Keywords: railway infrastructure, switch operation monitoring, power consumption curve, Fisher score, hidden process regression model, abnormality detection, diagnosis.

The remote monitoring of the railway infrastructure and particularly the switch mechanism continues to be of great interest for railway operators because its operating state directly impacts on the availability of the overall railway network. The problem then consists of earlier detection of the presence of anomalies (electrical, mechanical or civil engineering anomalies) in order to alert the maintenance services before a critical breakdown occurs. Moreover, the detection process has to be performed on-line by exploiting successive measurements acquired during the working process. In the context of switch monitoring, each measurement represents the electrical power consumption curve acquired during a switch operation, each new curve being recorded as soon as a switch is performed. It is then necessary to analyse each new curve to detect as soon as possible different type of defects.

For this monitoring purpose, a novel methodology for anomalous curve detection in the curves stream is introduced in this paper. Its main advantage is that it does not require a physical model of the system and can easily be adapted to other complex systems. This methodology consists in the following two steps.

The first set consists in fitting a probabilistic model so called "Hidden Process Regression Model (RHLP)" to a small number of curves corresponding to "normal" operations. This specific model allows the power consumption curves to be accurately represented. Indeed, due to physical characteristics of the considered switch mechanisms, these curves are subject to various changes in regime as a result of five successive mechanical movements of the physical components associated with the switch mechanism: starting phase, points unlocking, points translation, points locking and friction phase. In particular, the nominal probabilistic model makes use of polynomial regression to represent the regimes involved within the curves and logistic functions to model the transitions between the regimes. Therefore it captures efficiently the different phases of the curves. Furthermore, as a result of the hidden logistic process there is no need to define manually each phase, they are automatically recovered by the model.

The second step consists of detecting abnormal curves on the curve stream associated with successive switch operations, while keeping the nominal probabilistic model up to date. These tasks are performed on-line by using the Fisher score test to detect anomalies and by applying a stochastic gradient algorithm to recursively update the model parameters. The derivation of the test is detailed in the paper, together with its integration within the global tracking and testing algorithm.

The approach was applied to detect anomalies on a switch of the French high speed railway network, over a 28 months period. A sequence of 916 power consumption curves was recorded during this period, each curve being sampled at 100Hz over 5,5 seconds. Fifteen curves associated to mechanical misadjustments of the mechanism were correctly detected using the proposed approach. The paper gives more representative illustrations of the detection process.

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £110 +P&P)