Abstract
The advancement of artificial intelligence has enabled significant progress in data-driven fault diagnosis for rotating machinery. However, in real-world industrial environments, such machinery often operates under time-varying and non-stationary conditions, leading to continuous changes in data distributions. These variations pose a major challenge to conventional deep learning models, which are prone to catas-trophic forgetting when sequentially learning new diagnostic tasks. To address this, this paper proposes a novel Multi-head Self-attention Incremental Network (MSIN) for fault diagnosis across non-stationary operating conditions. MSIN employs a stack of multi-head self-attention encoder layers to effectively extract fault-relevant features from frequency-domain signals. To alleviate catastrophic forgetting, the model integrates a selective exemplar replay strategy and introduces a joint cross-head knowledge distillation loss, enabling effective knowledge retention across tasks. The proposed method is evaluated on two benchmark datasets under both stationary and non-stationary conditions. Experimental results demonstrate that MSIN achieves superior diagnostic accuracy and knowledge retention compared to existing Incremental Learning methods, offering a promising solution for continuous fault diagnosis in dynamic industrial settings. The source code of the MSIN model is available at: https://github.com/bankuaimianbao/MSIN-Non-stationary-Code.
Funding Information
  • National Natural Science Foundation of China (52505127)
  • Fundamental Research Funds for the Central Universities, Northwestern Polytechnical University (G2023KY05102)

This publication has 41 references indexed in Scilit: