Annual Conference of the PHM Society

Integration of LLMs for Multitasking Workload Prediction in Mixed Reality Environments

Safanah Abbas, Heejin Jeong, David He — Sun, 26 Oct 2025 00:00:00 +0000

Multitasking in mixed reality (MR) environments introduces unique cognitive demands, particularly in workload management. Accurate workload prediction is critical for optimizing user experience, safety, and performance in such settings. This study proposes a novel framework that integrates large language models (LLMs) with traditional workload assessment tools to enhance prediction accuracy in MR multitasking scenarios. A multitasking experiment involving 36 participants was conducted, combining real-world and virtual tasks, with workload evaluated using NASA-TLX. To address limited sample sizes, synthetic data was generated using generative adversarial networks (GANs), enabling robust model training. We employed a hybrid deep learning model that integrates LLM-generated text embeddings with numerical features in a feedforward neural network (FNN). Results show that integrating LLMs, specifically BERT and GPT-2, significantly improves workload prediction accuracy, with a root mean square error (RMSE) reduction from 6.82 (FNN-only) to 0.95 (BERT-integrated model). The findings underscore the potential of LLMs to augment cognitive workload assessment, supporting more adaptive and scalable human-machine collaboration in MR environments.

Mission Profile Clustering for Usage-Based Health Modeling of Flight Control Actuators Applied to a Fleet of Advanced Jet Trainers

Sun, 26 Oct 2025 00:00:00 +0000

This work introduces a mission profile clustering pipeline aimed at supporting usage-based health modeling of electro hydraulic flight control actuators employed in a fleet of Advanced Jet Trainer (AJT) aircraft. The study is part of a broader, high-level, modular Prognostics and Health Management (PHM) framework developed to predict Unscheduled Removals (URs) of the AJT horizontal tail flight control actuator. Operating in an industrial setting, this PHM effort specifically addresses the challenge of extracting prognostic information from a legacy fleet already in service, leveraging existing operational data to improve asset availability.
The overall project leverages an extensive real-world dataset that spans over ten years and more than 25,000 flight hours accumulated by a fleet of as many as 20 aircraft. This paper specifically focuses on the Flight Clustering Module within the Data Processing Layer of the PHM framework, which serves as a critical enabler for future feature projections.
Through an in-depth analysis of the underlying principles and a detailed overview of the main system interfaces, this work proposes a practical solution for categorizing and classifying mission profiles while highlighting the challenges of working with real operational data.
After a pre-processing pipeline, developed to standardize and align time-series flight data, the clean trends are then clustered via a Self-Organizing Map (SOM). In this work, a systematic SOM hyperparameter tuning pipeline is also introduced. The tuning routine employs a combined grid and random search strategy to optimize the SOM hyperparameters by simultaneously evaluating the topographic error, the quantization error, and the percentage of grid utilization. The result of the application of the trained SOM on the dataset is a set of Clustered Mission Types (CMTs), each linked to specific statistical distributions of actuator usage increments. These clusters are integrated into the broader PHM framework to simulate future aircraft behavior and estimate component degradation.
Placed in an operational industrial environment, this methodology effectively connects mission-specific usage patterns with predictive health modeling, improving the fidelity of PHM systems, and laying the foundation for smarter usage-based maintenance planning in aviation operations.

Event-Based Data in Prognostics and Health Management

Thi-Tinh Le, Phuc Do, Benoit Iung, Marcos E. Orchard — Sun, 26 Oct 2025 00:00:00 +0000

In the modern industrial context, Prognostics and Health Management (PHM) systems based on data-driven approaches have been widely and effectively developed to reduce maintenance costs. However, continuous data requires large memory capacity and high costs. Therefore, in recent years, the use of event-based data for PHM models has become prominent and increasingly attracts attention due to its cost-efficiency and effectiveness. This surge in data availability has opened new avenues for developing data-driven methods that leverage event patterns to enhance diagnostic, prognostic, and predictive maintenance capabilities. Building meaningful and interpretable patterns from raw event data is crucial for understanding system behavior, detecting faults early, forecasting future failures, and accurately estimating the Remaining Useful Life (RUL) of critical components. This review paper systematically surveys the state-of-the-art methodologies and frameworks for extracting, modeling, and utilizing event-based patterns in the context of diagnostic and prognostic applications. Furthermore, we analyze challenges related to event data heterogeneity, scalability, and interpretability, as well as the need for robust pattern extraction methods that can adapt to dynamic operating environments. The review further explores how these event-based patterns contribute to building reliable diagnostic models, enabling early fault detection, and supporting maintenance decision-making through precise prognostics.Finally, this paper identifies key research gaps and outlines future directions, emphasizing the need for explainable, adaptive, and scalable pattern mining approaches that effectively translate raw event data into actionable maintenance intelligence. To address these challenges, we propose a conceptual framework that integrates advanced pattern discovery techniques with domain knowledge and feedback loops, enabling continuous learning and decision support. This comprehensive survey aims to serve as a foundational reference for researchers and practitioners committed to leveraging event data for enhanced system reliability and the development of optimized, intelligent maintenance strategies.

Gear Teeth Parameter Identification in Helicopter Planetary Gearbox Using Tachometer and Vibration Signals

Changik Cho, Eric Bechhoefer — Sun, 26 Oct 2025 00:00:00 +0000

Accurate identification of gear tooth counts in planetary gearboxes is essential for condition-based maintenance and health monitoring in gearbox systems. However, direct inspection of internal gear components—especially planet gears—is often infeasible due to the gearbox's enclosed structure and lack of documentation in aging or mixed fleets. This paper presents a methodology to estimate the number of gear teeth in bevel, sun, planet, and ring gears with fixed ring gear planetary gearboxes using vibration and tachometer signals. By applying Time Synchronous Averaging (TSA) and Fast Fourier Transform (FFT), we isolate gear mesh frequencies from noise and harmonics. Three case studies— UH60 Black Hawk, AS350, and Bell 407GXi main gearboxes—demonstrate the application of the technique. For bevel gear estimation, TSA is performed using the input shaft rate, allowing dominant frequency peaks to reveal the gear mesh. In planetary gears, TSA is applied using the output shaft rate. Candidate mesh frequencies are identified, with harmonics and known gear interactions used to eliminate false positives. The most plausible planetary mesh frequency is used to estimate the ring gear tooth count, from which the sun gear tooth count is inferred using the total gear ratio.

Robust health indicator extraction and RUL prediction for PEMFCs under highly dynamic industrial conditions

Soufian Echabarri, Phuc Do, Hai-Canh Vu, Pierre-Yves Liegeois — Sun, 26 Oct 2025 00:00:00 +0000

Proton Exchange Membrane Fuel Cells (PEMFCs) are increasingly deployed in clean energy systems, such as GEH2 hydrogen generators, where they operate under highly dynamic and unpredictable load conditions. Accurate prediction of their Remaining Useful Life (RUL) is essential for ensuring reliable, cost-effective, and proactive maintenance strategies. However, conventional voltage-based Health Indicators (HIs) are highly sensitive to power fluctuations and fail to provide consistent degradation trends in real-world industrial scenarios, particularly when system usage varies significantly across different clients, as in the GEH2 case. In this paper, we propose a scalable two-stage framework for RUL prediction of PEMFCs operating under such conditions. First, we introduce a machine learning-based method to extract a degradation-specific Health Indicator directly from voltage measurements, effectively filtering out transient operational effects. Second, we develop a hybrid deep learning architecture that combines Transformer networks and Gated Recurrent Units (GRUs) to model temporal dependencies and provide accurate RUL predictions under dynamic conditions. The proposed approach is validated on a real-world industrial dataset collected from three PEMFC stacks deployed in GEH2 systems operating under highly variable conditions. Comparative results show that our method consistently outperforms baseline machine learning and deep learning models, achieving superior accuracy, robustness, and generalization across diverse mission profiles.

Evaluating Large Language Models for Turboshaft Engine Torque Prediction

Alessandro Tronconi, David He, Eric Bechhoefer — Sun, 26 Oct 2025 00:00:00 +0000

Recent advancements in deep learning have introduced new opportunities for quality management in manufacturing, particularly through transformer-based architectures capable of learning from limited datasets and handling complex, multimodal inputs. Among these, Large Language Models (LLMs) have emerged as a significant innovation, demonstrating strong capabilities in forecasting and representing the cutting edge of artificial intelligence (AI). Through transfer learning, LLMs effectively process and generate extended text sequences, and recent developments show their potential for multimodal integration, including text, images, audio, and video data.

Quality management is a critical area for industrial innovation, rapidly evolving as manufacturers seek to close the quality-manufacturing loop and achieve zero-defect production goals. While computer vision techniques based on deep learning have been widely implemented for visual inspection tasks, integrating multiple heterogeneous data sources offers the possibility for even greater improvements. Despite the success of LLMs in language tasks, their application to time series data remains relatively unexplored. Alternative statistical approaches and deep learning models have proven effective for time series forecasting. Nevertheless, LLMs could provide additional advantages in industrial contexts, offering opportunities to enhance in-line quality control, defect prevention, and predictive discarding strategies across various sectors.

This paper investigates the potential of applying LLMs to time series analysis by comparing the performance of an LLM (GPT-2), originally trained on textual data, with a model specifically designed for time series data (TimeGPT), and a more conventional transformer-based architecture. Our study includes a dedicated time series GPT model and a general-purpose LLM in a comparative evaluation. Through this analysis, we aim to better understand how language models can be effectively adapted to time series forecasting tasks and explore their transfer learning potential for enhancing quality management in manufacturing.

Prognostics of Rolling Element Bearings based on Cyclostationarity-based Indicators and Kalman filter under Varying Load and Speed

Sun, 26 Oct 2025 00:00:00 +0000

Rolling element bearings (REBs) are key components of rotating machines but the estimation of their remaining useful life (RUL) is still very challenging. First fault detection should be achieved as early as possible and then the RUL should be estimated as accurately as possible. Both steps require dedicated Health Indicators (HIs) which might not be the same when looking towards detection or prognostics. A key property of REB signals is cyclostationarity, as the statistical properties of their vibration behavior vary periodically over time. This characteristic has been effectively exploited to construct HIs for anomaly detection, and fault diagnosis in the field of condition monitoring (CM) achieving high performance. Although a plethora of methodologies have been proposed for RUL estimation, they usually are restricted in cases where the load conditions are assumed steady, reducing significantly their applicability and implementation in industry. Therefore there is a need for methodologies that are able to estimate the RUL of REBs operating under variable and/or varying load and speed conditions. The goal of this paper is the exploration of the performance of different vibration based HIs for fault detection, diagnosis and prognosis, including both time-domain and-order domain features. A dedicated bearing prognostics test rig was used to perform accelerated life tests of a self-aligned bearing, operating under varying load and speed conditions. The speed ranges from 0 to 3000 rpm and the load varies from 0 to 12 kN. The measurements lasted for around 400 hours and the bearing has an outer race fault in the loading zone. Different signals have been acquired during the tests, including accelerations, temperature and strain signals. The results indicate that the cyclic spectral coherence-based indicator is more sensitive to the change of states (healthy or damaged) and thus better for fault detection, while the correlation-based indicator is more sensitive to fault development, and therefore more suitable for the RUL estimation of REBs. Finally, to estimate RUL, different estimators, i.e., the Extended Kalman filter (EKF) and the Adaptive Kernal Kalman filter (AKKF), are used for RUL estimation.

Fault Identification Using System-Level Insights and Multi-Layered Classification

Ryan Aalund, Meenu Rajapandian, Vincent Paglioni — Sun, 26 Oct 2025 00:00:00 +0000

This paper presents a novel fault detection framework for IoT device fault detection, combining meta-algorithmic decision logic with a neural network-based classifier to enable efficient, scalable failure analysis. Leveraging system-level data, the methodology adopts a multi-layered architecture grounded in systems thinking to classify devices as failed or non-failed and then identify the root cause domain—hardware, software, or firmware.

The first layer implements a meta-classifier that integrates multiple lightweight algorithms weighted by application-specific criteria such as accuracy, precision, or recall. This ensemble approach capitalizes on the strengths of diverse classifiers to enhance fault detection performance using high-level system metrics. The second layer introduces a neural network trained on subsystem-specific features—such as power metrics, SoC diagnostics, and LTE module health—to infer the most probable root cause category. This structure not only boosts classification accuracy but also captures interaction effects across subsystems through derived features.

Demonstrated on real-world telematics devices that collect GPS and vehicle diagnostic data over cellular networks, the framework addresses the need for scalable diagnostic methods in high-volume, low-failure-rate environments. By minimizing unnecessary returns and streamlining corrective action workflows, this approach delivers practical value in field operations.

The modular nature of the two-tiered architecture enables adaptability to a range of device types and fault modes, while future work will explore model generalization across varied deployments. The integration of neural networks for subsystem-level classification offers a pathway to more nuanced root cause analysis, reinforcing the importance of structured, data-driven approaches to operational reliability.

