Faults in bearings are often represented as compound faults in practical engineering. It is far more difficult to extract compound faults than single ones, and the influence which is from component signals irrelevant to fault, and transient impulses caused by the defective bearing are likely to be overwhelmed; this challenges the exact judgment of compound faults of bearings. 1D local binary pattern (1D-LBP) can extract hidden information from the perspective of local features. However, the quantization of signals according to 1D-LBP is vulnerable to noise and the representation of information is not deep enough. To solve this problem, a bearing fault feature extraction method has been proposed, which combines envelope-cross-correlation and improved 1D-LBP. Firstly, the cross-correlation function of envelope signal of vibration signals detected by two sensors is calculated to blend signals; this solves the problem of incomplete fault information contained a single sensor while reducing the noise and highlighting fault features. Secondly, before quantization of the signal, the feature rule for vibration signals in bearing faults is considered and the local skewness (in place of local central values) of data in the window is taken as a criterion for quantization and the study of improved 1D-LBP is used to solve the problem of incomplete representation of feature information in classical 1D-LBP. Thirdly, through quantization of the cross-correlation function (rather than the original vibration signal) according to the improved 1D-LBP, the problem of inexact quantization of 1D-LBP caused by noise has been solved. Finally, a bearing fault is identified by spectrum analysis of autocorrelation function of reconstructed signals after quantization. Through analysis of bearing vibration signals from different compound faults and the comparison of proposed and other classical methods based on the same data, the effectiveness and advantages of the proposed method are verified.
