Rolling bearing fault diagnosis based on multiple wavelet feature fusion method

A feature extraction method called MWBF, based on the fusion of multiple wavelet bases, is proposed to address the issue of insufficient feature representation when applying a single wavelet base for wavelet packet transform on non-stationary rolling bearing signals. We constructed an initial feature vector by performing a multi-wavelet packet decomposition on the signal and extracting statistical characteristics. Subsequently, we optimized the feature selection using the Linear Discriminant Analysis (LDA) method. Finally, different machine learning models are employed for the classification of multiple fault modes in rolling bearings such as k-nearest neighbor (kNN) and logistic regression (LR). Experimental results on the XJTU-SY bearing dataset show that the diagnostic accuracy of the model using the MWBF method (99.75% ~ 99.83%) is significantly higher than that of the wavelet packet feature extraction method based on a single wavelet basis (95.22% ~ 97.46%). The MWBF-LR model achieved a classification accuracy of 99.83% when using LR as the subsequent classifier.