Impact of Environmental Temperature Variation on Vibration- Based Fault Detection for Air Compressors

Sherwood Polter, Colin Dingley, Neil Eklund, Adam Wechsler — Sun, 26 Oct 2025 00:00:00 +0000

Condition-Based Maintenance Plus (CBM+) aims to enhance operational readiness for U.S. Navy assets by using predictive models to forecast equipment failures. Applying CBM+ in the U.S. Navy faces a unique challenge: ships operate globally for extended periods, exposing machinery to a wide range of ambient air and seawater temperatures that alter their characteristic vibration signatures and can compromise model performance. This paper investigates the extent to which these seasonal temperature variations degrade the performance of a vibration-based fault detection model for a naval air compressor. Using data from controlled testing, vibration data was collected under healthy and various induced-fault conditions during both winter and summer to create two environmentally distinct datasets. Power Spectral Density analysis was used to extract features for training classifiers. Results show that models trained exclusively on data from one season performed poorly when tested against data from the other, confirming that environmental shifts significantly degrade predictive accuracy. In contrast, a model trained on a combined dataset incorporating data from both seasons demonstrated substantially improved and more generalized performance. These findings underscore that the development of robust, field-ready CBM+ systems is critically dependent on training ML models with comprehensive and environmentally diverse datasets that reflect the full spectrum of anticipated operational conditions.

A Contrastive Learning Approach for Anomaly Detection in Multi-View Scenarios

Sun, 26 Oct 2025 00:00:00 +0000

Quality control is a key task in smart manufacturing, since it ensures that processes consistently meet rigorous performance standards. The effective implementation of these mechanisms is crucial to ensuring both reliability and efficiency in modern manufacturing environments, where automation is increasingly integrated. Traditional anomaly detection algorithms typically rely on single-view data for each manufacturing product, overlooking relevant and complementary information available from multiple perspectives. Furthermore, cross-entropy-based loss functions are frequently adopted in the literature to train detection models; however, these approaches often struggle with imbalanced datasets or when detecting rare and subtle anomalies. In this work, a contrastive learning architecture for multi-view anomaly detection in industrial settings is proposed. The method performs a mid-level fusion to generate a structured representation of the input instances, thereby enhancing detection capabilities. The architecture was evaluated on the Real-IAD dataset, where it demonstrated better performance than traditional techniques. These findings highlight the potential of contrastive learning to improve anomaly detection performance, thus contributing to the construction of more reliable quality control systems in smart manufacturing environments.

Exploring LLM-based Agentic Frameworks for Fault Diagnosis

Xian Yeow Lee, Lasitha Vidyaratne, Ahmed Farahat, Chetan Gupta — Sun, 26 Oct 2025 00:00:00 +0000

Large Language Model (LLM)-based agentic systems present new opportunities for autonomous health monitoring of industrial systems in sensor-rich environments. This study investigates the potential of LLM agents to diagnose faults directly from raw sensor data, while producing inherently explainable outputs through natural language reasoning. Such explainability enables users to interpret and audit agent decisions and confidence levels with greater transparency. We begin by systematically analyzing how different agent configurations, such as centralized versus distributed agent setups, the ability to use computational tools for fault discovery, and the structure and scope of sensor input data impact fault detection accuracy and uncertainty estimation. Building on these findings, we then explore whether LLM agents can improve diagnostic performance over time through continual learning, calibrating their confidence based on historical ground truth outcomes. Through simulation-based experiments across varied degradation scenarios, this work aims to assess the feasibility of LLM-based agents as a foundation for transparent, adaptive fault diagnosis in real-world systems.

Detectability of Damages in Carbon Fiber Reinforced Plastics using Acoustic Emission

Jonathan Liebeton, Dirk Soffker — Sun, 26 Oct 2025 00:00:00 +0000

This contribution examines the usefulness of Acoustic Emissions (AE) as a non-destructive testing (NDT) method for detecting and distinguishing damages in carbon fiber reinforced polymer (CFRP) structures. Despite the widespread use of CFRP materials in various industries due to their favorable strength-to-weight ratio, the susceptibility to concealed internal damages necessitates advanced inspection techniques. Acoustic Emission, describing the use of ultrasonic waves emitted during deformation or damage events, is a proven and promising solution for real-time and reliable damage assessment. The study focuses on comparing two approaches: 1) a one-class Support Vector Machine (SVM) for initial damage detection, followed by detailed damage classification, and 2) a direct classification approach using five classes (four representing the material specific damage types and one for background noise). Both approaches undergo a systematic evaluation under diverse loading conditions to assess their reliability. A comprehensive experimental setup subjects CFRP specimens to controlled loading conditions, inducing various damage types and severities. Signal analysis reveals characteristic patterns associated with different damage modes, including matrix cracking, fiber breakage, debonding, and delamination. The investigation considers the influence of loading conditions on the detection and classification results to examine the robustness of the approach. The comparison between methodologies involves established metrics and analyses the posterior probability of the trained models, considering the impact of loading conditions on performance. The experimental results show AE’s effectiveness in detecting and classifying damages in CFRP structures, offering insights into technique sensitivity and specificity for different damage types. These findings contribute new knowledge to the NDT field, presenting a promising path for the advancement of CFRP structural health monitoring and maintenance practices in engineering applications. The study’s nuanced understanding of the strengths and limitations of the two classification approaches, considering loading conditions, contributes to the optimization of NDT strategies for diverse operation scenarios.

Bayesian Physics Informed Neural Networks for Reliable Transformer Prognostics

Ibai Ramirez, Jokin Alcibar, Joel Pino, Mikel Sanz, David Pardo, Jose Aizpurua — Sun, 26 Oct 2025 00:00:00 +0000

Scientific Machine Learning (SciML) integrates physics and data into the learning process, offering improved generalization compared with purely data-driven models. Despite its potential, applications of SciML in prognostics remain limited, partly due to the complexity of incorporating partial differential equations (PDEs) for ageing physics and the scarcity of robust uncertainty quantification methods. This work introduces a Bayesian Physics-Informed Neural Network (B-PINN) framework for probabilistic prognostics estimation. By embedding Bayesian Neural Networks into the PINN architecture, the proposed approach produces principled, uncertainty-aware predictions. The method is applied to a transformer ageing case study, where insulation degradation is primarily driven by thermal stress. The heat diffusion PDE is used as the physical residual, and different prior distributions are investigated to examine their impact on predictive posterior distributions and their ability to encode a priori physical knowledge. The framework is validated against a finite element model developed and tested with real measurements from a solar power plant. Results, benchmarked against a dropout-PINN baseline, show that the proposed B-PINN delivers more reliable prognostic predictions by accurately quantifying predictive uncertainty. This capability is crucial for supporting robust and informed maintenance decision-making in critical power assets.

Partial Domain Adaptation for Intelligent Machinery Fault Diagnosis

Sun, 26 Oct 2025 00:00:00 +0000

Accurate gearbox fault diagnosis across different operating conditions plays an important role in prognostics and health management. In real industrial scenarios, a common challenge arises when the source domain contains multiple fault classes, while the target domain includes only healthy samples during training. To address this issue, this study proposes a unified industrial fault diagnosis framework designed to handle the partial domain adaptation problem. Specifically, the overall framework involves: a unified data processing pipeline, a robust deep learning architecture for accurate fault classification, and integration of maximum mean discrepancy loss to align feature distributions between source and target domains. Experimental results demonstrate that our proposed partial domain adaptation-based deep learning model significantly outperforms benchmark models, achieving accuracy improvements exceeding 20% across multiple domain adaptation tasks. This study provides a practical solution for intelligent gearbox diagnosis under domain shift constraints.

Real-time Thermal Runaway Prognosis of a Lithium-ion Battery via Physics-informed Latent Ensemble DeepONet with Segmented Data

Jinho Jeong, Eunji Kwak, Jun-Hyeong Kim, Ki-Yong Oh — Sun, 26 Oct 2025 00:00:00 +0000

This study proposes a novel architecture of physics-informed latent ensemble deep operator network with segmented data. The proposed neural network aims to predict thermal runaway of a lithium-ion battery through prior temperature responses in real-time. The proposed neural network introduces three key features to provide an advanced control-enabling solution for battery thermal management systems (BTMS). First, the proposed neural network addresses the architecture of DeepONet as a surrogate model to effectively learn the internal temperature, chemical component concentration, and gas formation under supervision of complex and nonlinear multiphysics representing thermal runaway of lithium-ion batteries. This approach enables accurate and robust virtual sensing capability even with limited data by constraining the prognostic responses to follow the governing equation of the underlying multiphysics. Second, a dual-network architecture is introduced to extract valuable features from prior temperature responses, which inherently contain limited information in real-time scenarios. The network comprises two sub-networks; the first network extracts latent features from decomposed temporal domains across diverse local domains, and the second network ensemble these features to original features for mitigate concerns on overfitting and generalization. This approach ensures effective supervision by stiff governing equations in both local and global domains. Third, novel methods are employed to reduce the training complexity associated with integrating multiphysics equations including separate DeepONet, stan activation function, adaptive weights, and encoders. These methods enhance the expressiveness of temporal and spatial gradients that play an important role in physics-informed neural networks. Hence, this feature not only ensures convergence through a balanced learning but also improves the overall capability of the neural network. Extensive ablation studies validate the contribution of each feature, and thereby confirm the effectiveness of novel architecture and strategies in addressing failure issues in physics-informed neural networks. The proposed method enables real-time prognosis through prior thermal responses, offering a promising pathway toward artificial intelligence transformation in BMS to ensure the safety and efficiency of lithium-ion batteries.

Bearing Spall Size Estimation Under Varying Speed Conditions

Cees Taal, Dang Ngo-The, Jérome Antoni — Sun, 26 Oct 2025 00:00:00 +0000

Accurate estimation of spall size in rolling element bearings is critical for effective diagnostics and prognostics in rotating machinery. Traditional methods often struggle with generalization due to noise and speed variability. This work addresses these limitations by proposing a novel approach that leverages trends in vibration measurements over time and introduces a speed-normalized condition indicator. Building on prior work, we model the bearing fault signal as a periodic pulse wave and derive a Fourier-based representation that links harmonic magnitudes to spall size. We then introduce a normalization technique using harmonic speed ratios to eliminate the influence of the system’s transfer function. Experimental validation using controlled lab data confirms the method’s ability to preserve signal extrema and improve generalizability over different speeds, offering a promising path toward scalable, real-world bearing health monitoring.

A Hierarchical Agentic Framework for Autonomous Drone-Based Visual Inspection

Sun, 26 Oct 2025 00:00:00 +0000

Autonomous inspection systems are essential for ensuring the performance and longevity of industrial assets. Recently, agentic frameworks have demonstrated significant potential for automating inspection workflows but have been limited to digital tasks. Their application to physical assets in real-world environments, however, remains underexplored. In this work, we propose a hierarchical agentic framework for autonomous drone guidance, focusing on visual inspection tasks in indoor industrial settings, such as interpreting industrial readouts or inspecting equipment. Our framework employs a multi-agent system comprising a head agent and multiple worker agents, each controlling a single drone. The head agent performs high-level planning and evaluates outcomes, while the worker agents reason over and execute low-level actions. Operating entirely in the natural language space, the framework follows a plan, reason, act, evaluate cycle, enabling drones to handle tasks ranging from simple navigation (e.g., flying forward 10 meters and land) to complex high-level task (e.g., locating and reading a pressure gauge). The head agent’s evaluation phase serves as a feedback and/or replanning stage, ensuring the actions executed align with user objectives while preventing undesirable outcomes. We evaluate the framework in a simulated environment with two worker agents, assessing performance qualitatively and quantitatively based on task completion across varying levels of task complexity and agentic workflow efficiency. By leveraging natural language processing for agent communication, our approach offers a novel, flexible, and user-accessible alternative to traditional drone-based solutions, enabling a more autonomous problem-solving approach to industrial inspection tasks without requiring extensive user intervention.

A Dual-Contrastive-Attention Transformer for Unsupervised Anomaly Detection in Lamb Waves Structural Health Monitoring

Jiawei Guo, Boshi Chen, Sen Zhang, Nikta Amiri, Ge Song, Lingyu Yu, Yi Wang — Sun, 26 Oct 2025 00:00:00 +0000

Lamb wave-based Structural Health Monitoring (SHM) is a promising technique for detecting defects in materials and structures. However, traditional methods often rely on computationally intensive signal processing and struggle to detect subtle anomalies wave patterns. In this work, we propose a novel transformer-based framework, called Dual-Contrastive-Attention Transformer (DCAT), for unsupervised anomaly detection in Lamb wave data. DCAT uses two attention branches during training: a Global-Context Attention (GCA) branch that captures long-range patterns, and a Local-Context Attention (LCA) branch that serves as a constraint. A contrastive loss is used to prevent the global branch from over-learning local features, encouraging it to focus on the overall structure. Both branches are trained to reconstruct the input, using a structural similarity (SSIM) loss that better reflects waveform patterns than traditional mean squared error. After training, only the global branch is retained for inference. Anomalies are detected by comparing the input and reconstructed output. Since the global branch cannot easily reproduce local defects, it produces a higher SSIM loss when anomalies are present. We test our model on a Lamb wave dataset with multiple types of defects. DCAT achieves 97.8% accuracy and a precision of 98.6%, outperforming other SOTA baselines. These results show that DCAT is well-suited for accurate Lamb wave-based SHM without the need for labeled data.

