Fault Diagnosis For Track Circuit Based On Machine Learning And Data Assimilation

In order to ensure driving safety,the track circuit is used to automatically control the signal or switch device by continuously monitoring whether the section is occupied by trains.It’s crucial to improve the accuracy and timeliness of tack circuit fault diagnosis because track circuit fault will seriously affect the safety of driving.Although there are many intelligence algorithms for track circuit fault diagnosis,degradation detection for aluminum electrolytic capacitors and multi-fault classification for track circuit have not been deeply investigated.Large delay for detection and miss classification are witnessed in these tasks.In view of the above problems,this paper makes the following research work:For the degradation detection of aluminum electrolytic capacitors in the tuning and matching unit,the traditional method monitors the deviation of main track output voltage.Because the main track output voltage is severely influenced by ambient factors like temperature and moisture,it is difficult to detect the incipient characteristics when the aluminum electrolytic capacitor begins to degrade.In order to reduce the detection delay,we introduce support vector regression to model the main track output voltage in normal condition,and then monitor the residual signal to detect degradation.As a result,the impact of ambient factors can be largely reduced.However,most of machine learning methods ignore the impact of observation error and the modeling error of the trained models cannot be neglected.We,then,propose an algorithm framework that combines machine learning and data assimilation to solve these problems.The experiments for the degradation detection of aluminum electrolytic capacitors demonstrate that this hybrid algorithm is superior to the single machine learning modeling method.The faults in the receiving cable are related to many factors,and it is difficult to classify them because the strong coupling among them.We solve the problem of fault detection and location by hybrid design of various methods.A piecewise-linear Fisher classifier is proposed such that accurate classification of faults at the receiving end can be achieved.Further,we consider the classification task for 96 classes of track circuit faults.As strong coupling exists in the large number of data sets involved,the miss classification ratio is very high for the traditional methods.We solve the above problem by adopting ensemble learning method to construct classifier.The effectiveness of the proposed method is proved by simulation experiments with data sets involving 96 classes of track circuit faults.