9 January 2023 in IEEE Transactions on Instrumentation and Measurement
For imbalanced bearing fault diagnosis, GAN (Generative Adversarial Networks) is a common data augmentation approach. Nevertheless, current GAN-based methods cannot update generator from time-frequency domain simultaneously, downgrading the authenticity of signal time-frequency character. In this paper, FTGAN (Fourier-like Transform GAN), a novel GAN method, is proposed by introducing Fourier-like transformer (FLT) based on autoencoder (AE) to improve synthetic data quality. FLT approximates discrete Fourier transform (DFT) by the neural network, learning a universal map from time to frequency domain during training. FTGAN with FLT can decouple input into time-frequency domain, fitting the distribution of time and frequency of data simultaneously. Multi-domain distribution is manipulated in FTGAN without introducing additional signal transformation means. Furthermore, TRTS and TSTR analysis of one-dimensional data is introduced to evaluate data quality. Real and synthetic data are applied as training or test sets of diagnostic classifiers by turns so that data quality can be analysed through diagnosis results. Experiment results show that the proposed method can generate bearing fault signals closer to real data in the time and frequency domain, effectively improving the performance under an imbalanced dataset.
13 February 2024 in IEEE Transactions on Neural Networks and Learning Systems
Transfer learning (TL) and generative adversarial networks (GANs) have been widely applied to intelligent fault diagnosis under imbalanced data and different working conditions. However, the existing data synthesis methods focus on the overall distribution alignment between the generated data and real data, and ignore the fault-sensitive features in the time domain, which results in losing convincing temporal information for the generated signal. For this reason, a novel gated recurrent generative TL network (GRGTLN) is proposed. First, a smooth conditional matrix-based gated recurrent generator is proposed to extend the imbalanced dataset. It can adaptively increase the attention of fault-sensitive features in the generated sequence. Wasserstein distance (WD) is introduced to enhance the construction of mapping relationships to promote data generation ability and transfer performance of the fault diagnosis model. Then, an iterative “generation-transfer” co-training strategy is developed for continuous parallel training of the model and the parameter optimization. Finally, comprehensive case studies demonstrate that GRGTLN can generate high-quality data and achieve satisfactory cross-domain diagnosis accuracy.
31 July 2024 in IEEE Transactions on Instrumentation and Measurement
Domain-adversarial learning based fault diagnosis methods have shown prominent application prospects in the field of cross domain fault diagnosis. However, the existing domain-adversarial methods can’t simultaneously consider domain-level discrepancy, inter-class and intra-class correlation of target domain samples, which result in poor feature representation ability of fault diagnosis model for ambiguous targets. For this reason, a fault diagnosis method is proposed by implicit discriminator domain adversarial residual network (DDARN). Firstly, a unified domain adversarial framework based on the Nuclear-norm Wasserstein discrepancy (NWD) is constructed to achieve domain-level feature alignment and adjust the inter class-intra class correlation for ensuring the prediction certainty. Additionally, improved deep residual network (IDRN) integrating the efficient activation function is designed to facilitate the propagation of information and enhance nonlinear representation ability of the network. Finally, case studies on public datasets and self-made KUST-SY bearing datasets indicate DDARN can achieve slightly better diagnostic performance over state-of-the-art methods.
24 December 2025 in IEEE Transactions on Instrumentation and Measurement
Cross-domain few-shot recognition aims to establish accurate prototypes and design effective domain alignment strategies, which still presents great challenges in fault diagnosis field. To tackle these issues, a novel graph convolutional domain adversarial network (GCDAN) is proposed for cross-domain few-shot fault diagnosis. First, an attentive pooling-based feature fine-tuning strategy is designed to complete feature pre-processing, and a kernel ridge regression is developed to reproject the prototypical feature. Meanwhile, the self-attention mechanism based prototypical feature reprojection alignment module is proposed to construct accurate prototypes. Besides the prototypical feature alignment, the graph convolution network with structure-aware alignment is further designed to build structural feature relationships in the instance graphs, which can effectively achieve inter-domain structural distribution alignment under class imbalance. Extensive experimental results on 2 rolling bearing fault diagnosis cases demonstrate that GCDAN exhibits obvious advantages against state-of-the-art cross-domain few-shot recognition methods.
11 December 2025 in IEEE/ASME Transactions on Mechatronics
In recent years, the superiority of graph neural networks (GNNs) in the aspect of relational modeling has been confirmed in rotating machinery fault diagnosis field. However, most GNNs based intelligent diagnostic methods struggle to construct high-precision and low-complexity graph and perform poorly without efficient graph update strategies. To solve these problems, a semisupervised high-order graph diffusion network (SHGDN) for cross domain fault diagnosis is proposed in this study. First, the discriminative subsequences called Shapelets are sufficiently extracted by diffusion model. Thereafter, fast dynamic time warping and sparse strategy are leveraged to construct dynamic graph with semantic perception ability. Second, a high-order graph diffusion network based on approximate personalized propagation of neural predictions is put forward, which can implement dual-stream aggregation modeling for local first-order neighbors and global high-order neighbors. Then, the pseudolabel guided dynamic graph optimization and Leiden algorithm-based subgraph exchange enhancement is developed to reduce the cumulative propagation of pseudolabel errors. Finally, the explicit theoretical connection between graph construction and graph update is established, which can break the limitation of parameter isolation and provide a unified framework for model optimization. The proposed method is demonstrated on datasets from three types, and the experimental results indicate the SHGDN achieves the average accuracy of 97.29$\%$, surpassing other state-of-the-art methods.