Mixed style network based: A novel rotating machinery fault diagnosis method through batch spectral penalization

Abstract

The unsupervised fault diagnosis of rotating machinery holds significant importance, but it still faces numerous complex challenges. For instance, traditional convolutional neural networks often overlook inter-channel relationships, resulting in poor generalization and requiring manual adjustment of architecture parameters for different tasks. Additionally, traditional domain adversarial transfer learning has insufficient research on feature discriminability, leading to less distinguishable features. To address these issues, this paper proposes a MixStyle network based on the SE attention mechanism. This method achieves dynamic weight allocation through the SE attention mechanism, which is simple in design and introduces few additional parameters. By employing the MixStyle method for probabilistic mixed-domain training, the diversity of the source domain is increased, thereby improving the model's generalization capability. Since the principal singular vector enhances feature transferability, this paper penalizes the largest singular value through Batch Spectral Penalization to enhance other feature vectors, improving feature discriminability and domain adversarial performance. Experimental results show that the proposed method demonstrates outstanding performance in the task of unsupervised fault diagnosis for rotating machinery.