Identification of underwater vehicle dynamics with neural networks

In this paper modeling of underwater vehicle dynamics using neural networks is investigated. Extensive knowledge is available on the behaviour of underwater vehicles. However, it may prove disadvantageous to use an off line established model of the dynamics due to the varying nature of the underwater environment. Neural networks offer interesting possibilities as they can be used to update the model of the underwater craft during deployment. The literature mainly reports on neural networks used in parallel to the whole system model. The system is thus regarded as a black box and the neural network is required to obtain the desired behaviour without any a priori knowledge. This can result in training problems due to e.g. the inclusion of time delays, necessity of larger (training) data sets and the necessity of larger networks. Rather than using one neural network to approximate a functional representing the whole dynamics, it is proposed to use several neural networks to represent certain model parameters. In this way better use of the model structure, of which detailed information is available, can be made. Some of the parameters, notably the damping parameters, are inherently varying, as they are a function of the velocity. Furthermore, at best only empirical models are available to model the damping. It may thus prove advantageous to approximate the damping using neural networks. The advantages of the proposed 'divide and conquer' approach will be demonstrated