Numerous unsupervised domain adaptation methods for bearing fault diagnosis rely on extracting domain-invariant features to mitigate the impact of domain shift interference. However, the lack of evaluation criteria results in limited interpretability of domain-invariant features. Additionally, current pseudo-labels prediction methods heavily rely on label information or computational resources, and the traditional Softmax function fails to capture valuable information. To address these problems, this paper proposes an unsupervised domain adaptation method based on domain-invariant features evaluation and knowledge distillation for bearing fault diagnosis. Firstly, mutual information and soft attention mechanism are integrated into the extraction of multivariate features to access the quality of domain-invariant features and enhance interpretability. Then, the concept of knowledge distillation is introduced to predict pseudo-labels in the target domain without relying on label information or computational resources. Furthermore, an asynchronous feature metric adaptive strategy is developed to adjust the feature alignment metric, considering the maturity and precision of pseudo-labels. The effectiveness and superiority of the proposed method are demonstrated through comparative experiments and ablation studies on two bearing datasets.
10 October 2022 in IEEE Transactions on Instrumentation and Measurement
Fault diagnosis of rotating machinery under fixed operating conditions has been developed considerably. However, the designed model suffers from a performance loss due to the internal covariance shift caused by the changes in operating conditions. How to mine the essential characteristics of faults by multiple related domains and then generalize well to unseen operating conditions is of great significance for the fault diagnosis of rotating machinery. To this end, we present a generic representation gradient muting paradigm (RGMP) to enhance the generalization ability of the diagnostic model. We first develop a deep convolutional network with skip connection and batch normalization. Then, RGMP guides the network to purposely discard a small number of dominant features activated by the training samples and forces the network to learn the remaining label-related representations. RGMP facilitates the network to activate more broad perceptual fields for out-of-domain samples without prior knowledge of the unseen conditions. Finally, the fine-tuned representations are fed into the classification layer, and the network parameters are optimized by backpropagation. Experiments on gear and high-speed train bearing verify the effectiveness of the proposed model under unseen operating conditions.
10 April 2024 in IEEE Transactions on Industrial Informatics
Fault diagnosis under specific working conditions has achieved remarkable success. However, due to variations in working conditions, the assumption that training and test samples are independent and identically distributed is often violated, which makes the diagnostic model brittle under unseen working conditions. To this end, a generic generalization strategy, namely, regularized domain adaptive weight optimization strategy (RDAWOs), is devised for fault diagnosis of rotating machinery. We first design the architecture of a 1-D convolutional neural network. Then, the hyperparameter regularization term and an adaptive pooling layer are designed to control the complexity and improve the adaptability of the overparameterized deep model, respectively. Finally, domain adaptive weight optimization is established to identify the working condition abundant in spurious label-related information and to mine the robust fault knowledge under various working conditions. Obtained results indicate the strong generalization ability for out-of-distribution samples, as well as relatively high diagnostic accuracy of the RDAWOs-based deep model under unseen working conditions.