Project Summary
Wheelset inspections on freight wagons currently depend heavily on manual checks, often performed in challenging conditions. This approach can introduce inefficiencies, human error and increase operational costs. Manual inspections are also time-consuming, which can delay fault detection and negatively impact safety and maintenance planning. Maintaining wagon wheelsets, including wheels, axles and brakes costs the UK rail freight sector around £60 million annually. Wheel wear and damage can cause significant service disruptions and pose risks to both safety and the environment. A low-cost, automated inspection system could improve asset management, reduce maintenance costs, and support a more proactive and data-driven maintenance strategy. Advancements in monitoring technology are facilitating the development of automated inspection systems that integrate innovative image-based dimensional measurement with machine learning for effective fault detection. These non-contact, automated solutions enhance accuracy and repeatability, reduce the need for personnel to enter hazardous environments, and enable more comprehensive and reliable data capture. A previous proof-of-concept study (TRL 2) demonstrated the feasibility of extracting key wheel dimensions from images taken in a maintenance facility. Building on this foundation, the current project aims to develop and validate a fully automated wheelset inspection system in a realistic operational railway environment (TRL 4/5).
Project Achievements
A key achievement of the project was demonstrating that accurate rail‑vehicle wheel profile measurements can be obtained from mobile phone imagery without specialist gauging equipment. The work advanced a previous semi‑automated prototype into a fully automated end‑to‑end system suitable for realistic railway conditions. The solution integrates automatic wheel detection, calibrated image‑scaling, and machine‑learning‑based dimensional prediction, enabling wheel measurements from short video sequences captured on standard mobile devices. A full processing pipeline was developed, including a trained wheel‑segmentation model, rim‑based calibration for converting pixels to millimetres, and regression models validated against reference measurements. A structured dataset of over 22,000 images supported model development and testing. A prototype mobile‑friendly interface was produced to show how the system could operate within maintenance workflows. Validation demonstrated sub‑millimetre accuracy, confirming that mobile‑based metrology can deliver reliable, repeatable measurements while enabling consistent digital data capture and supporting condition monitoring and predictive maintenance.
Conclusions
The project successfully delivered an automated, image based wheel measurement pipeline capable of estimating key profile dimensions from mobile phone imagery, meeting all primary objectives and advancing the system from TRL 2 to TRL 4/5. Validation showed strong agreement with industry reference measurements, achieving sub millimetre accuracy under representative maintenance conditions. Mean Absolute Error in primary testing was approximately 0.13 mm, with robustness trials on legacy wheels yielding errors of 0.72 mm for flange height and 1.58 mm for flange thickness, remaining within operationally meaningful limits despite reduced statistical fit on more varied image conditions. The work demonstrated three major advances: reliable low cost metrology using consumer devices; a hybrid, physics informed AI architecture combining segmentation, geometric calibration and regression; and evidence of real world feasibility through operational testing and a mobile friendly demonstration interface. The project established essential datasets, trained models and created a functional prototype, resulting in new intellectual property covering the algorithm design and system architecture.
Next Steps
A commercialisation and exploitation plan has been developed to advance the application along the TRL pathway, outlining the research, development and testing needed to unlock its commercial potential. The next phase will strengthen system performance, validate user requirements and prepare the software for wider market adoption. Key activities include expanding the image and video dataset to improve algorithm robustness and collecting additional data under varied conditions to understand how image quality variability affects performance. Ongoing engagement with end users will refine functional needs, with freight stakeholders expressing strong interest in further trials. Discussions with third party software providers will explore integration options, including links to existing tools and potential collaboration with other TRIG projects. Integration with asset management systems will support seamless data transfer, while features such as QR based wheelset identification will enhance traceability.
