A Novel Domain Incremental Learning Method for Bearing Fault Diagnosis Based on F&K

Existing fault diagnosis methods for rolling bearings are mostly applied in static domains and cannot adapt to continuous and dynamic industrial data in practice. In domain incremental scenarios, achieving continuous learning and avoiding catastrophic forgetting is a major challenge for fault diagnosis. In this article, a novel domain incremental learning method based on filtering columns and knowledge base (F&K-DIL) is proposed to achieve bearing fault diagnosis in dynamic domains. It consists of filtering columns, knowledge bases, and parameter update modules. In the filtering columns, two L1-norm are used to select representative samples, which can protect previously learned knowledge in each domain while learning new knowledge. The correctness of knowledge delivery is guaranteed. Sparsity parameters that can adjust the number of representative samples flexibly are introduced to improve storage efficiency. Representative samples in different domains are stored in the knowledge base to realize knowledge sharing. In the parameter update module, cross entropy and elastic weight consolidation loss function are used to realize constraints on important parameters in old domains. It can further avoid catastrophic forgetting. Through experiments on three rotating machinery datasets, F&K-DIL is used for fault diagnosis in domain incremental scenarios with an accuracy of up to 99.6% and 37% improvement in memory rate.