Terrain following control of an unmanned underwater vehicle using artificial neural networks

Abstract

Unmanned Underwater Vehicles (UUVs) are increasingly being used in the repair of offshore installations since the discovery oflarge quantities of oil and natural gas reserves and in applications such as explosive ordinance disposal and inspection and repair of submarine cables. UUV technology is being applied to areas that are too dangerous for, or beyond the physical capabilities of human divers or manned submersibles Due to their highly non-linear hydrodynamics and the unknown environmental conditions in which they operate, the control of Unmanned Underwater Vehicles (UUV) poses serious difficulties for most classical design methods. The complex control task means that a degree of autonomy would be advantageous both to reduce the load on the operator and to allow precision control of such mundane tasks as depth, altitude or position keeping. If this can be accomplished the pilot will be able to concentrate more effectively on the task in hand. Artificial Neural Networks(ANN) as described by Rummelhart et al (1986), offer some of the adaptive and robust properties required for effective control in this area. Some research has been undertaken into the use of ANNs to control UUVs over varying seabed terrain, but this has been restricted to rather complex two hidden-layer feedforward networks at a single operating point, using only one training strategy. In the work described here consideration is given to the merits of using both feedforward and recurrent single hidden-layer neural network controllers trained using two algorithms, Chemotaxis and Alopex. >