Automatic train control using neuro-fuzzy modeling and optimal control techniques

In rapid transit applications, it is often necessary to optimize the ride of the train for certain parameters based upon time of day, occupant density, and system-wide scheduling. Trade-offs have to be made between energy conservation, time minimization, and ride comfort. Typically the dynamics of the train are not well known (or not initially known at all), change over time, and are non-linear. In the past, a transit control engineer would typically use P-I control but could spend days or weeks on-site adjusting the P-I constants to obtain a ride that felt good and met the design constraints. This process was both time consuming and expensive. This paper presents a control scheme for a rapid transit train that uses optimal concepts coupled with fuzzy control and neuro-fuzzy modeling techniques. The optimal controller allows users to define different ride types by adjusting weights on the cost equation. The controller design is done almost automatically, with minimal control engineer effort needed, by post-processing data collected from the train. The post-processing process uses neuro-fuzzy modeling techniques to create a dynamic model for the train, which can be used with optimal techniques to obtain fuzzy control rules for controlling the train. Once the initial design is in place, the controller becomes adaptive and fine-tunes itself to match the dynamics of the particular train that it is on