Attention guided multi-wavelet adversarial network for cross domain fault diagnosis

Abstract

Adversarial transfer learning is an effective method for diagnosing rotating machinery faults. However, it does not extract features completely and underutilizes unlabeled target data. Hence, an attention-guided multi-wavelet adversarial network (AMWAN) based on the hybrid metric strategy of a multi-wavelet convolutional feature extractor is proposed. First, hidden fault features in the input data are extracted via multi-wavelet convolution and an attention mechanism. Subsequently, the extracted features are input into the domain discriminator and label classifier. A gradient reversal layer is introduced to map and mix the source- and target- domain samples for feature extraction. Specifically, the k-nearest neighbor is adopted to construct a label propagation model in the source-domain feature layer to obtain pseudo-labels for the target-domain feature data. This extends the labeling information for the target-domain feature data. A hybrid metric based on the maximum mean discrepancy distance and feature cosine metric strategy is introduced in the feature layer for adversarial training, which can significantly enhance the domain adaptive effect and improve the cross-domain fault diagnosis performance of the AMWAN. Experimental validation based on two sets of planetary gearbox datasets indicate that the AMWAN exhibits outstanding fault diagnosis performance under both constant- and variable-speed and -load conditions, i.e., it achieves an average accuracy rate of up to 96.85 %, thus surpassing other methods.