Fault diagnosis of wind turbine gearboxes based on SVM feature selection and CGAN data augmentation under imbalanced samples

Abstract

Deep learning algorithms are increasingly applied in equipment condition monitoring. However, scarcity of fault samples in wind turbine gearboxes significantly hinders diagnostic accuracy in practical industrial scenarios. To address this challenge, this paper studied a new method based on SVM feature selection and Conditional Generative Adversarial Network. The method extracted time-domain and frequency-domain features from original vibration signals. A multi-task SVM model for important feature selection was constructed to obtain low-dimensional features. A CGAN model was designed to enhance the number of samples for multiple fault types, addressing the issue of sample imbalance. The effectiveness was validated through experiments on a planetary gearbox test bench. The experiments shown that it retained important features and generated high-quality fault data, resulting in a 35.67% improvement of diagnosis accuracy and a 111.6% enhancement in classifier performance. Furthermore, compared to other representative data augmentation algorithms, this method demonstrates superior performance.