LLMs as Pre-trained Models for Time-Series Applications in PHM

Takanobu Minami, Dai-Yan Ji, Lee Jay — Sun, 26 Oct 2025 00:00:00 +0000

In industrial Prognostics and Health Management (PHM), the scarcity of sufficiently large, high-quality datasets remains a persistent challenge, which limits the practical deployment of machine learning-based approaches. Recent efforts to address this include few-shot learning, domain adaptation, and other data-efficient learning paradigms. However, while the use of pre-trained models has shown great promise in other fields such as natural language processing and computer vision, its application in PHM remains relatively underexplored. Although initial studies have begun to introduce foundation-style models for specific components—such as recent efforts on bearing health diagnostics using transformer-based architectures—their development is still in an early stage compared to the maturity and versatility of large language models (LLMs) in the NLP domain. LLMs continue to advance rapidly, offering generalization capabilities that could be highly beneficial in data-constrained PHM settings. While some preliminary research has explored the use of LLMs as intelligent agents for decision support in PHM workflows, their application as direct learners for time-series sensor data remains rare. In this work, we propose a novel framework that adapts pre-trained LLMs for time-series-based PHM tasks. Our approach involves mapping temporal sensor signals to a tokenized format compatible with transformer-based language models, enabling the application of LLMs as generic sequence learners. Building on recent pioneering concepts such as multimodal LLM-based health management systems and prompt-driven signal encoding, our framework is benchmarked on publicly available industrial datasets under low-data conditions. The results demonstrate that our LLM-based approach not only maintains robust performance in scenarios with limited labeled data but also outperforms traditional models in fault classification accuracy. This study contributes a new perspective to the PHM community by highlighting the untapped potential of LLMs as general-purpose, pre-trained models in industrial health monitoring applications. Our findings suggest that incorporating LLMs into PHM workflows can be a powerful and forward-looking strategy to overcome data scarcity and improve adaptability across diverse operational domains.

A Novel Methodology for Vision Backbone Network Fine-Tuning and Continual Learning in Optical Inspection Tasks

Kody Haubeil, Tarek Yahia, Alex Suer, David Siegel, Donald Davis, Xiaodong Jia — Sun, 26 Oct 2025 00:00:00 +0000

Roll-to-roll (R2R) manufacturing plays a key role in the
production of electronic devices, liquid-crystal films, and
other advanced materials. In these processes,
maintaining precise material thickness and surface
uniformity is essential to ensure product quality, process
efficiency, and yield. While optical inspection systems are
widely used to detect surface defects, using AI to
automatically segment and classify the defect remains a
challenging task. This paper investigates a new
methodology to establish an AI powered optical
inspection testbed that automatically handles multi-label
defect classification and semantic segmentation in R2R
manufacturing environments. The proposed system
integrates semantic segmentation and multi-label defect
classification into an automated optical inspection
pipeline. Using a Vision Transformer deep learning
architecture, our classification model achieved a high
accuracy across multiple surface defect categories in a
production-representative dataset. Our framework
consists of automatic image pre-processing, rapid fine
tuning of pre-train vision models, and a continuous
learning strategy after the model is deployed. The system
is intended to handle high-throughput image data with
minimal latency, enabling fast and accurate product
inspection while maintaining a light-weight model
architecture. The end-to-end framework is designed to
enhance in-line defect detection, reduce human
dependency, and improve overall process visibility. The
proposed technology has been successfully deployed to a
manufacturing plan and has achieved satisfactory
detection and classification performance in a real
operation environment.

A Lithium-Ion Battery Degradation Model Agnostic to Cell Chemistry with Integrated State-of-Charge and Temperature Dependence

Sun, 26 Oct 2025 00:00:00 +0000

Lithium-ion battery (LIB) lifetime prognosis remains a decisive enabler for safe, economical, and sustainable electrification across mobility and stationary energy sectors. Despite the maturity of this storage technology, the impact of capacity reduction due to battery aging remains a subject of interest to this day. This study presents a degradation model that computes the incremental loss of usable capacity for every equivalent cycle experienced by the battery. The model depends on the State of Charge values and the mean ambient temperature at which the cycle occurred. The methodology generalizes the results of a previous study where an initial degradation model was calibrated on long‑term cycling data at which multiple commercial cells were driven to their end of life under varied state of charge ranges in addition to a chemical analysis of cell's degradation due to different operating temperatures. Unlike traditional calendar-plus-cycling additives or chemistry-specific semi-empirical methods, the proposed model is formulated—and further validated—under the assumption that cells exposed to comparable operating conditions degrade at comparable rates, independent of their electrode chemistry. Consequently, only the data of a single degradation campaign at a given operational mode is needed, which is typically found in the cell's datasheet, eliminating the need for time-consuming aging tests. Validation is performed using a public experimental dataset provided by the University of Stanford dataset. The results show the model's ability to predict the end-of-life within one percent of the reported battery State of Health (SOH), even for an electric vehicle usage profile. By embedding the framework in battery management systems, fleet operators can estimate remaining useful life (RUL) based on historical operational data, optimize battery usage based on less degrading SoC values and cooling systems, schedule preventive maintenance before catastrophic capacity loss, and compare alternative uses without the need for degradation studies. These contributions align with conference themes on prognostics, health management, and data-driven energy system optimization, offering a practical pathway towards longer-lived and better-valued electrified assets.

A Novel 3D Sensing Framework for Safety Monitoring in Human-Robot Collaboration Work Cells

Tarek Yahia, Kody Haubeil, Alex Suer, Yongzhi Qu, Janet Dong, Xiaodong Jia — Sun, 26 Oct 2025 00:00:00 +0000

The demand for work safety protection in Human-Robot Interaction (HRI) work cells is rapidly increasing, driven by the projected 34.3% Compound Annual Growth Rate (CAGR) of the global Collaborative Robot (Cobot) market from 2020 to 2030 [1]. According to IRF-World Robotics 2023, it is reported that there are nearly 4 million industrial robots in operation worldwide, with approximately 10% of them being cobot [2]. A NIOSH report highlighted 61 robot-related fatalities between 1992 and 2015, with an expectation of further rising due to the increasing use of industrial robots and cobots in the US work environment [3]. A recent study in [4] delved into 355 robot accidents documented by KOSHA between 2009 and 2019, revealing that 95% occurred in manufacturing businesses. Pinch and crush incidents accounted for 52% of the accidents, while impacts and collisions accounted for 36%, and the remaining 12% involved falls, flying objects, trips/slips, cuts, burns, etc. These findings align with US data reported in [5].
The rising integration of cobot units among major manufacturers emphasizes the critical need for enhancing cobot safety in manufacturing. Owing to safety considerations and regulatory requirements, existing cobots frequently operate at significantly reduced speeds and are restricted from undertaking complex interaction tasks in shared workspace. This limitation has curtailed the full potential utilization and productivity of cobots in manufacturing. This paper introduces a novel 3D sensing framework designed to address these limitations by enabling safety assurance in workspaces requiring close human-robot interaction. The framework generates 3D human pose information and relays it to the robot for real-time safety monitoring. Our methodology begins with data collection from a single RGB-D camera capturing human-robot interactions in a manufacturing environment. Human shape and pose are predicted using deep neural networks, which then incorporate depth information and undergo 3D geometric transformations to deduce size, shape, and translation. This process produces a reconstructed 3D avatar with pose, size, and location. Following 3D human posture estimation, this data is then integrated into a virtual environment with a real robot for real-time monitoring. Results demonstrate successful reconstruction of 3D human geometry within human-robot collaboration settings. By integrating both the reconstructed mesh and real-time robot state into a unified virtual environment, we achieved real-time, offline, continuous monitoring of the critical distance between robot and human throughout operation. These distance measurements provide crucial data for developing collision detection, prediction, and avoidance capabilities when incorporated into the robot control feedback loop.

EV-sim

Sun, 26 Oct 2025 00:00:00 +0000

Existing open-source traffic tools accurately reproduce driver behavior and congestion for conventional internal-combustion vehicles. However, in the case of electric vehicles (EVs), they often fail to incorporate critical electrical variables, such as battery voltage, power demand, and State-of-Health, which limits their applicability in operational planning and decision-making. This paper introduces a simulation platform tailored for EVs that bridges the gap between traditional transportation models and the needs of the PHM community in electromobility. The proposed platform combines power and energy consumption profiles derived from Gaussian Mixture Models with physics-based representations of battery behavior. Model parameters are calibrated using a publicly available dataset collected in Ann Arbor, Michigan. Each trip is partitioned into segments based on abrupt changes in speed, ensuring uniform operating conditions within segments and enhancing model transferability across routes. The platform simulates vehicle speed, electrical power demand, State-of-Charge (SoC), terminal voltage, and incremental capacity loss at each simulation step. Battery degradation is estimated through an empirical model fitted to long-term cycling data. A case study demonstrates the simulator’s ability to compare route alternatives between a shared origin and destination. Results show that the shortest path is not always the most energy-efficient nor the least degrading, highlighting the value of health-aware routing. The platform will be publicly released to enable reproducible testing of SoC estimation, range prediction, and degradation forecasting without requiring extensive instrumentation or prolonged field testing.

MAINTAG - Multi-Agent-based Predictive Maintenance Dataset Tagging System

Oguz Bektash, Dorian JOUBAUD, Chetan. S. Kulkarni, Sylvain KUBLER — Sun, 26 Oct 2025 00:00:00 +0000

With the ongoing digitization of global activities, the number of predictive maintenance datasets has been steadily growing. These datasets are often manually classified – in literature review papers – to assess their relevance for predictive maintenance applications. However, this manual approach is increasingly unsustainable, as it is time-intensive, prone to errors, and the accelerating pace at which new datasets emerge in both scientific and industrial contexts. To overcome these challenges, there is a growing need for automated solutions to curate, analyze, and categorize (tag) datasets in the literature. To this end, we propose and evaluate MAINTAG (Multi-Agent-based Predictive Maintenance Dataset Tagging System), a novel multi-agent system designed to automate the classification of predictive maintenance datasets. MAINTAG is compatible with any criteria-based taxonomy and is assessed by benchmarking its tagging accuracy against recent state-of-the-art literature. This evaluation highlights MAINTAG’s ability to replicate expert-level dataset tagging.

MAINTAG resolves three critical challenges in predictive maintenance dataset tagging. The first one arises from the dramatic increase in the number of datasets. The recent growth in predictive maintenance applications has resulted in higher data volume. This makes manual tagging methods impractical and unscalable. Researchers, therefore, will struggle to manually process this flood of new data. The second challenge is related to the complexity of data classification. Each dataset consist of different data types, sources, and structures. These create additional classification difficulties. The third challenge is keeping data tagging consistent across different domains. Predictive maintenance datasets come from multiple industries, and each has its own set of standards and/or classification criteria. MAINTAG proposes a specialized multi-agent framework in response to these challenges. The framework divides the complex classification tasks into simpler and more manageable parts. The system uses specialized agents working in parallel to handle different classification tasks. These agents work together to analyze different aspects of the datasets. This system marks a major leap in autonomous data tagging. It also sets the foundation for a more scalable approach for handling surge in the predictive maintenance datasets.

Characterizing Surface-damage Progression of Spur Gears with Vibration and Oil Data

Sun, 26 Oct 2025 00:00:00 +0000

We present an empirical investigation of the gradual progression of surface damage in spur gears using two data sources: oil monitoring and vibration measurements. The test stand was equipped with a commercial magnetic filter, and a novel test process was developed to remove particles from the magnetic filter and suspend it in oil. In addition, oil samples were drawn periodically to analyze using LaserNet Fines. Both data-driven and classical vibration-based condition indicators were computed and compared to a simple, image-based feature quantifying of the surface condition with some of the indicators showing more than 80\% correlation. Oil analyses found relatively large particles in the particles collected from the magnetic filter.

