Neural Network Based Rail Flaw Detection Using Unprocessed Ultrasonic Data:Final rept. (Oct 2000-Apr 2003)

摘要

The purpose of this project was to develop and test neural network based methods to improve the reliability and speed of ultrasonic railroad rail inspection and rail flaw detection, to enable earlier detection of flaws and to detect certain heretofore undetectable flaws. The current rail flaw detection technology is limited by the human operator's ability to interpret the ultrasonic data stream. This limits the detection car operating speeds and allows some important and critical flaws to go undetected. The results from an ongoing project co-funded by Sperry Rail Service and AAR, showed that neural networks can improve the rate and reliability of rail flaw detection by using the same processed data that is used in the operator-based system. Further improvements are possible by using the unprocessed data, which contains more information than the processed data. Development and application of the proposed technology is of particular importance on lines that, in the future, will combine the high speed passenger rail operations with freight traffic and heavier axle loads will further complicate the situation. They are likely to contribute to a higher rate of initiation and growth of rail flaws. Failure to detect and repair them in a timely fashion could cause service reliability problems and will pose safety concerns as well.