Data Driven Fault Prognosis Method Of Dynamic Systems And Its Application On Train Operation

Modern industrial process and other complex dynamic systems are developing towards large-scale integration.Once the fault occurs,it will directly affect the safety and quality of the whole system.Generally speaking,it is not only necessary to make an accurate diagnosis after the fault,but also to make a diagnosis when abnormal signs appear and predict the future development of the fault,that is,to predict the time of the fault or judge the probability of the system failure at a certain time in the future according to the past and present operating conditions of the system.Fault prediction and residual life prediction of complex dynamic system are effective ways to ensure its healthy and safe operation.In this paper,based on major projects of national natural science fund project "is based on large data information and knowledge of the high-speed train control system fault modeling theory and method,in the complicated dynamic system on the basis of summarizing the fault prediction methods,fault mechanism modeling is difficult,complex dynamic system which is difficult to fault prediction method based on mechanism model,the research of data-driven multivariate complex dynamic system fault prediction method,and combining with the characteristics of train dynamic operation and train applied equipment failure prediction study,main achievements include three aspects:(1)In view of the existing static data modeling methods that cannot extract the dynamic relationship between the high-dimensional data of dynamic systems,a data-driven Dynamic inner principal component analysis reconstruction combined with the vector time series model is proposed for dynamic system fault prediction.Combined with the dynamic spatio-temporal correlation of multi-motor temperature,a fault prediction method for train motor is proposed,which combines the Dynamic inner principal component analysis reconstruction with the vector time series model.The proposed dynamic system fault prediction method combining data-driven Dynamic inner principal component analysis reconstruction with vector time series model consists of three parts:dynamic latent structure modeling,fault amplitude estimation based on reconstruction,and fault prediction based on fault vector autoregressive modeling.Firstly,a dynamic latent structure model based on Dynamic inner principal component analysis is established based on the normal operation data of the system history.Then,based on the established Dynamic inner principal component analysis model,the fault direction is extracted from the historical fault data,and the fault amplitude estimation method of dynamic system based on dynamic principal component reconstruction is proposed.On this basis,a fault amplitude prediction method based on fault vector time series model is proposed.Simulation results of Tennessee Eastman process show the effectiveness of the proposed method.A fault prediction method based on dynamic latent structure reconstruction is proposed based on the dynamic correlation of temperature of multiple motors in train operation.(2)A Markov chain based method for predicting the remaining life of the train motor is proposed for the state transfer characteristics of the train motor when the fault occurs.Firstly,relevant theories of Markov chain are used to calculate and obtain the transfer probability between the states of the traction motor.Next,according to the state transfer probability of the traction motor,the failure probability of the traction motor is obtained to grasp the variation law of the failure of the traction motor.On this basis,the effective life of the traction motor is predicted by combining the state transfer probability and failure probability of the traction motor with the bathtub curve theory.The experimental results were about 10 hours earlier than the actual motor replacement time.