Synthesis of a neurocontroller for a system containing an essentially nonlinear block

In some cases, in order to improve the quality indicators of the transients of the automatic control system, it is necessary to take into account in the model of the control object various mechanisms that drive the control object itself (DC motor, wheels, amplifier-converting devices). When calculating controllers by analytical methods, difficulties arise due to the presence of various kinds of irregularities in such systems, including "significant" ones ("backlash", "friction", etc.). In this case, the solution of this issue may be related to the use of artificial neural networks as part of the simulator. This paper shows the application of a neurofeedback scheme using a neuro-emulator and a neurocontroller. This allows you to form a training sample and train a neural network controller in the operating modes of the system that are beyond the control capabilities of a nonlinear model of the control object using a controller calculated by an analytical method. This scheme is considered as an addition to the algorithm of synthesis of neural network controllers with a deterministic way of choosing the architecture and weighting coefficients of a neural network using a scheme of imitating neural control. An example is given of improving the qualitative characteristics of the transients of a system by means of fine-tuning a neurocontroller for a nonlinear system "inverse pendulum on a movable base", taking into account the presence in the system of an inertial link containing a significant non-linearity of the "backlash" type. The purpose of the control was indicated, i.e. stabilization of the inverted pendulum in a vertical position and moving the mobile base to a set value. To achieve these goals, a neurocontrol scheme is used, which contains two neural networks: a neurocontroller (performs the function of forming a control effect on an object) and a neuroemulator (performs the function of simulating a model of the control object and is necessary to calculate the error back pass and adjust the weighting coefficients of the neurocontroller). As a result, it is possible to obtain an automatic control system capable of controlling the specified object.