The assessment of equipment health during field operation is at the basis of condition-based and predictive maintenance. However, in many cases, Health Indicators (HIs) can only be measured in controlled environments, such as laboratories. The difference in the conditions of field operation and the controlled environment creates a domain shift that complicates health state estimation. To address this, we propose a novel Physics-informed Domain Adaptation (PIDA) method that uses data collected in controlled environment (source domain) to estimate the equipment HI in the field (target domain). The method assumes that shifts in operating conditions induce localized perturbations in the degradation dynamics, while the physical laws governing the degradation process remain invariant. PIDA first utilizes Sparse Identification of Nonlinear Dynamics with control (SINDYc) to discover the degradation laws governing the source domain data. Subsequently, a model is trained using a composite loss function that combines a Maximum Mean Discrepancy (MMD) term for domain alignment with a physics-informed term that constrains the target domain HI estimates to conform to the discovered law. We validate PIDA on the Aramis Data Challenge and a cross-domain case study involving lithium-ion batteries. PIDA is shown to achieve an average Root Mean Squared Error reduction of 16.95% in HI estimation compared to the best-performing among the state-of-the-art methods considered.

For effective maintenance decisions on electronic components, operators need predictions of the Remaining Useful Life (RUL) that are not only accurate but also stable. To this aim, a novel prognostic model based on gradient-enhanced Physics-Informed Neural Networks (gPINNs) is developed. It is based on mathematically formulating the physical fact that the ground-truth RUL of a component is reduced by one time unit for every unit of time elapsed in its life and its incorporation into the loss function of a Physics-Informed Long Short-Term Memory (PILSTM) network. To recover from possible inaccurate RUL predictions, which can occur especially at the end of the component life, a gradient-enhanced PILSTM is developed by considering the second derivative of the RUL, which should ideally be null. Additionally, a novel ensemble strategy is proposed for automatically weighing the different terms of the loss function so as to eliminate the labor-intensive and error-prone process of manually tuning the weights and to further improve the accuracy of the predictions. The proposed method is applied to two types of electronic components: Insulated Gate Bipolar Transistors (IGBTs) and lithium-ion batteries. The results demonstrate that it outperforms other state-of-the-art methods in terms of accuracy and stability of the RUL predictions.

Data-driven methods for fault diagnosis in Photovoltaic (PV) arrays typically rely solely on either electrical or thermal signals, disregarding their inherent association, which leads to insufficient information exploitation. Furthermore, most methods treat compound faults merely as new classes of faults or employ conventional multi-label classification, thereby neglecting the interaction and coupling effects among faults. To overcome these limitations, this work develops an innovative decoupling diagnostic model based on thermo-electrical multimodal data fusion and a cross-attention mechanism to identify compound faults in PV arrays. The model adopts the current–voltage (I-V) characteristics and the temperature distribution of the array as complementary input modalities. Specifically, a dual-branch encoding network is employed to separately extract electrical and thermal fault features. The core contribution of the method is the proposed cascaded cross-attention architecture: first, a cross-modal attention module enables adaptive fusion of the two heterogeneous information sources; then, a cross-attention decoder based on learnable fault query vectors dynamically decouples and separates discriminative feature components relevant to different individual faults from the fused representation. Finally, independent classifiers operating in parallel output the presence probability of each fault class. The proposed method is validated considering a dataset containing both single and compound fault scenarios, generated by a verified thermo-electrical coupled PV simulation system. The obtained results show the superior performance of the proposed method in comparison to state-of-the-art single-modality and multi-label classification methods, achieving an average diagnostic accuracy of 98.3% for single faults and of 95.1% for compound faults.

