Non-contact detection of railhead defects and their classification by using convolutional neural network

Abstract

Railhead defects must be detected and classified intelligently in order for railway transportation systems to operate safely. Rail defect identification and categorization can be automated by using machine learning models to process rail image data (acquired using cameras). However, such an automated method has significant drawbacks: it cannot detect subsurface defects, picture data requires a high-end GPU with a long computational time, and machine learning model training can be influenced by image quality, which is dependent on light intensity and shooting altitude. Rayleigh waves are a potential candidate for rail inspection because they can detect both surface and subsurface defects and travel long distances on curved surfaces (like a rail) at high speed. This article looks into the possibility of combining fully non-contact laser ultrasonic technology (LUT) and a deep learning approach for intelligent detection and classification of railhead surface and subsurface defects. The fully non-contact LUT was used to actuate and capture laser-generated Rayleigh wave signals on railhead specimens in order to create a database of A-scan signals from healthy, surface, subsurface, and edge defect railheads. The classification capabilities of a support vector machine (SVM), a fully connected deep neural network (DNN), and a convolutional neural network (CNN) were examined after they were applied to the preprocessed signals without extracting any statistical/signal processing-based characteristics. The comparative analysis demonstrates that CNN is robust in classifying railhead defects. As a result, when combined with CNN, the laser ultrasonic technology may ensure automatic defection and classification of railhead surface and subsurface flaws.