Crack-it.

Network Rail has 20,000 properties that it must monitor once per year (and in more detail once every 5 years). Images are taken by third-party contractors of faults in the properties and a report is generated. This is checked by 100 Network Rail staff. It is then sent on to 100 more Network Rail staff for final checking and work order generation. The process takes between 7 and 12 weeks.

Crack-it is developing an Artificial Intelligence (AI) solution that can check the output of the examiner against its large database of historical fault data. It can achieve this in a few seconds and deliver its feedback to the examiner whilst he or she is on site. Once the AI bot and the examiner agree on the severity of the fault, a report is sent to Network Rail staff who decide what work needs to be done.

The AI bot has been developed through trials with 3 different deep learning methods. The most successful of these has been selected and is being trained further with the data set that it found most easy to analyse.

Initially, we weren't sure if we had enough data. We had 10,000 pieces of high-quality data but this was comprised of a number of sub-categories. As we weren't sure which sub-category would be most easily read by the AI, we didn't know if we had enough data within it. Network Rail kindly spent 15 hours downloading more data from their database (OPAS) and we have been able to run training and testing.

We have achieved an accuracy of 82% which is higher than our machine learning experts had hoped for. We have methods to improve this further.

We have learnt a lot about the current processes in place at Network Rail and that an opportunity exists to make significant efficiency improvements. We have also learnt that catering tables provide good sleeping places.

Crack-it are excited about the current results and are keen to keep working on it.