International Journal of Prognostics and Health Management http://www.papers.phmsociety.org/index.php/ijphm <p>The flagship publication of the PHM Society is the open online journal entitled the International Journal of Prognostics and Health Management (IJPHM). The Journal has established a fast paced, yet rigorous peer-review policy. The Journal intends to publish original papers within 8-12 weeks of initial submission, much faster than what is possible with traditional print media.</p> PHM Society en-US International Journal of Prognostics and Health Management 2153-2648 Dual-Channel Acoustic Temporal–Spectral Representation and Noise-Perturbation Learning for Mining Conveyor Fault Diagnosis http://www.papers.phmsociety.org/index.php/ijphm/article/view/4841 <div> <div>This study develops a robust acoustic fault diagnosis framework for mining conveyor idlers, addressing the challenge of detecting early-stage mechanical degradation in noisy and imbalanced industrial environments. A dual-channel temporal--spectral representation is constructed by combining multi-scale log-Mel spectrograms and raw waveforms to capture complementary spectral patterns and fine-grained temporal dynamics. A Dual-Stream Cross-Attention Convolutional Recurrent Neural Network (DSCA-CRNN) is proposed to model cross-stream dependencies and enhance feature fusion. Noise-perturbation augmentation and a triplet-based contrastive objective are employed to enrich minority fault samples and improve embedding discriminability. Experiments on a self-collected conveyor auscultation dataset with 2,495 segments across three health states show that DSCA-CRNN achieves an overall accuracy of 0.95 and a macro-F1 score of 0.90, outperforming representative machine learning and deep learning baselines. Severe-fault recognition reaches an F1-score of 0.80. Ablation studies and PCA visualization confirm the effectiveness of recurrent temporal modeling, cross-attention fusion, noise-perturbation learning, and contrastive representation shaping. The proposed auscultation-based framework provides a practical and deployable solution for safety-oriented conveyor condition monitoring.</div> </div> Zhenyu Wang Taixiang Li Copyright (c) 2026 International Journal of Prognostics and Health Management 2026-08-03 2026-08-03 17 2 10.36001/ijphm.2026.v17i2.4841 Generalizing Bearing Fault Diagnosis to Physically Unseen Bearings http://www.papers.phmsociety.org/index.php/ijphm/article/view/5264 <p>Reliable bearing fault diagnosis requires generalization to physical components unseen during model development. This study evaluates such generalization using strict 32-fold Leave-One-Bearing-Out (LOBO) testing on the Paderborn University Bearing Dataset. All four diagnostic classes are retained, and each physical bearing is treated as the unit of generalization. The input combines 53 dynamic vibration-derived features and 30 physics-informed features. A classical machine-learning baseline, a deep model without adversarial learning (No-GRL), and a class-conditional Gradient-Reversal Learning (GRL) model with four bearing-domain adversarial heads are compared under the same protocol. At the bearing level, the classical model achieves 81.25% Accuracy, 74.81% Balanced Accuracy, and 77.82% Macro-F1; No-GRL achieves 78.12%, 74.81%, and 77.75%; and GRL achieves 81.25%, 76.89%, and 79.85%, respectively. Relative to No-GRL, GRL yields numerical gains of 3.13, 2.08, and 2.10 percentage points. However, only one held-out bearing changed from an incorrect No-GRL prediction to a correct GRL prediction, with no observed regression; the exact two-sided McNemar test gives p=1.000. A post-hoc Bearing-Identity Representation Audit shows that mean linear Bearing-ID accuracy decreases from 99.22% for No-GRL to 86.27% for GRL across the folds. This remains well above the 31-class chance level, indicating reduced rather than eliminated bearing-specific information. Thus, GRL is associated with modest aggregate diagnostic gains and lower linear recoverability of bearing identity, but the paired results do not establish a statistically significant general improvement in fault classification.&nbsp;</p> Rashwan Ammon Copyright (c) 2026 International Journal of Prognostics and Health Management 2026-11-21 2026-11-21 17 2 10.36001/ijphm.2026.v17i2.5264 From Faulty Greenhouse Sensors to Usable Training Data http://www.papers.phmsociety.org/index.php/ijphm/article/view/5171 <p>Predictive models for smart greenhouses commonly assume that the low-cost sensor data they are trained on are correct.