Open AccessApplication of Neural Networks for Recognizing Rail Structural Elements in Magnetic and Eddy Current Defectograms
Author(s) -
E. V. Kuzmin,
Oleg E. Gorbunov,
Petr O. Plotnikov,
Вадим Александрович Тюкин,
Vladimir A. Bashkin
Publication year - 2018
Publication title -
modelirovanie i analiz informacionnyh sistem
Language(s) - English
Resource type - Journals
eISSN - 2313-5417
pISSN - 1818-1015
DOI - 10.18255/1818-1015-2018-6-667-679
Subject(s) - eddy current , artificial neural network , welding , track (disk drive) , process (computing) , eddy current testing , nondestructive testing , pixel , structural engineering , engineering , artificial intelligence , computer science , bevel , joint (building) , pattern recognition (psychology) , mechanical engineering , electrical engineering , physics , quantum mechanics , operating system
To ensure traffic safety of railway transport, non-destructive test of rails is regularly carried out by using various approaches and methods, including magnetic and eddy current flaw detection methods. An automatic analysis of large data sets (defectgrams) that come from the corresponding equipment is an actual problem. The analysis means a process of determining the presence of defective sections along with identifying structural elements of railway tracks on defectograms. This article is devoted to the problem of recognition of rail structural element images in magnetic and eddy current defectograms. Three classes of rail track structural elements are considered: 1) a bolted joint with straight or beveled connection of rails, 2) a butt weld of rails, and 3) an aluminothermic weld of rails. Images that cannot be assigned to these three classes are conditionally considered as defects and are placed in a separate fourth class. For image recognition of structural elements in defectograms a neural network is applied. The neural network is implemented by using the open library TensorFlow. To this purpose each selected (picked out) area of a defectogram is converted into a graphic image in a grayscale with size of 20 x 39 pixels.