Early fault alarm method of rolling bearing based on wavelet analysis and convolution neural network
In view of the problems in condition monitoring of aero-engine main bearing,such as the difficulty in obtaining the real fault samples,the limitation in defining the general alarm threshold under variable conditions and the difficulty to identify the early weak faults,a general alarm method for early faults of rolling bearings was proposed.This method only trained convolutional neural networks based on normal samples,constructed evolution state indicator by the characteristic distance between degraded data and normal data,and unified the alarm thresholds of different working conditions based on training labels;at the same time,the sensitivity of wavelet band envelope signal to early high-frequency fault was used to realize early warning;then,the evolution stages were divided based on the Pauta criterion,according to which the degradation and failure thresholds were determined;finally,the remaining useful life was predicted step by step based on particle filter.Three groups of test results showed that the degradation threshold and failure threshold of wavelet analysis and convolution neural network(Wavelet-CNN)based on different fault test data can be normalized around 0.6 and 1.0,and the predictions of degradation start time were 13.01％,12.33％and 13.70％earlier than those of non wavelet methods respectively.
rolling bearinggeneral diagnosisearly warningdeep learningwavelet analysis
NETL
NSTL
万方数据
