Algorithm for the Synthesis of a Neural Network Controller for a Multi-Channel System

The neural network approach to solving the problem of automatic control system synthesis has shown its effectiveness for plants containing nonlinearities of various nature. However, the use of such an approach can be difficult due to the lack of a formalized method of choice: sufficient, but not redundant neural network architecture and initial parameters of weighting coefficients. This paper presents a formalized procedure for the synthesis of a neural network controller for a nonlinear plant based on information about the controller calculated for a linearized model of the plant. This procedure allows you to obtain a non-redundant neural network architecture. In addition, the issue of initialization of the weighting coefficients of the neural network is being solved. As an example of the algorithm operation, the case of solving the control problem with a non-linearized model of an inverted pendulum on a cart is given. This problem is a classic test example of a multichannel unstable plant with a non-square matrix transfer function. An example of using a block diagram of the controller in a matrix form is given, which allows considering multi-channel plants in a compact form. Some structural transformations of the neural network controller are proposed to simplify the procedure of its training. The resulting neural network controller allows you to expand the possibilities of controlling a plant containing nonlinear parameters.