Research On Early Warning And Prediction Method Of Rolling Bearing With Interpretability

With the advancement of Made in China 2035,the process of digitalization of the man-ufacturing industry is getting faster and faster,which provides sufficient data for fault pre-diction and health management technology.Rolling element bearings are widely used and prone to failure,and the prediction and health management during the whole life cycle of rolling bearings can greatly improve the reliability and safety of its operation.In this disser-tation,an interpretable early warning and prediction method is proposed for the monitoring tasks in the whole life cycle of rolling bearings.The main contents are as follows:(1)Aiming at the problem of multi-source measurement signal of rolling bearing,a feature extraction method based on pulse amplitude difference and the degradation warning method based on the evolution comparison of likelihood probability are proposed.The am-plitude of the high frequency of vibration signal is obtained by discrete wavelet transform,and then the amplitudes of the low rate strong pulse and the high rate carpet pulse are ex-tracted.The difference between the two pulse removes the influence of flat.The exponential Weibull distribution is utilized to fit the pulse amplitude difference in the healthy stage and the fault stage of the bearings,so as to calculate and compare the likelihood probability that the test bearing located in healthy and failure stage,and achieve online degradation warn-ing of bearings.The interpretability of early warning results can be explained by likelihood probability evolution.The experimental results show that the method can achieve a good early warning effect on the bearing,and has a very low false alarm rate.(2)Aiming at the lack of interpretability of end-to-end learning for rolling bearing re-maining useful life prediction,the spatial-temporal graph neural network and graph evolution are introduced into the field of rolling bearing remaining useful life prediction.Combining the regression task properties for remaining useful life prediction and the shapelet tool for time series analysis,the regression shapelet is proposed.The shapelet is treated as a node of the graph structure,and the edges between nodes are represented according to the rela-tionship between the shapelets,so as to obtain the complete graph structure data.Then,the graph data is used as the input and the percentage of remaining useful life as output.The spatial-temporal graph neural network is trained to establish a real-time remaining useful life prediction model of bearings.The experimental results based on the bearing dataset show that the prediction accuracy rate of the spatial-temporal graph network is higher than other networks? at the same time,the graph evolution process shows that the graph structure tends to equilibrate during bearing degradation,which can explain the prediction results of the testing bearing remaining useful life.(3)Aiming at the problem of over-smoothing in deep spatial-temporal graph neural net-work,a spatial-temporal multi-scale graph neural network based bearing remaining useful life prediction method is proposed.The Fourier transform and energy integration is used to obtain the spectral energy distribution at different times.Then,based on the difference of the spectral energy at different time,the dynamic characteristics are obtained.The iterative window method is used to give an accurate health stage boundary,and realize degrada-tion warning.By saving the information of each hidden layer to the output,the multi-scale spatial-temporal network retains the data of different information transfer scales and avoids the influence of the over-smoothing of the final hidden layer.The experimental results show that the degradation warning time for slowly changing bearings is greatly advanced? and the problem of over-smoothing of deep graph networks is effectively solved,and the prediction effect of the proposed spatial-temporal multi-scale network is more accurate.(4)Aiming at the lacking of internal interpretability of the model,a physics-informed neural network is proposed for remaining useful life prediction.First,the classical crack growth model for bearings is simplified and transformed to obtain the relationship between degradation state and time,and the physical empirical degradation model is obtained.The random disturbance and unknown mutation information in the actual degradation process are extracted from the real data.By combining the physical model and the real data,the physics-informed neural network that includes both physical information and real data in-formation is trained.The experimental results based on the bearing dataset show that the proposed physics-informed neural network model achieves higher prediction accuracy than only using the data driven model.