Real-time Residual Life Prediction Of Systems Based On Kernel Density Estimation

In recent years,with the continuous development of mechanical systems towards mechatronics,high power and multi-function,the improvement of reliability and safety in the operation process is becoming more and more important.Sensing technology,information and communication technology,big data and cloud computing technology are gradually applied to the monitoring of various complex mechanical equipment,and perform fault prediction and health management based on the degradation state information monitored during the operation of the mechanical system.The improvement of reliability and operating efficiency is extremely important.In this paper,the real-time residual life prediction model is established based on the information that represents the degradation state characteristics of the mechanical system detected during operation.The main research works of this paper is as follows:(1)Remaining useful life prediction based on kernel density estimation.Due to the lack of fault sample data and similar fault sample data,many existing equipment life predictions often require model structure assumption and parameter estimation.Since these methods are not accurate enough,this paper proposes a real-time residual life prediction method based on kernel density estimation.In this method,continuous degradation characteristics were used to construct the kernel density estimation model of degraded distribution,and then get the probability distribution function of the residual life,and after acquiring the new degenerate feature data,the probability density function is continuously updated by kernel density estimation,and the distribution of residual life is updated.Finally,the effectiveness of the proposed method in residual life prediction is verified by an example analysis.(2)Remaining useful life prediction based on kernel Differential homeomorphic transformation.When the traditional kernel density estimation is used to estimate the probability density of bounded variables,the estimation deviation will occur at the boundary of the interval,which will affect the accuracy of the remaining life prediction.For the above problems,this paper proposes a real-time residual life prediction method based on kernel diffeomorphism transformation estimation.This method uses the diffeomorphism transformation to transform the bounded random variable to the whole real number domain,so the deviation of kernel density estimation at interval boundaries can be effectively reduced.Finally,the feasibility and effectiveness of this method is verified by the real-time monitoring data.(3)Remaining useful life prediction based on kernel density estimation considering abrupt change point detection.The existence of abrupt change points in mechanical equipment degradation leads to inaccuracies in the prediction of its residual life.We propose a real-time residual life prediction method based on kernel density estimation(KDE)considering the influence of abrupt change points.First,a non-parametric cumulative sum method is used to detect abrupt change points in the degradation process.Then,integral mean square error is used to determine the abrupt change in the sample number that affects the accuracy of KDE.The weight coefficient is adaptively allocated according to the change in sample numbers relative to the minimum sample number before and after the abrupt change point in real-time monitoring.This method considers abrupt change states in the degradation process and uses KDE,which does not make model structure assumptions or parameter estimations for the degradation process.Finally,the effectiveness and the feasibility of this method is verified using the gear wear degradation data.(4)Remaining useful life prediction based on time-varying weighted kernel density estimation.In practical application,mechanical systems are often timevarying systems,and the degradation distribution changes with time.Therefore,a real-time residual life prediction method based on time-varying weighted kernel density estimation is proposed.According to the different influence of different sample points on the incremental kernel density estimation of degradation features at the present moment,different time-varying weights were given to the prediction of the real time residual life.Finally,the feasibility and effectiveness of the method are verified by analyzing the wear degradation data of gear.