Unsupervised learning algorithm for high-speed defect detection in rails by laser/air-coupled non-contact ultrasonic testing

Recent train accidents and associated direct and indirect costs including cost of repair of equipment and infrastructure as well as delay and death/injury costs, have reaffirmed the need for developing rail defect detection systems more effective than those used today. In fact, rail defect detection has been identified as a priority area in the U.S. Federal Railroad Administration 5-year R&D plan. This paper proposes an unsupervised learning algorithm for defect detection in rails. The algorithm is used in a non-contact inspection system that is targeted to the detection of transverse-type cracks in the rail head (including transverse fissures and detail fractures), notoriously the most dangerous flaws in rails. The system uses ultrasonic guided waves that are generated by a pulsed laser and are detected by air-coupled sensors positioned as far away as 76 mm (3") from the top of rail head. The inspection ranges is at least 500 mm (20") for surface head cracks as shallow as 1 mm. Fast data output is achieved by processing the ultrasonic defect signatures by Wavelet Transform algorithms. The features extracted after wavelet processing are analyzed by a learning algorithm based on novelty detection. This algorithm attempts to detect the presence of damage despite the normal variations in ultrasonic signal features that may be found in a field test.