PHM Society Asia-Pacific Conference

PHM-Vibench

Tue, 13 Jan 2026 00:00:00 +0000

The Prognostics and Health Management (PHM) field faces significant challenges due to fragmented benchmarks, inconsistent evaluation protocols, and limited accessibility to comprehensive frameworks, particularly in the era of large-scale data and foundation models. To address these critical limitations, we introduce PHM-Vibench, a unified, extensible, and modular benchmarking platform for vibration-based PHM research. PHM-Vibench features a novel architecture that decouples the PHM pipeline into distinct data, model, task, and trainer factories, enabling flexible instantiation and customization of specific PHM workflows. The platform integrates comprehensive 20+ datasets with standardized protocols. It supports diverse PHM tasks including fault diagnosis, remaining useful life prediction, and anomaly detection. The framework addresses complex scenarios such as domain generalization, cross-system transfer, few-shot learning. Grounded in the Unified PHM Problem (UPHMP) framework with seven fundamental spaces: domain knowledge space (P), data space (D), task space (T), model space (M), loss function space (L), protocol space (Π), and evaluation metric space (E), PHM-Vibench enables systematic problem formalization and reproducible experimentation. The platform accommodates both traditional machine learning models and foundation models, with extensive experimental validation demonstrating superior cross-domain performance. PHMVibench addresses the standardization challenges in PHM research and provides a comprehensive solution for benchmarking and advancing the field. The platform is openly available at https://github.com/PHMbench/PHM-Vibench.

A Feature-Engineering-Based Machine Learning Approach for Cutter Flank Wear Prediction under Data-Scarce Conditions

Tue, 13 Jan 2026 00:00:00 +0000

Accurate estimation of tool wear in machining processes is essential to ensure product quality and optimize maintenance strategies. This work presents a Machine Learning methodology for the PHM-AP 2025 Data Challenge. The objective of the challenge is the cutter flank wear prediction in a CNC mill-turn machine using accelerometer, acoustic emission, and controller data. The training data consists of six datasets with a limited number of labeled samples, resulting in a few-shot learning scenario. To address these constraints, a manual feature extraction method is proposed. Features are computed by aggregating data from the controller and sensors in the time and frequency domains across five-cut intervals. In this way, the wear behavior is captured, and the sensitivity to missing data is reduced. Then, an optimization process is performed to select the most relevant features based on correlation values. These 14 identified features are used to fit a Multilayer Perceptron through a leave-one-dataset-out cross-validation process. Results reveal variability between training sets, with pronounced errors in the 17-21 cutting interval in four datasets. However, in the evaluation stage, the model achieved a competitive performance: RMSE of 11.486, MAPE of 8.518, and R² of 0.875, placing fourth in the challenge.

A Bidirectional Structure Constraint framework for Domain Generalization in Intelligent Fault Diagnosis

Wenjing Zhou, Liang Chen, Hong Zhuang, Qitong Chen — Tue, 13 Jan 2026 00:00:00 +0000

Achieving robust generalization in intelligent fault diagnosis under diverse industrial conditions remains challenging. Most domain generalization (DG) methods focus on either feature compactness or category separation, seldom addressing both in a unified framework. To overcome this, we propose a Bidirectional Structure Constraint (BSC) framework comprising Momentum Feature Alignment (MFA) and Category Anchor Separation (CAS). MFA employs a momentum-driven strategy to capture domain-invariant features for each category, while CAS encourages learnable class anchors to repel each other in latent space, enhancing class separability. These objectives are jointly optimized in a multi-loss framework, enabling the model to learn representations that are both intra-class compact and inter-class distinct. Experiments on the Shandong University of Science and Technology (SDUST) rotating machinery fault diagnosis dataset show that BSC significantly improves cross-domain generalization.

A CNN-Multi-Head Attention Framework for Gearbox Incremental Fault Diagnosis Under Non-Stationary Conditions

Hao Zhang, Shunuan Liu, Bin Luo, Konstantinos Gryllias, Chenyu Liu — Tue, 13 Jan 2026 00:00:00 +0000

Deep learning-based gearbox fault diagnosis approaches have demonstrated exceptional performance in achieving accurate fault identification across diverse industrial applications. Nonetheless, machines frequently operate under conditions characterized by time-varying speeds or loads, known as non-stationary working conditions. When a series of different non-stationary conditions tasks are sequentially input into the model for training, an issue arises where the model tends to forget previous tasks, a phenomenon referred to as "catastrophic forgetting". To address the challenge posed by task increments within non-stationary conditions, this paper proposes an incremental learning-based multi-task fault diagnosis framework under non-stationary conditions. This methodology enhances the model's diagnostic capabilities under non-stationary conditions by amalgamating convolutional neural network (CNN) with multi-head self-attention mechanisms. It employs exemplar replay and hybrid cross-head knowledge distillation techniques to preserve the model's understanding of prior tasks, thereby facilitating the incremental learning of multiple tasks. The efficacy of this proposed framework is substantiated through its application on the MCC5-THU fault diagnosis datasets of gearbox under time-varying speed working conditions. Experimental results demonstrate that this approach significantly mitigates the "catastrophic forgetting" effect, thereby offering a robust solution for multi-tasks increment fault diagnosis of gearbox operating under non-stationary conditions.

A Lightweight Neural Network for End-to-End Bearing Fault Diagnosis in Multi-Sensor Scenarios

Yichao Li, Yanfang Liu, Xiangyang Xu — Tue, 13 Jan 2026 00:00:00 +0000

Deep learning techniques have been widely applied in bearing fault diagnosis. However, their inherent reliance on historical offline training data and the large number of parameters pose considerable challenges in meeting the real-time requirements of online fault diagnosis applications, particularly in Industrial Internet of Things (IIoT) and edge computing environments. To address these challenges, this paper introduces a lightweight temporal feature fusion network
(LTFFNet) for processing multi-sensor signals to enable end-to-end bearing fault diagnosis. Instead of following the prevalent approach of converting one-dimensional vibration signals into two-dimensional images for feature extraction and classification, we designed the architecture directly from the perspective of temporal signals. Besides, the incorporation of the Squeeze-and-Excitation (SE) module allows the network to adaptively recalibrate channel-wise feature responses. We assessed the accuracy and real-time performance of the developed network on an embedded platform using the CWRU bearing dataset. The results demonstrate high diagnostic capability and low computational time, indicating its effectiveness and suitability for real-time multi-sensor bearing fault diagnosis in industrial settings.

A Multi-Sensor Fault Diagnosis Method for Aero-Engine Bearings Based on Complex-Valued Convolution and Dual Attention Mechanism

Shuquan Xiao, Xueyi Li, Tianyang Wang, Fulei Chu — Tue, 13 Jan 2026 00:00:00 +0000

Aero engines are widely used in modern aviation due to their high thrust-to-weight ratio, high efficiency, and high reliability, placing greater demands on the operational safety of key components such as bearings. Traditional bearing fault diagnosis methods typically rely on vibration signals collected by a single sensor, which makes it difficult to handle challenges such as incomplete information and noise interference in industrial settings. The paper proposes an intelligent fault diagnosis model called the Time-Frequency Attention Network, which is based on a time-frequency-aware convolutional layer and a fused attention mechanism. The goal is to fully exploit the time-frequency feature information from multi-sensor signals. First, a time-frequency-aware convolutional layer is designed using a kernel function constrained by the Short-Time Fourier Transform, leveraging a complex-valued convolution structure to effectively extract non-stationary features and local instantaneous frequency variations. Subsequently, a fused attention module is constructed, introducing a dual-attention mechanism in both channel and spatial dimensions to adaptively adjust the response intensity and frequency-domain focus areas of different sensor signals. The proposed network is experimentally validated on the Harbin Institute of Technology bearing dataset, achieving an accuracy of 99.54%. The results demonstrate that the proposed method outperforms existing benchmark models in terms of fault recognition accuracy and robustness, showcasing excellent diagnostic performance and generalization ability.

A Practical Hybrid Framework for RUL Prediction in Ion Mill Etching by Integrating Operational States

minamitu — Tue, 13 Jan 2026 00:00:00 +0000

Recent deep learning approaches for Remaining Useful Life (RUL) prediction in ion mill etching have achieved remarkable performance. State-of-the-art models, particularly those based on the Transformer architecture, have reached high prediction accuracy by exclusively training on data from the primary operational state. However, this strategy discards data from ancillary operational states, failing to address the critical challenge of providing the continuous RUL predictions required for practical applications.This research highlights the limitations of existing methods, which create "prediction gaps," and proposes a state-aware hybrid framework that considers the complete operational profile of the equipment. We apply a high-performance deep learning model for the primary operational state to ensure our predictive accuracy is competitive with state-of-the-art methods. The core contribution of our work, however, is the systematic design and validation of estimation logic for the previously ignored ancillary states. Specifically, we integrate the primary deep learning model with rule-based methods and simpler statistical models that activate during these ancillary states. This hybrid approach enables a shift from a competition based on accuracy under ideal conditions to the development of a robust and practical RUL prediction system. This paper argues for the importance of an integrated predictive framework that covers the full equipment lifecycle, rather than focusing on a single, complex model for a subset of data.

A Prediction of Thermal Stress Profiles in the Steam Turbine Startup Phase Using Fourier Neural Operator

LEE SoJung — Tue, 13 Jan 2026 00:00:00 +0000

Frequent startup and shutdown of steam turbines in recent operations have increased the importance of analyzing thermal stress induced by rapid temperature changes. In turbine startup, particularly during the synchronization phase, the surface temperature rises sharply while the core temperature lags behind, creating thermal gradients that lead to stress accumulation. However, the limited availability of measured data during these transient intervals poses a significant challenge for data-driven temperature prediction models, which typically require large-scale training datasets.

To address this issue, we propose a Fourier Neural Operator (FNO)-based framework to predict four temperature sequences during the synchronization phase using limited warm and hot startup data. The input consists of statistical features derived from two temperature-related and two steam-related variables observed during the preceding 3000 RPM holding phase. To ensure temporal consistency, all samples are padded to a unified sequence length.

The proposed FNO architecture leverages spectral convolution to capture global dependencies while maintaining local temporal resolution. Comparative evaluations with CNN, DNN, and LSTM models under identical training conditions demonstrate that the FNO consistently achieves higher predictive accuracy and robustness in five-fold cross-validation. These results indicate that the FNO-based framework is well-suited for modeling thermal dynamics in transient turbine operations where high-resolution data is scarce.

A System of Systems Architecture for Optimizing Aircraft Health Management in Civil Aviation

Takuro Koizumi, Nozomu Kogiso — Tue, 13 Jan 2026 00:00:00 +0000

This study proposes a system of systems (SoS) architecture for efficient aircraft health management (AHM) in civil aviation from the perspective of an aircraft manufacturer and formulates AHM as a multi-objective optimization problem. First, the SoS architecture is described to capture the interrelationship of the strategic capabilities required among the relevant stakeholders including airline customers, and the regulatory authority by using the Unified Architecture Framework (UAF). Parameters to measure strategic capabilities and operational activities are identified and the relationships between them are defined using parametric causal correlation. Next, AHM performance, effect, and amount of required data are formulated in terms of the identified variables in the SoS architecture description. Quantification enables the maximization of the effectiveness of AHM implementation by formulating it as a multi objective optimization problem, which allows for the quantitative assessment of the relationships between the context of AHM implementation and strategic capabilities. This formulation makes it possible to evaluate AHM effectiveness quantitatively, improving upon our previously proposed SoS architecture model, which only evaluated the relationship between stakeholders qualitatively.

A Transfer Learning Framework for Remaining Useful Life Estimation

Melanie Bianca Sigl, Klaus Meyer-Wegener — Tue, 13 Jan 2026 00:00:00 +0000

Training a robust deep learning (DL) model for remaining useful life (RUL) estimation or fault detection typically requires a large, high-quality labeled dataset. However, such datasets are often unavailable in practice. Transfer learning is a solution for smaller labelled datasets. Yet, the effectiveness of transfer learning heavily depends on selecting an appropriate source DL model; an unsuitable choice can result in negative transfer, where model performance deteriorates significantly.

To address this challenge, we introduce REAPER (Reusable Neural Network Pattern Repository), a framework designed to assist users in selecting the most suitable DL model for reuse in transfer learning scenarios. REAPER analyzes and compares the characteristics of available datasets and employs a learned ranking model to recommend the optimal source model. This paper presents the architecture, including its dataset characterization, ranking methodology, training procedure, and practical usage guidance.

