A reliable degradation indicator (DI) and an effective degradation model are essential prerequisites for accurate remaining useful life (RUL) prediction of rolling bearings. However, numerous existing DI construction methods and degradation models overlook the fatigue damage mechanism—a core physical mechanism governing rolling bearing failure—which hinders the effective extraction of key degradation information from raw monitoring signals. To address this gap, this paper proposes a physics-driven RUL prediction method for rolling bearings that incorporates Paris' fatigue damage mechanism. Specifically, a multi-scale stackable network model constrained by Paris' fatigue damage mechanism-based soft labels is constructed to resolve the issues of poor interpretability and lack of physical constraints in conventionally built DIs. Furthermore, a rolling bearing degradation model is established by integrating the Paris' fatigue damage mechanism with stochastic processes, which mitigates the problem of weak correlation between the DI and the degradation model. The proposed method is validated through two sets of experiments and compared with other state-of-the-art methods. The comparative results demonstrate that the DI constructed under the constraint of Paris' fatigue damage mechanism exhibits excellent comprehensive performance, and the physics-driven prediction method achieves high accuracy in rolling bearing RUL estimation.

Accurate identification of compound mechanical faults through vibration analysis remains a critical challenge in industrial condition monitoring, primarily due to the nonlinear interaction of multiple failure modes and combinatorial explosion of potential fault combinations. While contemporary data-driven diagnostic methods demonstrate feature extraction capabilities when abundant labeled compound fault data exists, such idealized data conditions rarely occur in practical engineering applications. This paper presents a novel mechanism-guided decomposition diffusion network (McDDN) for resource-efficient compound fault diagnosis requiring only single-fault labeled samples for training. The proposed framework incorporates a physics-informed decomposition UNet (McD-UNet) within a diffusion-based learning architecture to disentangle overlapping fault signatures through mechanism-constrained signal separation. Feature mode decomposition (FMD) principles are mathematically encoded as regularization terms in the network's optimization objective, enabling data-driven learning of decomposition patterns while preserving physical interpretability. The diagnostic pipeline performs hierarchical fault identification through mechanism-guided signal decomposition into constituent single-fault components by single-fault classification. Experimental validation on the PU bearing dataset and BJTU-RAO high-speed train bogie dataset demonstrates superior performance, achieving 89.3% diagnostic accuracy for simultaneous bearing-gear-motor faults in multi-component rotating systems. Comparative analysis reveals 12-15% accuracy improvements over conventional model-based and pure data-driven benchmarks, validating the hybrid approach's effectiveness in knowledge-scarce compound fault scenarios.

Domain adaptation (DA) enhances the scalability of remaining useful life (RUL) prediction technologies, providing a reliable foundation for maintenance decisions across diverse equipment within Industrial Internet of Things (IIoT). Traditional DA approaches typically necessitate simultaneous access to both source and target domain data, which often conflicts with data privacy concerns prevalent in IIoT. Furthermore, the substantial storage resources required for source domain data hinder the implementation of efficient DA on resource-constrained edge devices. To address these challenges, we propose a source-free DA (SFDA) framework leverages uncertainty estimation pseudo-labels to conduct cross-domain RUL prediction in the absence of source domain data. Initially, we develop a self-supervised knowledge distillation framework that adapts efficiently to domain shift based on pseudo-labeling. Building on this, we propose an uncertainty estimation pseudo-label guided loss reweighting strategy to mitigate the impact of pseudo-label noise. This strategy prioritizes highly reliable pseudo-labels by assessing discrepancies in feature distributions from different augmented views of the input samples. Additionally, a reconstruction-based training method is designed to align the feature distributions. Extensive experimental evaluations demonstrate that our proposed method outperforms current state-of-the-art techniques, ensuring reliable RUL predictions in scenarios characterized by domain shift and the absence of source domain data.

Generative Adversarial Networks (GANs) for generating realistic data, have substantially improved fault diagnosis algorithms in various Internet of Things (IoT) systems. However, challenges such as training instability and dynamical inaccuracy limit their utility in high-speed rail (HSR) bogie fault diagnosis. To address these challenges, we introduce the Pulse Voltage-Guided Conditional Diffusion Model (VGCDM). Unlike traditional implicit GANs, VGCDM adopts a sequential U-Net architecture, facilitating multi-steps denoising diffusion for generation, which bolsters training stability and mitigate convergence issues. VGCDM also incorporates control pulse voltage by cross-attention mechanism to ensure the alignment of vibration with voltage signals, enhancing the Conditional Diffusion Model’s progressive controlablity. Consequently, solely straightforward sampling of control voltages, ensuring the efficient transformation from Gaussian Noise to vibration signals. This adaptability remains robust even in scenarios with time-varying speeds. To validate the effectiveness, we conducted two case studies using SQ dataset and high-simulation HSR bogie dataset. The results of our experiments unequivocally confirm that VGCDM outperforms other generative models, achieving the best RSME, PSNR, and FSCS, showing its superiority in conditional HSR bogie vibration signal generation. For access, our code is available athttps://github.com/xuanliu2000/VGCDM.

