Research And Application On Methods For Constructing Prognostic Health Index With Complicated Degradation Signals

Predictive maintenance can effectively reduce the cost of railway maintenance and improve the availability of railway equipment and its core is remaining useful life(RUL).The purpose of RUL prediction is to evaluate the degradation status of equipment and predict the failure time of equipment using condition monitoring data,so as to provide a basis for making maintenance decisions.With the development of sensor technology,multiple sensors are usually deployed to monitor the condition of complex equipment from different aspects.In order to systematically describe the degradation state of equipment,multi-sensor data are usually integrated into a composite health index(HI).However,most existing HI construction methods assume that the data from the degraded device is a simple one-dimensional signal,which is not applicable in the scenarios with complicated degradation signals.To address this issue,this thesis mainly studies the methods for constructing prognostic health index based on complicated degradation signals.It focuses on the nonlinear relationship of signals,the coupling relationship among components in signals,the coupling relationship among components under the influence of covariates and missing values.By considering the domain properties and the characteristics of degradation signals,we design the objective functions and efficient optimization algorithms based on the techniques of deep learning,sparse learning and tensor analysis,respectively.The key challenges in HI construction with complicated degradation signals are addressed for accurately describing the state of the equipment and predicting its RUL.Finally,we focus on the applications of railway point systems.Specifically,this thesis has studied the following research topics,(1)The HI construction method based on deep learning is studied to capture the nonlinear relationship among signals.Firstly,a deep neural network is used as the data fusion model to describe the nonlinear relationship between degradation signals and HI.To address the issue caused by the unlabeled data,a novel architecture with a pair of adversarial networks is designed based on the properties from degradation modeling,to enable the procedure for training model parameters.To address the instability of the solutions,a RMSprop-based sampling algorithm is proposed to stabilize the performance of HIs.Through a simulation study and a classical case study,the superiority and robustness of the constructed HI in RUL prediction are verified.(2)The HI construction method for the structured signals with the coupling relationship among several components is studied.Based on a two-step decoupling strategy,i.e.,estimation of the stable background-extraction and fusion of the information associated with degradation,we first remove the stable background of each degradation signal and then,by considering the temporal and spatial characteristics of the degradation component,we design a method to identify and integrate degradation information based on sparse learning and isotonic regression(SLIR)to guarantee the quality of the constructed HI.Secondly,an optimization algorithm based on block coordinate descent is designed to solve this problem.Finally,a simulation experiment and the subsequent application study are used to verify the efficiency of the proposed method,the accuracy of identifying degradation-related regions and predicting RUL.(3)The HI construction method for high-dimensional degradation signal with missing values and under the influence of covariates is studied.To address the key challenge for the estimation of time-varying background with missing values,this thesis studies the modeling of tensor regression based on low rank approximation and Tucker decomposition and designs a framework for collaborative estimation of model parameters and missing values.Then,a large-scale numerical optimization algorithm based on block coordinate descent is proposed to solve this problem.Furthermore,the HI is constructed based on feature extraction and fusion of the residual signals.The effectiveness of the proposed method is verified by a simulation study and a case study.(4)On the basis of above theoretical studies,by considering the characteristics of railway point systems,we construct the HIs for prognostic analysis of two common degradation modes using condition monitoring data.For the oil-leakage mode of electrohydraulic point machine,a penalized convolution model is proposed to model the relationship between temperature and oil level by considering the features of temperature effects.After removing the influence of temperature,the oil-level residuals are used to construct the HIs and predict RUL under this degradation mode.For the mode of increasing resistance of sliding chair plates,we use the proposed SLIR method to analyze the power signals collected during the degradation processes of a railway point,for accurate location and fusion of degradation-related information.Based on the constructed HIs,we conduct RUL prediction under this degradation mode.