Research On Residual Life Prediction Of Rolling Bearings Based On Feature Fusion

Rolling bearings are the key parts of rotating machinery and equipment.Monitoring the condition of rolling bearings and making accurate estimation of the Remaining Useful Life(RUL)can not only reduce costs and increase efficiency,but also ensure the safety of workers.Combined with the XJTU-SY dataset and PHM2012 dataset,this paper studies the whole life cycle monitoring data of rolling bearings around the four key steps of RUL prediction.It includes feature extraction,feature selection and fusion,health state division,and RUL prediction.The research contents mainly include:It is difficult to effectively characterize the weak fault signal of rolling bearings,which leads to the low accuracy of subsequent RUL prediction.To solve this problem,a RUL prediction method of rolling bearing based on the fusion of time domain and spectral kurtosis features was proposed.In this method,the time domain features and spectral kurtosis features are optimally linearly fused,which enhances the ability of Health Indicator(HI)to characterize the degradation process of rolling bearings.The example analysis shows that the proposed RUL prediction method has high accuracy.This indicates that the spectral kurtosis feature has superiority in characterizing the performance of weak fault signals of rolling bearings.Methods based on linear feature fusion have limitations in the face of nonlinear and timevarying behavior in the degradation process of rolling bearings.Therefore,this paper continues to study from the direction of nonlinear feature fusion.The nonlinear feature fusion method based on Genetic Programming(GP)can adaptively establish an interpretable nonlinear feature fusion relationship without any professional knowledge about feature fusion.However,the massive data at the input will reduce the search efficiency of GP algorithm,and GP algorithm will produce complex solutions after multiple iterations.Therefore,this paper proposed a nonlinear feature fusion method based on improved GP algorithm,and further proposed a rolling bearing RUL prediction method.This method can efficiently construct a concise nonlinear feature fusion relationship according to requirements.The example analysis shows that the proposed feature fusion method has feasibility and generalization in the process of building HI for rolling bearings.Most of the existing feature fusion methods evaluate and select based on HI itself,which is difficult to measure the degradation process of rolling bearings well.And after getting HI,it is still necessary to consider how to accurately find the initial time of failure to facilitate RUL prediction of rolling bearings.Aiming at the above problems,the nonlinear feature fusion method based on improved GP algorithm is optimized.On the one hand,a comprehensive index evaluation function is set to measure the robustness,monotonicity and trend of HI.On the other hand,the Bottom-Up(BUP)time series segmentation algorithm can accurately determine the initial fault time of rolling bearing.In addition,the Wiener process model was established to better describe the changes of the dynamic operating environment and the failure mechanism of rolling bearings.On this basis,a RUL prediction method of rolling bearing based on optimized nonlinear feature fusion and Wiener process is proposed.The example analysis shows that the proposed RUL prediction method has good accuracy.