Multi-scale dynamic adaptive residual network for fault diagnosis

In industrial systems, the vibration signals of rolling bearings are influenced by changing operating conditions and strong environmental noise, therefore they are often characterized by high complexity. The multi-scale deep learning method can achieve bearing fault diagnosis under complex operating conditions, however, the importance of dynamic feature selection is neglected. To solve this problem, we propose a multi-scale dynamic adaptive residual network (MSDARN) fault diagnosis method. In the proposed method, we combine multi-scale learning and attention mechanism to construct a multi-scale dynamic adaptive convolutional layer (MSDAC). To learn vibration signal features, MSDAC can dynamically adjust the weights of different scale convolutional layers. In addition, we introduce a nonlinear function to adaptively determine the scaling rate parameter in MSDAC. Finally, in order to improve the feature learning ability of the proposed method, we use MSDAC and residual connections to construct residual blocks, and use multiple such residual blocks to construct MSDARN. The effectiveness of the proposed method is verified by noise, variable load and mixed fault experiments, and the proposed method has higher fault classification accuracy than other three deep learning methods.

Automated summary

In industrial systems, the vibration signals of rolling bearings can be complex due to changing operating conditions and environmental noise. This study proposes a multi-scale dynamic adaptive residual network (MSDARN) fault diagnosis method that combines multi-scale learning and attention mechanism to dynamically adjust the weights of different scale convolutional layers. The effectiveness of the proposed method is verified through experiments, showing higher fault classification accuracy compared to other deep learning methods.