Adversarial Domain Adaptation Fault Diagnosis Method Based on Self-attention Graph Convolutional Network

Bo Zhang, shuai su, Ning Ma, Yingxue Wang, Wei Li — Tue, 13 Jan 2026 00:00:00 +0000

Intelligent fault diagnosis has made significant progress with the advancements in deep learning and big data. However, the assumption of identical training and testing data distributions often fails in dynamic industrial environments, leading to performance degradation. To address this issue, we propose an Adversarial Domain Adaptation Fault Diagnosis Model Based on Self-attention Graph Convolutional Network (ADA-SAG). The model employs the k-nearest neighbors algorithm to construct graph structures that capture faultinstance relationships across source and target domains. A self-attention enhanced graph convolutional network extracts critical features, while a dual-classifier framework, combined with adversarial learning and maximum mean discrepancy regularization, ensures domain-invariant feature alignment. Experimental results on two benchmark datasets show that the proposed model achieves higher accuracy and robustness
compared to existing methods, making it suitable for diverse
operating conditions. Ablation studies further validate the
contributions of each component to the overall effectiveness
of the model.

AI-Driven Design Optimization of Engineering Systems: A Case Study on Turboshaft Engines

Satish Thokala, Peeyush — Tue, 13 Jan 2026 00:00:00 +0000

In a typical engineering design, there are often many design parameters to consider. Also, there are multiple competing requirements and objectives to meet. Manual approach of adjusting the parameters to achieve specific objectives is not optimal especially as the design becomes complex.

In the quest for optimizing complex engineering systems, the exploration of the design space becomes imperative, especially when dealing with multi-objective systems characterized by an array of independent variables. This paper presents a comprehensive study on the design space mapping of complex engineering systems, utilizing a turboshaft engine as a case study. The initial phase of our methodology employs a physics-based model to generate synthetic dataset, reflecting the intricate interplay of various system parameters underpinning the engine's operation. This synthesized data serves as a foundation for the subsequent development of a Machine Learning or Deep Learning based surrogate model. The surrogate AI model, will be crafted to encapsulate multiple inputs and outputs inherent in the turboshaft engine's functioning, thereby facilitating an efficient and accurate exploration of the design space.

Through this investigation, we will evaluate the efficacy of combining physics-based models with AI-driven techniques in mapping the design space of multi-objective systems. The core of our investigation revolves around the utilization of the AI surrogate model for achieving multi-objective optimization. This optimization process is not only focused on enhancing specific performance metrics but is also geared towards identifying a comprehensive family of feasible design solutions. Such an approach enables the delineation of the entire design space, offering invaluable insights into the trade-offs and synergies among different design objectives. Through this methodology, our goal is to uncover a wide spectrum of viable design alternatives, thereby providing a robust framework for decision-making in the engineering design process.

AI-Powered Runway Safety: YOLOv11-Based Detection of Foreign Object Debris

Srinivasarao Surapu — Tue, 13 Jan 2026 00:00:00 +0000

Runway safety is a critical aspect of aviation operations, with Foreign Object Debris (FOD) posing severe risks to aircraft during takeoff and landing. Incidents such as the Air France Flight 4590 (Concorde) crash have demonstrated the devastating impact of undetected FOD. Manual inspection methods remain the standard but are time-consuming, error-prone, and limited by environmental conditions. With annual global FOD-related losses estimated at over $22.7 billion, there is a clear need for automated, intelligent detection systems.

This study presents an AI-powered system using computer vision and deep learning to detect and classify runway FOD in real time. Leveraging the YOLOv11 model trained on a combination of the open-source FOD-A dataset and custom-collected images, the system achieves a mean Average Precision (mAP@95) of 89.3%. Data augmentation, class balancing, and annotation with CVAT further enhance model performance. The trained model is deployed via a web-based application with an Angular frontend and Flask backend, enabling rapid detection with high precision and user-friendly visualization.

While the current system is optimized for image-based detection, future work will focus on real-time video integration, edge device deployment, and airport safety system interoperability. Limitations include partial coverage of real-world conditions and small debris detection challenges, which are being addressed through ongoing dataset expansion and model refinement.

An Innovative Machine Learning driven approach to detect anomalous behavior of Dry Gas Seal Heaters for Centrifugal Compressors

Tue, 13 Jan 2026 00:00:00 +0000

Compressors are used in various industries to elevate the pressure of a gas to meet the process requirements. Dry Gas Sealing (DGS) system is used in compressors to contain the process gas within the casing with near zero leakage. Due to efficient sealing capabilities, Dry Gas Seals became key technology in industry to prevent the leakage to environment and to ensure compliance with health and safety regulations.

DGS is very sensitive to process gas conditions. To ensure health and integrity of DGS and to maintain required process gas conditions, Dry Gas Seal systems include various Auxiliary Equipment viz. Filters, Heaters, Control valves etc. Heaters maintain the Seal gas inlet temperature at a specified value to avoid presence of condensate, which can otherwise result in Seal failure. To maintain this temperature, Heaters operate in an on-off toggling pattern. Failure of these Heaters can lead to unit unavailability, Early detection of anomalous Heater operation can ensure timely action to avoid any possible negative impacts on Dry Gas Seal health which in turn can impact Compressor operation.

The paper makes a theoretical survey of existing pattern recognition algorithms for time series and examines their applicability in detecting anomalous operation of Seal Gas Heaters. Finding these methods not directly useful, the paper presents a state-of-the-art physics plus data-driven approach. The method is developed by combining LSTM type Neural Network with Periodogram and Auto-correlation monitor to detect any deviations from normally expected operating behavior of Seal Gas Heaters. The LSTM learns an exhibited pattern and looks for the similar patterns in Heater signals. Auto-correlation monitor coupled with Periodogram helps in determining the dominant frequencies and window-size required for the LSTM component. The method is tuned to accommodate different operating modes of Heater based on Compressor running conditions. If Heater deviates from working in an established toggling mode, user is alerted before Seal gas temperature is impacted.

Applied in real-time, the method alerts engineers for any anomalies observed in Heater behavior, thus enabling swift action to prevent any harm to the Dry Gas Seals, caused by temperature upsets. The paper demonstrates the performance of this method when applied on 50 compressors, thus validating the applicability and accuracy in prognostic health management of Dry Gas Seal systems of Compressors.

Anomaly Detection Framework for Rotary Equipment Using ContinuousWavelet Transform and U-Net Autoencoders

Mohamed Zamil Kanjirathingal Rafeek, Ulrich Schäfer — Tue, 13 Jan 2026 00:00:00 +0000

Recent advances in data-driven methods, particularly deep learning, have transformed predictive maintenance for rotary machinery. These methods enable intelligent, sensor-based condition monitoring from unlabeled operational data, even under rare-fault conditions. This study proposes an unsupervised anomaly detection framework for rotary equipment that utilizes continuous wavelet transform (CWT) to transform unlabeled, multichannel vibration signals into stacked time-frequency scalograms using complex Morlet wavelet. These scalograms are then processed by an enhanced U-Net deep convolutional autoencoder (CWT-U-Net CAE), which learns features of healthy operational conditions and detects anomalies by identifying significant deviations in reconstruction error. Coupled with its edge-compatibility, the framework enables scalable real-time condition monitoring in industrial environments. A custom test bench with an induction motor was used to obtain realistic vibrational signatures under normal operating conditions, assessing the effectiveness of the proposed approach.

Assessment of simulation-based data augmentation technique by Uncertainty Quantification for Spacecraft Propulsion System PHM

Shotaro Hamato, Himeko Yamamoto, Noriyasu Omata, Yu Daimon, Seiji Tsutsumi — Tue, 13 Jan 2026 00:00:00 +0000

One of the challenges of applying Prognostics and Health Management (PHM) in industrial systems is the lack of labelled training data including anomalies and faults. This study proposes training data generation by a physics-based numerical model and uncertainty quantification (UQ) considering input uncertainty and model form uncertainty, and demonstrates the proposed methodology in a spacecraft propulsion system. A one-dimensional numerical model of the spacecraft propulsion system has been developed in which ignition delay and trapped bubble dynamics are modeled. Sources of uncertainty originating in input variables of the numerical model are identified by domain experts. The probability distributions of them are modeled as uniform distributions, and training data are generated through the propagation of these probability distributions using a Monte Carlo approach. The generated training data were compared with available experimental data and showed good agreement in time-series and frequency-domain response. The 95% confidence interval (C.I.) of total uncertainty, integrating input uncertainty and model form uncertainty, was evaluated through UQ. The generated data enables the use of unsupervised methods for anomaly detection. The C.I. can be used as the normal space for anomaly detection.

Bivariate degradation modeling and reliability analysis based on a shared frailty factor with truncated normal distribution

Lu Li, Zhihua Wang — Tue, 13 Jan 2026 00:00:00 +0000

For a complex product, a single performance characteristic(PC) often fails to fully reflect its degradation process, making it essential to consider the joint degradation of multiple PCs. In this paper, we propose a bivariate degradation model based on a shared frailty factor with the truncated normal distribution, using Wiener processes to characterize the marginal distributions of the PCs. The assumption of the truncated normal distribution aligns better with the physical background where the degradation rates of PCs are non-negative during actual degradation processes. Furthermore, a method for inferring unknown parameters is developed by employing the expectation maximization algorithm. Under this modeling assumption, it became possible to obtain an analytical expression for the product's lifetime distribution on the basis of the concept of the first hitting time. Therefore, in this paper, we further extend the normal distribution integral lemmas to the case of the truncated normal distribution, and provide analytical expression for the cumulative distribution function of the product lifetime. Finally, the rational effectiveness of the proposed model and methods is validated through a numerical simulation example and a case study on wheel wear.

Causal Graph-Based Anomaly Detection for Battery Modules in Electric Heavy-Duty Vehicles

Yuantao Fan, Carlos Camacho, Sepideh Pashami, Slawomir Nowaczyk — Tue, 13 Jan 2026 00:00:00 +0000

Heavy-duty battery electric vehicles rely on large and complex energy storage systems (ESS), composed of multiple battery modules, whose individual health and reliability are critical to vehicle performance and safety. This study applies an unsupervised anomaly detection framework, COSMO (Consensus Self-Organizing Models), to a naturalistic real-world dataset collected during routine operations of in-service heavy-duty vehicles. We extend the baseline COSMO by incorporating causal discovery algorithms to help detect early signs of faults in ESS across heterogeneous missions and external conditions. On-board sensors data is collected as a multivariate time series, including information such as voltage, current, temperature, state of charge, etc. Given the wide range of applications of heavy-duty vehicles, these signals typically exhibit extreme variability even under normal operation, making anomaly detection challenging. Causal graph discovery allows us to acquire latent structures that capture the underlying relationships among these influential features. The resulting learned causal graphs, for each battery module, serve as a more consistent representation that captures each battery module’s usage and behavior over time. Since battery modules within the same ESS are expected to behave similarly under comparable operating conditions, COSMO models them as a homogeneous group. We then mark as anomalous modules that are identified to exhibit causal graph representations deviating markedly from the consensus.

Cleaning Maintenance Logs with LLM Agents for Improved Predictive Maintenance

Valeriu Ionut Dimidov, Faisal Hawlader, Sasan Jafarnejad, Raphael Frank — Tue, 13 Jan 2026 00:00:00 +0000

Maintenance logs serve as the backbone of data-driven Predictive Maintenance (PdM) systems by providing information that can be used to create and label datasets for training survival analysis and machine learning (ML) models. However, due to personnel manually entering information into maintenance logs and the various levels of flexibility that maintenance tracking systems allow, service records often contain errors. Currently, the cleaning of equipment maintenance records is performed manually by experts such as data scientists or reliability engineers. Nevertheless, this task is time-consuming and often does not entirely eliminate noise from the data. In this paper, we propose using large language model (LLM)-based agents to automate the cleaning of maintenance logs. We provide an implementation that allows the agents to perform data cleaning as well as metrics to assess agents' performance. Finally, we compare the performance of several LLMs on this task. Our empirical results indicate that LLM-based agents are a promising solution for improving the quality of the datasets used in PdM systems and ultimately developing predictive maintenance models that are more reliable and useful.

Combining Statistical Models and AI for Predictive Maintenance: RUL Estimation of Reactor Protection System Components

Jung Hwan Kim, Chang Hwoi Kim, Joon Ha Jung, Sangchul Park — Tue, 13 Jan 2026 00:00:00 +0000

Reliable operation of digital instrumentation in nuclear power plants depends heavily on accurate prediction of component degradation. This study proposes a hybrid framework for estimating the remaining useful life of photo-couplers used in reactor protection systems. Accelerated aging tests were performed under elevated thermal conditions to generate representative degradation data. Both statistical models and a neural network were developed to analyze long-term performance decline.

The AI model incorporates polynomial features and custom loss functions to reflect realistic monotonic and exponential degradation behavior. Its predictions closely matched those of the statistical models, with projected lifespans ranging from 22 to 24 years. A user-oriented software tool was also implemented to support real-time remaining useful life forecasting using field data, demonstrating the practical value of combining traditional and AI-based approaches for predictive maintenance in nuclear systems.