<br />This paper treats the health of the sensing system as the object of study, using an uncurated 34-day greenhouse irrigation deployment (947,682 raw records, six streams, no fault-management layer). Every stream carries a naturally occurring fault signature: dropouts on 26.4% of the span, an eight-day out-of-range soil-moisture episode, a zero-variance channel, a CO2 channel with a suspected calibration bias, stuck-at episodes, spikes, and, after the longest outage, a variance collapse that marks a change of operating regime. The pipeline characterizes these faults, benchmarks rule-based screening, Isolation Forest and an LSTM autoencoder on 162 injection scenarios and seven repair methods, and measures how training-data treatment changes a soil-moisture forecaster against persistence. Rules and the autoencoder are complementary, bias and drift stay hard for all detectors, and linear interpolation is the best repair or not significantly different from the best in nearly every condition up to 12 h. The original LSTM forecasters of the soil-moisture level do not beat persistence at 5 min to 6 h, mainly because they cannot extrapolate to the post-outage level; within the training range, repaired data cut their pooled 5-min error 3.9–4.9-fold against raw data when samples are grouped by the true target and 3.1–3.7-fold when grouped by the last observed value; at 5 min, with the long-gap fallback, rules-guided repair reaches the persistence level. In an exploratory analysis added after these results, level-free change models selected on training data only beat persistence by 14% at 5 min when trained on repaired data, against 5% for raw data; at 1 h validation selected persistence for raw data and for two of three repair variants, and at 6 h the selected repaired-data models are worse than persistence. A random-mask control does not show a benefit of targeted repair on the full test set.</p> Mohamed Akram Lamhour Mohamed Msalek Mohamed Kasbouya Soufiane Ardchir Mohamed Azzouazi Copyright (c) 2026 International Journal of Prognostics and Health Management 2026-11-21 2026-11-21 17 2 10.36001/ijphm.2026.v17i2.5171 From Prediction to Decision http://www.papers.phmsociety.org/index.php/ijphm/article/view/5107 <p>Accurate prediction of remaining flight cycles is the goal of most turbofan prognostics, but low aggregate error does not by itself guarantee low realized maintenance cost. On the PHM North America 2025 Data Challenge (three maintenance targets of very different event frequencies, scored under an asymmetric, time-weighted cost) we build a conventional prediction pipeline and then ask how reliable its accuracy is. We take a strictly nested temporal analysis, in which calibration data are untouched by fitting or tuning, as the primary result. Across two allocations of the limited pre-deployment history, a nominal 90% split-conformal interval achieves 90–93% coverage on the late-life deployment window for the frequent water-wash, but only 25–29% for the high-pressure-turbine (HPT) shop visit and 19–28% for the high-pressure-compressor (HPC) shop visit. Complementary interval and base-model checks, and all four subsets of the public NASA C-MAPSS benchmark, reproduce the separation between water-wash and the two shop-visit targets.<br />Converting predictions into scheduling decisions, a policy whose safety buffer is the conformal half-width reduces realized cost by 23–41% against trusting the point prediction under the two nested protocols, cutting unplanned failures from 31 to 9–13. Its comparison with a tuned constant buffer changes from a 13% advantage to a 3% disadvantage as the scarce pre-deployment data are apportioned, so four engines do not order the two buffer policies. A cost sweep places their crossover at a failure penalty of roughly 250 forfeited cycles of useful life. On HPC, constant-buffer policies are not robust: the original and Nested Variant I buffers leave all six failures, while the much wider Variant II buffer cuts failures to three at substantial early-service cost. A prediction-free age-based fallback and, under the secondary non-nested calibration, two instance-dependent triggers also reduce failures. Per-component reliability must therefore gate the maintenance decision rather than follow it.</p> Gaoyuan Du Ark Ifeanyi Anahita Khojandi Copyright (c) 2026 International Journal of Prognostics and Health Management 2026-11-21 2026-11-21 17 2 10.36001/ijphm.2026.v17i2.5107 Development of an Integrated Digital Twin Framework for Online Monitoring and Anomaly Detection http://www.papers.phmsociety.org/index.php/ijphm/article/view/5076 <p class="phmbodytext">Digital Twins (DTs) are emerging as an important enabling technology in the nuclear field, as their development leverages the ongoing digitalization of nuclear systems and offers many advantages for optimizing nuclear operations. Nuclear DTs can be used for predictive maintenance, enhanced operations, improved scheduling, cradle-to-grave lifecycle management, decision support, and more. While the advantages are clear, continued DT development is needed, as many gaps remain, including technological and regulatory standardization, validation and verification, and digital-to-physical communication. To help close these gaps, a forced-water-flow experimental system at the University of Tennessee, Knoxville (UTK) was used to develop a complete, integrated DT, offering an end-to-end use case in the nuclear sector and related disciplines. For this work, a Long Short-Term Memory (LSTM) network was used for forecasting, and an interquartile range (IQR) LSTM hybrid method was used for anomaly detection. This was all interfaced with LabVIEW, enabling real-time data analysis and forecasting. A GUI was then created to provide the operator with a clear indication when a system anomaly is detected, enabling better insights during operation. The resulting platform demonstrates an end-to-end proof of concept for integrating data acquisition, machine learning-based forecasting and anomaly detection, and operator interaction within an experimental DT for various nuclear and related engineering systems.