Traction Motor Fault Diagnosis Based On Adaptive Long Short-term Memory Q Network

Traction motor is the core equipment of energy conversion and power source of high-speed train,which is very important to the safe operation of high-speed train.However,with the passage of time,the components of traction motor will gradually age and the performance will decline,resulting in higher failure frequency,and even threaten the driving safety of high-speed train.Therefore,the research on fault diagnosis technology of traction motor of high-speed train has important theoretical research significance and practical engineering value.This thesis studies the fault diagnosis method of traction motor of high-speed train.The main research contents and innovative work are as follows:1.For the problem that the three-phase current signal of traction motor has time dependence and periodic change,which leads to the difficulty of fault detection,a long short-term memory Q network model is proposed,and a traction motor fault detection method based on MRP paradigm long short-term memory Q network is proposed.The long short-term memory Q network model is composed of agent and Q-learning.In this fault detection method,the three-phase current time series is processed into the characteristic state sequence satisfying the Markov reward process,and then the agent samples the characteristic state sequence to obtain the experience pool for training the model,learning the fault probability of all characteristic states to realize fault detection.Simulation results verify the effectiveness of the proposed fault detection method.2.For the problem that the dynamic characteristics of different types of traction motor faults are complex and there are noise and disturbance,which leads to the difficulty of fault diagnosis,a fault diagnosis method of traction motor based on MDP paradigm long short-term memory Q network is proposed.Based on the Markov reward process,the fault diagnosis method introduces decision action and strategy to form a Markov decision process.The three-phase current time series is processed into the characteristic state sequence satisfying the Markov decision process,and then the agent samples the characteristic state sequence to obtain the experience pool for training the model to learn the probability of all characteristic states in different fault types.Fault diagnosis is realized through behavior strategy and continuous trigger alarm mechanism.Simulation results verify the effectiveness of the proposed fault diagnosis method.3.For the problem that the distribution of fault characteristics of traction motor is different under different speed conditions,which leads to the difficulty of fault diagnosis,a multi condition fault diagnosis method of traction motor based on adaptive long short-term memory Q network is proposed.By introducing the distribution difference measure,an adaptive long short-term memory Q network model is constructed to analyze the distribution of working conditions in source domain and target domain.It can learn knowledge from the source domain condition data with fault type label information and migrate to the target domain condition data without fault type label information.Simulation results verify the effectiveness of the proposed fault diagnosis method.This thesis contained 42 figures,16 tables and 97 references.