A Novel Small-Sample Fault Diagnosis Method for Rolling Bearings via Continuous Wavelet Transform and Siamese Neural Network

Fault diagnosis is essential to guarantee the reliability and safety of rolling bearings. However, the current research mainly focuses on a large amount of data and neglects the problem of insufficient sample size of rolling bearings in actual engineering. Therefore, this study investigates a novel small-sample fault diagnosis method for rolling bearings based on continuous wavelet transform and Siamese neural network (CWT-SNN). First, each raw signal captured by the rolling bearings is segmented into samples by sliding window and slicing operations and converted into signal time-frequency maps by CWT. Second, a 2-D convolutional neural network (2DCNN) is constructed to extract features from all fault samples and generate embedding vectors. Finally, sample pairs are constructed to input the embedding vectors into the Siamese neural network (SNN) and measure their similarity, thus realizing rolling bearing fault diagnosis under small-sample conditions. Experimental results show that the diagnostic accuracy of CWT-SNN model is improved by more than 10% compared with other models when there are only 60 fault samples. Meanwhile, CWT-SNN diagnostic accuracy can reach more than 99% with sample sizes larger than 120 and shows better generalization performance in new categories and new operating conditions.