A synergistic distillation-based continual learning framework for health assessment of rotating machinery using multi-sensor data

Abstract

Continual learning paradigm, which allows networks to learn from incremental samples without laborious retraining, has been introduced into health assessment of rotating machinery and has enjoyed considerable focus recently. However, there are still some challenges in continual learning that need to be addressed within multi-sensor inputs. (1) How to integrate information from multiple sensor sources during continual learning. (2) How to solidify knowledge from different sensor sources at new learning stage. To address the above challenges, a synergistic distillation continual learning framework is proposed using multi-sensor data in this article. In the framework, a joint attention mechanism is designed to facilitate cross-sensor interaction and integrate information from various sensor sources. Then, the synergistic distillation continual learning framework mitigates catastrophic forgetting through attention distillation at each new learning stage. Furthermore, a sensor-balanced gradient recalibration technique is developed and incorporated into the continual learning framework to facilitate concurrent learning across multiple sensor sources. Accelerated degradation tests are performed and the multi-sensor data were collected, based on which the proposed framework is verified and benchmarked against some existing advanced methods. Experimentally, the proposed method exhibits enhanced health assessment performance and reduced forgetting within multi-sensor data inputs in continual learning scenarios.