Research On AUV Modelling Based On Neural Networks

On account of the requirement of ocean exploration and military affairs, the technology of Autonomous Underwater Vehicle (AUV) has been rapidly developed, which is an important tool in the completement of a good many of intelligent tasks in ocean. An AUV is a serious nonlinear system with strong coupling, so the problem of how to describe exactly the dynamical characteristics of AUV is very difficult and which is the precondition to control, prediction and fault diagnose of AUV. So the research on AUV modelling attracts many scholars' attention.After introducing the "Beaver" AUV which is developed independently in the Underwater Vehicle Intelligent Control Technology Lab., the paper studies the problems of how to establish the forward dynamic model, predictive model and how to adjust the on-line neural network model of AUV.In this paper, a Compound Input Dynamical Recurrent Neural Network (CIDRNN) is proposed to identify the AUV dynamic model after analyzing the identification methods of nonlinear system with neural network firstly, and then the differences in the nonlinear system identification with multilayer feedforward neural network and modified Elman neural network are compared in detail. The neural network models of AUV surge velocity and yaw rate are builded, at the same time an on-line training algorithm — simplified backpropagation through time (SBPTT) is proposed, which restricts the influenced time step length in a constant value comparing to the standard backpropagation through time (BPTT) algorithm. Simulation experiment results based on nonlinear function approximation and nonlinear system identification prove the SBPTT algorithm is valid with CIDRNN. Basing on the forward model of AUV, the multi-step predictive dynamic model of AUV is also established with neural networks, which is applied to build the neural generalized predictive controller for AUV.