Class-incremental continual learning model for plunger pump faults based on weight space meta-representation

Abstract

Deep learning (DL)-based fault diagnosis models often require full access authority of all fault modes, which are unrealistic. In industrial scenarios, the type of fault database samples is gradually increasing, requiring that the diagnostic model is updated repeatedly. However, DL models inevitably suffer from catastrophic forgetting during continual learning. To accommodate dynamic changes in fault types, we propose a continual learning model based on weight space meta-representation (WSMR) applied to class-incremental fault diagnosis of switch machine plunger pumps. The proposed Modified WaveletKernelNet (MWKN) demonstrates that improving base model structure can reduce forgetting of old knowledge. WSMR maintains the diagnostic performance of the model for both new and old faults in a cost-friendly manner by inferring about and preserving the meta-representation distribution of the task-specific base model. The results show that WSMR with MWKN effectively alleviates catastrophic forgetting and outperform other methods.