Federated Learning (FL) has emerged as a promising paradigm for preserving client data ownership and control over distributed Internet of Things (IoT) environments. While discriminative models dominate most FL use cases, recent advances in generative models – such as Variational Autoencoders (VAE), Generative Adversarial Networks (GAN), and Diffusion Models (DM) – offer new opportunities for unsupervised anomaly detection in time series analysis, with relevant applications in predictive maintenance (PdM) in critical industrial infrastructures. In this work, we present a comprehensive analysis of VAEs, GANs, and DMs in the context of federated PdM. We analyze their performance and communication overhead under both full and partial federation setups, where only subsets of model components are shared. Building on this analysis, the paper proposes a novel taxonomy for federated generative models that formalizes partial component sharing as a principled mechanism for model personalization. Our experiments over a real-world time series dataset reveal distinct trade-offs in model utility, stability, and scalability, especially in heterogeneous and bandwidth-constrained FL settings. For the evaluated GAN-based configurations, full federation improves training stability relative to independent local training, although the model remains less robust than the VAE- and DDPM-based alternatives. For DMs, however, partial federation – especially decoder sharing – can outperform full federation in bandwidth-constrained, non-IID settings.

As wind power installations continue to expand rapidly, ensuring reliable and cost-effective Operation and Maintenance (O&M) over the wind turbine lifetime has become increasingly important. With the development of Industry 4.0, predicting the health status of wind turbines and making informed maintenance decisions has become an urgent challenge that must be addressed to enable the next generation of O&M paradigms. This paper starts with presenting a comprehensive review of health prognostics for wind turbines. Existing approaches are generally divided into two main categories: (1) model-based methods, including physics-based and knowledge-based approaches, and (2) data-driven methods, which encompass statistical methods as well as Artificial Intelligence (AI)-based methods, including both traditional and emerging AI methods. Subsequently, the maintenance decision-making problem informed by wind turbine health information is systematically summarized, with a particular focus on the historical evolution, problem formulation, data challenges, modeling techniques, optimization objectives, and solving techniques. Finally, key open challenges in the context of future digital and intelligent O&M are highlighted, and potential research directions are outlined to address these challenges.

It is a practical situation that the data used to develop data-driven prognostic models are collected from run-to-failure experiments performed in laboratory (source domain), where the component degradation is measured, whereas the data to which the model is applied are acquired during in-field operation (target domain), where operating conditions are different and degradation cannot be directly measured. This problem cannot be effectively tackled using traditional deep learning approaches; in this paper, an innovative Heterogeneous Transfer Learning (HTL) method is developed for the estimation of the Remaining Useful Life (RUL) of degrading components. The missing information about the degradation of the target component is estimated by developing a first-stage Domain Adaptation (DA) model, which learns the mapping between the signals measured in both domains and the degradation indicator. Then, the RUL of the target component is estimated by developing a second-stage DA model, which leverages the prognostic knowledge in the labeled source domain data. Both DA models use an encoder based on a Long-Short Term Memory (LSTM) neural network for extracting features, and the Maximum Mean Discrepancy (MMD) metric for reducing the domain discrepancy. The proposed HTL method is validated considering the data of the Aramis Data Challenge and of the MIT and Stanford University battery benchmark. The proposed method achieves more satisfactory performances than other DA state-of-the-art methods, with an increase of the Cumulative Relative Accuracy (CRA) of 4% and 7% in the two application cases, respectively.

This work considers the problem of developing a data-driven anomaly detection method in the common situation in which a large dataset is available only for few systems of a fleet and the data therein are unlabeled, i.e., the equipment health state (normal/anomalous) is unknown. An innovative two-stage unsupervised Transfer Learning (TL) framework is developed. In the first stage, a Long Short-Term Memory Encoder-Decoder (LSTM-ED) is pre-trained to reconstruct the values expected in normal condition of a specific ("source") system of the fleet using a dataset of signal measurements collected during a long period of time. Since industrial systems are typically in normal condition during most of their operational time, this pre-trained model is expected to reproduce their behavior in normal condition. In the second stage, the network architecture and parameters of the pre-trained model are used to initialize the dedicated LSTM-ED-based signal reconstruction model for another ("target'') system of the same fleet, for which only a limited amount of data collected during a short period of time is available. Data of the Aramis Data Challenge and of traction systems of a fleet of trains have been used to validate the proposed anomaly detection method in two different case studies. The obtained results show the superior performance of the proposed method in comparison to other state-of-the-art methods.

