Multiscale Deep Features and Empirical Features Fusion-based Gearbox Fault Diagnosis

Abstract

The application of gearbox intelligent fault diagnosis techniques in engineering practice is still challenging. Based on a deep neural network, the accuracy and generalization ability of the diagnosis model cannot be guaranteed when the labeled data is insufficient. In addition, due to the complex internal structure and strong coupling of gearbox, it is unrealistic to obtain reliable and complete fault information with a single model structure. Therefore, this paper proposes a fault diagnosis method for gearbox based on the fusion of multiscale deep features and empirical features, which can extract available information from various scales. The method combines traditional signal processing methods, multiscale learning and classical convolutional neural network (CNN) structure, which has two advantages. 1) Taking full advantage of the feature extraction of different techniques, multiscale learning strategies can obtain complementary diagnostic information to improve the diagnostic effect. 2) Variational mode decomposition (VMD) can obtain relatively stable signals, effectively dealing with the data noise problem that helps mine the fault characteristics of vibration signals. The performance on gearbox datasets demonstrates the superiority and robustness of the proposed method in gearbox fault diagnosis.