AI-assisted Ultrasonic Testing of Rails - A Case Study

The inspection of railroads is a challenging task due to the large volumes involved and the wide variety of potential defects. Ultrasonic testing is a commonly used method due to its high sensitivity and ability to show internal defects. Multi-probe setups allow a more comprehensive analysis of the inspection target, but the larger data quantities and the complexity of the analysis make it difficult to reliably perform these inspections at high speeds. In this work, we studied the industrial applicability of an AI-assisted inspection system for ultrasonic rail testing. The system was designed for a five-probe (±70°/±38°/0°) unit mounted on a motorized cart. A reconstruction was designed to merge the input from the probes into a single B-scan-like merged view for easier interpretation. A deep learning model was trained to annotate defects, bolt holes and rail joints. The model was evaluated for two use cases: real-time assistance during inspection, and the automatic labelling of archived data.