Complete Life Cycle Degradation Trajectory Learning Utilizing Time-Transformer and Denoise Auto-Encoder for Remaining Useful Life Prediction

xianpengqiao, Veronica Lestari Jauw, Tiyamike Bnada, Chin Seong Lim — Tue, 13 Jan 2026 00:00:00 +0000

Deep Learning (DL) has substantially expanded its role in Prognostics and Health Management (PHM), particularly by enabling automated feature extraction for Remaining Useful Life (RUL) prediction. Despite this progress, existing DL models such as Long Short-Term Memory (LSTM) networks still face challenges in accurately capturing complete life-cycle degradation trajectories. To address this limitation, this study introduces a hybrid semi-supervised model that integrates a Time-Transformer (TT) with a Denoising Autoencoder (DAE), termed TT-DAE. The DAE first extracts spatial features and suppresses noise through signal reconstruction from degraded inputs. These extracted features are then separated into source and target domains and normalized to a uniform sequence length using a padding strategy. Subsequently, the TT module leverages both source features and a Sliding Variable-Length Window (SVW) mechanism to learn full degradation trajectories. A comprehensive experimental evaluation conducted on the C-MAPSS dataset demonstrates the effectiveness of the proposed approach, achieving an average Pearson Correlation Coefficient (PCC) of 0.89 between the predicted and actual target signals.

Damage Localization in a CFRP Beam via Modal Frequency Shifts and Nearest Neighbor Classification

Ömer Dehan Özboz, Özkan Altay, Murat Özbayoğlu, Özgür Ünver — Tue, 13 Jan 2026 00:00:00 +0000

Vibration-Based Structural Health Monitoring (SHM) systems offer significant potential for damage detection due to their non-destructive nature and real-time capabilities, while reducing maintenance costs for aerospace and automotive applications. This study investigates the effect of damage on the modal parameters of a Carbon Fiber Reinforced Polymer (CFRP) fixed-free beam, with the goal of identifying damage location and severity. The lamina material properties of the CFRP were evaluated using composite lamination theory (CLT). By altering the location and depth of the damage, numerical analyses were conducted on the CFRP beam, and discrepancies between the intact and damaged models were examined. Modal frequency shifts were quantified using Relative Natural Frequency Change (RNFC), and RNFC-based mapping surfaces dependent on damage location and severity were generated for first four transverse vibrational modes of the beam. The model was validated through experiments on the intact and damaged CFRP specimens. The beam was excited with an impact hammer near the fixed-end, and responses were collected by piezoelectric sensors placed along the beam and laser vibrometer focused on the free end of the beam. The modal parameters were extracted using Ho-Kalman’s subspace method and experimental RNFC results of damaged samples were calculated. Then Nearest Neighbor search algorithm was successfully employed to estimate the damage location and severity by comparing experimental results to generated RNFC-based mapping surfaces.

Data-driven Anomaly Detection for Quadcopter UAV Indoor Flight Platform

Gengyu Li, Chun Fui Liew, Naoya Takeishi, Takehisa Yairi — Tue, 13 Jan 2026 00:00:00 +0000

Ensuring the safe operation of unmanned aerial vehicles (UAVs) requires timely and accurate detection of anomalies that may indicate system faults or external disturbances. In this study, we propose a data-driven approach for unsupervised anomaly detection in UAVs, leveraging a newly developed multimodal dataset that includes synchronized telemetry, sensor measurements, motion capture data, and pilot inputs. Our method learns representations of normal UAV behavior from healthy flight records and is applied to fault-injection scenarios to identify potential anomalies. Preliminary results on experimental data suggest that the approach can capture subtle deviations from expected behavior across multiple data modalities, including flight dynamics and environmental feedback. This work lays the foundation for data-driven UAV health monitoring through unsupervised learning. It complements our publicly released dataset and analysis tools and aims to facilitate broader research on autonomous anomaly detection, early fault diagnostics, and the development of resilient UAV systems in safety-critical applications.

Development and Sharing of a Multi-Modal Indoor UAV Dataset for PHM Research

Tue, 13 Jan 2026 00:00:00 +0000

With the rapid expansion of unmanned aerial vehicle (UAV) applications, ensuring reliable and safe operation has become a pressing challenge. In this paper, we present a modular quadrotor-based data acquisition platform designed to capture rich, high-fidelity data under motion capture guidance. Our system integrates conventional flight telemetry with detailed vibration and temperature measurements, user input logs, and precise 6D motion tracking. This offers an comprehensive view into the UAV’s physical and control state. We describe our systematic process for data cleaning, organization, and exploratory analysis, laying the groundwork for robust prognostics and health management (PHM) research. To illustrate the platform’s potential, we implement supervised and semi-supervised models for anomaly detection and fault identification. We release the dataset, synchronized flight videos, and analysis code to accelerate UAV health-monitoring research and collaboration.

Development of a Hierarchical Anomaly Detection System for Steelmaking Processes and Proposal of a Model Update Method

Yohei Harada, Masafumi Matsushita, Takehide Hirata — Tue, 13 Jan 2026 00:00:00 +0000

In an integrated steelmaking process, equipment failures can significantly impact overall operations. Therefore, predictive detection and prevention of failures are critical. In this study, we developed and implemented a three-layer hierarchical predictive detection system to utilize large-scale, multivariate operational data. This system is designed to identify overall trends by leveraging big data, detect correlation breakdowns through domain knowledge, and detect shifts in single-signal levels. We demonstrated its effectiveness using real plant data from the steelmaking process. In addition, general anomaly detection models, including our system, rely on quantifying deviations from a normal state as an anomaly score. In manufacturing settings, data drift often occurs due to factors such as equipment part replacements or changes in operational conditions. When data drift occurs, it becomes necessary to redefine the normal state. However, in manufacturing environments, temporary runs or experimental operations mean that the data following a drift is not necessarily guaranteed to normal data. Therefore, it is necessary to evaluate whether the data distribution is normal before and after the drift on a case-by-case basis. Current approaches do not provide a quantitative means to make this decision, leading to the issue that model updates depend on the judgment of experts. To address this, we propose a method that utilizes similar equipment conditions to guide the timing and procedure for model updates. By applying Jensen–Shannon divergence to measure differences among four data distributions—derived from two machines and two distinct periods— we provide appropriate guidance for model construction based on a table of potential anomalies. Through validation using real data from two adjacent continuous casters, we confirmed that identifying abnormal equipment and time periods enables us to propose appropriate normal operating windows. From these validation results, Verification results indicate that the proposed system allows for comprehensive predictive maintenance, integrating domain knowledge and thereby contributing to stable operations in steelmaking facilities.

Development of a HUMS for UAV hybrid power system using digital twin and AI techniques

Chiara Sperlì — Tue, 13 Jan 2026 00:00:00 +0000

Traditional aeronautical power systems, typically based on fossil fuels, present a series of limitations, including: i) the added weight associated with onboard fuel storage, ii) limited endurance due to fuel consumption, and iii) the emission of atmospheric pollutants. These constraints become particularly critical in application scenarios where extended endurance and environmental sustainability are key requirements. A notable example is represented by high-altitude, long-endurance (HALE) unmanned aerial vehicles (UAVs), whose deployment is rapidly expanding due to their suitability for a wide range of missions, including surveillance, environmental monitoring, and long-range communications. To enable the technological advancement of such platforms, alternative power generation architectures must be explored. In this context, hybrid electric power systems, integrating solar panels, lithium-ion batteries, and fuel cells, offer a promising solution. Solar and fuel cell subsystems ensure stable and continuous power generation over extended periods, including during night-time operations, while lithium-ion batteries provide high-power bursts during transient phases such as take-off, landing, or auxiliary system activation. Nevertheless, the use of these hybrid systems introduces unique challenges in terms of safety, reliability, and system complexity. They must operate in harsh environmental conditions and often require remote monitoring capabilities that enable condition-based intervention without interrupting critical missions. To address these challenges, this paper presents a Health and Usage Monitoring System (HUMS) for a hybrid power system composed of a solar panel, a lithium-ion battery, and a fuel cell, developed through the integration of digital twin modeling and artificial intelligence (AI) techniques. In particular, AI data-driven methods provide a powerful and flexible framework for monitoring the complex system composed of multiple energy sources. However, to achieve reliable performance, they require large and representative datasets, which are often unfeasible to obtain experimentally. To overcome this limitation, a digital twin of the hybrid power system is developed in the MATLAB/Simulink environment and used to simulate system behavior under both healthy and faulty conditions. The resulting synthetic data are then employed to train diagnostic/prognostic algorithms. This approach offers an efficient and scalable solution for implementing intelligent health monitoring in hybrid power systems, enhancing reliability, autonomy, and operational availability in long-endurance UAV applications.

Development of an Interactive Twinverse System

Yongho Lee, Huichan Park, Seongbin Choi, Sang Won Lee — Tue, 13 Jan 2026 00:00:00 +0000

Digital twin (DT) platforms are increasingly used for PHM, yet most systems still lack real-time, bi-directional control between physical assets and virtual models, and a unified semantic layer that grounds natural-language commands in plant constraints. To address this gap, this research present Twinverse, an interactive metaverse environment that integrates a ROS/Kafka-based bi-directional DT, a knowledge-graph (KG) semantic backbone, and an LLM-powered agent. The KG encodes structural/operational constraints (e.g., kinematics, limits) and is serialized into a vector store to support RAG-based intent interpretation, while a constraint-aware execution pipeline verifies workspace, joint limits, and speed bounds prior to dispatch. Implemented on an industrial robot cell in Unity, the system provides real-time synchronization and multi-user operation within a single immersive interface. In evaluation, the platform maintained tight virtual–physical tracking and stable latency under increasing user load, and it enabled PHM-oriented functions such as anomaly interrogation and explainable, context-aware action generation. Our contribution is a cohesive DT–KG–LLM architecture that (1) grounds language-to-action in machine-readable plant constraints, (2) closes the loop from natural-language intent to verified execution, and (3) operationalizes PHM analytics inside an immersive DT environment. This work demonstrates a practical path toward interactive, explainable, and real-time PHM decision support.

Development of Bearing Fault Diagnosis Model Using Low Frequency Data Based on Knowledge Distillation

Yongjae Jeon, Secheol Yang, Sang Won Lee — Tue, 13 Jan 2026 00:00:00 +0000

Bearings are critical components for ensuring smooth rotational motion in mechanical systems, and reliable operation requires continuous condition monitoring for fault diagnosis. Recently, there has been growing interest in diagnosing bearing conditions using artificial intelligence, particularly deep learning-based approaches. However, in real industrial environments, limitations such as high sensor cost and restricted data storage often lead to the use of low sampling frequency sensor data, which poses challenges in developing accurate diagnosis models. To address this issue, this paper proposes a bearing fault diagnosis method based on knowledge distillation to enhance the utility of low sampling frequency data. High-frequency acceleration data were collected under both normal and faulty conditions and subsequently downsampled for knowledge distillation. A 1D CNN-based teacher model was trained using high-frequency data, and multiple loss functions were designed to distill both final predictions and intermediate features into a student model trained on low sampling frequency data. The performance comparison between models with and without knowledge distillation verified the effectiveness of the proposed approach. The results demonstrate the feasibility of developing fault diagnosis models using low sampling frequency data in real industrial settings and suggest an effective knowledge distillation strategy.

Development of Virtual Thermal Sensor based on Multivariate Time Series Prediction for Estimating Internal Contact Temperature of High-Voltage Relays

Sewoong Gim, Jaephil Park, Sanghoon Lee — Tue, 13 Jan 2026 00:00:00 +0000

In power electric (PE) systems of electric vehicles (xEVs), the function of switching power between the battery and inverter is essential for safe operation. This function is performed by the Power Relay Assembly (PRA), which consists of two high-voltage relays (HV relays) that connect the battery’s positive terminals and continuously conduct high current during driving. Unlike low-voltage relays, HV relays are designed with arc-suppression structures and timing control to minimize arc damage, resulting in relatively rare arc erosion failures. However, the passage of high current through the contact interface generates significant heat, leading to material degradation and structural deformation, which have been identified as major failure modes. Such failures can cause abnormal open or short circuits in the high-voltage system, potentially resulting in loss of power control or fires—posing critical safety risks. Therefore, early diagnosis of these failures is essential. HV relays have a sealed design without dedicated cooling, causing heat to accumulate internally under high-current conditions. For effective failure diagnosis, internal thermal monitoring is crucial. However, due to constraints such as limited design space, the need to maintain airtightness, and cost considerations, the number of sensors that can be installed for relay state monitoring is severely limited. As a result, accurately capturing the internal terminal temperature, which is a critical indicator of failure risk, remains challenging. This study proposes the development of a multivariate time-series prediction-based Virtual Thermal Sensor (VTS) [1, 2] model capable of estimating the internal terminal temperature of HV relays using only limited external sensing data available in actual vehicles—such as relay voltage, current, and ambient temperatures. Because the internal temperature evolution depends both on past history and current operating conditions, a Bidirectional Long Short-Term Memory (Bi-LSTM) [3, 4] network was employed to effectively capture these dependencies.diagnostics and remaining useful life (RUL) prediction of high-voltage relays. Future work will focus on refining the VTS model for improved accuracy and reduced computational complexity, supporting the development of practical RUL prediction systems for HV relays.

