Automatic Detection of Rail Features Using Signals from Eddy Current Sensors

Abstract

Rail damage caused by rolling contact fatigue (RCF) poses a massive challenge for rail companies worldwide. Without appropriate maintenance to address its influence, the safety of the network cannot be guaranteed; however, if the maintenance is not carefully administered, the life of the rail can be unnecessarily reduced. In an effort to address these concerns, a number of systems have been developed to locate and quantify the depth to which surface cracking penetrates into the rail. This study concerns one such system: the Surface Crack Detection (SCD) System designed by Sperry Rail Ltd. The work documented in this thesis can be divided into two independent investigations. The first establishes a foundation for the theory underpinning the system, while the second concerns itself with the development of algorithms to automatically identify features in the data that the system generates. The first investigation commenced by detailing the functionality of the hardware and signal processing used to convert the impedance of the eddy current probes into in-phase and quadrature components. Finite element models of the probes were then constructed along with models of samples of steel containing simple defects of various geometries. In parallel, physical samples of rail steel were prepared and defects of similar geometry were electro-discharge machined into them. A comparison of the simulation results to the response of the system to the machined steel samples showed considerable agreement which gave confidence that the models were sufficiently accurate. The modelling was then extended to look at defect geometries that are difficult or impossible to physically fabricate. The second investigation employed artificial neural networks to identify features in the raw data generated by the system, focusing on thermit welds and rail crossings. In order to train the networks, sections of raw data containing features of interest were collected into libraries. While generating these libraries, many interesting discoveries were made which could be leveraged to extend the capabilities of the system in the future. The trained neural networks were evaluated on a number of validation runs and encouraging results were obtained. In both the weld and crossing case, a correlation between the performance of the network and the size of the training dataset was observed. The final part of this investigation consisted of a case study that used the positions of the welds and crossings identified by the neural networks to determine the location of the test vehicle within the rail network.

Date of Award	31 Dec 2019
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Anthony Peyton (Supervisor) & Wuliang Yin (Supervisor)