| Rolling bearings are one of the most used components in modern industrial equipment,but they are the most vulnerable to strong shocks to the point of damage.The performance condition of bearings determines the state of the whole industrial operation,and it is necessary to predict the performance degradation trend of them.Therefore,signal noise reduction,feature fusion and performance degradation stage and performance trend prediction are studied for the problem of rolling bearing performance degradation prediction as follows.Firstly,to address the problems such as endpoint effects in signal decomposition,the complete ensemble empirical modal decomposition（CEEMDAN）-wavelet semi-soft thresholding（WSST）based noise reduction methods are proposed.The vibration signal is decomposed by CEEMDAN to obtain multiple eigenmode components（IMF）;Pearson correlation coefficient method is used to correlate different IMFs,and the IMFs with high noise content are separated and used for secondary noise reduction by WSST;finally,the signal is reconstructed.By comparing the simulated data,it is verified that the method can retain the effective components of information to a great extent,with the largest signal-to-noise ratio and the smallest mean square error.Secondly,to address the problems that it is difficult to completely characterize the performance change trend of rolling bearings by a single feature and the gradient disappearance caused by the traditional activation function,the fusion of rolling bearing performance degradation feature indicators based on stack autoencoder（SAE）and the stage division method are proposed.The noise reduction data are pre-processed to obtain multiple time domain and frequency domain feature indicators;the SAE feature fusion method is used to fuse multiple time domain and frequency domain feature indicators to obtain the corresponding fused feature indicators;multiple evaluation indicators are used to comprehensively evaluate the alternative feature indicators and the fused feature indicators;the fixed window averaging method is used to divide the rolling bearing performance stages;by comparing with the By comparing and analyzing with the conventional fusion algorithm,the fused feature indicators are better than other feature indicators and have a clearer division of each performance stage of rolling bearings.Then,to address the problems of low prediction accuracy and reliance on manual experience of conventional prediction methods,the performance degradation prediction method of time-series convolutional neural network（TCN）-Attention mechanism（Attention）model is proposed.A network junction model for TCN-Attention performance degradation prediction is constructed,and a weight scoring function is also designed to study the way the Attention model is combined with the TCN model,to improve the weight assignment to sensitive information,and to select reasonable network structure parameters.Through experimental verification,compared with other prediction models,TCN-Attention has the highest prediction accuracy,the best curve fit,and the smallest root mean square error and average absolute error.Finally,the data collected from the bearing experimental bench built by Xi’an Jiaotong University are used to verify the signal noise reduction,feature fusion and performance degradation prediction methods. |