Wind energy is a cornerstone in the global transition toward carbon neutrality, with its long Operation and Maintenance (O&M) phase playing a significant role in affecting overall profitability, efficiency, safety, and sustainability. Digital twin (DT) technology has emerged as a key frontier in the wind energy sector due to its potential to construct comprehensive virtual representations of physical wind turbines and enable a digitalized loop to enhance performance across the entire life cycle. While research on DT technology in wind energy O&M is rapidly gaining visibility, there remains a substantial gap between current academic developments and practical implementation across methodological, technical, and operational aspects. In order to address this issue, this paper begins by briefly summarizing recent trends in wind turbine technology and identifying the most critical components that deserve DT technology. The existing DT capability levels and modeling approaches are then reviewed, and a prospective DT framework specifically tailored for the O&M of wind energy systems is proposed. The proposed framework encompasses a closed-loop process from the physical to the virtual domain and back again. The physical-to-virtual loop includes data acquisition, data management, virtual model construction, and adaptive operations. Conversely, the virtual-to-physical loop involves diagnostics and prognostics, maintenance decision-making, resource planning, and maintenance execution. Each stage is analyzed in terms of its enabling technologies and representative methodologies. By comparing the state of the art with the envisioned DT-enabled O&M paradigm, this paper identifies key research gaps and outlines promising directions for future investigation.

Burn-in testing of semiconductor devices is performed to ensure product quality by identifying and removing early-life failures. Given the cost and time required by burn-in testing, this work proposes a framework to predict the quality of a production batch of semiconductor devices before burn-in. Unlike traditional methods for quality prediction that rely solely on statistical data, this framework incorporates production data to improve prediction accuracy. The framework combines statistical methods for feature extraction (Piecewise Aggregate Approximation and Principal Component Analysis) and quality estimation (Clopper-Pearson Estimator) with a modified Probabilistic Support Vector Regression (PSVR) to predict early-life failures. The PSVR hyperparameters are set by a Bayesian Optimization (BO) technique. The framework is validated on a synthetic case study designed to emulate the BI process of semiconductor devices and, then, applied to real data collected during semiconductor production. Results from a synthetic case study and real-world semiconductor production data demonstrate the accuracy of the proposed method in predicting the quality of production batches. The quality predictions can, then, be used to inform efficient burn-in test planning in terms of the number of devices to undergo burn-in and the type of burn-in tests to perform.

We consider the problem of optimizing the drilling scheduling in Field Development Planning. The objective is to identify the well drilling sequence and well types that maximize the project Net Present Value (NPV) properly handling uncertainty on our knowledge of the reservoir geological parameters. This Sequential Decision Problem (SDP) is modelled as a Markov Decision Problem (MDP) and solved using Deep Reinforcement Learning. Specifically, we develop an approach based on Deep Q-Networks (DQN), where an online Neural Networks (NN) learns and selects the most suitable drilling action, and a target NN estimates the expected future NPV (Q-value). A soft-update mechanism is adopted to gradually blend the weights of the target NN with those of the online NN with the objective of improving training stability. The proposed method is tested on a synthetic case study that simulates a real drilling case. It is shown to outperform a state-of-the-art DQN implementation, achieving a larger NPV and solutions more robust to the uncertainty on the reservoir geological properties.

