Joint adaptive transfer learning network for cross-domain fault diagnosis based on multi-layer feature fusion
Traditional intelligent fault diagnosis models are required to be trained and tested under an identical probability distribution. However, the shift in data distributions is inevitable due to changes in environmental and operational conditions, which results in diagnostic performance degradation. Currently, transfer learning has been successfully applied to learn a discriminative diagnosis model in the presence of a shift. But conventional transfer learning approaches encounter obstacles without adequately considering the feature interactivity and transferable ability at different layers. In this study, a joint adaptive transfer learning framework based on multi-layer feature fusion for reliable cross-domain diagnosis is presented to address these issues. Firstly, the multilinear map is employed to implement a novel multi-layer feature fusion. This fusion is key to realizing a substantial improvement of feature representation capability and effectively embedding joint distribution of multi-layer features. Furthermore, a novel joint adaptive transfer learning (JATL) framework is devised to facilitate reliable cross-domain adaption by making utmost use of cross-domain-invariant features with a small amount of data. Experiments with different transfer scenarios on two benchmark datasets have been conducted, and experimental results demonstrate the superiority of the proposed approach.