Challenge
A contractor responsible for maintaining railway infrastructure monitors the condition of the rails and their connections using cameras mounted on a train. All footage is then reviewed manually, with staff assessing when a maintenance intervention is required. This is extremely time-consuming and monotonous work. We were asked to investigate whether the visual scanning and interpretation of the images could be automated using image recognition, with the output being a prediction of where and when locations would need maintenance to prevent failures. This involved collecting data on completed inspections to create a large test set of observations.
Approach
- Gather data on inspections executed to create a large test set of observations.
- Develop a deep learning algorithm (based on neural networks) to automate the inspection.
- Minimise false positives (judged okay but in fact not okay) and false negatives (judged not okay but in fact okay).
- Test and validate the software in the live environment.
Results
- The software has proven itself in live operation and is now fully integrated into the contractor’s way of working.
