Structural and parametric synthesis of a neural network emulator of an autonomous unmanned underwater vehicle dynamics

Abstract

Currently, developers of information and control systems for marine underwater equipment are faced with new opportunities for using modern high-performance technologies to improve the quality of control processes and the accuracy of operations. For example, approaches based on predictive models (PM) are actively studied to ensure the synthesis of motion control algorithms. In this case, modern machine learning methods, including artificial neural networks (ANN), can be used to synthesize PM. A method for constructing a PM as part of the algorithmic support of information and control systems of an autonomous unmanned underwater vehicle using a neural network emulator of dynamics is proposed. The main disadvantages of the traditional approach to the synthesis of PM in the form of a system of differential equations are analyzed, a sequential structural and parametric synthesis of the neural emulator is performed. In particular, the issues of initial initialization of the neural network parameters and the formation of a training sample are considered, the structure of input and output data is determined. A feature of the proposed ANN structure is the use of pretraining based on a cascade of autoencoders. Results of pretraining the neural network emulator are presented, justifying the choice of the ANN architecture. Also, to check the adequacy of the PM in the form of a neuroemulator, verification is performed with respect to a known nonlinear dynamic model during statistical simulation modeling.