For model training and validation, datasets were collected through high-temperature accelerated life tests of HV relays. The dataset includes time-series measurements of ambient temperature, relay surface temperature, bus-bar temperature, operating voltage, and load current, along with specially instrumented measurements of internal terminal temperature for ground truth validation. To improve prediction accuracy, key influencing variables were selected through correlation analysis, and advanced data preprocessing steps were applied to handle irregular cycles, long-term trends, and saturation regions—ensuring the data’s suitability for modeling. Up to 50% of the total dataset will be reserved for validation, with prediction accuracy to be statistically assessed using metrics such as Root Mean Squared Error (RMSE). This research provides a foundational technology for indirectly estimating the internal terminal temperature—an otherwise difficult-to-measure parameter—enabling degradation diagnostics and remaining useful life (RUL) prediction of high-voltage relays. Future work will focus on refining the VTS model for improved accuracy and reduced computational complexity, supporting the development of practical RUL prediction systems for HV relays.

Enabling Condition Based Maintenance Strategy for Radar Systems – Data Driven Approach

Rafik HADJRIA — Tue, 13 Jan 2026 00:00:00 +0000

Systems of a radar could encounter major failures that lead to a complete stop of functions. There exist warning signs before that event, and with learning algorithms, a developed model can predict that occurrence and give the opportunity to prevent it or to plan maintenance before. The challenge is to understand systems and define rights features in order to create the best predictive model possible with available data.

In this paper, we will address the concept of a condition-based maintenance strategy in the context of an electromechanical system based on a data-driven approach. Technically speaking, we will address the block function presented in Figure 1 and show some promising results regarding anomaly detection (Figure 2).

Enabling Model-Based RAMS Through LLM-Driven Legacy Data Transformation

You-Jung Jun, Navid T. Zaman, Derek Kim, Stecki Yanek, Raphaël Chagnoleau — Tue, 13 Jan 2026 00:00:00 +0000

The rapid digital transformation in engineering, coupled with the development of increasingly complex systems, is pushing industries to develop smarter and more efficient methods for system development. Major stakeholders/ industries are moving towards a model-based framework for systems engineering, RAM, and safety analysis to manage growing system complexity while maintaining data consistency and traceability.

The convergence and consolidation of previously document-based engineering approaches allows for the standardization and streamlined capture of knowledge across engineering disciplines. In this framework, data availability and interoperability can easily become a bottleneck without comparable innovation to tooling and processes. In more recent times Artificial Intelligence (AI) has been identified as a powerful enabler on this front. AI can assist engineers in developing RAMS models more efficiently by leveraging legacy data, such as historical FMECAs, and aligning it with standardized taxonomies to automatically and rapidly develop system models for downstream analysis requirements.

Enhanced Fault Isolation and Part Recommendation for Airplane Health Management with Hybrid Probabilistic Modeling

Tue, 13 Jan 2026 00:00:00 +0000

Aircraft maintenance plays a crucial role in ensuring the safety and reliability of aircraft operations. Effective fault isolation and accurate part recommendation are essential tasks in the maintenance process. The accuracy of existing fault isolation solutions in complex situations (e.g. having multiple fault code scenarios) needs improvement. In this paper, we propose a novel approach of Hybrid Probabilistic Modeling based Fault Isolation Framework combining two solutions. One of the solutions is Pattern Similarity-based Probabilistic Modeling (PSPM) which leverages historical maintenance data to build a probabilistic model that captures patterns of faults and their associated parts replacement. By comparing the current fault symptoms to these patterns, this solution enables more accurate fault isolation and suggests suitable parts for replacement compared to legacy methods. On the other hand, the Physics Informed Probabilistic Modeling (PIPM) employs a Bayesian network to leverage system knowledge in terms of schematics, particularly in scenarios where historical data is sparse or non-existent. Both probabilistic modeling-based solutions complement each other, address gaps, and enhance the efficiency and effectiveness of aircraft fault isolation.

In this paper, we will first provide an introduction that outlines the significance of aircraft maintenance and the challenges associated with fault isolation. Following this, we will present a survey of existing fault isolation techniques, highlighting their strengths and limitations. We will then discuss the proposed hybrid solution and its advantages in improving fault isolation. Next, we will delve into the Pattern Similarity-based Probabilistic Modeling (PSPM) methodology, detailing its benefits and showcasing a case study that highlights its effectiveness. We will also explore the Physics Informed Probabilistic Modeling (PIPM) approach, presenting an overview of its theoretical foundations and a case study that illustrates its practical application. Finally, we will conclude with a summary of our findings and their implications for future research and practice in the field of aircraft maintenance.

Enhancing Machine Reliability in Industrial Plants Leveraging Diagnostic and Prognostic Approach to measure reliability improvements

Pranay Mathur, Carlo Michelassi, Leonardo Vieri, Gilda Pedoto — Tue, 13 Jan 2026 00:00:00 +0000

In the dynamic and demanding environment of industrial plants, the reliability of machines is paramount. Ensuring that machinery operates reliably and efficiently is crucial for profitability of the plant. Reliability in industrial plants is beyond preventing failures and also about enhancing performance and extending the lifespan of equipment. By focusing on the most failure-prone components or systems, maintenance teams can prioritize their efforts and resources effectively, leading to significant improvements in overall reliability and total cost of ownership. This abstract delves into the critical role of reliability in industrial environments, emphasizing the importance of employing reliability growth models to systematically validate the effectiveness of solutions implemented to address machine reliability issues.

For every unplanned events(trips), remote real-time data gathering and analysis conducted to identify the components or systems responsible for the trip. All the events and contributors are tracked and trended to identify top offenders. Identified top offenders are deeply investigated to find the solution & opportunity to develop the automatic diagnostic and prognostic tool based on remotely acquired time-series data. Based on outcome of Diagnostic and prognostic tools, identifying the degradation of equipment. Once a malfunction is identified, we analyze root causes, extract learnings, and develop targeted improvements. These improvements are first validated in controlled environments (e.g., lab or test bench), then implemented incrementally across the fleet. Each implementation cycle is tracked using reliability growth models to statistically measure the reduction in failure rates and validate the effectiveness of the solution over time.

This process allows us to Diagnose and isolate malfunctions using real-time analytics, Generate and refine new analytics based on observed failure modes, Quantify reliability growth through Mean Time between failure (MTBF) decrease / Mean Time Between Trip (MTBT) improvements, Scale validated solutions from individual assets to the entire fleet.

By integrating reliability growth models into our reliability process, we ensure that each improvement is measured and also predictively estimate the reliability improvement on any other unit in the fleet. This methodology has already demonstrated success, with MTBT improvements from 1,000 to 8,000 hours, showcasing the power of structured reliability growth modelling in complex, distributed systems.

Evaluating Failure Time Probabilities for Compound Degradation with Linear Path and Mixed Jumps

Shihao Cao, Zhihua Wang, Xiangmin Ouyang, Pengjun Zeng — Tue, 13 Jan 2026 00:00:00 +0000

Many devices may experience nature degradation and mixed jumps simultaneously whose types can be divided into positive jumps and negative jumps, while these complicated performance rules also bring difficulties in lifetime analysis within the concept of the first hitting time. To address this issue, this paper first proposes a compound degradation process, which is characterized by linear path and mixed jumps. Then, by adopting the idea that transforms the positive jumps into the threshold, an approximate lifetime solution is derived. Given the realistic application of furnace wall, numerical verification shows that the proposed method can maintain consistency with Monte Carlo simulation, while conspicuous errors exist for existing methods, demonstrating that the proposed method can be regarded as theoretical support for the future studies.

Explainable and Trustworthy AI for Fault Classification in the Tennessee Eastman Process: A Step Toward Industrial Autonomy

Jayanth Balaji Avanashilingam, Bijuraj Pandiyath Velayudhan — Tue, 13 Jan 2026 00:00:00 +0000

Achieving higher levels of Industrial Autonomy (IA) requires fault diagnostic systems that combine predictive accuracy with transparent decision-making. In safety-critical process industries like petroleum refinery, black-box AI models often face adoption barriers due to limited interpretability. The work introduces a glass-box fault classification framework for the Tennessee Eastman Process, comparing a baseline direct-modeling approach with a novel dual-branch architecture. The proposed method decomposes process parameters into trend and cyclic components, trains dedicated classifiers on each and fuses their probabilistic outputs. The proposed design improves sensitivity to both gradual drifts and oscillatory anomalies. In the present work SHAP explainability is incorporated to provide global, local, and class-wise feature attribution, enabling operators to trace model reasoning and align diagnostics with process knowledge. A strong industrial AI platform, purpose-built for domain engineers, emerges as essential for operationalizing such capabilities, empowering process experts to directly harness AI for decision-making. The present work serves as a steppingstone toward realizing such an Industrial AI platform, demonstrating how interpretable AI can bridge the gap between advanced analytics and domain expertise. The experimental evaluation of the proposed technique demonstrates that 35% of the fault classes achieved improved accuracy, with an average accuracy gain of 4.34% over the baseline, with pronounced gains in cyclic-dominated faults. The approach demonstrates a pathway toward Level 5 IA by delivering interpretable, high-performance fault diagnostics ready for real-time deployment.

Fault-Induced Signal Distortion in FMCW Automotive Radar: A Simulation-Based Analysis

Sheriff Murtala, Ingyu Lee, Soojung Hur, Gyusang Choi — Tue, 13 Jan 2026 00:00:00 +0000

Faults in automotive radar subsystems distort the radar signal and compromise system performance, especially in frequency-modulated continuous wave (FMCW) radar architectures. However, the signal-level consequences of such faults remain underexplored. This paper presents a simulation-based analysis of signal distortions caused by five representative fault behaviors across three critical FMCW radar subsystems: the waveform generator, transmitter, and receiver. We examine the effects of each fault on complex baseband signals and range estimation accuracy, providing both qualitative and quantitative evaluations. The results reveal distinct distortion patterns and demonstrate that range errors and false negatives can occur independently, highlighting the need for diagnostic and fault-aware processing strategies. This work offers a foundational perspective on fault-induced anomalies in radar signal processing and supports the development of more robust FMCW radar systems.

Flare Gas Flow Rate Estimation Using Multimodal Deep Learning

Yu Watanabe, Kento Ishii, Nana Tamai, Takehisa Yairi, Naoya Takeishi — Tue, 13 Jan 2026 00:00:00 +0000

In refinery operations, flare gas is generated as a byproduct. It is harmful to human health and the environment and causes secondary issues such as noise and unpleasant odors. Flare stacks are commonly used to combust and neutralize flare gas before releasing it into the atmosphere. Accurate monitoring of flare gas flow rate is essential for flare gas reduction and recovery, but installing flow meters is costly. This study proposes a method to estimate flare gas flow rate using the flare images and suppression steam flow rate. Flare images are processed with a convolutional neural network (CNN) to extract spatial features, while suppression steam time-series data are processed with a long short-term memory (LSTM) network to capture temporal dynamics. These features are fused and passed through fully connected layers to regress the flare gas flow rate. To address data imbalance due to the infrequent occurrence of flare events, we designed a custom loss function that assigns higher weights to high-flow samples while penalizing overestimation when low-flow samples are incorrectly predicted as high flow. Furthermore, we employed data augmentation, preprocessing techniques, and feature engineering to improve prediction accuracy.

Fractal-based Satellite Health Monitoring

Tue, 13 Jan 2026 00:00:00 +0000

Satellites and space systems are crucial to the success of modern space exploration, particularly as the scale and complexity of missions continue to increase. Given the high investment costs and the impossibility of physical intervention once in orbit, designing reliable and fault-tolerant platforms is crucial for success. Nevertheless, the extreme and unpredictable conditions of the space environment frequently lead to anomalies that threaten mission success.
Telemetry data is therefore indispensable for real-time and predictive monitoring of system health. However, its complexity, multidimensionality, and the presence of noise pose significant challenges to traditional analytical techniques.

In this context, fractal analysis provides a robust set of tools for uncovering hidden patterns in telemetry signals, enabling the early detection of system degradation and anomalies. Unlike conventional threshold-based approaches, fractal methods are sensitive to changes in signal regularity and complexity, making them suitable for pre-failure diagnostics and trend forecasting.

This work investigates the use of fractal-based techniques for satellite health monitoring. The methods are applied to the Mission 1 dataset of the ESA Anomaly Detection Benchmark (ESA-ADB) database, enabling performance evaluation under realistic operational conditions. A comparative analysis is conducted to assess the diagnostic capability, robustness, and computational efficiency of each method, with a focus on identifying subtle anomalies and facilitating proactive decision-making.

