Intelligence series: Neural Networks and Rail
The key feature of neural networks is the structure: an input layer, output layer, and a variable number of hidden layers (deep learning, a form of machine learning, utilises multiple hidden layers). Each layer is made up of many nodes working in parallel to make a small transformation of the data they receive, before passing it along to the next layer. The connections between the nodes are ‘weighted’, i.e. some are more influential than others. The network modifies the weighting of the connections between nodes based on whether the transformation they have made has brought the network output closer or further from the target.
Neural networks could be used in image-recognition applications such as rail defect detection and classification. For example, data could be collected and analysed to detect cracks within the rail or the condition of bolt holes. Consequently, this could ensure unscheduled maintenance issues are identified and addressed quickly, thus reducing maintenance costs and delays. Image-recognition can also be used to identify fare-evaders, therefore protecting revenue. Furthermore, deep learning networks can spot patterns in historical datasets, which can be used for rail safety level prediction.