Active Control of Multiple Neural Networks for Oscillating Combustion

Abstract

A multiple neural network controller is proposed and demonstrated to suppress the pressure oscillation of the Rijke tube acoustic network. This controller consists of three modules including two separate neural networks, i.e., the neural network of controlled object that is pretrained before control and the neural network of controller that is trained in real time during the control process. This controller can identify the characteristics of oscillating combustion, achieve adaptive output, and extend the applicability and expansibility. Results show that multiple neural network controller can suppress the pressure oscillation in different oscillating stages using fuel valve or loudspeaker as actuators. When the exact mathematical model of the controlled object is difficult to obtain, it is effective to take a zero-dimensional simplified model with similar oscillating characteristics for prototype controller, and then actively control the real controlled object through parameter migration. It is further demonstrated that this controller is insensitive to noise. In addition, delay correction is achieved by adding sensor and actuator delay modules to the original controller to offset the effects of system delay.