Rolling bearing fault diagnosis based on ST-CNN-SVM

Rolling bearing fault diagnosis under complex operating conditions forms the essential foundation for the predictive maintenance of rotating machinery. However, traditional methods are often overwhelmed by strong noise, and constrained by the empirical risk minimization (ERM) principle, leading to significant overfitting in small sample learning scenarios. To address the aforementioned limitations, a lightweight diagnostic model integrating S-Transform (ST), convolutional neural network (CNN), and support vector machine (SVM) is proposed in this paper. Time-frequency features are extracted by leveraging the multi-resolution characteristics of the ST, deep feature mapping is performed through a customized CNN, and SVM is introduced to construct the maximum-margin classification hyperplane based on the structural risk minimization (SRM) principle. The experimental results illustrate that the method exhibits exceptional diagnostic accuracy under intense noise and small sample sizes. Randomized subset cross-validation confirms that this architecture effectively eliminates the interference of sampling randomness. Consequently, the ST-CNN-SVM model demonstrates high statistical stability.