Sound Modeling Using Artificial Neural Networks for Virtual Mine Detection Training

In real field demining, soldiers can only judge the landmine location and type by the sound generated from a landmine detector. Therefore, a virtual landmine detection training system can replicate the sound in a realistic manner is imperative. In this paper, several sound datasets for various targets have been collected. Each dataset contains about 500 instances, each representing a different radius and height of the detector head away from the target. To study the characteristics of different landmine targets and to devise a mathematical model for sound generation, a multilayer perceptrons (MLP) artificial neural network (ANN) utilizing back propagation (BP) is implemented to represent the sound model. Neural networks including a particle swarm optimization (PSO) based neural network and a genetic algorithm (GA) based neural network are also applied to the datasets to obtain a good mathematical model for sound estimation and generation. The mean squared error (MSE) resulted from the different methods is compared with each other, and it is shown that PSO based neural network has the least MSE.