Application of Neural Networks for Recognizing Rail Structural Elements in Magnetic and Eddy Current Defectograms

Abstract

Nondestructive testing of rails is regularly conducted using various approaches and methods, including magnetic and eddy current testing methods, to ensure railroad safety. Automatic analysis of large data arrays (defectograms) from the corresponding equipment is an important problem. The analysis is the process of detecting defective pieces and identifying structural elements of a railroad track using defectograms. This article considers the problem of recognizing patterns of structural elements of railroad rails in defectograms of multichannel magnetic and eddy current defectoscopes. Three classes of structural elements of a railroad track are investigated: (1) a bolted joint with a straight or beveled rail connection, (2) a flash butt rail weld, and (3) an aluminothermic rail weld. Patterns that cannot be assigned to these three classes are conventionally considered as defects and are attributed to a separate fourth class. Patterns of structural elements are recognized in defectograms using a neural network based on the TensorFlow open library. For this purpose, each defectogram area selected for analysis is converted into a grayscale image with a size of 20 by 39 pixels.