</p> Lauren Bailey Jamie Coble Xingang Zhao Copyright (c) 2026 International Journal of Prognostics and Health Management 2026-09-30 2026-09-30 17 2 10.36001/ijphm.2026.v17i2.5076 Rolling Bearing Fault Diagnosis using Time-frequency Image-based Deep Learning Networks http://www.papers.phmsociety.org/index.php/ijphm/article/view/5068 <p>Deep learning approaches have shown strong potential for vibration-based fault diagnosis; however, their evaluation is often affected by similarity bias caused by inappropriate data splitting strategies. In such cases, highly similar samples may be simultaneously present in training and testing sets, leading to overly optimistic performance estimates and limited generalization assessment. This paper investigates the use of time–frequency signal representations and transfer learning under a fault diagnosis framework that avoids similarity bias conditions. Raw vibration signals are transformed into two-dimensional representations using the Short-Time Fourier Transform (spectrogram) and the Continuous Wavelet Transform (scalogram), which are used as inputs to ImageNet-pretrained convolutional neural networks.<br />To reduce the impact of similarity bias, a controlled data splitting strategy is adopted to prevent highly similar samples from being shared between training and testing subsets. The framework is evaluated using four publicly available benchmark datasets widely adopted in intelligent fault diagnosis research, enabling consistent comparison with existing studies. Different fine-tuning strategies and network architectures are considered to analyze the influence of signal representation and model adaptation under similarity bias.</p> Rodrigo Corrêa Fardin Marcus Lourical Neves Filho Thomas Walter Rauber Berilhes Borges Garcia Flávio Miguel Varejão Copyright (c) 2026 International Journal of Prognostics and Health Management 2026-11-21 2026-11-21 17 2 10.36001/ijphm.2026.v17i2.5068 Leaf It to Renewal http://www.papers.phmsociety.org/index.php/ijphm/article/view/5033 <p>We propose a hybrid planning method for deriving prognostics-based predictive maintenance policies. The method accounts for the available decision options, the information on the future state of the system provided by a prognostic model, and the costs of the underlying renewal-reward process. It results in policies defined by only a few parameters, which can be determined based on theoretical considerations or by optimization from run-to-failure data. We demonstrate the potential of the method in two separate predictive maintenance decision settings: preventive replacement and preventive ordering. Numerical investigations show that the derived policies rival the performance of optimized benchmark policies, while being significantly more efficient and robust against overfitting.</p> Daniel Koutas Daniel Straub Copyright (c) 2026 International Journal of Prognostics and Health Management 2026-11-21 2026-11-21 17 2 10.36001/ijphm.2026.v17i2.5033 Simulation-Based Fault Detection and Diagnosis for Electric Servomotor Control Rod Drives in BWRX-300 http://www.papers.phmsociety.org/index.php/ijphm/article/view/4974 <p>GE-Hitachi’s BWRX-300 Small Modular Reactor design adopts innovations from the Advanced Boiling Water Reactor (ABWR), including the use of servomotor-driven fine motion control rod drives (FMCRDs) instead of the hydraulically-driven locking piston control rod drives (LPCRDs) commonly used in current boiling water reactors (BWRs). The FMCRDs provides more precise power control during normal operation and functional backup to the hydraulic SCRAM system during emergency scenarios. To address the challenge of monitoring the novel servomotor-driven FMCRDs in the BWRX-300, for which no operational data exist, this paper develops and validates a simulation-based, two-stage diagnostic framework. A model of the FMRCDs was created and verified in MATLAB/Simulink to generate comprehensive datasets for healthy operation and key fault modes, including stator short circuits and mechanical load failures. From these simulations, diagnostic features were extracted from motor currents and control response signals. The proposed framework first uses Auto-Associative Kernel Regression (AAKR) and a Sequential Probability Ratio Test (SPRT) for robust anomaly detection, achieving an 85.9% true positive rate with a 2.15% false positive rate. Upon detection, a second-stage multi-channel SPRT classifier diagnoses the specific fault, successfully identifying 100% of load faults and 75% of short faults. This study validates that a simulation-driven approach can establish an effective and interpretable predictive health monitoring system for critical reactor components before they are physically deployed, enabling condition-based maintenance from the outset.