Rethinking RUL Prediction

Mariana Salinas-Camus, Kai Goebel, Nick Eleftheroglou — Sun, 26 Oct 2025 00:00:00 +0000

Prognostics and Health Management (PHM) plays a key role in predicting the Remaining Useful Life (RUL) of systems, which is essential for enabling decision-making for Predictive Maintenance (PdM) and operations. While most research has traditionally focused on improving the accuracy of RUL predictions, this paper argues that four essential characteristics, uncertainty, robustness, interpretability, and feasibility, are key for real-world PHM applications. This study explores these characteristics through a comparative analysis of two data-driven models (DDMs): the probabilistic Bidirectional Long Short-Term Memory (BiLSTM) model and the Adaptive Hidden Semi-Markov Model (AHSMM). Deep Learning (DL) models such as the BiLSTM often achieve high prediction accuracy but struggle with uncertainty quantification and adaptability across varying operating conditions. In contrast, stochastic models like AHSMM offer stronger robustness and feasibility, performing well even with limited or noisy data. Using the C-MAPSS dataset, the models are evaluated through the lens of the four proposed characteristics. This more holistic approach clarifies each model’s strengths, limitations, and practical trade-offs in PHM settings. The findings highlight that while accuracy remains important, focusing solely on it can overlook critical factors that affect model performance in real operational environments. Balancing all four characteristics is essential for deploying reliable and effective decision-making for predictive maintenance and operations.

Health Monitoring and Drift Detection of Bearing Using Direct Density Ratio Estimation

Sun, 26 Oct 2025 00:00:00 +0000

Prognostics and health management (PHM) has been widely employed for condition monitoring, fault diagnosis and failure prediction in mechanical systems. However, the presence of uncertainty and transient fluctuations in condition monitoring data makes the precise detection of degradation challenging. This paper presents a novel direct density ratio estimation (DDRE) method that computes the change score of the health indicator to detect degradation. The approach continuously computes the change score between two sliding windows using noise-assisted relative unconstrained least-squares importance fitting (NARuLSIF). This study does not rely solely on magnitude of the DDRE-based dissimilarity score; instead, it analyses the dynamic behaviors of the change score to categorize degradations into steady and progressive types. Additionally, this research identifies the onset of runaway failures, referred to as the initial degradation point (IDP), which is used as the starting point for remaining useful life (RUL) estimation. To validate the proposed approach, a publicly available rolling-element bearing dataset is utilized. Experimental results demonstrate the effectiveness and robustness of the proposed DDRE method for both degradation detection and selection of the IDP.

Postprocessing of Autoencoder Reconstruction Error for Detection and Diagnostics of Faults in Infrequently-driven Ground Vehicles

Matthew Moon, Ethan Kohrt, Michael Thurston, Nenad Nenadic — Sun, 26 Oct 2025 00:00:00 +0000

We investigated the detection and classification of engine and transmission faults in infrequently-driven ground vehicles using data-driven methods based on neural network autoencoders. The data came from seventeen vehicles, each with an engine-related or a transmission-related maintenance event. The vehicles had months to years of sensor controller area network (CAN) bus data sampled at 1Hz. Separate autoencoder models were trained for each vehicle to improve detection sensitivity. The paper investigates several condition indicators (CIs) derived from autoencoder reconstruction error, each computed from a sequence of the reconstruction’s mean absolute error (MAE). These CIs were compared using a performance metric computed as the area under the Pareto front with respect to normalized detection horizon and normalized baseline-relative CI margin. A novel detection procedure, consistent detection, effectively filtered out short-duration isolated spikes, likely false positives, while also increasing sensitivity to more plausible anomalies. In addition, the initial development of data-driven diagnostics, based on a novel approach of classifying full reconstruction error vectors associated with the fault state, showed promise but failed our robustness checks.

Data-Efficient and Uncertainty-Aware RUL Prediction Using Physics-Informed Neural Networks

Sun, 26 Oct 2025 00:00:00 +0000

This study presents a prognostic framework that integrates Physics-Informed Neural Network (PINN) with uncertainty quantification (UQ) techniques to enable probabilistic prediction of the Remaining Useful Life (RUL) of rubber components subjected to degradation. The framework utilizes data acquired from thermal Highly Accelerated Life Testing (HALT), replicating long-term material aging behavior under elevated temperature conditions within a shortened time frame. To address the high cost and time consumption of HALT experiments, the proposed approach aims to ensure accurate and reliable predictions even with limited data availability. An empirical degradation model is embedded within the PINN structure, enabling physically consistent and data-efficient estimation of degradation model parameters. The framework employs uncertainty quantification techniques based on Bayesian inference, in which data-driven approaches (e.g., Gaussian Process modeling, Bayesian neural networks) and physics-based methods (e.g., Markov chain Monte Carlo, particle filtering) are separately applied to quantify variations arising from material properties, experimental conditions, and measurement noise. These methods generate posterior distributions from which failure time and probabilistic RUL estimates are derived based on a predefined degradation threshold. Compared to deterministic optimization methods, the proposed approach improves prediction robustness and interpretability, offering a cost-effective and scalable solution for prognostic modeling in engineering systems.

Semantic Framework for IT-OT Integration in Industrial Environments

Anand Todkar, Dr Mrinmoy Sarkar, Jitendra Solanki — Sun, 26 Oct 2025 00:00:00 +0000

This paper presents a semantic framework to bridge the gap between IT-OT integration in industrial environments. The proposed solution addresses fundamental challenges of PHM (prognostics and health management) by providing contextualized semantic information from the shop floor to enterprise IT systems. Built upon an OPCUA (Open Platform Communications Unified Architecture) aggregation server architecture, the framework leverages OPCUA Information Models and companion specifications as its foundation for semantic representation. By transforming these models into knowledge graphs stored in RDF format, the system enables sophisticated semantic information retrieval through SPARQL-based semantic queries that can traverse complex relationships between equipment, processes, and operational parameters. The framework further implements GraphQL to automatically generate a Type schema derived from OPCUA types, creating a unified query interface that facilitates IT-like interaction with industrial data. This semantic approach significantly improves fault diagnostics, predictive maintenance, and anomaly detection by preserving contextual relationships that are often lost in traditional data integration methods. Furthermore, the GraphQL schema provides a structured foundation for generative AI applications to formulate contextually appropriate queries, extract relevant maintenance insights, and generate human-interpretable explanations of equipment health patterns, all while maintaining semantic fidelity across the IT-OT boundary. The vertical integration capability ensures that domain-specific models remain coherent across organizational levels such as line, area, floor, etc., enabling PHM practitioners to implement more effective condition-based maintenance strategies with improved visibility into causal factors affecting equipment reliability and performance.

An Evaluation Framework for Fault Diagnosis Using Technical Manuals in Retrieval-Augmented Large Language Models

Sarah Lukens, Matthew Bishof, Nadir Siddiqui, Destiny West — Sun, 26 Oct 2025 00:00:00 +0000

Fault diagnosis is a time-intensive maintenance task often reliant on the expertise of senior technicians. As this workforce ages and demand grows for digital tools, there is a growing need to capture and automate this knowledge while maintaining the precision required for technical applications. This study introduces an evaluation-driven framework for fault code recommendation, applied to a ground vehicle diagnosis system. Two tasks were designed to reflect potential system configurations: (1) a chat-style task simulating large language model (LLM) interaction, and (2) a label-constrained task using structured fault codes from technical manuals. Multiple retrieval-augmented generation (RAG) configurations were compared against LLM-only and retrieval-only baselines. Results showed that retrieval-based methods outperformed LLM-based ones for label-matching tasks, while the chat task showed challenges in linking observations to fault codes from the manual. These results highlight the importance of aligning task design with evaluation goals and considering retrieval-first approaches as viable alternatives to LLMs in technical language processing (TLP) applications. Beyond experimental findings, we outline industrial lessons learned: the importance of aligning system design to use case goals, adopting evaluation-first validation, and the need to pilot LLM-based systems under realistic conditions. These lessons provide practical guidance for developing effective diagnostic support systems in industrial contexts.

The Impact of Sensor Faults on Condition Monitoring of a Hydraulic Actuator

Sun, 26 Oct 2025 00:00:00 +0000

While supervised machine learning is prevalent in prognostics and health management applications, the success of these models is dependent upon being trained on accurate data. Training data collected with faulty sensors can degrade the performance of these models when deployed in an operational environment. This study investigates the impact of faulty data and the robustness of feature extraction methods and tree-based classifiers. This study also provides an open-source software package for injecting faults into time series data. The numerical experiments are performed on an open-source hydraulic actuator data set and demonstrate that certain features are robust to certain types of faults and that more complex models, such as ensemble techniques, are more robust to sensor faults than simple models. This work suggests that more complex models and larger (and possibly redundant) feature sets may be preferred in situations where sensor faults are likely. Furthermore, certain feature extraction techniques may be selected if certain faults are more likely than others.

Study on Optimal Design and PHM Methods for New Electrification Systems

Hong Suk Chang — Sun, 26 Oct 2025 00:00:00 +0000

This study will describe the development of PHM method for electrified systems in mass-produced vehicles. The results demonstrate that the method is not limited to the research subject but can also be applied to newly developed electrified systems, demonstrating continued applicability even when the target is changed. Next, we will provide an overview of the development of PHM method that can be used universally across electrified systems. In Phase 1, we identified key failure modes that could occur in electrified systems using FMEA, based on data collected from design, analysis, and testing. In Phase 2, we explored appropriate diagnostic methods for each failure mode. For gear failures, we developed rule-based indicators and verified their validity through experiments. For bearing failures, we also developed a rule-based approach to determine the presence of a fault. However, due to limitations in predicting the location of the fault, we re-evaluated the method based on data to confirm its validity. For failure modes, we used CAE analysis models to identify differences between normal and fault signals for eccentricity and demagnetization faults. Similar signal differences were also observed in the test results of the target product. Based on this, we were able to build a robust diagnostic model using only a small amount of experimental data. In Phase 3, we developed a device capable of data collection and edge computing capabilities capable of analyzing and diagnosing signals from actual vehicles, enabling the collection and analysis of the necessary data.

Deep Reinforcement Learning for Airplane Components Failure Prognostics Full Cycle Automation

Baoqian Wang, Changzhou Wang, Denis Osipychev — Sun, 26 Oct 2025 00:00:00 +0000

As airplane components degrade over time, airplane service organizations (e.g., Boeing Global Services) and their airline customers need to collaborate on airplane components failure prognostics and replace/maintain components proactively to improve operation efficiency and reduce cost. In particular, airplane service organizations analyze various sensor data that captures the operational states of airplane components to predict possible component failures. Upon identifying an impending component failure, the service organization promptly sends alerts to the airline maintenance team. In response, the airline maintenance team conducts inspections and maintenance on the component and replaces it if necessary.

In this airplane components failure prognostics procedure, machine learning or engineering-based models can be used to make predictions of components failure on each flight. However, it is crucial for airplane service organizations to determine when to send alerts to airlines given the predictions of the full history of flights. Late alerts may cause schedule interruptions or even grounding of the airplane waiting for parts. Early alerts can bring unnecessary inspections that lead to significant cost to airlines. Current solutions rely on heuristics and/or manual engineering reviews to make decisions on sending alerts, which requires significant manual efforts and is difficult to scale.

To improve efficiency of airplane components failure prognostics, we applied deep reinforcement learning (RL) to automate the prognostics procedure while enhancing accuracy of alerts timing. Specifically, we used Long Short-Term Memory (LSTM) neural network model to represent alert policy that outputs alerts decisions based on flight sensor data and interaction history with airlines. To train the alert policy, we built a prognostics environment by using probability models to simulate airplane component state transitions over time and the airline’s feedback to alerts. With this environment, the parameters of alert policy are updated to minimize costs for airlines during the simulated prognostics procedure. This is achieved through the Deep Q-Network algorithm with memory prioritization to mitigate reward sparsity issue. Once learned, the alert policy is deployed to make decisions on sending alerts automatically by consuming incoming flight records and parsing current interactions with airlines. Moreover, we can fine-tune alert policy parameters to incorporate new airplane component features and airline operation changes. We conducted a case study on Boeing 787 air cycle machine (ACM) prognostics, which demonstrated the feasibility and effectiveness of our approach.

IntelliMaint

Ramesh Krishnamurthy, Shweta S, Rachana Sreedhar, Archana Chandrashekar — Sun, 26 Oct 2025 00:00:00 +0000

Predictive maintenance of complex mechanical requires robust health monitoring capabilities that can generalize across diverse components and operating conditions. We present a novel component-agnostic framework that unifies Health Indicator (HI) generation and Remaining Useful Life (RUL) prediction through an integrated pipeline comprising: (1) advanced feature engineering, (2) unsupervised health baseline modelling, (3) monotonicity and trendability learning (4) probabilistic degradation detection with confidence-aware RUL estimation.

We validate our framework on two distinct Industrial Case Studies: Firstly, Tool wear monitoring in CNC machine using vibration and spindle current data collected from the real production machine. Our framework achieves early degradation detection of tool life with RUL prediction within ±15% of actual failure time. Flank wear (VB) was measured as a standard parameter for evaluating tool wear. Secondly, bearing degradation assessment using the IMS bearing dataset. This validation demonstrates fault detection 40% earlier than traditional threshold methods with 90% confidence intervals.

