A novel convolutional sparse filtering adversarial network for cross domain fault diagnosis

Abstract Over recent years, adversarial transfer learning (TL) has gained extensive application in fault diagnosis of rotating machinery. However, traditional adversarial TL is relatively generalized in feature extraction, which makes it difficult to capture the features that are crucial for fault diagnosis. This will cause the model to be inaccurate in identifying fault categories, especially when faced with intricate fault features in the signal. Therefore, a novel convolutional sparse filtering (SF) adversarial network is proposed in this paper. Initially, convolutional SF combined with an attention mechanism is utilized to extract sparse features from the input data. Subsequently, these features are fed into both the domain discriminator and the label classifier. To achieve domain invariance, a gradient reversal layer is incorporated, which maps and blends samples from the source and target domains. Additionally, the joint maximum mean difference is incorporated into the adversarial training process to significantly improve domain adaptation (DA). The combination of convolutional SF and adversarial network can effectively perform feature extraction and DA in the scene of cross-domain TL, thus enhancing the generalization of the model in the target domain. Experimental results on two planetary gearbox datasets show that the method achieves excellent fault diagnosis performance under varying conditions, with an average accuracy of 96.85%.