</p> Tran C.H. Nguyen Ark Ifeanyi Daniel Dos Santos Nicholas Fullilove Abhinav Saxena Jamie Coble Copyright (c) 2026 International Journal of Prognostics and Health Management 2026-10-19 2026-10-19 17 2 10.36001/ijphm.2026.v17i2.4974 Monotonic Latent Manifold Bayesian Inference for Robust Industrial Predictive Maintenance http://www.papers.phmsociety.org/index.php/ijphm/article/view/4874 <div>Industrial predictive maintenance systems require not only accurate fault prediction but also reliable uncertainty estimation and robustness under noisy operational conditions. However, conventional deep learning approaches often produce temporally inconsistent failure trajectories and exhibit poor probabilistic calibration when exposed to stochastic sensor disturbances. This study proposes a Monotonic Latent Manifold Bayesian Neural Network (MM-BNN) framework for robust industrial predictive maintenance. The MM-BNN models degradation progression as a constrained non-decreasing latent probabilistic trajectory, thereby ensuring physically consistent temporal evolution while preserving Bayesian uncertainty quantification. A monotonic latent manifold update mechanism is introduced to enforce structured degradation dynamics across sequential observations. In addition, Bayesian posterior inference is performed using the No-U-Turn Sampler (NUTS) to capture uncertainty in latent degradation states. The proposed MM-BNN framework is evaluated against multiple baseline models, including multilayer perceptron networks, gated recurrent units, random forests, support vector machines, XGBoost, and standard Bayesian neural networks. Experimental results demonstrate that the MM-BNN achieves superior discriminative performance with an average Area Under the Curve (AUC) of 0.772, outperforming competing approaches, most of which remain close to random classification performance. Reliability analysis further demonstrates improved probabilistic calibration, with a low Expected Calibration Error (ECE approx 0.034) and stable reliability behavior under increasing levels of stochastic noise. Ablation studies confirm that the monotonic degradation constraint is the primary factor driving predictive stability and robustness improvements. Furthermore, robustness experiments show that the MM-BNN maintains stable predictive performance even under severe noise perturbations. Overall, the findings indicate that the proposed MM-BNN framework provides a reliable, interpretable, and uncertainty-aware approach for predictive maintenance in safety-critical industrial environments.</div> Oyetayo Oyebisi Copyright (c) 2026 International Journal of Prognostics and Health Management 2026-08-23 2026-08-23 17 2 10.36001/ijphm.2026.v17i2.4874 Channel-Attentive Temporal Convolutional Networks (CA-TCN) for Interpretable Fault Diagnosis in the Tennessee Eastman Process http://www.papers.phmsociety.org/index.php/ijphm/article/view/4859 <p>Fault detection and diagnosis (FDD), Fault classification in the Tennessee Eastman Process (TEP) involves strongly coupled, nonlinear multivariate trajectories. The present research evaluates a residual Temporal Convolutional Network (TCN) with Squeeze-and-Excitation (SE) channel attention, denoted CA-TCN. We compare CA-TCN with plain TCN, Long Short-Term Memory (LSTM), and Convolutional Neural Network-LSTM (CNN-LSTM) baselines. CA-TCN and TCN+SE refer to the same implemented architecture.</p> <p>In the primary five-seed comparison, plain TCN achieved a mean accuracy of 99.8756 ± 0.0807%, compared with 99.8511 ± 0.1104% for CA-TCN, 99.1711 ± 0.9467% for LSTM, and 99.8444 ± 0.0820% for CNN-LSTM. A separate matched five-seed TCN/CA-TCN ablation yielded mean accuracies of 99.8000 ± 0.1199% and 99.7756 ± 0.1823%, respectively.</p> <p>The mean paired CA-TCN-minus-TCN difference was −0.0244 percentage points. The exploratory paired analysis did not identify a statistically detectable difference in accuracy. Adding the SE module introduced 4,192 additional trainable parameters and increased the measured training time while providing explicit reweighting of 64 learned latent feature channels.</p> <p>These coefficients are interpreted as internal latent-channel recalibration rather than sensor-importance scores. The results therefore characterize the performance–complexity trade-off associated with adding SE attention to the residual TCN backbone.</p> <p>The conclusions are limited to offline whole-trajectory classification on the evaluated TEP benchmark; cross-domain generalization and online fault-detection performance were not assessed.