Research On Fault Detection For Traction Systems Of High-speed Trains Based On Just-in-time Learning

With the booming transport industry,high-speed trains have become an essential means of transport,combining punctuality and excellent carrying capacity.As a key component of high-speed trains,traction systems provide inexhaustible power for their efficient operation.The traction system has a complex structure,and there is intricate coupling between the internal components.Consequently,the system is prone to failure due to friction between internal mechanical structures during operation and natural aging of components.Additionally,the complex and changing environment in which trains operate means that the traction system is constantly subjected to harsh external conditions and overloads,accelerating wear on system components.In other words,the fault susceptibility of the traction system is determined by the internal structure and the working environment.Therefore,to maintain the health of the traction system and ensure the safe and reliable operation of high-speed trains,timely detection of faults in the traction system is an important requirement.This thesis takes the traction system as the research object and builds a strategy for detecting faults in the traction system under actual operating conditions.The strategy is based on a just-in-time learning modeling approach.The work in this thesis focuses on the following areas:(1)The structure and working mechanism of traction systems are presented and analyzed.The observed variables,faults,and disturbances in the system are expressed in a signal-based form.Simultaneously,the principles and advantages of the just-in-time learning method for fault detection in traction systems are analyzed from the perspective of system characteristics.(2)A fault detection scheme,combining just-in-time learning and slow feature analysis,is proposed for the incipient fault detection of traction systems.The causes and manifestations of incipient faults in the traction system are qualitatively analyzed from the actual operating conditions of the traction system.Firstly,a local just-in-time learning model is used to predict the health state output of the traction system,and then a residual signal is calculated that can be applied to the static system analysis method.Secondly,the slow feature analysis method is employed to extract the slow-varying feature variables from the high-dimensional observation data,effectively filtering out noise and other disturbing factors.Finally,the excellent detection capability of the proposed method for incipient faults in traction systems under actual operating conditions is demonstrated by injecting various types of incipient faults into a traction system platform.(3)A just-in-time manifold learning fault detection model is proposed to address the problem that it is challenging to distinguish between faults and normal mode switching during multi-mode operation of traction systems.Mode switching is accompanied by sudden changes in system parameters,and the jumps in monitoring signals make conventional fault detection methods prone to false detection.Therefore,based on the idea of decomposing global problems into local problems,this thesis integrates local manifold learning into the just-in-time learning framework.The proposed scheme avoids the loss of feature information of the global data structure by extracting the feature information of the local data structure.Using the feature structure of the local data as the basis for modeling reduces the computational complexity of modeling and improves recognition accuracy.Finally,the effectiveness and superiority of the proposed scheme are verified through a series of experiments on the traction system platform.