Fault Classification Method for Rotating Machinery Based on Hybrid Model of CWT and CNN-DOA-LSSVM

Traditional signal processing methods for rotating machinery fault diagnosis rely heavily on human experience, while deep learning models often suffer from unstable classification boundaries and poor generalization under complex operating conditions. To address these issues, this paper proposes a hybrid fault diagnosis method based on CWT and CNN-DOA-LSSVM. Firstly, CWT is employed to convert one-dimensional vibration signals into high-resolution time-frequency maps, fully highlighting the transient impact features of faults. Secondly, CNNs automatically extract deep discriminative features, avoiding the cumbersome process of manual feature engineering. Thirdly, LSSVM replaces the Softmax classification layer in traditional CNNs to overcome the deficiency of the Softmax classifier in nonlinear classification. Finally, by leveraging the two-stage separation mechanism of exploration and exploitation in DOA, along with its unique forgetting-supplement and dream-sharing strategies, an adaptive optimal configuration of the key parameters of LSSVM is achieved. Validation results on the Southeast University gearbox dataset and the Huazhong University of Science and Technology bearing dataset show that the proposed method achieves average classification accuracies of 99.59% and 99.50%, respectively, demonstrating good performance in both classification accuracy and stability.