Novel methods of object recognition and fault detection applied to non-destructive testing of rail's surface during production

摘要

A series of rail image inspection algorithms have been developed for Tata Steels Scunthorpe rail production line. The following thesis describes the contributions made by the author in the design and application of these algorithms. A fully automated rail inspection system that has never been implemented before in any such company or setup has been developed. An industrial computer vision system (JLI) already exists for the image acquisition of rails during production at a rail manufacturing plant in Scunthorpe. An automated inspection system using the same JLI vision system has been developed for the detection of rail‟s surface defects during manufacturing process. This is to complement the human factor by developing a fully automated image processing based system to recognize the faults with an improved efficiency and to allow an exhaustive detection on the entire rail in production. A set of bespoke algorithms has been developed from a plethora of available image processing techniques to extract and identify components in an image of rail in order to detect abnormalities. This has been achieved through offline processing of the rail images using the blended use of different object recognition and image processing techniques, in particular, variation of standard image processing techniques. Several edge detection methods as well as adapted well known Artificial Neural Network and Principal Component Analysis techniques for fault detection on rail have been developed. A combination of customised existing image algorithms and newly developed algorithms have been put together to perform the efficient defect detection. The developed system is fast, reliable and efficient for detection of unique artefacts occurring on the rail surface during production followed by fault classification on the rail imaging system. Extensive testing shows that the defect detection techniques developed for automated rail inspection is capable of detecting more than 90% of the defects present in the available data set of rail images, which has more than 100,000 images under investigation. This demonstrates the efficiency and accuracy of the algorithms developed in this work.