A Novel Deep Transfer Adversarial Dictionary Learning Strategy for Bearing Cross-Domain Fault Diagnosis

Dictionary learning (DL) has gradually demonstrated its unique advantages in many fields with its powerful feature extraction and data representation capabilities. However, it still has some problems. For example, DL is susceptible to the time-shift properties of vibration signals, which is very common in industry equipment. Secondly, due to the lack of effective transfer learning strategies, the performance of DL in the field of cross-domain diagnosis is very limited. To overcome these drawbacks, a novel deep transfer adversarial dictionary learning (DTADL) strategy is proposed in this paper. First, a sample convolution module is constructed to extract shift-invariant features, and then a new deep dictionary module is designed, in which the iterative soft thresholding and the gradient descent method are used to train the dictionary for extracting the class-specific representations from the convolution module further. Besides, an adversarial domain predictor module, which includes a gradient flipping layer is designed for predicting samples from source or target domains and obtaining domain adversarial losses, which can be used to encourage domain confusion in the sparse representation space. The effectiveness of DTADL is verified on two bearing datasets, which achieved recognition rates of 99.60% and 97.10% in the experiment of transferring diagnosis between two datasets, respectively. In addition, DTADL is also compared with other traditional transfer learning methods, which also demonstrates the superiority of the proposed method.