Neural Network Analyzer of Processes in Automatic Control System

The analyzer of the operating modes of an automatic system, based on the neural network, is considered. The processes in the system are investigated by the movement of the depicting point in the state space of three variables of the regulating device: system error, time derivative of the error signal and time derivative of the control signal. To classify the state of the system, the space of possible trajectories of the depicting point is divided into areas corresponding to normal operating modes and processes with various deviations. To analyze the trajectory of the depicting point of the regulating device in the state space and the formation of training sequences (learning patterns) of the neural network, the authors take not the controlled variables themselves, but logical functions defining their belonging to the specified ranges. In this case, the trajectory of the depicting point is represented as a vector. The artificial neural network is trained in such a way as to assign to the input vector the values from the state space a certain class in the set of normal operating modes of the regulating device as well as modes with different deviations from the normal process. For this purpose, multilayer perceptron is used. The number of input and output elements of the neural network is determined by the problem conditions. In this case, the input layer contains 12 neurons (by the number of components of the input vector), and the output layer - k neurons, by the number of diagnosed faults. The case of k=2 corresponds to the recognition of 2 abnormal modes: oscillations in the system and system operation with a mismatch, caused the actuator limited power.