The results highlight the potential of fractal techniques to enhance the interpretability and autonomy of satellite prognostics and health management (PHM) systems. By enabling more sensitive and timely diagnostics, this approach contributes to improving the operational resilience and life-cycle management of future space missions.

From Engineering Drawings to Assembly Instructions: A Vision and Language Model Approach

Shokhikha Amalana Murdivien, Minji Kim, Kyung Wan Choi, Jumyung Um — Tue, 13 Jan 2026 00:00:00 +0000

Engineering drawings such as CAD draft sheets are widely used in manufacturing to document product structure, part geometry, and dimensional specifications. While these doc- uments contain valuable information, they are not typically organized to support step-by-step assembly tasks, which can present challenges for non-expert technicians during installa- tion, maintenance, or repair. This paper presents a system that automatically generates structured and human-readable assembly instructions from CAD drafts by combining a vi- sion model, an OCR model, and a language model. The vision model, trained on a constructed synthetic dataset, was able to detect mechanical components with an average precision score of 95.2% on real CAD sheets, while the OCR model suc- cessfully extracted dimensional information. These outputs, together with existing description text, were processed by a language model to produce clear and interpretable assembly steps. A synthetic dataset was used to train the vision model, addressing the lack of publicly available CAD annotations. The results demonstrate that the proposed system improves the interpretability and usability of engineering documentation in assembly-related tasks.

From state observer to deep neural network: design, optimization, and application in bearing dynamics modeling

Yiliang Qian, Yan Wang, Diwang Ruan, Zhaorong Li, Jianping Yan, Clemens Gühmann — Tue, 13 Jan 2026 00:00:00 +0000

Neural networks have been widely applied in system dynamics modeling. However, traditional neural networks still encounter limitations in capturing long-term dynamics, nonlinear modeling, and interpretability. To address these challenges, this study proposes a novel neural network architecture, Deep Stacked State-observer based Neural Network (DSSO-NN). Firstly, the state-space representation is introduced, integrating discretized state-space equations into the neural network design to leverage both system state information and deep learning capabilities. Subsequently, two optimization measures are employed to enhance the network's nonlinear modeling ability with activation functions and the state observer, respectively. Finally, DSSO-NN is validated using the Case Western Reserve University bearing dataset. Experimental results demonstrate that activation functions provide minimal improvement to model performance. In contrast, the incorporation of the state observer significantly enhances the DSSO-NN's ability to capture system dynamics behaviors and improves modeling accuracy. DSSO-NN exhibits higher precision and greater stability, offering a novel perspective on using the state observer as an alternative to traditional activation functions.

Graph-Based Adaptive Anomaly Detection Framework for Dual-Fuel Marine Engines

Jaewoong Choi, Yoojeong Noh, Young-Jin Kang — Tue, 13 Jan 2026 00:00:00 +0000

Dual‑fuel (DF) marine engines, capable of operating on both diesel and LNG, face significant monitoring challenges due to frequent mode switching, dual valve timing, and load variability, which create nonlinear, time‑varying dependencies among sensors. Such dynamics undermine conventional time‑series anomaly detection methods that overlook structural relationships. To address this, we propose a graph‑based anomaly detection framework tailored for DF engine monitoring. Sensor readings are modeled as nodes, with edges encoding domain‑informed physical or functional dependencies. A multi‑head Graph Attention Network (GAT)–based overcomplete autoencoder captures both local sensor behavior and global structural patterns; the expanded latent space preserves fine‑grained features and heightens sensitivity to subtle deviations. The encoder aggregates context‑aware features, and the decoder ensures graph‑consistent reconstruction. Anomalies are scored using a λ‑weighted combination of node‑level reconstruction error (RMSE) and graph‑level structural inconsistency from Graph Laplacian Smoothness (GLS). The λ parameter is optimized post hoc on validation data via F1‑score, balancing sensitivity and precision. Evaluation on ten months of DF engine data demonstrates interpretable, real‑time fault detection and sensor‑level localization, supporting practical, condition‑based maintenance.

Hierarchical Anomaly Detection and Model Update Framework for Steel Manufacturing

Yohei Harada, Masafumi Matsushita, Takehide Hirata — Tue, 13 Jan 2026 00:00:00 +0000

In an integrated steelmaking process, equipment failures can significantly impact overall operations. Therefore, predictive detection and prevention of failures are critical. In this study, A three-layer hierarchical predictive detection system has been developed in order to utilize large-scale, multivariate operational data. This system is designed to identify overall trends by leveraging big data, detect correlation breakdowns through domain knowledge, and detect shifts in single-signal levels. The effectiveness of the proposed system has been confirmed through its application to actual operational data from the steel manufacturing process. In addition, general anomaly detection models, including our system, rely on quantifying deviations from a normal state as an anomaly score. In manufacturing settings, data drift often occurs due to factors such as equipment part replacements or changes in operational conditions. When data drift occurs, it becomes necessary to redefine the normal state. However, in manufacturing environments, temporary runs or experimental operations mean that the data following a drift is not necessarily guaranteed to normal data. Therefore, it is necessary to evaluate whether the data distribution is normal before and after the drift on a case-by-case basis. Current approaches do not provide a quantitative means to make this decision, leading to the issue that model updates depend on the judgment of experts. To address this, we propose a method that utilizes similar equipment conditions to guide the timing and procedure for model updates. By applying Jensen–Shannon divergence to measure differences among four data distributions—derived from two machines and two distinct periods— we provide appropriate guidance for model construction based on a table of potential anomalies. Through validation using real data from two adjacent continuous casters, we confirmed that identifying abnormal equipment and time periods enables us to propose appropriate normal operating windows. These validation results indicate that the proposed system allows for comprehensive predictive maintenance, integrating domain knowledge and thereby contributing to stable operations in steelmaking facilities.

Improved LSTM-Based Battery SOH Estimation with Differential Evolution Hyperparameter Optimization

Karthickumar Ponnambalam, Sivaneasan Bala Krishnan, Anurag Sharma, Sze Sing Lee — Tue, 13 Jan 2026 00:00:00 +0000

Reliable estimation of battery State of Health (SoH) is essential for the safety and longevity of lithium-ion battery systems. Previous work introduced a PSO-aided LSTM (PA-LSTM) model for SoH prediction using NASA’s battery degradation dataset. Building on this, our earlier GA-LSTM model utilized a Genetic Algorithm (GA) for LSTM hyperparameter tuning, achieving RMSE reductions of 12.4% to 76.79% using 70% of the discharge cycle data.
In this study, we propose a novel approach for SoH prediction by integrating Long Short-Term Memory (LSTM) networks with Differential Evolution (DE), a more efficient and scalable metaheuristic optimizer. Leveraging DE’s robust convergence behavior and global search capability, the proposed DE-LSTM model is evaluated on the same NASA dataset. Experimental results demonstrate that DE-LSTM outperforms our previous GA-LSTM model, achieving further RMSE reductions and highlighting DE’s effectiveness for hyperparameter optimization in data-driven battery health prognostics.

Improving Virtual Metrology Predictions via Transfer Learning and Active Learning

Swee Kuan Loh — Tue, 13 Jan 2026 00:00:00 +0000

Unlike traditional metrology in semiconductor manufacturing, it uses physical methods to measure wafers that are both resource-intensive and time-consuming, increasing possibilities of causing defects to production of wafers. Virtual metrology (WM) predicts wafer measurement using sensor data, enabling real-time, non-intrusive monitoring of process performance. Our study introduces a smarter approach to virtual metrology by combining regression modeling with transfer learning to enhance model generalization under varying manufacturing conditions. The framework consists of two stages: first we fine-tuning a base model using a limited amount of labeled data from the target domain, then followed with iterative refinement via active learning in which the most uncertain predictions are identified and incorporated into the training set. This method improves prediction in the target domain, especially in cases where standalone models do not perform well. Experimental results demonstrate proposed framework significantly outperforms models trained solely on target domain data. There is significant improvement in the refined model, achieving 77.80% in Mean Absolute Error (MAE) and 58.095% in Root Mean Squared Error (RMSE) compared to the original model. In addition, improvements in recall and reductions in false positive rates were observed, indicating the method is more effective at identifying abnormal wafers. Active learning helps to select the most appropriate sample for labelling to reduce the need for extensive datasets. The purposed method is advantageous in high-mix, low-volume (HMLV) manufacturing industry settings, where some stage or products are produced in small amounts. This innovative approach to virtual metrology aims to streamline semiconductor manufacturing, minimize defects, and optimize resource utilization by delivering strong, adaptable predictive capabilities.

Intelligent Bearing Fault Diagnosis Under Various Load Conditions Using Bias Mitigation

Seungyun Lee, Sungjong Kim, Minjae Kim, Heonjun Yoon, Byeng D. Youn — Tue, 13 Jan 2026 00:00:00 +0000

Intelligent bearing fault diagnosis with domain adaptations has accomplished remarkable performances under various operating conditions. However, especially for different load conditions, the model bias due to the physical characteristics of bearing signals has not been considered. In the absence of handling bias, the root cause for generalization errors cannot be clarified under various load conditions. This paper thus demonstrates that certain bias exists in diagnostic models for different loads of bearings, and the main factor of bias is impulsiveness. The existence of bias is shown with quantitative analysis by applying fairness criteria to diagnostic models. Also, qualitative analysis is conducted with gradient-weighted class activation mapping (Grad-CAM) for vibration signals of bearings, which proves that the large amplitude of impulse can be the source of bias. To correct this impulsiveness bias, a framework of fairness approach is newly proposed for bearing fault diagnosis under various loads. The process of correcting bias contains two steps: categorizing samples based on impulsiveness and training models with fairness criteria. Different from the previous domain adaptation-based approaches, the proposed method can achieve superior diagnostic performances by correcting bias that causes generalization errors. The effectiveness of the proposed method is validated with public-bearing datasets with various loads. The results show that the fairness approach can be the mainstream solution for fault diagnosis of rotary machines under different load conditions.

Interpretable Sensor Importance-Based Multi-Sensor Integration for Condition Monitoring of Rotating Machinery

Sungjong Kim, Seungyun Lee, Minjae Kim, Heonjun Yoon, Byeng D. Youn — Tue, 13 Jan 2026 00:00:00 +0000

Accurate condition monitoring of rotating machinery requires integrating multi-sensor data to capture fault-related information distributed across sensing locations. While attention-based deep learning models can assess sensor importance, their lack of transparency limits industrial adoption. This study proposes an interpretable sensor importance-based multi-sensor integration framework combining a CNN-inspired kernel sharing strategy, a Transformer encoder for local and global feature extraction, and a channel attention mechanism for dynamic sensor weighting. Attention weight in Transformer encoder was analyzed in frequency domain to reveals spectral components influencing sensor importance evaluation. Validation on a pump testbed with various speeds conditions shows superior fault diagnosis accuracy, robustness to unseen conditions, and clear alignment between high-weight sensors and known fault frequencies, supporting trustworthy AI-driven condition monitoring in practice.

Irregular Time-Series Hybrid Model for Enhanced Prognostics of Engine Degradation and Failures

Rohit Deo, Aman Yadav, Shruti Bharti, Dr. Nilesh Powar — Tue, 13 Jan 2026 00:00:00 +0000

Classification-based prognostics aims to predict the Remaining Useful Life (RUL) of components in diesel engines by identifying failure and degradation stages. This is critical for industries such as automotive, aviation, and manufacturing. Traditional methods rely on classification models trained on historical data from multiple engines to forecast failures based on current engine parameters. However, these global classifiers often struggle with generalization when applied to unseen engines, resulting in poor precision and recall. Moreover, they fail to capture the temporal dependencies inherent in engine degradation, which are crucial for accurate failure prediction. We propose a hybrid model that integrates predictions from global classifiers with time-based memory units to address these limitations, effectively building irregular time-series models. Our approach demonstrates a significant performance improvement, with precision and recall metrics doubling compared to traditional global classifiers.

Leveraging Few-Shot In-Context Learning for Scaling Railway Log Anomaly Detection

Tue, 13 Jan 2026 00:00:00 +0000

This paper presents a scalable, data-driven approach for anomaly detection in railway signaling logs using Large Language Models (LLMs) and in-context learning. By classifying log keys — the structural templates of log messages — instead of individual messages, the method dramatically reduces the number of required model calls, thereby lowering computational costs (in terms of energy or monetary resources). Expert-labeled log keys are incorporated into LLM prompts to help the models differentiate between normal and abnormal log messages. Multiple state-of-the-art LLMs are evaluated on this task, revealing that performance increases as more labeled examples are added to the prompt, although the improvement gain diminishes with each additional label. Further analysis indicates that GPT-4.1 offers the best balance of monetary cost, response time, and F1 score for this application. The study highlights both the advantages and limitations of in-context learning for railway log anomaly detection, notably its ability to leverage expert-labeled examples without additional model training, but also its sensitivity to data imbalance and exclusion of parameter values. It further discusses avenues for future improvement, such as model fine-tuning, prompt enrichment with additional contextual information, and the potential use of Retrieval-Augmented Generation (RAG) or self-feedback strategies to enhance classification performance.

