A Novel Fault Identification Method for Rolling Bearing

Abstract One-dimensional local binary pattern (1D-LBP) identifies the running state of bearings by describing local textural features of vibration signals. However, the 1D-LBP algorithm is rather sensitive to noise and intrinsic periodic features of rotating machinery which is not included in consideration. This makes it difficult for the 1D-LBP algorithm to comprehensively dig fault information of bearings. To solve this problem, a new method has been brought forward by combining the intrinsic periodicity of rotating machinery and extraction of textural features. Before extraction of local features of signals, vibration signals are subjected to first-order difference operation to highlight local impact features of bearing failure. Meanwhile, to reduce the noise and further strengthen periodic features, differential signals are segmented by a complete period, and the signals after segmentation are subjected to cross-correlation successively to acquire the corresponding cross-correlation functions. Second, 1D-LBP is employed on cross-correlation functions (not raw signals) to extract local features and local textural signals (LTS) are obtained. Third, feature parameters and vectors have been built according to LTS and the running state of bearings has been described from the perspective of textural feature. Finally, according to the established feature vector and classification algorithm, running state of bearing was determined. A comparative analysis is given to the proposed method and other methods through public data sets and data from an experimental equipment of bearing. The results of comparative analysis indicate that with the two kinds of data sets, the identification rate of the proposed method for unknown samples is over 97% and it can judge the running state and the type of faults more precisely.