An event-based machinery fault diagnosis method via incremental retention and condition-aware alignment

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Abstract An event camera,as an emerging dynamic visual sensor,provides a new test method for non-contact fault diagnosis in complex industrial environments.However,most existing studies follow the offline static assumption.Aiming at the continual learning challenges brought by the constant change of operating conditions and dynamic evolution of fault categories in industrial online monitoring,this paper proposes a continual learning framework based on event-based dynamic vision.Specifically,a Condition-aware MixStyle Encoder module is designed to enhance the robustness of event representations to condition style changes.In addition,an asymmetric class conditional manifold discrepancy strategy for shared classes is proposed to mitigate the conditional distribution drift of shared classes between old and new conditions.Finally,an incremental retention learning strategy is constructed to mitigate catastrophic forgetting by combining sample playback and history logits persistence constraints under small buffer constraints.A dynamic visualization dataset of planetary gearboxes is constructed based on the event camera,and experiments are carried out under the incremental scenarios of continuous operating conditions.The experimental results show that the proposed method outperforms the compared methods in terms of average diagnostic accuracy and stage stability.