Prediction of future train states using neural networks

Developments such as LEADER and TrainStar have pioneered the concept of on-board train state estimation. TrainStar also provides the capability to predict future train velocity. Both LEADER and TrainStar technologies rely on the use of a stepwise simulation that keeps pace with real time. Following the idea of providing estimates of velocity in the future, research at the Centre for Railway Engineering has been progressing toward a system that provides estimates of in-train forces in future time. The Centre for Railway Engineering has developed neural network models which can predict the longitudinal coupler force at a selected position for a given train type in future time. The systems were developed with a capability of providing 50 seconds of future predicted data. Inputs to the neural network include hypothetical future control sequences, measured locomotive control parameters and Global Position System information that is linked to a database of track grade and curvature. The neural network model is 'trained' using a combination of measured and simulated data. A new network is required for each wagon position and each train type. Trained networks are stored as matrices of weights in data files and can be quickly loaded for use as required. The networks, when trained, offer a mathematical model that can be used to take on-board instrumentation a step further. The paper presents early results and details of progress in the developments of both software and in-cabin hardware.