Data Driven Fault Detection And Forecast For Maglev Trains

With the maglev train after more than ten years of development.The electromagnetic Suspension(EMS)middle-low-speed maglev train has independent intellectual property rights and is moving towards the commercial operation.Due to the safety,environmentally friendly,economy,low noise,and strong climbing ability of middle-low-speed maglev trains,Beijing and Changsha already have two commercial middle-low-speed maglev lines.It will be promoted nationwide,such as Chengdu,Wuhan,and other cities are in the planning.Because EMS magnetic levitation technology is more sensitive to the environment such as the track line,especially the middle-low-speed maglev trains will vibrate during operation.Vibrate will affect the air gap and current feedback signals,resulting rapid changes in the levitation electromagnetic force,and causing resonance to occur in cycles.Resonance will affect the safe operation and ride comfort of the train.There may even be suspension failures that could cause the vehicle to fall off or suck.In order to improve the ride comfort and safety of the train and ensure that the suspension height is stable.It is necessary to strengthen the detection and prediction of resonance to improve the stability of the suspension control system.Therefore,timely and accurate resonance detection and prediction is of great significance to improve the stability and safety of suspension control.This thesis takes the middle-low-speed maglev train from Changsha south railway station to Huanghua airport as the engineering background.Based on the data of air gap,current and acceleration,a data-driven resonance detection and prediction method for the maglev train is proposed.The main research contents and achievements are as follows:(1)Analysis of operational data collected by sensors.According to the change of the statistical characteristics of the signal,such as the large fluctuation range of the signal and the frequent deviation from the mean when the resonance occurs,a method for detecting resonance with statistical variance and change rate threshold is proposed,and an index of resonance coefficient is proposed to measure the degree of resonance occurrence evaluation,has good robustness.(2)By extracting the signal spectral information and the time and frequencydomain parameter features combined with effective Intrinsic Mode Function(IMF)components to increase feature coverage.The resonance prediction model of the Elman neural network based on multidimensional features was proposed.Through real data simulation,the prediction accuracy of resonance within 1 minute reached93%.To solve the problem of small data sample size.The improved Elamn-Adaboost prediction model based on multi-dimensional features was proposed based on the idea of integrated learning to increase the sample utilization.The simulation results show that the resonance prediction accuracy is 88.72% within 5～10 minutes.