implicit thermal field reconstruction and differentiable heat-transfer simulation. Third, it addresses practical sensing constraints by enabling multi-modal reconstruction from limited and unsynchronized thermal observations.

Progressive Physics-AI Hybrid Methodology for Unsupervised Anomaly Detection in Electromechanical Systems

Melvin FERNANDES NOVO — 2026-07-03

Physics-informed machine learning has emerged as a promising
paradigm for industrial health monitoring, yet practical
guidance on when and how to integrate domain knowledge
into detection pipelines remains limited. This paper proposes
a structured methodology for progressive physics integration
in unsupervised anomaly detection, organised into three levels
of increasing depth: data-driven baselines (Level 0), operational
conditioning (Level 1), and structural physics injection
(Level 2). The methodology is designed for systems
where qualitative expert knowledge is available but no quantitative
degradation model exists. It is applied systematically
across three method families—statistical envelopes, principal
component analysis, isolation forests, deterministic autoencoders,
and variational autoencoders—for the monitoring of
medium-voltage circuit breaker coil currents. At the highest integration
level, a physics-informed conditional variational autoencoder
(PicVAE) incorporates domain knowledge through
phase-segmented inputs, FiLM-conditioned architecture, and
a phase-weighted reconstruction loss. Validated on real operational
data with expert-labelled anomalies, the results reveal
two findings: operational conditioning at Level 1 consistently
improves detection across all method families, while structural
physics injection at Level 2 has a method-dependent impact,
yielding clear gains for phase-aware representation learning
while introducing trade-offs for simpler models. The PhD
plan extends this work along three axes: quantitative physics
integration on a second use case, cross-domain validation on
public benchmarks through a human-in-the-loop knowledge
acquisition protocol, and consolidation of the methodology
[Melvin Fernandes Novo] et al. This is an open-access article distributed under
the terms of the Creative Commons Attribution 3.0 United States License,
which permits unrestricted use, distribution, and reproduction in any medium,
provided the original author and source are credited.
into a transferable deployment framework.

Physics-Informed Machine Learning for Robust Industrial Diagnostics: A Systematic Investigation Using Heat Pump Systems

Savvas Eftychis — 2026-07-03

Industrial prognostics systems must operate reliably under real-world constraints: limited labeled data, shifting operational conditions, and deployment across heterogeneous units. While Industrial Internet of Things (IIoT) -enabled systems generate vast sensor data, purely data-driven approaches lack the robustness to exploit it effectively, as they tend to overfit to training distributions and fail when conditions change. Physics-Informed Machine Learning (PIML) offers a principled solution by grounding learned models in physical laws, making them more transferable and interpretable. This thesis investigates whether physical knowledge can provide a robust foundation for industrial diagnostics, using heat pump systems to study generalization across operating conditions and units.

Time-Varying Mesh Stiffness Modelling for Multi-Fault Spur and Helical Gear Diagnostics: An Integrated Analytical, Numerical, and Experimental Framework

Maruti Patil — 2026-07-03

Gear Time-Varying Mesh Stiffness (TVMS) is the primary internal excitation governing gear dynamic response. It is the physical bridge between fault severity and measurable vibration signatures. That is why its accurate modeling is important. An accurate model of TVMS may lead to accurate prediction of the gearbox system’s response. Most existing TVMS formulations are restricted to spur gears with single, idealized faults and limited experimental validation. This research aims to develop a physics-consistent, slice-coupled analytical TVMS framework for spur and helical gears, incorporating compound multi-location faults (pitting, tooth cracking), along with elastohydrodynamic (EHL) lubrication corrections, profile shift, and tooth lead deviations to accurately represent real gears used in practice. A three-phase methodology, comprising improved analytical model development, multi-source validation (Finite Element Analysis (FEA), Multibody simulations (MBS), precision test rig), and signal processing for fault diagnosis, establishes an experimentally grounded framework for vibration-based gear health monitoring, directly relevant to PHM practice.

Towards Early and Reliable Detection of Thermal Degradation in High-Precision Machine Tools via Hybrid Condition Monitoring

Darío Fernández — 2026-07-03

Thermal Condition Monitoring (TCM) provides a means to monitor the operation of precision machine tools on a continuous
basis to maintain micron level accuracy and allow for the identification of the early stages of component degradation.
Gradual changes in the mechanical aspects of friction, lubrication, pre-load and wear all contribute to variations in both
heat generated and transferred which directly affect the location of the tool center point (TCP). While conventional TCP
correction models are highly effective at compensating for instantaneous positioning errors in real time, their residuals are
heavily influenced by reversible operational and environmental factors. Consequently, these residuals alone do not allow
reliable differentiation between reversible variations and irreversible, slow-evolving degradation processes. This research
will propose an innovative hybrid model combining physical based thermal modeling for increased interpretability, with
data driven methods to improve the sensitivity of progressive changes. Data-driven components will analyze the machine’s
full operational history to identify how the system evolves over time and the normal patterns and relationships between
variables. Key innovations include (i) the systematic determination of residuals, and (ii) a dual-time scale methodology
that isolates fast transient thermal responses from slower degradation processes. In this manner, the framework utilizes
TCP correction models as baseline diagnostics to extract physically meaningful degradation parameters that can
be monitored along with residuals. Preliminary modeling results demonstrate the effectiveness of the hybrid model separating the reversible and irreversible effects.

Toward Intelligent Prognostics and Health Management for Floating Offshore Wind Turbines

Son Hai Nguyen — 2026-07-03

Floating offshore wind turbines (FOWTs) enable the exploitation of deep-water wind resources where conventional fixed-bottom foundations become technically or economically infeasible. While this technology significantly expands the potential of offshore renewable energy, it also introduces new challenges for reliable operation and maintenance due to harsh marine environments, complex aero–hydro–servo–elastic dynamics, and limited operational data availability. Although AI-driven Prognostics and Health Management (PHM) has achieved substantial progress for conventional wind turbines, its application to FOWTs remains relatively limited. This doctoral research proposes an intelligent PHM framework specifically developed for floating wind systems, addressing key challenges related to limited operational data and domain knowledge, heterogeneous and unreliable monitoring data, and dynamic environmental complexity. Domain adaptation combined with knowledge graph construction, multimodal learning for heterogeneous data integration, and physics-informed machine learning for structural dynamic modeling are investigated as complementary methodological contributions. These components are progressively integrated into a unified PHM lifecycle pipeline supporting fault detection, remaining useful life prediction, and maintenance decision support for FOWTs.

PHME 2026 Management Team and Publisher Information

DO — 2026-07-03