Abstract
Safety is the principal concern of the railway industry, and rail internal defects can pose significant risks to safe and efficient railroad operations. Effective rail flaw detection, especially for transverse defects, is critical to prevent accidents and derailments due to broken rails. Both industrial and research communities have invested much effort in solving this problem. For example, the rail industry relies heavily on ultrasonic bulk waves and, more recently, began exploring ultrasonic guided waves for rail internal defect detection. This study developed an improved semi-supervised learning algorithm based on a deep autoencoder (DAE) for ultrasound-based rail flaw detection. The DAE algorithm identifies observations in a dataset that deviate significantly from the remaining observations. First, the team trained a DAE to reconstruct ultrasonic signals obtained from clean rail segments. To improve the robustness for defect detection, we then optimized the architecture and hyperparameters of the DAE models. Also, we adopted mean squared error (MSE) as a feature to highlight rail defects. Lastly, the team fed the test set of ultrasonic signals into the trained DAE model and evaluated its capability for rail defect detection. We found the proposed DAE can support a superior and robust ultrasonic rail defect detection capability compared to conventional knowledge-driven approaches.
