Traction control using artificial neural networks

Modern traction control techniques manage driven wheel speed from data obtained through a small number of sensors from around an automobile. This data is then fed into an on-board computer that has been specially programmed to attempt to maximise driven wheel traction from its information base. Problems can arise, however, when abnormal conditions are encountered. The small number of sensors used as controller inputs and the various assumptions made during the conventional controller programming can lead to erroneous correction commands — with an increased risk of spinout and other undesirable situations. The incorporation of artificial neural networks into the traction controller command logic has the scope to bring a two-fold advantage. The first stems from a neural network's ability to learn. It can be programmed through experimental data, and thus removes the need for tedious mathematics programming of the vehicle dynamics, which can be used to reduce controller cost and also brings about the possibility of continual reprogramming as the vehicle's condition alters due to wear, etc. The second is its ability to cope with data from a large number of sensors. With more sensors in place, the NN has a large base of information on which to make decisions on how to optimise traction. The result, hopefully, will be a traction control response that will provide superior traction solutions and be readily programmed to suit different applications. This study aims at conducting preliminary research into the accuracy and application limitations of artificial neural networks in the interests of determining their feasibility for use in closed loop traction control applications. This includes selecting parameters to be predicted, as well as model input variables, network architectures and the collection of network training data for off-line control. To this end the accuracy of back-propagation and general regression neural networks will be compared using 14 input, single output architectures in the interests of predicting longitudinal acceleration, lateral acceleration and yaw angle, which are identified as crucial factors in vehicle dynamics control. The study also includes the complete construction of a test vehicle, in which the sensors are mounted and data acquisition systems used to gather the network training and testing data, from which this investigation is based. This work is a step towards establishing the applications of artificial neural networks to automobile applications on a sound mathematical and quantitative basis