Neural-network-based adaptive control systems for AUVs

Abstract

In this paper, two types of learning systems, the supervised learning system and the unsupervised learning system, are introduced to construct neural-network-based control systems. Both approaches are applied to longitudinal motion control of the free-swimming vehicle “PTEROA”. The supervised learning system is based on the simple concept of learning the behavior of the supervisor controller. It is implemented along with a fuzzy controller as the supervisor, and evaluated through numerical simulations and experiments. It is shown that the characteristics of the neural networks, such as flexibility of the I/O selection and saturation of the outputs, provide favorable performance to the control system for AUVs. The unsupervised learning system, which is called “SONCS”, is introduced as an adaptive control system. The subsystems and the organizing process of the controller are described in detail. The SONCS is applied to the control problem of the untethered test-bed vehicle PW45, and its performance is evaluated through free-swimming tank tests. It is shown that after several times of adaptation, the SONCS succeeds in organizing an appropriate controller for horizontal swimming at a desired depth.