Both case studies show strong HI monotonicity (>85%) and reliable uncertainty quantification, establishing the foundation for scalable, explainable predictive maintenance solutions. The framework's component-agnostic design enables rapid deployment across heterogeneous assets without extensive reconfiguration, while its interpretable architecture facilitates root cause analysis and maintenance decision support. These results demonstrate significant advances in scalable, explainable predictive maintenance, offering practitioners a unified solution for diverse industrial health monitoring challenges.

Potential of Generative AI in Knowledge-Based Predictive Maintenance for Aircraft Engines

Meriem Hafsi — Sun, 26 Oct 2025 00:00:00 +0000

In recent decades, predictive maintenance has become a strategic solution to detect anomalies and anticipate failures in industrial equipment and sophisticated machines. This strategy relies on the continuous collection of multi-sensor data and performance indicators to feed machine learning algorithms capable of identifying early signs of malfunction, thereby enabling preventive interventions and reducing downtime. In the literature, three main approaches are described: physics-based, data-driven, and knowledge-based. Physics-based methods require accurate mathematical modeling of the physical degradation processes involved, but they are often difficult to apply to complex systems where physical laws are hard to formalize. Data-driven methods dominate current implementations due to their ability to learn complex patterns from large datasets. However, they suffer from several limitations, such as a lack of interpretability, dependence on large amounts of labeled data, and poor generalization to new operating conditions. Knowledge-based approaches, on the other hand, are more explainable and rely on expert-defined rules, but they struggle with adaptability and scalability in dynamic or uncertain environments.

This study investigates the potential of generative artificial intelligence (Generative AI) to address these limitations by supporting the creation of hybrid predictive maintenance models that combine the strengths of both data-driven and knowledge-based approaches. Generative AI offers new capabilities to simulate realistic failure scenarios, augment limited datasets, and extract structured knowledge from unstructured technical sources. It can also support the construction of domain-specific knowledge graphs or ontologies by identifying relevant concepts and semantic relationships, as well as generate logical reasoning rules based on expert input or technical documentation. Additionally, Generative AI shows promising potential in assisting the physical modeling of complex systems by proposing plausible approximations or surrogate models when traditional analytical modeling proves difficult. To support this investigation, the study will address a use case focused on Remaining Useful Life estimation for aircraft engines using the C-MAPSS benchmark dataset.

Digital Twin-based IVHM for Predictive Maintenance

Sun, 26 Oct 2025 00:00:00 +0000

This paper proposes a Digital Twin-based Integrated Vehicle Health Management (IVHM) approach to enable predictive maintenance in aviation and aerospace industry. Predictive maintenance enables the identification of potential failure before they occur, improving operational efficiency, safety, and cost management by reducing downtown and optimizing maintenance scheduling. However, conventional approaches face three key challenges: lack of reliable run-to-failure data, uncertainties in system behavior and predictions, and fragmented processes between design and maintenance activities. This article introduces the concept of Authoritative Hybrid as-operated Digital Twin to overcome the current limitations. The proposed solution brings three main technical advancements: the integration of physics-informed Artificial Intelligence (AI) architecture reusing design artifacts into an IVHM system; the implementation of a comprehensive Validation, Verification, and Accreditation (VVA) process to support certification; and the enhancement of Model-Based Systems Engineering (MBSE) methods to ensure digital continuity across the different processes. This supports the development of advanced predictive maintenance capabilities, aligned with the vision of Type III IVHM systems, ultimately enabling more resilient, informed, and cost-effective operations in aerospace domain.

Development of a Scalable Digital Twin for Tram and Light-rail Infrastructure based on Open Data for Early Prediction of Rail and Track Defects

Jannik Goersch, Philipp Leibner, Benedikt Neubauer, Thomas Hempel, Raphael Pfaff — Sun, 26 Oct 2025 00:00:00 +0000

Due to an increasing passenger demand in rail-based transportation and a desire for sustainable mobility, rail infrastructure is nowadays confronted with increased loads requiring timely and efficient maintenance regimes (Holzfeind et al., 2025). To improve maintenance scheduling and give rail infrastructure operators better insights into the state of their networks, a comprehensive digital twin based on open data has been developed. The digital twin allows to connect sensor data from vehicles to railway assets and enables the development of custom algorithms for condition-based maintenance of railway tracks. For practical tests and validation of the digital twin, smartphones were placed in various trams and lightrail vehicles in the city of Frankfurt (Main) to record vibration and geolocation data over a period of more than a year. The results demonstrate that infrastructure quality changes can be automatically detected and monitored through the developed digital twin framework using a low-cost measurement set-up. Hereby, new capabilities for proactive maintenance scheduling and resource allocation emerge, and infrastructure operators can prioritize interventions effectively and ensure safe and comfortable railway operations.

RailNet

Jiawei Guo, Boshi Chen, Yu Qian, Yi Wang — Sun, 26 Oct 2025 00:00:00 +0000

Railroad inspections are critical for ensuring operational safety, as track defects such as missing clips and surface damages can lead to catastrophic failures. Traditional inspection methods are often labor-intensive, time-consuming, and prone to inconsistency. Although deep learning approaches have been introduced for track monitoring, they typically focus on a single task, require extensive retraining for multi-tasking, and suffer from deteriorated performance when adapted to new tasks. There is a growing need for lightweight, real-time, and multi-functional solutions that can simultaneously detect rail components and segment rail surface defects without compromising accuracy or speed.

To address these challenges, this paper presents RailNet, a lightweight and modular transfer learning framework tailored for real-time railroad inspection. RailNet integrates a frozen pre-trained component detection model and a new trainable module for surface defect segmentation. The trainable module refines feature maps from the frozen backbone and FPN through targeted correction and attention, introducing three key innovations: a Context Rebalancing Module (CRM) to offset pre-trained biases, a Selective Channel Attention (SCA) mechanism to highlight important channels to minimize computational costs, and a single-step Upsample Block for efficient high-resolution reconstruction. This design enables independent segmentation training without affecting upstream detection, achieving rapid adaptation, high accuracy, and efficient multi-tasking.

RailNet was evaluated on a custom rail defect dataset using only a lightweight trainable component (~5 MB, 0.96 GFLOPs). It achieves 93.20%-pixel accuracy and 92.59% recall for surface defect segmentation while preserving upstream component (e.g., spikes and clips) detection performance (a mAP@0.5 98.7%), as shown in Table 1. Compared to benchmark models like SegFormer, YOLOv12, DINOv2, and MobileSAM, RailNet demonstrates superior accuracy and faster inference speed on edge devices (Nvidia AGX Orin). Ablation studies further confirm the critical roles of the CRM, SCA, and the Upsample Block in enhancing overall performance. As shown in Figure 1, RailNet can simultaneously detect rail components and segment surface damage. These results highlight RailNet’s potential as a robust, real-time, and energy-efficient solution for multi-task railroad inspection and related industrial applications.

Maintenance, Engineering, and Operational Decision-Making Metrics Derived from Simple Maintenance and Aircraft Datasets

Paul Bordush — Sun, 26 Oct 2025 00:00:00 +0000

Using in-service maintenance data, it is possible to predict and forecast the propulsion system contribution to aircraft and fleet level unavailability and to identify sub-system degraders of overall engine reliability. While more complex means of assessing reliability exist, increased layers of complexity can lead to increasing difficulty when used to convince a military commander or fleet manager of the appropriate action to take. Furthermore, increased complexity increases the time required to produce, to analyze, and to assess results of reliability assessments. In a time-critical situation, when faced with the need for an immediate maintenance or engineering decision, the best information is that which is the simplest and easiest to understand, quickest to produce, and fastest to apply. In this work, a minimum list of data requirements will be developed with an associated means of analyzing these data to produce meaningful indicators to predict and to forecast unavailability and mission abort rates that can be used to plan for deployed or sustained operations. Analysis of the same data set can produce a prioritized listing of sub-system reliability degraders to drive engineering decisions for component improvement. The Royal Canadian Air Force’s CT114 Tutor aircraft will be the basis for analysis demonstrating that sophisticated sensors and data systems are not required to be able to produce meaningful data suitable for significant fleet level decisions. Statistical methods and appropriate data filtering were applied to the engine system to derive rates for overall mission aborts, aircraft unavailability and aircraft unreliability for the top sub-system degraders. Conclusions drawn include that this information, if calculated correctly, can provide decision makers with the critical information required to make significant fleet wide decisions. Recommendations and methodology are presented that are applicable to any military or civil aircraft fleet at the sub-system, aircraft, and fleet level.

CBM/PBM Operational Performance

Franck Dessertenne — Sun, 26 Oct 2025 00:00:00 +0000

Condition-Based Maintenance (CBM) and Predictive-Based Maintenance (PBM) have emerged as pivotal strategies in the
aircraft industry, offering the potential to revolutionize maintenance operations. These approaches aim to optimize maintenance schedules, reduce operational costs, increase aircraft and assets availability, and enhance safety by leveraging Health Indicators (HIs). These indicators enable the monitoring of the aircraft health throughout its lifecycle, supporting the prediction of potential failures and proactive maintenance interventions. Consequently, a critical aspect of realizing the full benefits of CBM/PBM lies in the ability to accurately measure the performance of both predictive models and the operational performance of the equipment itself. However, the widespread and effective implementation of CBM/PBM faces challenges. Difficulties often arise from the inherent complexity of these systems and the need for efficient collaboration among stakeholders with diverse backgrounds and business objectives. In particular, the concept of "performance" can be interpreted differently by various stakeholders, leading to misunderstandings and hindering effective decision-making.
This paper addresses the limitations of relying solely on conventional operational metrics, such as the MTBF (Mean Time Between Failures), within a CBM/PBM context. It recognizes the necessity to move beyond traditional concepts and metrics associated with reactive maintenance and embrace a more holistic view. To bridge the gap between different perspectives, this work proposes an adapted framework that facilitates the reconciliation of diverse viewpoints. The proposed approach encompasses a detailed understanding of internal failure modes and the establishment of appropriate correspondence laws at equipment level. Ultimately, it seeks to foster stronger communication and understanding within the broader ecosystem of stakeholders, including suppliers, Maintenance, Repair and Overhaul (MRO) providers, and end-users.
To achieve this, this paper introduces a set of new operational metrics as a paradigm shift, specifically designed for CBM/PBM environments. The MTBPR (Mean Time Between Predictive-based preventive Removals), which quantifies the time between preventive removals based on predictions; the NDF (No Degradation Found) rate, which measures the rate of early removals where no degradation was detected; the MLR (Mean Lifetime Reduction), which accounts for the reduction in equipment lifetime due to early removals. Importantly, the paper establishes clear analogies between these new metrics and the legacy metrics used in unscheduled maintenance strategies, namely the MTBUR (Mean Time Between Unscheduled Removals) and the NFF (No Fault Found) rate.
Furthermore, this paper delves into the crucial relationship between the performance of predictive models and operational performance. It demonstrates how typical data analytics metrics, such as Recall and Precision, can be effectively linked to classic equipment-level reliability metrics. By establishing these formal connections, this paper provides a generalized and robust approach to support the definition of performance objectives for predictive monitoring, and the way to properly monitor in-service performance (feedback loop). This approach fosters alignment and collaboration among all stakeholders involved in the CBM/PBM domain, ultimately leading to more effective and harmonized maintenance strategies and improved operational outcomes.

Real-Time Fatigue Risk Assessment of BHA Connectors Using Combined Physics and Data-Driven Approach

Dmitry Belov, Yaou Wang, Wei Chen, Jennifer Gilliam — Sun, 26 Oct 2025 00:00:00 +0000

Drilling operations depend not only on controlling surface parameters but also on keeping bottom-hole assembly (BHA) components structurally sound. The BHA is the lower portion of the drill string in a drilling operation – the part that actually contacts the wellbore and guides the drilling process. Failures, especially at the connection between the flow diverter and the drive shaft behind the mud-motor power section, can cause major non-productive time (NPT), high costs, and poor performance. These failures are often linked to combined surface and downhole rotational speeds and high bending moments, which are common during directional drilling. To reduce this risk, we present a new method for real-time health monitoring and remaining useful life (RUL) estimation of these connections. The method combines physics-based fatigue modeling with machine-learning estimators, making it possible to track connector use across time and jobs using serialized component data. The system processes real-time drilling parameters to estimate downhole rotational speed (RPM) and bending moment. When measurement-while-drilling (MWD) data are available, direct RPM values are used; otherwise, a predictive model based on temperature, flow rate, and differential pressure is applied. Bending moment is inferred from drilling parameters and BHA design. The framework then calculates fatigue damage with connector-specific S–N (stress–number of cycles) curves and updates both current and cumulative RUL values. This helps operators make proactive decisions and lowers the risk of expensive failures. Tests with historical drilling data show strong agreement between predicted damage and observed connector failures, proving that the approach works in the field. The solution is already integrated into a commercial platform and used by field teams. Case studies show it reduces unexpected failures, cuts non-productive time, and improves the efficiency of directional drilling.

