Tianwei Wang

Industrial fault diagnosis frequently confronts the challenge that both the data and targets are not independent and identically distributed (Non-IID). Conventional fault diagnosis often requires tailoring the model to specific task scenarios, thereby suffering from high costs and limited generalization capabilities. An industrial large-scale diagnostic model (ILDM) is proposed, with lightweight learning and customized deployment for distributed multiple Non-IID fault diagnosis tasks. ILDM can acquire knowledge from large public datasets that completely irrelevant to downstream tasks, and has the advantages of low learning cost, robust task adaptability, and high quality deployment customization. Firstly, a large-scale foundation model is built utilizing time-series embedding and 1D-2D-1D variation, enhancing feature extraction capability while preserving the original temporal information. Secondly, through designed pre- and post-processors, the constructed foundation model is pre-trained on numerous public Non-IID datasets, ensuring that the model acquires generic classification capabilities. Finally, the pre-trained large-scale model is tailored for deployment at different low-computing-power terminals to conduct various process/system specific Non-IID diagnostic tasks, with lightweight customization achieved through fine-tuning and label smoothing based on terminal datasets. The performance of ILDM is evaluated across three representative Non-IID industrial diagnostic tasks in few-shot, and full-shot experiments, as well as in experiments with varying sample lengths. Additionally, the paper discusses the potential of larger-scale model and longer data transmission chain. Code is available at this repository: https://github.com/JMu-Jia/ILDM.

In fault diagnosis for high-speed trains (HSTs), data-driven methods are becoming increasingly popular. However, diagnosing unknown faults remains challenging due to limitations of existing GZSL methods, such as prediction confidence shifts from data imbalance, and domain shift between known and unknown classes, making it an open problem in engineering applications. This paper addresses this issue by proposing a stacked generalized zero-shot learning (Stacked-GZSL) framework which combines attribute-based GZSL strategies and generative modelbased GZSL strategies. The former leverages the correlations among data attributes and categories to tackle feature distribution differences across classes, while the latter addresses weight drift issue caused by the imbalance within known data and between known and unknown classes. Then, stacked model techniques are employed to resolve compatibility issues and further enhance accuracy. Experimental results confirm that the proposed framework can effectively address the challenge of unknown compound faults in train bogies under data imbalance. The framework achieves over 90% accuracy for known faults and 70% for unknown faults, outperforming single GZSL strategies and other classic ensemble models.

The intelligent fault diagnosis in engineering scenarios frequently struggle to achieve the desired results due to the presence of data imbalance. Meanwhile, since the data used may come from multiple sensors, and there are compound faults coupled by multiple faults, conventional imbalanced fault diagnosis methods are difficult to facilitate accurate fault diagnosis. To address this issue, from the perspective of data augmentation and feature processing, this paper proposes AttGAN-DPCNN, an imbalanced fault diagnosis scheme that leverages an attention-based generative adversarial network (AttGAN) and dynamic perceptual convolutional neural network (DPCNN). Firstly, the original GAN is enhanced by incorporating an attention-based multi-level information fusion structure and a class-conditional loss function, thereby facilitating the generation of high-quality multi-sensor signals while effectively balancing the data distribution. Secondly, in order to further improve the feature recognition ability of the model for multi-sensor complex signals, a lightweight and high-performance dynamic perceptual network is constructed to adaptively mine the deep features of signals via four-dimensional dynamic convolution. Finally, the performance of the method is validated on the bogie dataset, achieving accuracies of 97.1%, 98.0%, 97.1%, and 96% for the multi-class extreme imbalance and three single-class extreme imbalance cases, respectively. Meanwhile, its robustness is evaluated on the standard bearing dataset, which reaches an accuracy of 97.4% in the multi-class imbalance case and consistently surpassing 99% in all single-class imbalance cases. This study provides a reliable solution to the problem of extremely imbalanced data for complex signals in engineering scenarios.

The high-speed train bogie has complex structures, various conditions, and high maintenance demands. As a result, an extremely lightweight network (ELN) is proposed in this paper for HST fault diagnosis and may be employed in hand-held terminals. Firstly, a lightweight architecture based on blueprint separable convolution (BSConv) is designed as a feature extractor. Secondly, the mask mechanism is introduced to temporarily deactivate low-weight neurons in order to further diminish network size. Finally, the optimal network is fixed and deployed after learning and evaluation. Experiments prove that ELN has an extremely lightweight structure as well as high failure recognition accuracy.

Based on variational mode decomposition and dual model of convolutional recurrent neural network (VMD-DCRNN), a novel diagnosis method is proposed to discover the faults of insulated gate bipolar transistors (IGBTs), diodes, and series resonant circuit in a single-phase pulsewidth modulation ac–dc rectifier, an important part of traction power supply system of high-speed railway. By virtue of the combination of signal processing and deep learning schemes, the multiscale feature information for fault diagnosis is extracted by regarding the input current of the rectifier and the voltage of dc side as the original signal. More clearly, VMD is adopted to decompose the original signal to a series of intrinsic mode function components (IMFs). Then, the submodel of current is designed to detect the fault types of IGBTs and diodes. Meanwhile, the submodel of voltage is established to identify the fault types of series resonant circuit. These IMFs are fed into DCRNN consisting of the two submodels for training and testing. Finally, the features extracted from the two submodels are integrated and converted to label space to complete fault diagnosis. Experimental results show that the proposed VMD-DCRNN algorithm is superior to the single-channel models, reaching 96.27%.