Deep Learning (DL) models processing images to recognize the health state of large infrastructure components can exhibit biases and rely on non-causal shortcuts. eXplainable Artificial Intelligence (XAI) can address these issues but manually analyzing explanations generated by XAI techniques is time-consuming and prone to errors. This work proposes a novel framework that combines post-hoc explanations with semi-supervised learning to automatically identify anomalous explanations that deviate from those of correctly classified images and may therefore indicate model abnormal behaviors. This significantly reduces the workload for maintenance decision-makers, who only need to manually reclassify images flagged as having anomalous explanations. The proposed framework is applied to drone-collected images of insulator shells for power grid infrastructure monitoring, considering two different Convolutional Neural Networks (CNNs), GradCAM explanations and Deep Semi-Supervised Anomaly Detection. The average classification accuracy on two faulty classes is improved by 8 % and maintenance operators are required to manually reclassify only 15 % of the images. We compare the proposed framework with a state-of-the-art approach based on the faithfulness metric: the experimental results obtained demonstrate that the proposed framework consistently achieves F1 scores larger than those of the faithfulness-based approach. Additionally, the proposed framework successfully identifies correct classifications that result from non-causal shortcuts, such as the presence of ID tags printed on insulator shells.

This study presents an analysis of hydrogen incidents in the HIAD 2.1 dataset using two methodological approaches: a multi-task learning (MTL) classifier for joint prediction of cause and consequence, and Retrieval Augmented Generation (RAG) applied both to find missing labels (RAG#1) and to classify the incidents (RAG#2). After applying quality gating (Quality Seal $\geq 4$), cleaning, and normalization, the dataset comprises 304 events. Multi-label targets are adopted, with cause encoded as [human, non-human] and consequence encoded as [injuries, fatalities]. Performances are evaluated on a fixed 90/10 split (273 train / 31 test). The impact of data completion is validated by comparing the performances of the MTL classifier for the accident causes with and without data completion, where the former $(F1=0.8049)$ outperforms the latter $(F1=0.7907)$. RAG is then compared with the MTL classifier (with and without data completion) on the task of classifying causes and consequences. Overall, results support a hybrid strategy: use RAG to find missing data and then develop an MTL classifier on the complete dataset.

Fault Tree Analysis (FTA) is a systematic deductive technique for identifying causal relationships that link component failures to undesired system-level events. It is used to build a logical model of the undesired event, enabling the evaluation of system reliability. Building fault trees typically requires considerable modeling effort and is subject to analyst bias. To address these challenges, this work proposes a method based on Large Language Models (LLMs) to support safety analysts in building FTs. The method is evaluated on an artificial case study of a liquid mixing subsystem of a chemical plant.

In new generation aircraft, traditional hydromechanical and electrohydraulic actuators of flight control systems are replaced by Electro-Mechanical Actuators (EMAs). Ensuring the functionality of an EMA requires to monitor the health state of its components and promptly detect anomalies. This work develops an anomaly detection method for power inverters, which are among the most critical components of EMAs. It is based on a signal reconstruction model trained to reproduce the values of the signal expected under normal conditions. The cumulative Z-Score of the residuals between reconstructed and measured signals is used as anomaly indicator. The signal reconstruction model combines a Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) cells. The CNN enables the extraction of features representative of the system health state from the multidimensional time-series of the measured signals of voltage, motor angular speed and motor position. The LSTM cells allows capturing the complex, non-linear temporal dynamics of the extracted features. The anomaly detection method is validated by considering sensor faults and the degradation of the inverter Metal-Oxide-Semiconductor FieldEffect Transistor (MOSFET).

This work proposes a systematic procedure for analyzing maintenance reports to support maintenance decision-making for a fleet of similar systems. The proposed procedure allows achieving three objectives: 1) grouping maintenance interventions, 2) identifying common characteristics in the maintenance interventions, and 3) recognizing occurrences of rare events of maintenance intervention. Specifically, the attention mechanism of Bidirectional Encoder Representation from Transformer (BERT) and the Density Based Spatial Clustering Applications with Noise (DBSCAN) methods are combined to group maintenance interventions according to their similarity of stated features. A taxonomy of the words used in the textual reports to state the maintenance interventions is developed to systematically identify common features of the clusters, such as the involved components, their working state, the occurred failures or malfunctions, the performed maintenance actions and the personnel that has performed the intervention. The proposed procedure is applied to a repository of reports of maintenance interventions performed on mechanical and electric components of traction systems of a fleet of trains. The obtained results show that it can effectively support decision-making on the maintenance of traction systems.