Feature extraction method of rolling bearing based on adaptive divergence matrix linear discriminant analysis

Status feature extraction is crucial to bearing fault diagnosis and the maintenance of rotating machinery. There are many challenges in extracting the effective status features from vibration signals for bearing fault diagnosis. A linear discriminant analysis (LDA) based on an adaptive divergence matrix (ALDA) is proposed to extract the status features of rolling bearings in this paper. The main idea of the method is that the sample clustering evaluation index (SI) is used to adjust the weight of the within-class divergence matrix of the LDA algorithm to reduce the cross or overlap among different types of samples, especially for the marginal samples. In the method, vibration signals of the rolling bearing under different running conditions are acquired, and the original features, such as time domain and frequency domain, are extracted from the vibration signals. Then, the optimal exponential weight of the within-class divergence matrix of the LDA is selected with the maximum SI. The optimal fusion status features of the bearing under different conditions were extracted by the ALDA algorithm from the original features. Finally, the fusion features were identified by the support vector machine classifier to verify the effectiveness of the proposed method. The experimental results show that the bearing status features extracted by ALDA can be used to identify the bearing status effectively.