With the improvement of intelligence and integration, automatic supervision of large-scale systems is a current challenge in guaranteeing the high-reliability of edge devices. Hence, fast & accurate anomaly detection (AD) has become an urgent need via the edge computing of the industrial Internet of Things (IIoT). For this purpose, this paper creatively proposes a dual agents based on two-phase adversarial training strategy (2P-DAs) to perform rapid, stable and unsupervised AD for large-scale IIoT-edge devices. It integrates the superiorities of deep autoencoder (AEs) and generative adversarial network (GANs), utilizing normal multivariate time-series as inputs, 1-Encoder vs. 2-Decoders architecture as backbone, and two-phase unsupervised adversarial learning to make it isolate anomalies while providing efficient training. On the one hand, this allows the inherent limitations of AEs to be overcome by training a model capable for recognizing non-anomalies and thus performing a good reconstruction. On the other hand, dual structures allow for stability in adversarial training, thereby solving the issues of collapse and non-convergence encountered in GANs. Two practical industrial data, as cloud & edge data, are used to verify the robustness, inference speed and high detection performance of 2P-DAs in IIoT-edge AD, which demonstrates an impressive performance under multiple evaluation indexes.

Developing a trustworthy framework for intelligent fault diagnosis (IFD) of machines has two major challenges: confidently recognizing known faults and precisely detecting novel faults. However, current IFD frameworks are typically based on the closed-world assumption and tackle the two issues independently, making it hard to meet the expectations of reliable diagnosis in complicated working situations. In this paper, aProbabilistic framework forMechanical fault diagnosis (ProMo) is proposed to integrate misidentification and out-of-distribution (OOD) detection for high-speed train bogie. ProMo uses variational parameters to capture uncertainties at both the feature and prediction levels, in this way the misclassified and OOD samples with high predictive uncertainty are differentiated. To begin, a Bayesian deep neural network with a hierarchical classifier is built, offering diverse predictions and enabling ensemble uncertainty estimate. Then, a probabilistic null space analysis technique for post hoc OOD detection is presented, in which the magnitude of feature projections indicates the OOD-ness. Additionally, a novel metric assessing classification correctness and prediction reliability simultaneously is proposed. Extensive experiments revealed that ProMo outperforms state-of-the-art methods and achieves reliable fault diagnosis, even in the presence of covariate shift in monitoring data.

Deep graph neural networks (GNNs) have demonstrated their exceptional expressive capability in detecting fault features from multisensor signals for fault diagnosis. However, the excessive depth of these models often hinders their deployment on Internet of Things (IoT) systems. In order to facilitate the implementation of graph-based models on IoT devices for fault diagnosis of HSR bogie, this paper proposes a novel compression technique called GraKD. GraKD distills the latent knowledge from the teacher model to a lightweight student model in an adversarial manner. The discriminator of GraKD comprises a representation identifier and a logit identifier. The former effectively distinguishes between the node representations of the student and the teacher by discerning the local affinity of connected node patch-patch pairs and the global affinity represented by the patch-global pairs. The latter employs a residual multi-layer perceptron block to differentiate between the logits of the student and the teacher. The effectiveness of GraKD is validated using data collected from test rigs of bogies. A series of experiments substantiate the broad applicability of GraKD in compressing various GNN-based models.

Class imbalance issue has been a major problem in mechanical fault detection, which exists when the number of instances presenting in a class is significantly fewer than that in another class. This article focuses on the problem of zero-shot fault detection of rolling bearing, which is the extreme case of class imbalance. Aiming at this problem, a two-stage zero-shot fault recognition method is proposed. First, inspired by the conditional generative adversarial network, a novel feature generating network which is composed of a feature extractor, a discriminator, and a generator is designed to capture the potential distribution of normal samples. Then, the generator will generate abundant pseudofault features by adding an additional sequence to the condition. Second, an improved deep neural network is trained with these synthetic pseudofault features as the classifier. Specially, a condition index is designed to represent different fault classes so that it can recognize the unseen fault samples. Finally, the effectiveness of the proposed method is verified by three datasets and a comparison method is also given to show the superiority. Results show that the feature generation network can effectively detect the typical faults even though the fault data are unavailable during training, which is practical for industrial application.

This paper discusses contact discontinuity of electromechanical switches, and presents a model that considers the effect of vibration-induced inertial force on operational reliability. Using this model, the operational reliability of a switch with a specific contact assembly can be assessed for given vibration conditions. Under the vibration conditions, a minimum contact spring force is necessary for proper functioning of the switch, while the magnitude of contact uncertainty determines the need of an additional force to ensure operational reliability. With reduced contact uncertainty, the operational reliability of switches approaches either 0 or 1, solely depending upon the design & operational conditions. In this model, the parameters of c', b, & /spl sigma///spl mu/ need to be determined either empirically or experimentally before the model can be used for reliability assessment. To theoretically determine those parameters, Hertz contact with randomly distributed surface asperities needs to be considered. Finally, although this model starts from a log-normal distribution of electrical contact, the approach can be applied to other distributions, such as the inverse Gaussian and Weibull distributions.