</p> Pratyush Kumar Pal Sumona Ray Narottam Behera Somnath Chowdhury Abhiram Hens Sandip Kumar Lahiri Copyright (c) 2026 International Journal of Prognostics and Health Management 2026-10-20 2026-10-20 17 2 10.36001/ijphm.2026.v17i2.4859 A Hybrid Particle Filter-Weibull Approach for Probabilistic End-of-life Prediction of Industrial Machines Using Vibration Data http://www.papers.phmsociety.org/index.php/ijphm/article/view/4735 <p>Accurate degradation modeling is essential for ensuring the reliability of industrial machinery operating under uncertain conditions. This paper presents a hybrid prognostic framework that integrates Particle Filters (PF) with a Weibull Failure Rate Function (WFRF) for probabilistic End-of-Life (EOL) prediction using vibration data. The PF is employed to estimate hidden degradation states and update model parameters sequentially, while the WFRF provides a statistical representation of failure behavior. The proposed framework enables joint degradation tracking and uncertainty quantification. The method is validated using vibration datasets from a mining dumper operating under real conditions and from an induction motor operating under controlled laboratory conditions. Results demonstrate that the hybrid PF-WFRF approach achieves accurate EOL prediction with small deviation from actual failure thresholds (within ~120 seconds) and provides meaningful uncertainty bounds through probability density functions. Unlike existing hybrid approaches that rely on machine learning, the proposed framework provides a computationally efficient, interpretable probabilistic solution. The framework is robust to both smooth and noisy degradation signals, making it suitable for real-world prognostics applications.</p> Nagesh Dewangan Amiya Ranjan Mohanty Copyright (c) 2026 International Journal of Prognostics and Health Management 2026-09-11 2026-09-11 17 2 10.36001/ijphm.2026.v17i2.4735 Multiscale Ordinal Complexity Analysis for Wind Turbine Bearing Fault Diagnosis http://www.papers.phmsociety.org/index.php/ijphm/article/view/4835 <p>Permutation-entropy methods for bearing fault diagnosis in wind turbines are typically validated on laboratory test rigs without enforcing specimen-level separation in cross-validation or quantifying axis-specific diagnostic contributions. This study addresses these limitations using the Fraunhofer LBF operational wind turbine bearing dataset, applying multiscale permutation entropy (MPE) to triaxial front-bearing accelerometer signals across 44 bearing specimens (18 healthy, 10 inner race, 4 outer race, 12 roller element), with two corrupted files excluded following data quality screening. GroupKFold cross-validation with unique specimen-level group identifiers prevents the data leakage that arises when temporally correlated analysis windows are split without regard to bearing identity - a limitation present in all ten studies identified in a systematic literature search. MPE achieves 97.67% +/- 2.10% window-level and 97.73% specimen-level accuracy using 12 features across three accelerometer axes, outperforming weighted permutation entropy (WPE, 91.83% window-level, 95.45% specimen-level) and matching a physically-augmented hybrid (MPE+Physical, 21 features) that contributes no additional specimen-level accuracy despite 33.0% feature importance within the combined set. The X-axis accelerometer (brng.f.x) accounts for 40.7% of classification importance, consistent with the primary radial load direction. Permutation entropy reveals a monotonic complexity hierarchy across fault types - Healthy (mu=0.702) &lt; Roller Element (mu=0.889) &lt; Inner Race (mu=0.981) &lt; Outer Race (mu=0.992) - with Cohen's d &gt; 0.9 for all pairwise comparisons. Roller element faults exhibit bimodal permutation entropy distributions, explained by load-zone-dependent impulsive generation. These results demonstrate that multiscale ordinal pattern analysis captures physically meaningful fault signatures in operational wind turbine data when evaluated under methodologically sound cross-validation protocols.</p> Paulo R. L. Almeida Thyago L. V. Lima Alisson V. Brito Abel C. Lima Filho Copyright (c) 2026 International Journal of Prognostics and Health Management 2026-08-29 2026-08-29 17 2 10.36001/ijphm.2026.v17i2.4835 Estimating Sufficient Amount of Training Data for Data-Driven Fault Diagnosis http://www.papers.phmsociety.org/index.php/ijphm/article/view/4834 <p>Data-driven approaches for diagnosing and predicting the health of engineering systems require sufficient training data to achieve reliable performance. However, despite its high practical relevance, the question of how much data are actually required has not yet been systematically investigated.