A multibranch residual network for fault-diagnosis of bearings

Time-frequency domain analysis methods are used to diagnose faults in bearings by extracting the features of fault signals. Given that a fault signal is also a form of audio signal, we extracted the characteristics of the mel spectrum from the original signal and applied it to a convolution neural network proposed in this paper. Focusing on the residual structure in the residual neural network (ResNet), we solved the gradient disappearance problem and accelerated the training of the model. The importance of each feature channel could be estimated adaptively using the squeeze-and-excitation network (SENet) considering the relationships between the channels. We examined the feature map of each layer using a multibranch residual network (MB-ResNet) to characterize the bearing fault signal. We used the multibranch residual structure to reduce the sense field of each residual and added a parallel local sensing module to train the model to recognize the weight of each input feature to either increase or reduce the influence of local features. Our experimental results show that the MB-ResNet is very good at extracting features, is robust, and capable of generalization.

Automated summary

The paper proposes a fault diagnosis method using time-frequency domain analysis, which extracts features from fault signals and applies them to a convolution neural network. Experimental results show that the proposed method can effectively extract fault features with strong robustness and generalization ability.