The railway track circuit performs different tasks (occupation of a specific section, track/vehicle information transmission, detection of broken rail) and its regular diagnosis is necessary to ensure a high level of safety and availability. It consists of a transmitter connected to one of the two section ends, a receiver that is connected to the other end of the section, a transmission line constituted by the two rails and trimming capacitors that are connected between the two rails at constant spacing. The aim of the diagnosis is to automatically detect the change of electrical track circuit characteristics .

The inspection is achieved by a specific car, equipped with coils located in front of the first axle. The carrier current level is measured and recorded at each position of the train , while the track circuit is shunted by the inspection train itself. This recorded signal has a specific signature due to the presence of electrical components on the track (trimming capacitors, electrical separation joints?) .

The aim of the presented study is to both detect and localize defects in a context of predictive maintenance. The first phase of the project has been dedicated to the electrical modelling of the track circuit while the second phase is focused on pattern recognition approaches.


A complete electrical modelling of the system has been achieved . It allows simulating a large variety of dysfunctions of the system. Thus, it constitutes a useful training aid to complete the existing expertise and to compensate the lack of available in-situ labelled defects . Simulated signals are generated by the model in different working conditions and are used as database for pattern recognition algorithms. This model (~100 electrical parameters per track circuit) includes also a model of noise based on the expert knowledge of electrical conditions during the inspection runs.




Different diagnosis systems have been studied. A global diagnosis method based on pattern recognition approach is performed. In this case, a multilayer neural network is used as a classifier to assign the measured signals to one of the two classes (no major defect / major defect), according to their estimated posterior probabilities. A suitable signal pre-processing is optimised in order to transform the data representation space and to make the classification easier. A reduced set of relevant features are extracted to approximate the signal. A detection module was implemented under Matlab?. Performances estimated on simulated signals reach 97%. Reliable results are also obtained on the few labelled measurements.

An iterative physical diagnosis is designed to detect the defects on the trimming cells. Instead of the global electrical model that requires the adjustment of 100 parameters, the iterative approach considers a sliding physical model with no more than 4 parameters to optimize that correspond to one trimming cell characteristics. Once the parameters of one cell are obtained, we study the next cell, and so on, until the whole track circuit is analysed. Then the diagnosis task consists in comparing the optimised parameters to the standard real ones corresponding to a reliable track circuit.


A diagnosis based on «  mixture of experts   ». The idea is to exploit the spatial dependency between trimming cells into the decision phase. One classifier is built per trimming cell. Each of them has its own representation space. The method rests on the following observation: when a defect occurs, the signal is entirely modified between the defect and the receiver. Hence the output of each elementary classifier gives information about the presence of defect on its own trimming cell and also on the cells located after it. The fusion of all classifiers responses is performed using the transferable belief model (TBM) theory to assign an observation to a given class
+ نوشته شده توسط علی در پنجشنبه چهاردهم آبان 1388 و ساعت 10:5 |