A new collaborative lifelong learning framework based on channel distillation and pseudo-feature replay for class-incremental bearing fault diagnosis

Abstract

Deep learning has achieved notable advancements in mechanical fault diagnosis and health monitoring. However, in real-world industrial settings, machinery may develop new fault types that models fail to recognize, resulting in diminished performance and compromising the reliability and safety of mechanical systems. Lifelong learning mitigates this challenge by enabling models to acquire new knowledge while preserving previous knowledge, effectively preventing catastrophic forgetting. Replay-based lifelong learning has gained recognition for its effectiveness. However, replay-based methods utilize stored samples from past tasks when learning new tasks, leading to an imbalance in training data distribution, and causing the model to favor fault classes with larger sample sizes. To address this issue, a new collaborative lifelong learning method integrating channel distillation and pseudo-feature replay for class-incremental bearing fault diagnosis is proposed. It incorporates two key technologies: channel distillation and pseudo-feature replay. Channel distillation transfers channel attention information from the previous model to the current model, enhancing the retention of prior knowledge. Pseudo-feature replay alleviates the imbalance between current task data and exemplars by generating synthetic features based on historical class prototypes and variances. Concurrently, it employs a momentum-update mechanism to dynamically adapt prototypes, thereby mitigating the prototype shift issue encountered in lifelong learning scenarios. In the comparative experiments, the proposed method achieved improvements in diagnostic accuracy of 8.96% and 6.83%, respectively, in the final phase compared with the optimal comparison methods, demonstrating its effectiveness in alleviating imbalanced training data distribution and mitigating catastrophic forgetting.