LLM-based multi-agent system for autonomous maintenance process of machine tools

Jongsu Park, Seongwoo Cho, Yoonji Chae, Sena Nur Durgunlu, Jumyung Um — Tue, 13 Jan 2026 00:00:00 +0000

This paper presents a large language model (LLM)-based system for autonomous maintenance in manufacturing facilities. While many machine alarms are interpreted with existing manuals, understqanding and acting on these instructions of all facilities remains a challenge for operators. The proposed system processes user inputs including error codes, identifies corresponding procedures from manuals, and decomposes them into structured action sequences. These sequences include action, user interface target, preconditions, and expected outcomes, and are executed by agent capable of interacting with Human–Machine Interfaces (HMIs). The proposed system is built on an LLM-powered multi-agent framework comprising four agents: a chatbot, solution_finder, actor, and supervisor. Each agent operates based on role-specific prompts that define their responsibilities and decision rules. Instead of relying on predefined rule sets, the system interprets unfamiliar or previously unseen alarms by reasoning over machine manuals and context, enabling flexible and scalable maintenance. The system was implemented on a HMI system of CNC machine tools and successfully performed automatic responses to selected alarms. Prompt-based control ensure adaptability to other machines, and the use of a local LLM maintains data security. This approach enables general-purpose, self-directed maintenance with minimal operator intervention.

Methods and Systems for Hybrid Digital Twin Driven Health Predictions for Aircraft Sub-systems

Partha Pratim Adhikari, Deepu Vettimittathu Mathai, Avik Sadhu, Darren Macer — Tue, 13 Jan 2026 00:00:00 +0000

In the aerospace industry, modern aircraft are increasingly equipped with a growing number of sensors, which enable the development of predictive maintenance solutions utilizing data-driven diagnostic and prognostic (D&P) techniques to enhance operational availability and reduce maintenance costs. However, constructing a purely data-driven D&P solution requires a substantial amount of run-to-fail sensor data, which is often unavailable for highly reliable and safety-critical aircraft systems. This limitation restricts the applicability of purely data-driven D&P solutions for aircraft subsystems. To address this limitation, we developed a novel Hybrid Digital Twin framework that integrates physics-based subsystem models with sensor data, enabling enhanced feature generation for improved fault diagnostics and prognostics. Our approach simultaneously estimates both design and health-related parameters, facilitating accurate model calibration even when some of design data is not available. Sensor features enhanced with estimated health-related parameters enable more accurate data-driven diagnostics and prognostics solutions of a sub-system or a component. The framework is demonstrated on key subsystems of the aircraft Environment Control System (ECS), including the Heat Exchanger and Centrifugal Compressor. Various parameter estimation techniques including nonlinear least squares, particle swarm optimization, and extended Kalman filter, Unscented Kalman filter, Physics-Informed Neural Networks, etc., are evaluated. This Hybrid Digital Twin approach offers a promising pathway for more accurate, robust and scalable health management of aircraft subsystems having limited operational data.

MLOps Framework for Fault Diagnosis in Air Conditioners Using Field Noise

SangUk Son, Yoojeong Noh, Sunhwa Park, Jangwoo Lee — Tue, 13 Jan 2026 00:00:00 +0000

Fault diagnosis of heating, ventilation and air‑conditioning (HVAC) equipment relies increasingly on data‑driven models. However, real‑world after‑service recordings captured by technicians are noisy, imbalanced and often contain meaningless segments. These are labeled by domain experts but sometimes mislabeled. This paper proposes an initial noise‑aware machine learning operations (MLOps) framework that enables robust classification, calibration as a prerequisite to uncertainty estimation and continuous improvement of air‑conditioner sound diagnostics. The framework performs data preprocessing, uncertainty-based identification of label noise, systematic relabeling through gradient-based class activation maps (Grad-CAM++, hereafter referred to as CAMs), and clustering. A comprehensive metrics tracking facilitates reproducible experiments. Experiments on field recordings demonstrate that removing label noise leads to better generalization, as the learned representation forms more distinct clusters in the logits space, reducing the presence of mislabeled samples within each cluster. The proposed approach yields better generalization and provides a scalable pathway toward automated labeling and open‑set recognition.

Multi-Branch Joint Time-Frequency Transformer for Domain Generalization Fault Diagnosis of Rotating Machinery

Qitong Chen, Liang Chen, Hong Zhuang, Qi Li, Wenjing Zhou — Tue, 13 Jan 2026 00:00:00 +0000

Conventional time-frequency Transformers primarily focus on the global features of signals in the time-frequency domain while neglecting the local features in both the time domain and frequency domain. This limitation hinders the ability of the model to effectively capture the shared features among time, frequency, and time-frequency representations. To address this issue, a Multi-Branch Joint Time-Frequency Transformer (MBJTF-Transformer) is proposed for domain generalization (DG) fault diagnosis of rotating machinery. Specifically, a time-branch Transformer is designed to extract temporal features, while a frequency-branch Transformer captures frequency-domain information. In addition, a time-frequency Transformer is employed to learn the shared representations across time, frequency, and time-frequency domains. Finally, a multi-decision fusion strategy of MBJTF-Transformer is adopted to enhance the generalization capability of the model. Experimental results on both the SCARA (Selective Compliance Assembly Robot Arm, SCARA) dataset and the PU (Paderborn University) bearing dataset demonstrate that the proposed MBJTF-Transformer achieves superior DG performance compared to multiple state-of-the-art sequential models.

Multi-Class Gearbox Fault Diagnosis via Pre-Trained Model-based Domain Adaptation with Healthy-Only Target Data

Tue, 13 Jan 2026 00:00:00 +0000

Accurate gearbox fault diagnosis under varying operational speeds is critical for industrial predictive maintenance. A significant challenge is domain shift, where models trained under one condition fail to generalize to another, especially when only healthy data from the target domain is available for training. This study proposes a novel domain adaptation framework, CDANet, that directly leverages raw sensor data to perform multi-class fault classification without manual feature engineering. The model combines a lightweight CNNbased temporal feature extractor with a frozen DistilBERT encoder to capture transferable, domain-invariant representations, combined with a maximum mean discrepancy loss to align the feature distributions between the source and target domains using only healthy samples. Experimental results demonstrate that our proposed model significantly outperforms conventional deep learning approaches, achieving high classification accuracy across six domain adaptation tasks. This work validates the effectiveness of applying pre-trained models in domain adaptation for gearbox fault diagnosis under real-world domain shift constraints.

Multi-source Variable Decoupling Network for Compound Fault Diagnosis of Train Bogie

Qitao Yin, Zhibin Guo, Tiantian Wang, Jingsong Xie, Jinsong Yang — Tue, 13 Jan 2026 00:00:00 +0000

When compound faults occur in rotating machinery, the mutual coupling and interference among different fault sources make it extremely difficult to directly isolate individual faults from the observed signals. Therefore, fault decoupling is essential prior to diagnosis. In this study, we propose a semi-supervised multi-source variable decoupling network (MVD-Net) that enables blind separation of unknown compound fault signals using only single-fault samples for training. First, low-dimensional features are extracted from the mixed signal through an encoder. These features are then mapped to multiple independent latent spaces corresponding to different fault sources via variational inference, while the number of sources is adaptively estimated using the evidence lower bound (ELBO). Subsequently, each source-specific decoder generates an estimated source signal from its corresponding latent representation. To ensure that each decoder focuses on a distinct fault component, a source-selective activation mechanism is incorporated into the decoding process, effectively mitigating the random assignment issue commonly encountered in traditional blind source separation methods. Finally, based on the estimated source signals, a separation mask is derived to extract individual sources from the original mixed signal. Two compound fault decoupling and diagnosis experiments were conducted on the BJTU-RAO dataset. The results demonstrate that compared with other methods, the proposed approach yields cleaner separated signals with more distinct time-frequency fault features and achieves higher diagnostic accuracy.

Multilevel fault diagnostics for railway applications using limited historical data

Tue, 13 Jan 2026 00:00:00 +0000

This study proposes a fault diagnostics methodology that addresses the challenges posed by highly imbalanced datasets typical of railway applications, where faulty conditions constitute the minority class. Fault diagnostics is performed from the component level upward, considering each sensor’s proximity to its respective critical component. Advanced signal analysis, feature engineering, and automated data-driven model generation techniques were explored to achieve comprehensive diagnostics, such that the model development process accounts for variations in the operating conditions and differing levels of information availability. The proposed methodology is evaluated on datasets from the MONOCAB, for scenarios with limited faulty instances and on the Beijing 2024 IEEE PHM Conference data challenge, which focused on fault diagnostics of railway systems under various fault modes and operating conditions.

Nonlinear and Trend-Aware Industrial Time Series Anomaly Detection with Federated Learning

Zhiqing Luo, Yan Qin — Tue, 13 Jan 2026 00:00:00 +0000

Industrial anomaly detection aims to identify significant data deviations. However, it is hampered by the complex dynamics of time series, distributed data silos, and data heterogeneity. To overcome these challenges, we introduce a novel federated learning framework (FL) with two core modules: Multiple Definition Operators (MDO) to capture intricate temporal dynamics, and Temporal Trend Convolution (TTC) to extract interpretable trend patterns. FL enables multiple clients to collaboratively train a robust global model without centralizing raw data, thereby boosting generalization and preserving privacy. Critically, a tailored data-sharing strategy is implemented within the framework to mitigate the challenge of non-independent and identically distributed data. Experiments conducted on the Skoltech Anomaly Benchmark and other real-world datasets validate the efficacy of the MDO and TTC modules as well as confirm that the proposed framework significantly improves anomaly detection performance, demonstrating its practical potential for industrial applications.

ONGOING: A Human-readable, Model-enriching, Continuous Technician Knowledge Modeling Framework

Adrien Bolling, Sylvain Kubler — Tue, 13 Jan 2026 00:00:00 +0000

Industry 5.0 reframes manufacturing around human-centric concerns: resilient operations, safe work, and decisions people can understand and contest. For PHM, that means elevating human-based features: competency, recency of practice, mentoring links, explainability, and fair exposure, rather than relying only on sensors or opaque models. Today those signals sit in ticket logs and massive databases, making them hard to audit, transfer, or reuse at scale.
We suggest ONGOING, a representation layer framework that turns unstructured maintenance text into a human-auditable Knowledge Grid and a complex but modular feature vector, independent of any particular embedding model or projector. At its core, the grid tracks technician experiences by incrementing a part of the Knowledge Grid whenever tickets are resolved. Two mechanisms capture more advanced dynamics: knowledge transfer between people (e.g., mentorship) via a convex blend of Knowledge Grids, and neighborhood propagation that diffuses experience increases to semantically adjacent tasks through a Gaussian kernel. From each grid we derive interpretable features, such as hypervolume, sparsity, or maximum knowledge, that summarize knowledge distribution more accurately for better downstream use (e.g., dispatching optimizer models, LLMs, production forecast models).
We implement the framework on a partner company's data, and deploy an instance at-scale (50000 tickets, 100 technicians) in real-time, using a multilingual sentence encoder and a toroidal SOM for ticket embedding.
On our deployed instance, we designed a technician recommendation use-case. A maintenance expert study with human feedback over 55 real tickets found that grid-based recommendation were judged more pertinent than a scalar-based and a vector-based knowledge modeling approaches. Crucially, dispatchers could articulate rationales from visible grid neighborhoods and feature attributions, preserving interpretability.
Beyond dispatch support, the Knowledge Grid enables training planning (identify coverage gaps), fairness monitoring (avoid single-point failure through over-reliance on “heroes”), and promotes workload balancing.

Physics-Informed Transformer with ODE-Guided Joint Modeling for Fault Classification and RUL Prediction in Collaborative Robots

Yingjun Shen, Kang Wang, Zhuoxin Chen, Junkai Huang, Yifan Zhu, Zhe Song — Tue, 13 Jan 2026 00:00:00 +0000

Accurate early diagnosis of fault types and Remaining Useful Life (RUL) prediction are critical for predictive maintenance in collaborative robotic systems, especially under limited labeled data conditions. This paper proposes PhysODE-Joint, a physics-informed deep learning framework that unifies Transformer-based temporal modeling with fault-specific ODE-guided degradation dynamics for joint fault classification and RUL estimation. The method incorporates domain knowledge of mechanical wear and thermal degradation into the feature learning process and employs a cascade architecture to ensure physical plausibility and class-aware prediction. A hybrid training strategy is introduced, integrating limited real-world sensor data with synthetic degradation sequences generated from physics-based models. Experimental results on real-world robotic datasets demonstrate that PhysODE-Joint significantly outperforms conventional data-driven models, particularly in small-sample scenarios, offering a robust solution for health monitoring and maintenance scheduling.