<br />This paper presents and evaluates two complementary approaches to answering this question. Approach I fits parametric functions to the initial segments of learning curves and extrapolates them to estimate the amount of training data required to achieve a target classification accuracy. Approach II introduces a training-free, entropy-based stopping criterion that uses the gradient of normalized entropy (GNE). The information gain from the data collected over time is monitored to determine whether further data acquisition is necessary. Both approaches are evaluated using a dataset for fault classification in rolling bearings, employing a multi-layer perceptron classifier and randomly drawn, in-distribution test sets. The results show that learning curve extrapolation enables predictions of the required amount of data and that GNE can serve as an effective online stopping rule during data acquisition due to the strong correlation between GNE and classification accuracy. The two approaches complement each other: entropy analysis can monitor the progress of data acquisition in real time without model training, while learning curve extrapolation provides quantitative estimates of the remaining data requirements for specific accuracy targets.</p> Moritz Neu Marcel Braig Fabian Mauthe Peter Zeiler Copyright (c) 2026 International Journal of Prognostics and Health Management 2026-10-11 2026-10-11 17 2 10.36001/ijphm.2026.v17i2.4834 Remaining Useful Life Prediction across Datasets via Semi-Supervised Domain Adaptation http://www.papers.phmsociety.org/index.php/ijphm/article/view/4804 <p>Remaining useful life (RUL) prediction has become a critical task, as accurate prediction enables effective maintenance planning and minimizes unplanned downtime. Complex systems frequently operate under diverse operating conditions, with distributional differences and limited RUL labels. To address these issues, domain adaptation (DA) methods have attracted increasing attention. Mainstream RUL prediction DA methods mainly adopt an unsupervised setting (UDA), relying solely on unlabeled target-domain data. However, prediction performances of UDA methods are far inferior to Target-Only (TO) models trained on full labeled target data, and it fails to effectively transfer degradation-related knowledge for RUL prediction when source-target distribution differences are large. Additionally, existing research mainly focuses on in-dataset adaptation across operating conditions, lacking systematic exploration of cross-dataset RUL prediction. Therefore, this work proposes a semi-supervised domain adaptation (SSDA) workflow for RUL prediction between different datasets. We design an alignment pipeline to address feature discrepancies between datasets and sampling frequency differences. During SSDA training, the original RUL labels of the source domain are treated as noisy labels with respect to the target domain and are progressively modified throughout the adaptation process. We conduct two case studies, one on turbofan engines and the other on rolling bearings. The experimental results demonstrate that the proposed approach can effectively realize cross-dataset RUL prediction.</p> Gengyu Li Takehisa Yairi Copyright (c) 2026 International Journal of Prognostics and Health Management 2026-07-17 2026-07-17 17 2 10.36001/ijphm.2026.v17i2.4804 Data-Driven Fault Diagnosis of Rolling Bearings Using Neural Networks Optimized via Bayesian, Particle Swarm, and Genetic Algorithm Methods with Time–Frequency Features http://www.papers.phmsociety.org/index.php/ijphm/article/view/4768 <p class="phmbodytext">This study presents an applied comparative evaluation of automated rolling bearing fault identification using Artificial Neural Networks (ANNs) optimized through Bayesian Optimization (BO), Particle Swarm Optimization (PSO), and Genetic Algorithm (GA). Vibration signals were collected from bearings operating under five health conditions (healthy, outer ring fault, inner ring fault, ball fault, and combined faults), at three rotational speeds, and along three measurement directions. The acquired signals were preprocessed using filtering, normalization, and segmentation. Time-domain and Fast Fourier Transform (FFT)-based frequency-domain features were extracted and used to train ANN models. The ANN architectures, including hidden layers, neurons, and activation functions, were optimized using BO, PSO, and GA, resulting in six configurations. Since the problem is a multi-class classification task, performance was assessed using F1-score, Accuracy, precision, and recall. The optimized ANN models were also benchmarked against Support Vector Machine (SVM), K-Nearest Neighbors (kNN), and Random Forest (RF) classifiers using the same feature sets. Results show that FFT -based features consistently outperformed time-domain features, and ANN-PSO with FFT-based features achieved the best performance, with F1-score = 0.982, Accuracy = 0.998, precision = 0.982, and recall = 0.982. This work contributes a systematic applied comparison of ANN optimization strategies rather than a fundamentally new machine-learning architecture, highlighting optimized ANNs as competitive and computationally efficient solutions for bearing fault diagnosis.