Fault diagnosis using transfer learning with dynamic multiscale representation

A critical problem for fault diagnosis is caused by the feature shift under different working conditions, which significantly degenerates the diagnosis accuracy in practice. Aiming to solve this problem, this paper proposes a novel Transfser Learning (TL) framework with Dynamic Multiscale Representation (DMR) for fault diagnosis. This model draws the inspiration from the shared learning and transfer learning, processing information captured and exploited by multiscale signal factors. In particular, a novel multi-path merging network is proposed to generate dynamic weights for fusing multiscale factors. To drive this generation, and to control the extent of the shared fusion, the Multi-gate Mixture-of-Experts (MMoE) is introduced to model the tradeoff between scale-specific representation and inter-scale correlation. A transfer learning backend is also introduced to align cross-domain features, which enables proposed method to diagnose faults across distinct working conditions. Experiments evaluate the fault-diagnosis performance. Our primary, ablation and interpretation evaluations comprehensively indicate the robustness and flexibility of the proposed method to diverse fault diagnosis applications. Especially, the proposed method achieves 4.71% and 3.86% improved to the second best one (MSSLN) on the PHM2009 and MCP datasets, respectively.