Bearing fault diagnosis method based on improved deep residual Siamese neural network

Abstract

Aiming to address the problem that faults in rolling bearings make effective fault diagnosis difficult under small-sample and varying working conditions, this paper proposes a new fault diagnosis method for rolling bearings that monitors their vibration signals and is based on an improved deep residual Siamese neural network, called a WDRCNN. Firstly, the Siamese neural network is applied to extract features with shared weights to achieve an expansion in the number of fault samples. Then, multiple residual blocks are used to extract deeper feature information and effectively alleviate the problem of overfitting. In addition, the attention mechanism is employed to assign weights to the feature information to reduce the interference of redundant features. Finally, the Euclidean distance between the sample pairs is calculated to determine the similarity of the sample pairs and to classify bearing faults for end-to-end bearing fault diagnosis. The experimental results demonstrate that the WDRCNN achieves an average accuracy of 96.31% under different operating conditions. Even when only 90 training samples are available, the WDRCNN achieves an accuracy of over 93%.