Adaptive frequency attention-based interpretable Transformer network for few-shot fault diagnosis of rolling bearings

Abstract

In recent years, deep learning-based approaches have demonstrated superior performance in few-shot fault diagnosis. Nevertheless, many of these methods lack explicit interpretability, making it difficult to intuitively understand their diagnostic logic. To tackle this issue, an interpretable deep learning model called the adaptive frequency attention-based interpretable Transformer network is proposed for few-shot fault diagnosis of rolling bearings. From a frequency interpretability perspective, the standard Transformer network architecture has been innovatively improved. First, an adaptive frequency attention mechanism is developed that quantifies the importance of various frequency components during the diagnostic process, adaptively identifying and emphasizing key frequency components closely associated with fault modes. This boosts both diagnostic performance and model interpretability. Second, to enhance the diversity of fault features under limited sample conditions, a multiscale convolutional architecture is developed to replace the linear projection layer in input embedding. This architecture employs parallel multiscale convolution kernels to extract both local and global fault features, enabling a comprehensive capture of fault information and further supporting the interpretability of the diagnostic model. Finally, Experiments on interpretable few-shot fault diagnosis are carried out on three rolling bearing datasets, and the diagnostic results further validate the effectiveness and interpretability of the proposed method.