Automatic Rail Flaw Localization and Recognition by Featureless Ultrasound Signal Analysis

Abstract

Ultrasound testing is a popular technique to find some hidden rail damages. In this paper we focus on the modern Russian railway flaw detectors, such as AVICON-14, which produce the results of ultrasound testing in the form of B-scan signals. We propose an approach simple enough to do fast automatic localization of B-scan signal segments, which could contain rail flaws. In order to recognize the selected segments as flaws of some kind or not flaws we apply SVM classifier jointly with DTW-based dissimilarity measure, specifically adapted by us to B-scan signals. To improve rail flaw localization and recognition quality we preprocess B-scan signals by applying some filter and making their convergence. Fast localization procedure jointly with CUDA implementation of B-scan segments comparison possesses the possibility to process big amounts of data. The experiments have shown that all rail flaws have been localized correctly and cross-validation ROC-score = 0.82 for the rail flaw recognition has been reached.