Predicting Maintenance Actions from Historical Logs using Domain-Specific LLMs

Aman Kumar, Ahmed Farahat, Chetan Gupta — Tue, 13 Jan 2026 00:00:00 +0000

Maintenance logs of complex specialized equipment capture problem–action records that are essential for building predictive maintenance solutions but remain difficult to utilize due to their terse, abbreviation-heavy style. This work provides the first systematic benchmark and domain-adaptation study of large language models (LLMs) for predicting maintenance actions from free-text problem descriptions in the MaintNet aviation dataset. We evaluate a range of proprietary and open-source LLMs under zero-shot and few-shot prompting and additionally fine-tune selected open models for supervised evaluation. Experiments are conducted on both raw-abbreviation and expanded datasets, using both lexical (ROUGE, BLEU) and semantic (cosine similarity, BERTScore) metrics. Results show that GPT-4o achieves the strongest semantic alignment, while the instruct version of Gemma-3-4B leads in lexical overlap. Few-shot prompting boosts weaker models disproportionately, narrowing the gap with stronger baselines. Fine-tuning delivers the most significant gains, with instruct versions of Gemma-3-4B, LLaMA-3.2-3B, and Phi-4-mini, improving BLEU by up to 90% and ROUGE-2 by 30%. Notably, the fine-tuned Gemma-3-4B surpasses GPT-4o across multiple metrics, demonstrating the effectiveness of domain-specific adaptation. These findings highlight the potential of fine-tuned LLMs to utilize unstructured aviation logs for building reliable maintenance systems.

Predictive Prioritization of Railway Bearings Using Acoustic Similarity of NOISY(RS1) Alarms from Wayside Monitoring Systems

leandro_rocha — Tue, 13 Jan 2026 00:00:00 +0000

The operational reliability of heavy haul railways, such as the Carajás Railway (EFC), depends on early detection of failures in critical components like bearings. This study proposes a predictive prioritization approach based on acoustic similarity analysis of NOISY(RS1) alarms from the RailBAM® wayside monitoring system. Traditionally discarded due to suspected interference, these alarms have shown statistical overlap with confirmed failures. By applying multivariate similarity analysis using Mahalanobis distance and acoustic parameters—ERS DB, ERS Neighbors DB, and ΔERS DB—the methodology identifies patterns indicative of real defects. A new rule was developed to reclassify NOISY(RS1) alarms based on statistical thresholds and repetition criteria, enhancing failure detection accuracy. Experimental validation revealed previously unprioritized bearings with physical damage, demonstrating the rule’s potential to complement existing predictive matrices. The approach improves maintenance planning, reduces undetected failures, and supports the integration of data-driven strategies in Prognostics and Health Management (PHM) for railway assets.

Probabilistic graphical models for diagnosing defectivity patterns

Leonardo Barbini, Peter Kruizinga, Micha Lipplaa, Alvaro Piedrafita — Tue, 13 Jan 2026 00:00:00 +0000

In the high-tech sector, diagnosing performance issues often involves analysing a variety of defectivity patterns on products. High-tech systems perform numerous processes – e.g. product handling, light projection, jetted ink application, and thermal treatments – all of which affect the quality of the product itself. Their potential malfunctioning can contribute to defects with characteristic patterns. Often there is not a one-to-one mapping between root causes of these malfunctions and the resulting observable defectivity patterns. Consequently, identifying the root cause of these patterns is a challenging and an intrinsically probabilistic task. This paper proposes a framework to relate these patterns to the underlying root causes and employs Probabilistic Graphical Models (PGM) to reason about these relations. We find that PGMs ability to contain arbitrary graph topologies and jointly reason across all root causes empowers the modeller to adapt the models to the system at hand and include domain specific knowledge that would be hard to account for using more data-driven approaches. When provided with data from an operational system in the field, the PGM identifies the underlying root causes of product quality issues. We demonstrate the methodology with a real use case from the production printing industry.

Radar Health Monitoring Using Anomaly Detection

Jean-Marc Divanon, Thomas Lavigne, Theo Cornu, Teck Yoong Chai — Tue, 13 Jan 2026 00:00:00 +0000

Predictive maintenance has emerged as a crucial strategy in complex systems management, leveraging machine learning and data-driven health monitoring to anticipate failures and optimize operational uptime. While significant progress has been made in developing general-purpose models for anomaly detection and condition-based maintenance, their effectiveness often diminishes when applied to highly specialized systems such as radar platforms. These systems exhibit unique operational behaviours and failure modes, necessitating tailored monitoring solutions. This paper presents a methodology for anomaly detection tailored to radar systems, addressing the inherent challenge of limited labeled data and the ambiguity surrounding the definition of anomalies. We employ a reconstruction-based approach using autoencoders in conjunction with Mahalanobis distance to generate anomaly scores, enabling the detection of subtle deviations from normal system behavior without requiring explicit failure labels. The proposed approach has been applied to real sensor data collected from multiple radar units, specifically from sensors located on the antenna mast. For confidentiality, the data has been anonymized. Experimental results demonstrate that the method effectively highlights outliers and identifies the contributing features responsible for anomalies. Furthermore, the model reveals interpretable abnormal patterns and provides early indications that condition-based monitoring can be a viable strategy for identifying potential issues in radar operations.

RAVEN: Unsupervised Anomaly Detection in Multivariate Jet Engine Time Series using Residual Learning on Real Test Data

Nouf Almesafri, Mohamed Ragab, Zahi Mohamed, Abdulla Alseiari, Salama Almheiri — Tue, 13 Jan 2026 00:00:00 +0000

Jet engines operate under demanding conditions, subjecting critical components to gradual wear and degradation over time. Early identification of incipient faults is essential for maintaining performance, safety, and reliability. Detecting incipient faults early is essential but remains difficult due to two major challenges: the scarcity of faulty data and the strong variability in operating conditions that obscure fault signatures. Most existing anomaly detection approaches rely on simulated datasets or assume the availability of labeled faults, limiting their applicability to real-world engine monitoring. In this work, we introduce RAVEN, a fully unsupervised anomaly detection framework designed for jet engine monitoring under real test conditions. RAVEN integrates (i) a regression-based residual model to normalize sensor responses against varying operating regimes, with (ii) a deep LSTM autoencoder that captures subtle deviations in time-series behavior without requiring fault labels. By explicitly addressing operational variability, sensor noise, and label scarcity, RAVEN provides a robust pathway for early fault detection. We validate RAVEN on real jet engine test data, demonstrating its ability to detect anomalies under diverse operating conditions. Results show that our approach delivers reliable detection performance in scenarios where conventional approaches struggle, offering a practical and scalable solution for propulsion system health monitoring.

Real-time Sensor Data Streaming for deployment in Edge AI for Health Index Construction and Remaining Useful Life Prediction

Salama Almheiri, Zahi Mohamed, Mohamed Ragab, Abdulla Alseiari, Nouf Almesafri — Tue, 13 Jan 2026 00:00:00 +0000

This paper introduces a real-time predictive analytics framework that integrates edge artificial intelligence for remaining useful life estimation and health index construction using turbofan engine sensor data. A MATLAB/Simulink model was designed to stream 14 critical sensor signals, derived from the NASA C-MAPSS dataset, into an Opal-RT OP5707XG simulator for real-time emulation. These signals were output as analog voltages, converted into digital values using ADS1115 converters, and processed on an Nvidia Jetson AGX Orin edge-computing platform. A CatBoost regressor, trained on a feature-rich time-series dataset and refined through SHapley Additive Explanations-based feature selection was employed as the predictive model. System performance was benchmarked on two hardware platforms: a mid-tier desktop computer and the Jetson AGX Orin. The mid-tier desktop computer completed training in 18 minutes, while the Jetson required around 3 hours. Inference speed was also faster on the computer at 2.8 ms versus 7.5 ms, though both satisfied the 33 ms requirement for real-time processing of a 30 Hz data stream. The Jetson demonstrated a significant efficiency advantage, consuming 20—40 W compared to 250-350 W for the computer. The framework achieved high accuracy with strong generalization and transparent explainability through SHapley Additive Explanations-based feature selection confirming the feasibility of deploying advanced prognostics on edge AI hardware for real-time health monitoring.

Remote monitoring system for detection of faults in drive motors of electric vehicles

Amiya Mohanty, Nagesh — Tue, 13 Jan 2026 00:00:00 +0000

Electric vehicles (EVs) rely on electric motors (EMs) for drive, offering an eco-friendly alternative to conventional internal combustion engines. However, EMs in EVs are prone to multiple defects, such as bearing faults and load torque fluctuations, induced by electromagnetic interference (EMI), mechanical misalignments, and variable loading conditions arising from dynamic driving environments and controller-induced torque ripple. The resulting external mechanical load on the electric motor, which in turn modulates the stator current, produces distinct fault-related frequency components in the motor stator current spectrum. This study presents a system for remotely monitoring the health of such EMs which are used to drive EVs. A non-invasive fault detection methodology using Motor Current Signature Analysis (MCSA) which has come of age in present day to detect and characterize bearing-related faults and load torque fluctuations is used. The proposed approach is examined and validated on permanent magnet synchronous motors (PMSM), which are predominantly used as drive motors in EVs. A hall effect current sensor in one situation and a current transformer (CT) in another have been used to measure the current waveform of the stator current in the PMSM motors, which is then analyzed using the principles of MCSA. MCSA identifies the fault frequencies associated with bearing defects and torque fluctuations without requiring motor disassembly or additional vibration sensors. By implementing MCSA into a standalone monitoring system, this study demonstrates a reliable means of detecting bearing and load torque-related faults, ultimately improving the durability, efficiency, and operational safety of electric vehicle drivetrains. Future work can explore scaling this approach with cyber-physical system (CPS)-based architectures for real-time monitoring of EVs, enabling centralized analytics and smart decision-making as has been showcased in the present work.

Robust Fault Diagnosis of Electric Vehicle Drivetrain Using Amplitude Adjustment Techniques

Tue, 13 Jan 2026 00:00:00 +0000

Most electric vehicle drivetrain fault diagnosis methods have been validated only under constant load and rotational speed conditions, showing limited performance in real driving environments where load and speed continuously vary. This study proposes a novel vibration signal normalization method that combines order tracking with physics-based amplitude adjustment techniques to improve diagnostic accuracy under variable operating conditions. Order tracking addresses the problem of frequency variation of vibration signals that vary with speed over time. The proposed method converts vibration signals under variable speed conditions into pseudo-stationary signals of equivalent levels by adjusting amplitudes through factors that consider both centrifugal and tangential forces acting on rotating components in the drivetrain. To validate the effectiveness of the proposed technique, experiments were conducted using actual electric vehicles equipped with drivetrains at various degradation levels. Drivetrain vibration data were collected and evaluated across multiple operating scenarios. Experimental results demonstrate that the proposed method reduces variability across different speed conditions compared to raw signals. The proposed method shows promise for robust drivetrain diagnosis applications even under variable speed conditions, addressing a significant limitation of existing diagnostic approaches.

Rotary Encoder-Based Health Indicators for Early Detection of Gear Pitting in Commercial Gearboxes

Toby Verwimp, Rui Zhu, Hao Wen, Achilleas Achilleos, Konstantinos Gryllias — Tue, 13 Jan 2026 00:00:00 +0000

This paper investigates the diagnostic capabilities of rotary encoders for condition monitoring of commercial gearboxes, focusing on early detection of gear pitting. Within a multi-sensor framework, high-resolution rotary encoder data were collected alongside accelerometer signals, both mounted on a multistage commercial gearbox subjected to gear pitting. Encoder-based health indicators were developed from the square envelope (SE) of the transmission error (TE), synchronously averaged (SA) to isolate the targeted gear component within the gearbox. By automatically capturing images of the targeted gear during an accelerated life test, these indicators were evaluated for their ability to detect the onset of gear pitting and compared against conventional vibration-based metrics. Results show that encoder-based indicators can provide earlier and more consistent detection of gear pitting. These findings highlight the potential of rotary encoders as a complementary or standalone sensing solution in advanced diagnostic frameworks for commercial gearboxes.

SAE-FSC

Karkulali Pugalenthi, Van Tung Tran, Ang Shiming, Doan Ngoc Chi Nam — Tue, 13 Jan 2026 00:00:00 +0000

Rolling element bearings are critical components in rotating machinery, where failures can cause severe downtime and safety risks. Existing fault diagnosis methods are predominantly supervised, requiring large amounts of labeled data across multiple operating conditions. However, in realistic industrial scenarios, such labeled datasets are scarce, and models trained on one regime often fail to generalize to others. To overcome this cross-domain generalization challenge, we propose a Siamese Attention Encoder–based few-shot cross-domain fault diagnosis (SAE-FSC) framework. The key novelty of this work lies in an attentionaugmented Siamese encoder that extracts highly discriminative and transferable time-series features, coupled with a composite objective function that jointly optimizes supervised cross-entropy, pairwise binary cross-entropy, and domain adversarial loss. This combination enforces intra-class domain invariant feature learning across multiple operating conditions. Extensive experiments on the Case Western Reserve University (CWRU) dataset under leave-one-fault-out (LOFO) and leave-two-fault-out (LTFO) protocols demonstrate robust generalization across unseen fault types, load conditions, and fault severities, achieving a prediction accuracy of 87% for 5 shot learning.

