Rolling Stock Diagnostics using RBF networks

Abstract

Modern automated control systems are created as cyberphysical production systems, when information is collected from several information systems for making management decisions. During maintenance and repair of traction rolling stock, automated control systems must receive information from on-board diagnostic systems to determine the pre-failure conditions of the equipment and plan an individual repair volume. The affiliation of the diagnosed event is determined by proximity to the class vector. The article is based on examples from the Internet of calculating the belonging of the diagnostic object to one of the possible classes according to the training sample data, based on which a proprietary program in the algorithmic language Visual Basic for Applications in the MS Excel environment has been developed. The program has been verified using test examples. The program automatically determines the size of the training sample and the number of classes, and allows to perform many iterative calculations. It is proposed to control the percentage of belonging to each class. Using the program, calculations were performed for a specially compiled training sample of the currents in four traction electric motors of one section of the locomotive: 10 events for each class. There are 5 classes in total: serviceable condition («normal») and failures of each of the four engines. The data has a random spread of values within the class. Simultaneous failures and pre-failure conditions of one, two and three engines are simulated. Calculations have shown the high efficiency of the method, especially in terms of determining the belonging to the «norm» class. The conclusion is made about the effectiveness of the proposed algorithm. It is proposed to use it as encapsulated in software when planning an individual volume of maintenance and repair with constant training by replenishing the sample.