Bearing Fault Diagnosis Based on Auto-Encoder Combined with CNN

Abstract

To address the problem that traditional bearing fault diagnosis methods rely on professional knowledge and are tedious, this paper proposes an end-to-end CNN-based bearing fault diagnosis model to achieve automatic fault recognition. In addition, considering the problem that noise exists in the actual working conditions, a bearing fault diagnosis model based on Auto-encoder (AE) combined with CNN is proposed (AE-CNN). The noisy signal is coded and decoded by the designed AE, and the de-noised result is used as the input of the designed CNN to achieve the bearing fault diagnosis under noisy conditions. Experiments on CWRU have proved the effectiveness of the designed CNN and AE-CNN. The designed CNN achieves 99.83% fault diagnosis accuracy under noise-free condition. The AE-CNN achieves 97.14% fault diagnosis accuracy under –4db signal-to-noise ratio (SNR) noise condition, which is 2.31% higher than the CNN with the same noise, and compared with the results of other advanced methods, it has achieved competitive results.