Fault identification for rolling bearing based on ITD-ILBP-Hankel matrix.

Abstract

When a failure occurs in bearings, vibration signals are characterized by strong non-stationarity and nonlinearity. Therefore, it is difficult to sufficiently dig fault features. 1D local binary pattern (1D-LBP) has the advantageous feature to effectively extract local information of signals. Unexpectedly, it is vulnerable to the influence of noise when directly applied which led to quantization is inaccurate. To improve the accuracy of bearing fault diagnosis and solve the problem of imprecise quantization, the paper has studied the quantization criterion of 1D-LBP and proposed a combined method of improved 1D-LBP and intrinsic time-scale decomposition (ITD) and Hankel matrix (ITD-ILBP-Hankel). Firstly, a new signal pretreating strategy is proposed to further highlight feature information of bearing failure. Original signals are subjected to first-order difference operation to further highlight the impact feature of bearing failure and differential signals (non-original signals) are decomposed by ITD to obtain proper rotation components (PRCs). Secondly, to correctly quantize signals, a new quantization criterion is applied to 1D-LBP. Classical 1D-LBP is likely to be affected by individual extreme values or strong noise when quantizing signals inside window with median are threshold; meanwhile the root mean square (RMS) of signals can reflect the distribution of energy and represent the impact feature of signals in bearing fault. Therefore, RMS of signals is taken as threshold (in place of local median) to improve traditional quantization criterion of 1D-LBP in order to improve the accuracy of 1D-LBP quantization signals. Thirdly, the strategy of quantizing component signals, PRCs, rather than the whole original signals, according to improved 1D-LBP is taken to reduce interference among signals and correctly represent fault information. Fourthly, covariance matrix of Hankel matrix of local textural signal (LTS) corresponding to each component is constructed and signals are reconstructed to reduce noise interference and dig out hidden feature information in low-dimension space. Finally, fault feature frequencies of bearings are extracted through power spectrum of reconstructed signals and the type of fault is judged. The efficiency and advantage of proposed method is verified through the comparative analysis of simulation signals, tester signals and classical methods.