Under the Header “ locomotive AI” we are working on a development project applying possibilities of current machine learning strategies, like support vector machines, deep learning, fraud detection systems and multivariate regression to real-time monitor the performance of the railway system and inform about abnormalities within the system behavior as possible indicator of upcoming failures.
Many current monitoring systems rely on supervised learning or expertsknowledge to indicate future failure. For example carbon levels in oil samples or resistance and ampere measurements in electrical circuits. But also the error messages in the vehicle system. Also more sophisticated indicators exist such as temperature development graphs in diesel engines or pressure of injector pumps. Or speed versus load versus engine rotations versus gear ratios.
There remains a group of failures that is very hard to find and determine the root cause because they are hidden in the vast amount of data and possible unexpected correlations between events.
We were inspired by one example of a specific version of the Br189 that very frequently came to an emergency halt in the Stuttgart area. After 1,5 years of evaluating it became clear that 1 bailies on the track was slightly outside frequency and combined with a very tight frequency setting of this type of 189 (partly due to specific country specifications and requirements of other safety stems) it came to this emergency stop/halt.
We are experimenting with developing an unsupervised abnormally (abnormality?) detection system that provides feedback on the real-time systems performance.
Contact us if you are interested in learning more.