</p> Khoualdia Kaaïs Khoualdia Tarek Mourad Nahal Copyright (c) 2026 International Journal of Prognostics and Health Management 2026-08-01 2026-08-01 17 2 10.36001/ijphm.2026.v17i2.4768 Physics-Informed Multi-Scale Network with Loss-Guided Curriculum Learning for Robust Fault Diagnosis http://www.papers.phmsociety.org/index.php/ijphm/article/view/4766 <div> <div>Reliable fault diagnosis of rotating machinery is critical, yet early weak fault impulses are frequently buried in severe compound interference from mechanical harmonics and environmental noise. To address this, a novel Physics-Informed Multi-Scale Network (PI-MSN) is proposed. Variational Mode Decomposition (VMD) is first employed to decouple raw signals into distinct physical frequency bands. Subsequently, a Physics-Informed Channel Attention (PICA) module jointly evaluates the Kurtosis and Root Mean Square of each channel to autonomously highlight fault impulses and suppress harmonic interference. A Multi-Scale Feature Extractor then captures comprehensive fault characteristics. Furthermore, a closed-loop Loss-Guided Smooth Interference Scheduler (LGSIS) dynamically regulates injected interference during training based on real-time loss, fundamentally eradicating catastrophic forgetting. Extensive experiments on the CWRU and HUST datasets demonstrate the framework's exceptional robustness. The highly lightweight PI-MSN achieves state-of-the-art diagnostic accuracy, sustaining over 98\% accuracy even under severe -4 dB compound interference, proving that physical interpretability effectively eliminates the reliance on massive parameter stacking.</div> </div> Panfeng Bao Wenjun Yi Yue Zhu Yufeng Shen Haotian Peng Copyright (c) 2026 International Journal of Prognostics and Health Management 2026-07-05 2026-07-05 17 2 10.36001/ijphm.2026.v17i2.4766 An Enhanced Lightweight Multi-scale CNN with Resonance-dominant Frequency-band Learning for Robust Bearing Fault Diagnosis http://www.papers.phmsociety.org/index.php/ijphm/article/view/4757 <p>Rolling bearings are essential components in rotating machinery, and reliable early fault detection is crucial for avoiding unexpected failures and reducing maintenance costs. While deep learning techniques have achieved notable progress, most prior studies have emphasized feature enrichment without thoroughly examining vibration spectral characteristics or establishing mechanisms to select the most informative time–frequency patterns. To exploit this physical insight, this study proposes an enhanced lightweight multi-scale convolutional neural network (CNN) with adaptive frequency-band learning for robust bearing fault diagnosis. Vibration signals are first band-limited to the resonance-dominant range and transformed using the Continuous Wavelet Transform with the Morlet wavelet, improving signal-to-noise ratio while preserving impulsive fault components. The proposed architecture integrates depthwise separable convolutions, multi-scale parallel branches, Squeeze-and-Excitation channel recalibration, and Atrous Spatial Pyramid Pooling to capture discriminative resonance-sensitive features with reduced computational complexity. Grad-CAM is employed to visualize the resonance bands contributing to classification decisions. Experiments on the CWRU dataset achieve accuracy above 99% with only 7.69 M parameters. Robustness under additive noise and cross-domain validation on the HUST dataset confirm strong generalization and domain invariance.</p> Thanh Tung Luu Duy An Huynh Thanh Hai Nguyen Copyright (c) 2026 International Journal of Prognostics and Health Management 2026-11-07 2026-11-07 17 2 10.36001/ijphm.2026.v17i2.4757 Regularization of Divergent Series via Classical Summability Methods with Applications to Reliability and Engineering Systems http://www.papers.phmsociety.org/index.php/ijphm/article/view/4750 <p class="phmbodytext"><span lang="EN-US">Accurate estimation of remaining useful life (RUL) from noisy and oscillatory degradation data remains a fundamental challenge in prognostics and health management (PHM), especially in safety-critical and resource-constrained settings. While modern data-driven methods achieve high predictive accuracy, they often require extensive training, lack interpretability, and provide limited guarantees on reliability. This study addresses these challenges by proposing a Cesàro-based regularization framework for constructing stable, interpretable, and conservative health indicators from degradation signals. The main objective is to develop a parameter-free and computationally efficient smoothing method that ensures safety-aware degradation tracking. The novelty of this work lies in adapting classical summability theory to PHM and establishing an explicit finite-sample bound that guarantees the smoothed estimate remains conservatively biased relative to the most recent observation, even under mild non-monotonic conditions. The method applies Cesàro averaging to a univariate health index derived from sensor fusion and is evaluated on NASA C-MAPSS FD001 and FD004 datasets using a consistent experimental framework. It is compared with standard smoothing techniques, including SMA, EWMA, Kalman filtering, and isotonic regression. Results show that the proposed approach achieves competitive or improved performance across RMSE, MAE, PHM score, and prognostic horizon, while significantly reducing variance and computational complexity. The Cesàro regularization provides a theoretically grounded, lightweight, and interpretable solution for degradation modeling. Its conservative nature enhances reliability, making it particularly suitable for real-time predictive maintenance applications in aerospace, energy, and industrial systems.