Rail track is a critical component of rail systems. Accidents or interruptions caused by rail track anomalies usually possess severe outcomes. Therefore, rail track condition monitoring is an important task. Over the past decade, deep learning techniques have been rapidly developed and deployed. In the paper, we review the existing literature on applying deep learning to rail track condition monitoring. Potential challenges and opportunities are discussed for the research community to decide on possible directions. Two application cases are presented to illustrate the implementation of deep learning to rail track condition monitoring in practice before we conclude the paper.

An Mgorithm for finding the characterization of a class of objects on the basis of a randonfly ordered sequence of labeled individual objects--some members of the class, some not--is described. The class is characterized as a disjunction of terms, each term being a conjunction of attributes. "All red, round objects or all square, small objects" is an example. Mechanisms based on this algorithm are described in terms of such properties as the amount of storage available for recording instances and the number of instances which had to be examined until the class was first guessed.In searching for classes which are conjunctions of attributes only, it is found t h a t : (a) the average number of trials to the first correct "guess" is significantly less than the average number to first finding a class description which is logically determined by all the data; (b) certain properties of a weight attached to each guess are reliable indicators of the presence of a "close guess"; (c) limiting the number of instances to be recorded to about the preceding 25 does not significantly increase the time to first guess over the case where all instances up to the remote past are recorded, independently of the number of attributes.The algorithm has been extended and programmed for disjunctions, and monte carlo simulation experiments are being conducted. It is planned to further refine the algorithm and explore its significance in p a t t e r n identification training, dynamic classification problems, and successive approximations to minimal Boolean expressions used in relay circuits.

No abstract