Self-Adaptive RUL Prediction of Power Electronic Devices with Package Failure

Chao Guo, Zhonghai Lu — Tue, 13 Jan 2026 00:00:00 +0000

Power electronic devices vary in lifetime due to intrinsic device characteristics and extrinsic operational environments, which pose significant challenges in lifetime prediction. Traditional Deep Learning methods often directly map precursor signals to the Remaining Useful Lifetime (RUL), lacking the health state information needed to adapt dynamically to device characteristics. To address this limitation, we propose a stateful, self-adaptive RUL prediction method for package failure of power diodes. It utilizes junction temperature signals as inputs, representing thermal-mechanical fatigue influenced by external operational environments, to adjust the algorithm states, which contain the device characteristics and health state information. The proposed method combines two models, a stateful-LESIT (SLESIT) model and a Kalman Filter (KF). The SLESIT model dynamically adjusts its state using current junction temperature signals to estimate the RUL. The produced estimation is then used to rectify the predictions from an intuitive RUL propagation model in KF, providing a statistically optimal RUL estimation at each cycle. Validated through online simulation with accelerated aging data from power diodes that exhibit significant lifetime variability (68.1%), our approach reduces Mean Absolute Error (MAE) from 44.17% to 84.52% compared to popular Deep Learning methods.

Shared Representation Learning for Generalizable SOH Estimation Across Multiple Battery Configurations

Tue, 13 Jan 2026 00:00:00 +0000

Battery health monitoring is essential in applications such as electric vehicles and energy storage systems, where the lifespan and health state of batteries directly impact the safety and operational costs. However, existing works have demonstrated promising performance in predicting the state of health (SOH) of batteries within the same type under certain working conditions. However, batteries are produced with different types and work under different conditions in real applications. Existing methods fail to leverage the inherent correlations between related battery types and overlook the various working conditions, resulting in suboptimal robustness and prediction accuracy. To address this limitation, we propose SRSE: a novel Shared Representation learning framework that jointly learns shared representation (hidden knowledge) across multiple battery configurations for robust and generalized SOH Estimation. In particular, an adversarial training scheme is utilized to eliminate task-specific contamination in the shared feature space. SRSE captures both feature-level and logit-level shared knowledge and subsequently transfers it from the shared layer to task-specific layers, enhancing the adaptability and efficiency of each task. Extensive experiments on three large-scale battery health datasets demonstrate that our proposed method significantly improves SOH estimation performance across diverse battery types and operating conditions.

Study of the degradation of rolling element bearings with artificial dents

Theo Tselonis, Valentius Wirjana, Thomas Hughes, Zhongxiao Peng — Tue, 13 Jan 2026 00:00:00 +0000

Contact fatigue in rolling element bearings (REBs) is a common surface degradation and failure mechanism. A better understanding of this process, especially the evolution of tribological features and resultant dynamic responses, is critical for developing effective tools to monitor degradation and predict remaining useful life. This paper investigates the relationship between the geometry of seeded faults – specifically their shape and size – and bearing performance under grease-lubricated conditions. Two bearings with different fault shapes and sizes were tested. To study the degradation process, moulds and images of the fault area were collected at different intervals to capture the gradual progression of wear. The acceleration root mean square (RMS) value of vibrations was used as a real-time indicator to detect significant changes in bearing dynamics throughout the degradation process. This study provides valuable experimental data and insights into the degradation processes of rolling element bearings when they are subjected to initial defects related to manufacturing or contamination, under grease-lubricated conditions. This knowledge is crucial for developing effective real-time techniques for monitoring the degradation process.

Towards Autonomous PHM: An Application to Turboshaft Engine Torque Prediction

David He, Eric Bechhoefer, Miao He — Tue, 13 Jan 2026 00:00:00 +0000

Autonomous Prognostics and Health Management (Autonomous PHM) refers to the capability of a system to independently monitor, diagnose, predict, and manage its own health status without human intervention. It combines traditional PHM functions with autonomy and intelligent decision-making to enable self-sustaining operation, especially in complex or remote environments. The key characteristics of an autonomous PHM system include: (1) self-monitoring: continuous collection and analysis of sensor data to assess system health in real time; (2) self-diagnosis: identification of faults, anomalies, or degradations using AI, machine learning, or model-based reasoning; (3) self-prognosis: prediction of remaining useful life (RUL) or time to failure based on current and historical data; (4) autonomous decision-making: autonomous selection and execution of maintenance or mitigation actions (e.g., reconfiguration, load reduction); (5) adaptability: adapt pre-trained models (e.g., for fault detection or RUL estimation) from one system or component to another with limited new data; (6) minimal human oversight: designed to function reliably with little to no manual input, particularly useful in inaccessible or high-risk settings (e.g., space missions, underwater robotics, military systems). A few challenges remain for developing an effective autonomous PHM system: (1) learning with limited labeled data: limited availability of failure data for training ML models; (2) cross-platform autonomy: autonomous PHM systems often operate in varied conditions or on different equipment types. PHM functions should be adapted from one system or component to another to reduce the need to retrain models from scratch in every new setting. (3) scalability: autonomous PHM systems should scale to large, complex systems (e.g., fleets of aircraft or satellites). A model trained on one unit can be transferred to other units in the fleet to scale autonomous PHM capabilities efficiently. In this paper, the development of an autonomous PHM system by integrating self-supervised learning and large language models (LLMs) is presented. The effectiveness of the autonomous PHM system is demonstrated with an application to turboshaft engine torque prediction.

Towards Adaptive and Robust Unsupervised Anomaly Detection in Satellite Telemetry

Lorenzo Brancato, Alessandro Lucchetti, Francesco Cadini, Marco Giglio — Tue, 13 Jan 2026 00:00:00 +0000

The role of satellites and space systems in space exploration is of paramount importance, particularly given the growing interest in this field. The considerable financial outlay and commitment of resources that are necessary for space missions, coupled with the inherent difficulty of directly intervening in the systems in question in the event of faults or anomalies, give rise to the imperative of designing and developing highly reliable platforms. However, the harsh environment in which these systems operate means that unexpected and undesired events may occur, with the potential to have a catastrophic impact on mission objectives. To address this issue, it is of paramount importance to use telemetry data for the prompt detection of anomalies. This facilitates the implementation of corrective or mitigating actions, although this is challenging due to the remoteness of space platforms and the limited range of intervention options. A proactive approach can extend the duration of the mission, thereby preventing the loss of data and functions that are scientifically or commercially valuable. However, the intricate and multifaceted nature of telemetry data, which encompasses both sensor readings and commands, poses considerable analytical challenges, reinforcing the ongoing necessity of experts to determine system integrity.

Most existing anomaly detection algorithms depend excessively on extensive hyperparameter tuning and fixed thresholds, which renders them less robust when applied to different signals or evolving scenarios. Furthermore, numerous algorithms necessitate voluminous amounts of labeled data and protracted training phases, which results in increased deployment latency and operational costs, thereby limiting their utility throughout the duration of a mission.

To address these limitations, a novel framework for intelligent time-series anomaly detection (TSAD) is proposed. The solution has been designed to be plug-and-play, thereby minimizing the necessity of hyperparameter tuning and ensuring robust performance across diverse scenarios without the need for optimization. The system incorporates an adaptive thresholding mechanism that automatically adapts to evolving time-series behavior, ultimately enhancing detection accuracy. Furthermore, it requires only a statistically significant volume of unlabeled data, drastically reducing deployment latency and enabling rapid operational integration.

Towards Open-Set Fault Diagnosis for Reactor Coolant Pumps under Unknown Fault Conditions

Jonghyeok Kim, Jeongmin Oh, Jueun Lee, Minseok Choi, Hyunseok Oh — Tue, 13 Jan 2026 00:00:00 +0000

The reactor coolant pump – vibration monitoring system (RCP-VMS) ensures the safe operation of nuclear power plants by detecting anomalies in the shaft and bearing components of reactor coolant pumps. While effective for known fault modes, conventional AI-based diagnostic models often fail to detect unseen faults, especially when labeled data are limited. To address this limitation, an open-set recognition approach is proposed based on class-specific semantic reconstruction. Vibration signals collected from RCP-VMS are processed into orbit plot and recurrence plots, which serve as multi-channel image inputs to the model. The reconstruction errors are then used to distinguish both known and unknown fault conditions. Experimental results demonstrate that the proposed method achieves competitive closed-set accuracy while significantly enhancing open-set fault detection performance compared to baseline models. This approach enhances the reliability and robustness of fault diagnosis in safety-critical rotating machinery such as RCPs.

Towards Systematic Reliability Assessment: A Multi-Criteria Decision Framework for Modeling Heat Pump Systems

Ahmed Qarqour — Tue, 13 Jan 2026 00:00:00 +0000

Reliability assessment is essential to ensure the performance, availability, and safety of heat pump systems. This requires modeling strategies that capture both component-level behavior and system-level interactions. A wide range of reliability modeling approaches exists—including physics-based, data-driven, and hybrid methods—each offering distinct strengths suited to specific operational conditions and system architectures. Modern heat pump systems introduce added complexity: technically, through tightly coupled components; and organizationally, through fragmented supply chains and varying supplier inputs. These factors lead to heterogeneous levels of physical insight across components—from well-understood to poorly characterized. In parallel, the growing adoption of IoT technologies enables operational data collection, though such data often remains unstructured and lacks consistent failure labeling. Together, these challenges hinder the integration of appropriate modeling strategies and create a practical gap in applied reliability methods. To address this, we present a structured, scalable, and adaptable multi-criteria decision-making framework for reliability modeling in complex heat pump systems. The framework begins with component prioritization at the system level, followed by a structured evaluation of risk components using four key indicators: forecast granularity, physical understanding, data availability, and cost efficiency. This decision process is demonstrated on a real-world air-to-water heat pump use case. The proposed approach provides practitioners with a systematic pathway for selecting reliability modeling strategies tailored to varying levels of system complexity, available resources, practitioner expertise, and data constraints.

Unsupervised Modeling of Progressive Wear in Aircraft Engines for Predictive Maintenance

Abdellah Madane, Jérôme Lacaille, Hanane Azzag, Mustapha Lebbah — Tue, 13 Jan 2026 00:00:00 +0000

Predicting progressive wear in aircraft engines is critical for enabling condition-based maintenance and ensuring operational reliability. A persistent challenge lies in the discrepancy between benchmark datasets and real-world engine data. Although simulated datasets offer controlled and labeled conditions for model development, they often fail to represent the full complexity, noise characteristics, and operational irregularities observed in actual flight environments. This leads to models that perform well in simulation but degrade significantly when applied in practice. To address this limitation, this work introduces a data-driven framework to simulate realistic wear-and-tear effects using high-resolution timeseries data collected over sequences of engine missions. The method infers long-term degradation patterns in an unsupervised manner, without relying on explicit wear labels, while accounting for variability introduced by mission conditions.

Unsupervised Health Indicator Construction via Deep Reinforcement Learning with Terminal-Dominant Reward

Zeqi Wei, Zhibin Zhao, Ruqiang Yan — Tue, 13 Jan 2026 00:00:00 +0000

In industrial intelligent maintenance, the construction of a reliable health indicator (HI) is crucial for accurate degradation assessment and fault prediction. However, existing methods face two major limitations: fusion-based approaches often suffer from low-quality or irrelevant features that degrade the discriminative capability of the HI, while reconstruction-based approaches rely heavily on high-quality healthy data, which is difficult to obtain in real-world scenarios. To overcome these challenges, this paper proposes an Unsupervised Terminal-Dominant framework for HI construction (UTD-HI). The method does not rely on remaining useful life (RUL) labels or pre-defined thresholds. Within a deep reinforcement learning (DRL) paradigm, UTD-HI learns an adaptive feature-weighting policy that suppresses irrelevant features and enhances informative ones. A reward mechanism integrating monotonicity, smoothness, and a sparse terminal constraint is designed, while hindsight experience replay (HER) is introduced to address reward sparsity. Furthermore, by employing different reward strategies in normal and abnormal stages, the framework can automatically and accurately distinguish between healthy and degraded operating conditions. Experimental results on the XJTU-SY bearing dataset demonstrate that the proposed method constructs HIs with superior trendability, monotonicity, and robustness across different operating conditions, thereby offering a practical solution for HI construction in real-world environments.