This thesis explores the use of a magnetic sensor array and artificial neural networks for determining the position of a permanent magnet on a 2D plane. One of the major motivations for this research is to track the stylus of a magnetic haptic interface that uses electromagnets to provide haptic feedback to a permanent magnet on the stylus. Hence our method must be able to sense the magnet position in the presence of disruptive materials, such as the iron cores of the actuating electromagnets. As these electromagnets could saturate the sensors in the direction normal to the array, we also investigate the network’s effectiveness when only using the magnetic field components in directions parallel to the plane of the array.

The networks used are multilayer perceptron networks consisting of a hidden layer and an output layer. The effectiveness of four different training methods are compared to determine the most effective method for training such a network and the best network parameters for that method. The accuracy is then compared to that of a traditional tracking method, both with and without a disruptive steel bar and/or magnetic field present.

The neural networks are found to have much better accuracy than the traditional method in the presence of the interfering material and solve for position more than 6 times faster. Their solution speed is also much less variable than that of the traditional method, making them more suitable for real-time tracking applications. Of the training methods investigated, networks trained using Bayesian Regularization were found to be most accurate, with several networks achieving mean position errors of less than 1 mm. The Bayesian Regularization method was also found to be less susceptible to premature termination of training.