Technology Digests / Edge Computing Using Ultrasonic A-Scans for Rail Inspection

TD25-014

August 22, 2025
Authors
  • Anish Poudel
  • Ryan Alishio
  • Survesh Shrestha
Share

Key Findings

  • MxV Rail is exploring the potential to utilize UT A-scan data to develop a machine learning system to accurately identify internal defects in rail.
  • Phase 1 demonstrated the capability of a CNN model to utilize UT A-scan data to accurately discern different types of rail defects. Phase 2 focused on identifying hardware, software, and configuration methods to provide the necessary speed for real world deployment.
  • The CNN model implemented in the edge device achieved an overall accuracy of 89 percent for the 10-class problem.
  • Through the optimization of multiple parameters, the team was able to increase the processing rate from 200 FPS to 2,500 FPS. This translates to an approximate processing time of 0.4 ms per A-scan and would allow a scan to be collected and analyzed approximately every 3/16 inch (4.5 mm) at a detector car speed of 25 mph.

Edge Computing Using Ultrasonic A-Scans for Rail Inspection

In 2024, MxV Rail continued collaborating with the Project JeZero team to deploy the convolutional neural networks (CNN) model developed in Phase 1 on an edge computing device for rail ultrasonic testing (UT) using A-scans. The intent of this collaboration was to increase the accuracy and speed of ultrasonic rail flaw inspection and open up the possibility of using ultrasonic amplitude scans (A-scans) for future defect trending and data fusion work. This Technology Digest highlights the technical approach and implementation of Phase 2 work, which includes edge hardware selection, optimization of CNN model for edge deployment, data processing pipelines, and evaluation of the deployed system’s performance. MxV Rail’s work is not aimed at developing new products; rather, it is centered on demonstrating the feasibility of refining and optimizing existing rail UT technology using edge processing with A-scans. The edge computing device chosen for this test was the NVIDIA® Jetson Orin™…