Lifecycle and best practices for predictive maintenance alert lifecycle management.

Justin Sindewald, Ryan Latini, Joseph Rice — Sun, 26 Oct 2025 00:00:00 +0000

The predictive maintenance alert lifecycle is a critical topic in the aviation industry.

Stakeholders, including operators, suppliers, and Original Equipment Manufacturers (OEMs), require effective frameworks to

support the value proposition of predictive maintenance products and services.

However, defining alert effectiveness is challenging due to the lack of industry

standards for the end-to-end lifecycle of predictive maintenance alerts. Adding to the

challenge, different stakeholders may want to optimize on different objectives. Often,

alert performance is measured prematurely or not at all. To ensure high-quality alerts,

all alerts should be managed through their entire lifecycle until obsolescence. This

whitepaper outlines a clear lifecycle and best practices for predictive maintenance alert

lifecycle management.

Guidance for the Certification and Continued Airworthiness of PHM System for Aviation

Ravi Rajamani — Sun, 26 Oct 2025 00:00:00 +0000

One of the barriers for the wider use of Prognostics and Health Management (PHM) systems in regular usage in commercial aviation has been the need to certify all components and functions related to PHM. Because PHM systems are not entirely like other systems on an aircraft, little guidance has been provided by the authorities or standards development organizations (SDO) in regards to such certification. Additionally, there is a lack of guidance for the continued airworthiness of the PHM system, i.e., rules for monitoring, maintaining, and updating them. We are not even touching on the ever sensitive topic of the need for significant monetary investment in the development, testing, manufacture, and operations of the PHM system, which OEMs are loath to do. However, there is some good news in the offing. One piece of this complex puzzle has recently been solved and in this paper, we will review and report on three events of significant progress which will help with the development and deployment of PHM systems for commercial aircraft.

The MPIG (the Maintenance Programs Industry Group, which develops the maintenance guidance commercial aircraft – MSG) published guidance on how a PHM task can replace an approved scheduled maintenance. Next, the FAA (Federal Aviation Administration) published an advisory circular laying out the requirements for what an end-to-end PHM system needs to comply with to be deployed on aircraft certified in the US. But even more critically, the SAE’s E-32 (Propulsion Health Management) technical committee published a short guidance on how to certify a PHM system – and any required ground support equipment – on an engine. The HM-1 (Integrated Vehicle Health Management) committee later updated this guidance to include the entire vehicle. With these three recent developments, one part of getting PHM systems on aircraft is made easier. Other challenges – such as justifying them financially – still remain, but it will be harder to argue that the certification and continued airworthiness authorities are not in favor of employing PHM systems in commercial aviation.

Enhancing Nuclear Safeguards with Time Series Sketching-Based Nuclear Material Loss Tracking

Hao Huang, Scott Evans, Philip Honnold, Nathan Shoman — Sun, 26 Oct 2025 00:00:00 +0000

Detecting nuclear material loss or leak events is a critical challenge for nuclear safeguards and material accountancy in the recycling of used nuclear fuel. The MAYER (Multi-Sensor Assimilation Yielding Enhanced Reliability) project, part of the ARPA-E CURIE (Converting UNF Radioisotopes Into Energy) program, aims to develop a framework for tracking material loss in nuclear facilities by integrating multisensor data with predictive modeling. In this paper, we introduce MAterial Loss Tracking via Time Series Sketching (MALTS), a deep learning-based method designed to detect material loss events across the nuclear fuel recycling system. MALTS enhances the accuracy and robustness of nuclear material loss tracking by employing time series sketching to capture essential patterns while filtering out sensor noise, resulting in more stable predictions despite sensor noise effects. This approach also improves the time efficiency of material loss tracking by reducing the dimensionality of highfrequency sensor data, thereby enhancing computational scalability and enabling real-time inference. To further provide insights into the leak, MALTS ranks anomalous channels by post-processing results with a pretrained vision-language model (VLM) that considers the system flow diagram, generating a sorted list of anomalous channels from upstream to downstream. The initial leak location is identified as the first upstream channel. Experimental results demonstrate MALTS’s effectiveness and efficiency in accurately identifying unseen nuclear material loss events and pinpointing initial leak locations, making it suitable for deployment within the MAYER digital twin framework for nuclear material safeguards.

Achieving Three Nines for Product Health Data Monitoring Systems

Raghothama M. Rao, Travis Mann, Bill Edinger — Sun, 26 Oct 2025 00:00:00 +0000

Effective data and health management are critical throughout the lifecycle of engineered systems. When implemented correctly, product health management tools and processes can help lower total product ownership costs, improve safety, maximize availability and utilization rates, thereby delivering value to system operators and maintainers. Central to achieving this capability is robust data management that includes data acquisition, secure transmission, efficient storage and timely processing. These steps ensure that health insights can be delivered to stakeholders in support of diagnostics, prognostics, and informed decision-making.

In sectors such as aerospace, defense and power utilities, the demand for high availability - targeting 99.9% uptime or “three nines” – places stringent requirements on product health monitoring systems. Hardware infrastructure, software, engineering processes and procedures are an integral part of achieving this target. This paper presents a focused exploration of the software, automation strategies, and best-in-class engineering processes that support high-reliability health data monitoring, with an emphasis on commercial aircraft engine applications. Drawing from Belcan Engineering’s experience, we highlight key software architectures, process improvements, and practices that enable scalable and maintainable solutions. Additionally, we discuss how early integration of health management capabilities into the development cycle enhances product value and reduces lifecycle costs. The insights presented are based on real-world implementations and are intended to guide practitioners seeking to build resilient, high-availability monitoring systems.

Atomic Whispers to System Health Diagnosis and Prognosis: First-Principles-Based Degradation Modeling of 2D Materials in Next-Generation Bioelectronics

Yi Cao, Paulette Clancy — Sun, 26 Oct 2025 00:00:00 +0000

Our results demonstrate that Cr intercalation into 2D transition metal dichalcogenide materials will significantly elevate the interlayer shear resistance, acting as an atomic-scale "glue" that mitigates delamination and structural failure—a key degradation mechanism. We uncover dynamic stabilization mechanisms and quantify the energy barriers that retard lateral sliding, which are crucial inputs for physics-of-failure models. We combine ab initio density functional theory (DFT) and machine-learned-force-field Molecular Dynamics (MLFF-MD) for this study. Leveraging MLFFs allows us to extend our simulations to larger length- and time-scales and hence capture long-term dopant dynamics and degradation evolution. MLFF-MD has the advantage of combining a bigger scope with near-DFT accuracy, enhancing predictive capabilities for materials design.

This work provides mechanistic insight into transition metal intercalation for 2D material's stabilization and offers a physics-informed computational framework for assessing material longevity and reliability. Such predictive capabilities are critical for proactive Prognostics and Health Management (PHM), enabling the rational design of robust 2D heterostructures, guiding synthesis strategies, and informing maintenance protocols for advanced electronic and spintronic devices.

Novel Segmentation Methodology for Robust Feature Engineering of Time Series Data in Prognostics and Health Management

Dai-Yan Ji, Jay Lee — Sun, 26 Oct 2025 00:00:00 +0000

Time series segmentation plays a critical role in feature engineering for prognostics and health management (PHM), yet most existing approaches rely on domain-specific rules or fail to preserve meaningful transient patterns. This research proposes a segmentation-driven framework that leverages a greedy Perceptually Important Point (PIP) algorithm to identify informative structural regimes in sensor signals without prior domain knowledge. A global reference signal is constructed from class-level Euclidean-barycenter averages, and consistent segment boundaries are applied across all samples. Segment-level statistical features are then extracted and used for classification. Evaluation on a chemical gas sensor dataset demonstrates that the proposed method significantly outperforms traditional whole-signal summary statistics, achieving improved robustness to drift and unit variability. Future work includes parameter optimization of the PIP algorithm, exploration of class-sensitive segmentation strategies, and extension of the framework to remaining useful life (RUL) prediction and anomaly detection tasks.

Gear Diagnostics Based On Transfer Learning Methodologies and Digital Twinning

Henrique Duarte Vieira de Sousa, Konstantinos Gryllias — Sun, 26 Oct 2025 00:00:00 +0000

This paper outlines the motivation for the research, reviewed the relevant SOTA in TL and CM, and identified some current research gaps. Moreover a dedicated test rig that will be used for methodological development and experimental validation has been described in detail. Finally, a structured research plan has been proposed, with the ultimate objective of developing a robust and scalable methodology combining ML and DTs for fault diagnostics of WT gearboxes, thereby contributing meaningfully to the field of PHM.

A Method for Self-Healing System Integration within Human Habitation Systems

Marc Koerschner, Michael Balchanos, Dimitri Mavris — Sun, 26 Oct 2025 00:00:00 +0000

This document serves as a summary of the creation of a method that supports and enables the development, integration, evaluation, and selection of self-healing systems within deep space habitats. We begin with developing the inherent motivation and desire for the inclusion of self-healing systems within space habitats: the problem with traditional methods of fault recovery and how they fail in a space habitat context. The intrinsic links between the concept of the self-healing system, Prognostics and Health Management, and affiliated areas of research are then shown to prove the relevance of this thesis’ work to the PHM Society. The research plan is then laid out fully, beginning with the initial architecting of the baseline system of interest, generation of promising self-healing architectures, development of an analytical model to investigate faults, failures, and promising self-healing architectures, approaches to trade said self-healing architectures, and finally, showcasing how a selection process could be handled for a user-defined optimal architecture to be integrated within a habitat of choice. Preliminary results will be presented here. Finally, the novelty of the problem and approach is showcased specifically, along with contributions to the field of PHM, but also self-healing system development, and resilience engineering as a whole.

PhD Symposium Stochastic_Optimisation_of_Tail_Assignment_and_Maintenance_Task_Scheduling_with_health_aware_models

Benno Kaslin, Marta Ribeiro, Dimitrios Zarouchas, Manuel Arias Chao — Sun, 26 Oct 2025 00:00:00 +0000

Efficient maintenance management is essential—not only to reduce costs but also to maximize aircraft availability and uphold safety standards. This requires balancing maintenance scheduling (MS), which drives downtime, with tail assignment (TA), which governs aircraft utilization. While recent research has explored the integration of MS and TA, these efforts have largely neglected the role of Condition-Based Maintenance (CBM) and the uncertainty inherent in prognostic models. This research proposes a novel, unified framework that jointly optimizes MS, TA, and CBM using stochastic programming and health-aware models. The approach leverages sensor-derived prognostic information to forecast component degradation and incorporates its probabilistic nature directly into the planning process. By accounting for uncertainty in remaining useful life (RUL) predictions, the model produces robust flight and maintenance schedules that reduce the risk of unplanned disruptions. Preliminary experiments using real-world airline data demonstrate that explicitly modeling health uncertainty leads to more reliable scheduling outcomes, while improving operational efficiency and reducing maintenance costs. Compared to current industry practice, the integrated framework enables data-driven, future-oriented decision-making at the interface between fleet operations and maintenance planning. This work advances the state-of-the-art by holistically addressing TA, MS, and CBM within a scalable and interpretable optimization model—closing a critical gap in the practical deployment of CBM strategies in civil aviation.

PhD Symposium - Interpretable and Uncertainty-Aware Hybrid Prognostics Using Multimodal Knowledge for RUL Prediction

Dario Goglio, Dimitrios Zarouchas, Manuel Arias Chao — Sun, 26 Oct 2025 00:00:00 +0000

Unforeseen technical failures contribute significantly to airline delays, highlighting the need for predictive maintenance. However, developing reliable prognostic models in aviation is challenging due to strict safety requirements, limited labeled data, and the need for interpretable and trustworthy predictions. This research proposes a hybrid framework for remaining useful life (RUL) prediction that integrates multimodal domain knowledge available to airlines, such as sensor data, contextual information and reliability insights, into interpretable and uncertainty-aware algorithms. To this end, the proposed framework resorts to unsupervised degradation extraction with knowledge-informed autoencoders and supports extensions for failure mode segmentation. Initial experiments on a benchmark dataset show promising results, and application to real-world commercial aircraft data is planned to further validate the approach.

