Multi-source information fusion meta-learning network with convolutional block attention module for bearing fault diagnosis under limited dataset
Shanshan Song, Shuqing Zhang, Wei Dong, Gaochen Li, Chengyu Pan
Abstract
Applications in industrial production have indicated that the challenges of sparse fault samples and singular monitoring data will diminish the performance of deep learning-based diagnostic models to varying degrees. To alleviate the above issues, a multi-source information fusion meta-learning network with convolutional block attention module (CBAM) is proposed in this study for bearing fault diagnosis under limited dataset. This method can fully extract and exploit the complementary and enriched fault-related features in the multi-source monitoring data through the designed multi-branch fusion structure and incorporate metric-based meta-learning to enhance the fault diagnosis performance of the model under limited data samples. Furthermore, the introduction of CBAM can further assist the model to trade-off and focus on more discriminative information in both spatial and channel dimensions. Extensive experiments conducted on two bearing datasets that cover multi-source monitoring data fully demonstrate the validity and superiority of the proposed method.