Using an Artificial Neural Network in a DC Motor Electromechanical Speed-Control System

This paper considers the synthesis of a neural controller with training of an artificial neural network on a control object model with swizzling of loss function calculation. An electromechanical system with a dc motor is considered as a control object. Hyperparameters of the neural network and control criteria for the synthesis of an angular velocity controller of the motor are determined. The neural controller is implemented using the DeepLearning4j (DL4j) library in Java. The neural network is trained, and the results of modeling the speed control system are presented, confirming the operability of the neural controller. Once trained on one given sequence, the neural controller is tested on other sequences, with the quality of regulation not deteriorating. To test the possibility of using the controller in real conditions, random noise is superimposed on the measured state of the control object, while the controller trained on a noise-free model successfully copes with speed regulation even on a noisy control object. The proposed approach provides the developer with almost unlimited possibilities for fine-tuning the controller to the required quality of transient processes. Unlike traditional controllers, the neurocontroller also has predictive properties and preemptively issues a control action preventing an error.