Deep Learning Based Remaining Useful Life Prediction of Lithium-Ion Batteries Using Early Cycle Degradation Features

Kyutae Park, Heung Soo Kim — Sun, 26 Oct 2025 00:00:00 +0000

The failure of a lithium-ion battery (LiB), which is used as an energy storage system (ESS) in the mobility industry, such as electric vehicles and aircraft, can lead to substantial loss of life and property, thereby causing significant problems. Therefore, it is essential to monitor the capacity degradation of the mobility battery and accurately predict the remaining useful life (RUL) from the early cycle stage. Particularly, RUL prediction is the main objective of the Battery Management System (BMS) and is important for guaranteeing the safety of the mobility system (Wu et al., 2016). This research introduces a hybrid deep learning model for RUL prediction, using LSTM-attention and Multi-Layer Perceptron (MLP) methodologies. The proposed model uses statistical degradation features and domain knowledge-based features as input data acquired from the early 100 cycles of charge/discharge data of a lithium-ion battery. The model's performance evaluation was divided into two phases: primary and secondary, providing root mean square errors of 158.4 and 168.67, respectively. This study's results aim to contribute to the advancement of Prognostic and Health Management (PHM) technology, Condition-Based Maintenance (CBM) strategies, and BMS-based life prediction technology for mobility battery systems.

A Single Carbon Nanotube-paper Composite Electrode Sensor Based Brake Oil Degradation Detection

Yubin Cheon, Changwoo Lee, Jae-Hyun Chung, Heung Soo Kim — Sun, 26 Oct 2025 00:00:00 +0000

Brake oil is essential to the performance and safety of hydraulic braking systems, but its degradation—primarily caused by water absorption—can lead to reduced boiling points, corrosion, and brake failure. This study presents a non-invasive method for real-time monitoring of brake oil degradation by detecting changes in water content using a single-electrode capacitive sensor based on a carbon nanotube paper composite (CPC). The sensor operates on the principle of fringing electric fields, enhanced by high-aspect-ratio carbon nanotube fibers that increase local field intensity and dielectric sensitivity. Unlike conventional two-electrode designs, this configuration offers structural simplicity and is well-suited for embedded automotive platforms. Experimental testing was conducted using two types of brake fluids, with incremental additions of deionized water (0.5% by volume) to simulate moisture-induced degradation. The sensor exhibited a strong linear response to increasing water concentration, with consistent slopes across fluid types, enabling a generalized calibration model for real-time water content estimation. The CPC sensor demonstrated high sensitivity, fast response, and excellent repeatability, making it an effective solution for in-situ brake fluid monitoring. This work supports the development of predictive maintenance systems aimed at improving vehicle safety and operational reliability

Integrating Few-Shot Learning and Pre-trained Models into Similarity-Based PHM using Small Data in Complex Engineering Systems

Takanobu Minami, Jay Lee — Sun, 26 Oct 2025 00:00:00 +0000

Prognostics and Health Management (PHM) is vital for complex engineering systems, yet its data-driven solutions are often hampered by the "small data problem"—a scarcity of labeled fault data in industrial settings. This limitation restricts the training and generalization of machine learning models and is compounded by varying operational conditions that reduce the relevance of historical data and pre-trained models. This research introduces a research framework to tackle these small data challenges in PHM. The primary objective is to develop a robust and adaptable PHM methodology by enhancing and synergistically integrating similarity-based Few-Shot Learning (FSL) with large-scale pre-trained time-series models. The research will focus on two main thrusts. First, it aims to improve the generalization capabilities of FSL frameworks by addressing limitations such as noise robustness, domain shift adaptability, and generalization to novel faults across diverse PHM domains. This involves developing noise-robust feature extraction, integrating domain adaptation techniques, and exploring expressive similarity metrics. Second, the study will investigate the effective adaptation of state-of-the-art pre-trained time-series models (e.g., TimesNet) for PHM tasks under data scarcity, focusing on efficient fine-tuning and synergistic integration with the enhanced FSL approaches. The author's prior success in a PHM data challenge using a similarity-based method for spacecraft systems provides preliminary validation. This research is expected to deliver an enhanced PHM framework for high-accuracy diagnostics with limited data, contributing generalized FSL models, systematic methods for leveraging pre-trained models in PHM, and advancing the practical deployment of intelligent PHM solutions.

Neural Counterfactual Reasoning for Interacting Systems: Bridging Physics-Informed Learning and Reasoning for PHM

Amaury Wei, Olga Fink — Sun, 26 Oct 2025 00:00:00 +0000

Over the past decade, advances in sensing and information technologies have enabled industries to collect large amounts of data. Yet, decision-making often remains driven by the intuition of domain experts who rely on simplistic analyses and short-term considerations. This frequently leads to suboptimal decisions that fail to account for long-term effects, particularly in complex, interconnected systems. Current data-driven strategies typically focus on immediate objectives, overlooking relational structures and longer-term impacts. There is a growing need for more transparent, generalizable models that can simulate system behavior, reason about alternative future scenarios, and extrapolate to unseen conditions—capabilities that are essential for decision-making in Prognostics and
Health Management (PHM). This research aims to advance reasoning and decision support in PHM through three novel contributions: (1) a physics-informed surrogate model for simulating rigid body interactions, enabling the exploration of ”what-if” scenarios, (2) an object-centric visual reasoning model for dynamics prediction in sensor-limited environments, supporting visual inspection tasks, and (3) a neuro-symbolic framework for interpretable root-cause analysis in time series, improving diagnostic transparency and providing actionable insights.

A Health Monitoring Framework for Thermal Degradation Mitigation in Solar Power Plants

Nadia N. Sanchez-Pozo, Jon Olaizola, Erik Vanem, Jose I. Aizpurua — Sun, 26 Oct 2025 00:00:00 +0000

Prognostics and Health Management (PHM) of photovoltaic (PV) systems requires integrated approaches that link temperature forecasting with physical degradation modeling under thermal stress. This study addresses key limitations of existing PHM frameworks, such as the lack of high-resolution climate projections and limited coupling with degradation models, by proposing a unified PHM methodology tailored for high-temperature scenarios. The framework consists of three main components: (1) a formal problem definition of PV performance loss during extreme temperature and high cooling demand periods; (2) high-resolution spatiotemporal forecasting of temperatures; (3) probabilistic modeling of PV thermal degradation. The proposed approach integrates two
innovations, including a Gaussian copula–based risk assessment for capturing joint distributions of environmental stressors (e.g., air temperature, solar irradiance, and wind speed) and a Spatiotemporal Graph Neural Network (ST-GNN) architecture for accurate prediction of extreme temperature
events. Accelerated aging tests and ERA5 reanalysis data (1974–2023) have been used to parameterize the probabilistic aging models. Preliminary results from forecasting experiments achieved root-mean-square errors of 5.1–5.5°C across three representative Spanish climate zones. Future work will focus on enhancing the expressiveness of spatial dependencies through dynamic graph structures with learnable edge weights, as well as propagating predictive uncertainty from temperature forecasts into degradation models using uncertainty quantification techniques.

Deep Learning for Robust Manufacturing: From Quality Control to Predictive Maintenance

Paula Mielgo, Anibal Bregon, Miguel A. Martinez-Prieto — Sun, 26 Oct 2025 00:00:00 +0000

In today’s global manufacturing landscape, companies are required to balance speed, precision, and sustainability, thereby making intelligent, data-driven solutions a necessity. The convergence of Industry 4.0, cyber-physical systems and artificial intelligence technologies is leading to a new paradigm known as smart manufacturing, where the effective use of collected data can increase productivity, efficiency, and quality. This research explores the potential of deep learning to enhance industrial productivity by leveraging automated quality control of manufactured components and predictive maintenance of systems. Thus, this thesis focuses on two main objectives within an industrial context: (OB1) the development of models to enhance the quality control process, and (OB2) the development of models to implement a predictive maintenance strategy. These objectives are approached through three expected contributions. First, a quality control model for thermal images aligned with factory requirements (C1). Second, a contrastive learning model for anomaly detection in multiview images (C2). Third, a predictive maintenance model for die-casting molds (C3). Initial results are starting to show the advantage of these contributions to improve productivity in smart factories.

Large Language Model Accelerated Maintenance Insights

Noah Getz, Xiaorui Tong — Sun, 26 Oct 2025 00:00:00 +0000

Maintenance operations play a crucial role in the efficient functioning of manufacturing plants across various industries. Legacy systems that record maintenance operation data have long been utilized to facilitate day-to-day activities, conduct investigations, and maintain records. Previous studies have leveraged this data to explore insights and patterns related to interruption causes. However, the application and scalability of these studies have been impeded by issues such as data quality, text inaccuracies, and a lack of common natural language processing tools compatible with legacy systems. The rapid advancement of generative artificial intelligence, particularly large language models (LLMs), presents opportunities to address these challenges and enable quicker insights. This paper proposes a technical architecture for expediting insights into maintenance operation data through LLM-enabled data augmentation, summarization, and extraction, as well as embedding-based feature extraction and downstream clustering. A use case is presented based on maintenance data from an aerospace manufacturing plant, with user interviews conducted to evaluate the generated insights and system feedback. This work pioneers the adoption of LLMs in accelerating insights from under-utilized maintenance operation data, paving the way for an LLM-powered maintenance co-pilot.

Robust Fault Detection with One-Class Training

Ark Ifeanyi, Jamie Coble — Sun, 26 Oct 2025 00:00:00 +0000

Anomaly detection is a critical capability in modern industrial systems, particularly in the energy sector where early fault identification can prevent catastrophic failures, minimize downtime, and reduce maintenance costs. However, the scarcity of labeled fault data in real-world applications makes traditional supervised learning approaches infeasible. This motivates the need for methods trained using only healthy data, a paradigm known as one-class training. One-class approaches are especially relevant for deployment in safety-critical domains such as nuclear power generation, grid monitoring, and process control, where failure data is rare, diverse, and expensive to collect. This study evaluates the performance and generalization capabilities of four data-driven methods trained exclusively on healthy data. The first method uses Principal Component Analysis to reduce data dimensionality and leverages reconstruction error for anomaly scoring. The second approach applies sequence modeling via a Long Short-Term Memory forecasting model, predicting future time steps based on past behavior and flagging sequences that deviate significantly from predicted values. The third is a one-dimensional convolutional autoencoder designed to reconstruct multivariate time-series inputs, with deviations in reconstruction used to identify potential anomalies. The fourth method, termed Deep Center Encoding, employs a neural network encoder trained to map healthy data to a compact region in latent space centered around a learned centroid, with outliers identified based on distance from this center. All methods are evaluated on sensor data from a real, operating nuclear power plant and tested for their ability to detect previously unseen fault distributions. Our results highlight trade-offs in sensitivity and generalization across the approaches, with Deep Center Encoding showing promising robustness to distribution shifts. These findings reinforce the feasibility and importance of one-class training frameworks for generalizable, fault-agnostic condition monitoring in industrial environments, supporting broader efforts in reliable artificial intelligence and predictive maintenance.

Remaining Useful Life Prediction via Computation of Physical and Material Properties

Navid Zaman, Milo Cooper, Frank M. Juarez — Sun, 26 Oct 2025 00:00:00 +0000

Remaining useful life (RUL) prediction is a crucial aspect of predictive maintenance, enabling industries to optimize asset longevity and minimize unexpected failures. RUL becomes increasingly difficult with system complexity. Such systems are built up of components exhibiting known material properties. When degradation occurs, these materials are predictably affected and can indicate the remaining life of the system, both directly and indirectly. Changes reflected from damage, ageing and wear are particularly detectable in stiffness of materials for example. This paper aims to utilize the physics and domain knowledge of systems along with material degradation indicators such as Young’s Modulus and the interdependencies between such metrics, to accomplish RUL with a deeper understanding of wear and fatigue as compared to purely data-driven methods. Experimental studies and computational simulations demonstrate the effectiveness of this approach, offering a novel perspective on predictive maintenance strategies.

Bearings Fault Detection via Physics-Informed Convolutional Neural Networks

Sun, 26 Oct 2025 00:00:00 +0000

The reliability of rotating machinery is essential in industrial environments, where early fault detection can prevent significant losses. In this scenario, Condition-Based Maintenance (CBM) strategies benefit from combined signal processing and machine learning techniques. Although Deep Learning (DL) based models present good results in automatic fault classification, their exclusive dependence on data can generate inconsistent predictions and non-compliance with physics. To overcome this limitation, this work proposes an approach that uses physical principles based on Physics-Informed Deep Learning (PIDL) for fault classification in bearings, using vibration data obtained in an experimental bench. The dataset covers three operating conditions — healthy, light damage and heavy damage — with vibration signals captured by two sensors, one always healthy and the other with or without damage. The methodology involves the application of the Hilbert transform (HHT) for envelope analysis and defining amplitude thresholds that reflect the physical levels of degradation. These thresholds are incorporated into the learning process, guiding the classification and promoting greater interpretability and robustness. Initial results show that, in the multiclass classification problem, traditional DL outperformed PIDL, achieving a balanced accuracy of 97.3% compared to 84.78% for PIDL. However, in the binary classification scenario — distinguishing between healthy and unhealthy conditions — the PIDL model achieved performance comparable to DL's, with balanced accuracies of 95.17% and 95.71%, respectively. These findings highlight that incorporating physical constraints into DL models can enhance the robustness of predictions, particularly in simpler classification contexts.