</span></p> Suresh Kumar Sahani Kameshwar Sahani Aashiq Mahato Rishav Jha Copyright (c) 2026 International Journal of Prognostics and Health Management 2026-09-29 2026-09-29 17 2 10.36001/ijphm.2026.v17i2.4750 Bayesian CNN-LSTM for Battery State of Health (SoH) Estimation under Randomized Usage Conditions http://www.papers.phmsociety.org/index.php/ijphm/article/view/4743 <p>The accurate assessment of State of Health (SoH) of battery plays pivotal role in enabling the safe and dependable operation of lithium-ion battery systems. In this work, we propose a hybrid deep learning architecture integrating Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks within a Bayesian learning paradigm to improve SoH prediction accuracy considering complex charge-discharge patterns. The CNN module automatically extracts temporal and local degradation features from sequential charge-discharge profiles, while the LSTM component captures long-term dependencies in the cycling behavior. Bayesian inference is incorporated to quantify predictive uncertainty; posterior sampling yields 90% prediction intervals (S = 100; μ ± 1.64σ), which support uncertainty-informed assessment under irregular and noisy inputs. The model is trained and evaluated on the NASA randomized battery usage dataset; labels are computed from capacities at consecutive reference discharge cycles as the incremental change in SoH (dSoH), with absolute SoH reconstructed by accumulation at inference. Experimental results demonstrate that the proposed Bayesian CNN-LSTM architecture achieves high prediction accuracy, with an R² score of 0.932 and an RMSE of 0.022, while providing uncertainty estimates with reasonable empirical coverage under diverse operational conditions. These findings indicate potential applicability in battery management workflows.</p> Saurabh Jaywant Desai Muralikrishan Ramanujam Venkataramana Runkana Copyright (c) 2026 International Journal of Prognostics and Health Management 2026-07-17 2026-07-17 17 2 10.36001/ijphm.2026.v17i2.4743 A Systematic Review of AI-Based PHM Frameworks for Ship Systems and Equipment http://www.papers.phmsociety.org/index.php/ijphm/article/view/4764 <p>Maritime operations require high levels of availability and safety under changing cargo, operational, and environmental conditions. Therefore, anticipating degradation and failures in critical ship systems is an important operational requirement.<br />Although predictive maintenance has advanced rapidly through data-driven approaches and artificial intelligence (AI), the available literature remains fragmented and does not consistently establish the relationships among the ship systems analyzed, the methodological approaches employed, and the conditions required for operational implementation.<br />This article presents a systematic literature review guided by the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) methodology to examine the development of AI-based predictive maintenance for ship systems from a Prognostics and Health Management (PHM) perspective. Scopus and Web of Science were searched, and predefined screening, eligibility, inclusion, and exclusion criteria were applied.<br />The findings show that research is strongly concentrated on propulsion systems, classified under Ship Work Breakdown Structure (SWBS) Group 200, and electric-plant systems, classified under SWBS Group 300, with most studies focusing on detection and diagnosis. Prognostic and prescriptive applications remain comparatively less developed. Methodologically, the literature is dominated by machine-learning and deep-learning approaches using operational and condition-monitoring variables associated with different operating regimes.<br />Overall, maritime PHM exhibits uneven levels of functional and implementation maturity. Important challenges remain in model comparability, independent validation, and sustained onboard deployment, while opportunities persist for extending PHM to underrepresented ship systems and validating these approaches in real-world fleet operations</p> Luis Fernando Mendoza Cardona Joan Suarez Edwin Alexander Puertas Del castillo Edwin Paipa Sanabria Juan Carlos Martinez Santos Copyright (c) 2026 International Journal of Prognostics and Health Management 2026-10-18 2026-10-18 17 2 10.36001/ijphm.2026.v17i2.4764