Bearing Fault Diagnosis Based on Lifelong Learning Under Cross Operating Conditions

Rolling bearing, a widely used core component in industry, will bring a serious threat to the safety of the machine and staff when it fails. At present, the time-varying operating conditions and catastrophic forgetting have brought great challenges to bearing fault diagnosis. One of the reasons is that good performance can only be maintained if the model is kept under the same conditions as the offline training phase. If the model is directly trained by using the data acquired from new operating condition, the model will suffer from catastrophic forgetting, resulting in poor performance of previous operating condition. In order to solve the above problems, a bearing fault diagnosis method based on lifelong learning is proposed in this paper, which is implemented based on Residual Network with Convolutional Block Attention Module(Res-CBAM) and Elastic Weight Consolidation (EWC). As the basic fault diagnosis model, Res-CBAM can adaptively extract fault features. The introduction of elastic weight consolidation can make the model retain the feature extraction ability of the past condition when learning the fault features of the new condition, so as to solve the catastrophic forgetting problem. The experimental results show that the proposed method has good performance in fault diagnosis under cross conditions.

Keywords: Bearing fault diagnosis, Lifelong learning, Elastic weight consolidation.