A Physics-informed Multi-fidelity Neural Network Framework for Virtual Sensing in Rotating Machinery

Seho Son, Dayeon Jeong, Kyung Ho Sun, Ki-Yong Oh — Sun, 26 Oct 2025 00:00:00 +0000

This study proposes a novel integrated framework of physics-informed machine learning for virtual sensing in rotating machinery systems. The proposed framework aims to overcome the limitations of sparse physical measurements and enable comprehensive system monitoring. The proposed framework leverages a multi-fidelity data fusion strategy and physics-informed surrogate networks to achieve accurate and physically consistent predictions of dynamic responses across the entire domain under diverse operational conditions. The proposed framework comprises three key characteristics. First, a physics-constrained multi-agent diverse generative adversarial network (PC-MAD-GAN) is proposed to synthesize high-fidelity synthetic data. This architecture of a generative neural network effectively fuses extensive low-fidelity simulations datasets from finite element model (FEM), which provide full-field data across the system, with limited high-fidelity experimental measurements obtained from physically accessible regions. The multi-agent structure and physics constraints ensure that the generated synthetic data is both diverse and physically plausible, bridging the fidelity gap between simulation and reality. Second, a surrogate modeling scheme is introduced in the consideration of an adversarial domain adaptation architecture and a physics-informed domain-adversarial deep operator network (PI-DADON). This architecture is specifically designed for operator learning, enabling accurate interpolation and extrapolation of system dynamics, including responses under various rotating speeds, without requiring extensive retraining for unseen conditions. PI-DADON is trained on both the high-fidelity synthetic data and the limited real measurement data. Third, both the PC-MAD-GAN and PI-DADON architectures are rigorously supervised by the physics of rotating machinery. This strategy for physics-informed regularization is crucial to ensure that the model's predictions remain physically consistent and robust, even in unmeasured regions or under untrained operational conditions. The effectiveness of the proposed framework is comprehensively validated using dynamic response datasets obtained from an induction motor, including experiments under diverse operating conditions. Systematic analysis on experiments confirms that the proposed framework with physics-informed strategies significantly enhances accuracy, robustness, and generalization capability compared to purely data-driven approaches. The proposed framework facilitates the development of AI transformation for intelligent mechanical systems by enabling reliable virtual sensing in inaccessible areas, providing rich and full-field information critical for advanced condition monitoring and diagnosis.

Causal-Aware LLM Agents for PHM Co-pilots

Rajarajan Kirubanandan — Sun, 26 Oct 2025 00:00:00 +0000

Large language models (LLMs), while capable of generating plausible diagnostic plans from sensor data, inherently lack true causal reasoning capabilities and are prone to hallucinations. To address this limitation, we propose a hybrid AI framework that integrates LLMs with structured causal inference to enable robust, interpretable decision-making in predictive health monitoring (PHM) for complex systems. Our architecture positions the LLM as a planning agent that infers candidate failure modes and troubleshooting steps, while delegating causal evaluation to an external inference model grounded in formal causal principles. The system constructs a localized causal knowledge graph (KG) by retrieving top-k similar historical traces based on the current sensor context, and uses this graph to simulate and evaluate the impact of potential actions. We explore three strategies for handling multi-step diagnostic plans: step-wise decomposition, compound treatment modelling, and sequential intervention chains. Recommendations are ranked based on their estimated effect on resolution likelihood and further validated by a dedicated Evaluator Agent using counterfactual reasoning. Our results demonstrate that augmenting LLM-generated plans with external causal inference significantly improves relevance, consistency, and safety—offering a deployable blueprint for high-stakes PHM scenarios where LLMs alone cannot be trusted to reason reliably.

An Autonomous Multimodal System for Intelligent Railway Inspection

Boshi Chen, Jiawei Guo, Qian Zhang, Yi Wang — Sun, 26 Oct 2025 00:00:00 +0000

We propose an autonomous aerial inspection system to address growing safety concerns of railway infrastructure degradation. Unlike conventional labor- and sensor-intensive methods, our quadrotor integrates a depth camera, monocular inspection camera, Global Positioning System (GPS) module, and onboard computing unit. Combining visual-inertial fusion with GPS, it achieves robust localization even in GPS-denied environments. A lightweight deep learning model built on You Only Look Once v12 (YOLOv12) enables real-time detection of key components such as spikes and clips. To enhance autonomy, we introduce Railway Autonomous Navigation Guided by Embedded Recognition (RANGER), a novel algorithm that reconstructs 3D world coordinates from 2D detections using only onboard sensing, without requiring prior global maps. By fusing detection with localization data, RANGER enables precise track following and stable altitude control in complex or GPS-denied conditions. This reduces hardware demand while ensuring accurate navigation. Our system reduces operational costs, enhances scalability, and enables accurate, real-time inspections in complex, unstructured environments.

Autonomous Inspection Strategy for Insulator Strings Using a Drone

Munsu Jeon, Junhyeok Moon, Junsoo Kim, Ki-Yong Oh — Sun, 26 Oct 2025 00:00:00 +0000

This study proposes a novel autonomous inspection strategy for insulator strings using a drone. The proposed method not only optimizes the viewpoint of an optical camera for acquiring high-quality images of insulator strings but also detect anomalies of insulator strings from the acquired images. The proposed method features three key characteristics. First, an adaptive flight strategy is proposed based on the spatial configuration of transmission facilities. Specifically, the type of transmission tower is classified as either suspension or strain by analyzing the orientation of the insulator strings detected from optical images. Key structural features of transmission facilities are then extracted from point cloud data by addressing effective signal processing methods including random sample consensus, Euclidean distance clustering, and probabilistic downsampling. This feature enables the drone dynamically adjust the heading, altitude, and camera tilt to acquire optimal images of insulator strings. Second, a novel architecture of a deep neural network is proposed to detect defects in insulator strings based on the acquired images of insulator strings. Specifically, the architecture of the proposed network combines a multi-scale variational autoencoder and a lightweight classifier for anomaly detection. The variational autoencoder reconstructs normal insulator images at multiple scales to acquire hierarchical features, and the classifier distinguishes between normal and defective patterns by utilizing the extracted multi-scale features. Third, synthetic images of insulator strings are generated to mitigate a concern on the data imbalance between normal and abnormal images of insulator strings. Specifically, 3D models of insulator strings are constructed by using computer-aided design tools, and fault patterns are embedded to generate abnormal samples. 2D synthetic images are then rendered under varying viewpoints, lighting conditions, and backgrounds. Additionally, a generative adversarial network is addressed to produce realistic defect images to enhance the diversity of abnormal samples. These synthetic images contribute to improving the robustness of the proposed anomaly detection network. Systematic analyses conducted in both virtual and real-world environments show the effectiveness of the proposed method. The adaptive flight mission was successfully completed to acquire high-quality images of insulator strings without visual overlap between adjacent insulator strings. The proposed network achieves classification accuracy of 95.0% in distinguishing between normal and abnormal insulator strings for anomaly detection. The proposed strategy not only improves the performance of autonomous inspection but also enhances operational safety by reducing reliance on manual inspection in hazardous environments.

Minimizing Unplanned Downtime in Rotary Vacuum Drum Filters for Iron Ore Mining through Image-based Analysis

Sameer Prabhu, Naveen Venkatesh, Amit Patwardhan, Ramin Karim — Sun, 26 Oct 2025 00:00:00 +0000

In iron mining the processing phase broadly consists of sorting, concentrating, and pelletizing of the iron ore, this is to increase the iron content in the final product. In pelletizing, the filtering stage which controls the moisture content in the iron cake plays a crucial role. A Rotatory Vacuum Drum Filter (RVDF) is one of the mining equipment for removing excessive moisture by separating solid iron cake from slurry. A supporting wire which holds the cloth mounted on the frame of the RVDF is one of the critical components. During operation, recursive compression and stretching due to variation in pressure may lead to wire failure. This failure significantly impacts the integrity and efficiency of filter cloth that affects the filtration performance. If the wire failure is not detected promptly, it can lead to prolonged maintenance time, substantial maintenance cost and unplanned downtime, consequently affecting system availability. This work demonstrates health monitoring of filtering system in mining, designed to alert the operators about the emerging failure, to take appropriate maintenance action and minimize further damage, and unplanned downtime.

This paper introduces a computer-vision based monitoring approach that leverages image data of the drum filter during operation. The proposed approach identifies wire-induced degradation pattern on the filter cloth. Extracted video frames from the drum filter are processed to isolate the region of interest. Using Hough transform horizontal sections of the drum are detected followed by a sliding window analysis to evaluate the variations in pixel intensity. For normal surface, the average intensity variations remain low, typically ranging from 5 to 10. However, it spikes up to around 40 when irregular patterns are detected. The focus of this work is on detection and diagnostics, a transition towards prognostics is envisioned by incorporating pressure sensor data. Integrating multi-modal data may enhance the capability of predicting failure and improve the system availability.

A Two-Stage Framework for Small-Sample RUL Prediction on Structurally Complex Time-Series Data

Peng Gao, Fanyu Qi, Yizhang Zhu, Jianyu Zhang, Wenfei Li, Jianshe Feng — Sun, 26 Oct 2025 00:00:00 +0000

Remaining Useful Life prediction for high-value assets, such as aero-engines, presents a formidable challenge, compounded by sample scarcity, complex data structures and knowledge dilution. This paper proposes a two-stage framework designed to decouple representation construction from temporal pattern learning. Stage I mitigates data complexity by transforming multi-phase snapshot streams into standardized cycle-level sequences through hierarchical aggregation. Stage II addresses data scarcity and multi-target prediction using a Multi-task Shared Transformer. Furthermore, the model is optimized via a risk-aligned loss function that penalizes tardy predictions. The effectiveness of the proposed framework was validated by its strong generalization on PHM 2025 Data Challenge dataset, which ultimately secured a first-place result.

Extended abstract: Remaining Cycle Estimation based on a Maintenance Cycle Model

Kirin Inoue, Koji Wakimoto, Kosei Ozeki, Toshiyuki Kuriyama, Takahiko Masuzaki — Sun, 26 Oct 2025 00:00:00 +0000

This paper presents a remaining cycle estimation method for aircraft engines, developed during our participation in the PHM 2025 Data Challenge Competition.

The features of our method are as follows:

Physics-informed Feature Exploration: Through exploratory data analysis utilizing physical insights in the field of aircraft, we found good features that reflect performance degradation.
Maintenance Cycle Model: We developed a model that describes cycles of performance degradation and recovery by a weighted composite of health value for each maintenance type. The model fits well with our designated features that reflect the engine performance degradation.
Estimation Optimization: Taking the scoring rules into account, we optimized the estimated results by assuming probability distribution of the true values. The optimization enabled precise and stable estimation.

Maintenance Service Events Prediction Modeling of Aircraft Gas Turbine Engines

Peeyush Pankaj, Shyam Joshi, Xiaomeng Peng, Reece Teramoto, Taylor Hearn — Sun, 26 Oct 2025 00:00:00 +0000

This work addresses the PHM North America 2025 Conference data challenge for multi-event Remaining Useful Life (RUL) estimation on aircraft gas turbine engine modules, predicting the time-to-event for three maintenance actions: High Pressure Turbine shop visits (HPT SV), High Pressure Compressor shop visits (HPC SV), and Water Wash (WW).

We present a comprehensive workflow that integrates snapshot data quality control, virtual sensing for missing sensors (P₂₅ and T₅, and event-specific modeling with consensus mechanisms. Long short-term memory (LSTM) regression models are trained for HPC and WW using a custom loss function adapted from the competition, which heavily penalizes errors on early and near-term events. HPT RUL is produced by a confluence of an Artificial Neural Network (ANN) regressor and a linear degradation prior to stabilize extrapolation. A profile registration algorithm reconstructs temporal ordering in shuffled test/validation files, preserving health indicator (HI) monotonicity and degradation physics, proving a vital sanity check and building trust on the submitted results.

The MathWorks team achieved 1^st place in the public test phase with the best submission score of 0.3528, proving the high quality of predictions. The functionalities and tools demonstrated in our work are generally applicable aircraft fleet maintenance services RUL predictions.

Estimating Cycles to Maintenance Events For Jet Engines Using Engine-specific Measurement Residual

Peihua Han, Qin Liang — Sun, 26 Oct 2025 00:00:00 +0000

This paper introduces a data-driven method for predicting remaining cycles to major maintenance events in commercial jet engines, developed for the PHM North America 2025 Data Challenge. The method leverages measurement residuals that capture sensor deviations from expected values after accounting for operating conditions with simple linear models. These residuals serve as interpretable indicators of engine health. Health indices are constructed for High Pressure Turbine and High Pressure Compressor visits, while Compressor Water Wash events are estimated through linear extrapolation.