The Research And Improvement Of Multivariable PID Neural Network Algorithm

PID controller is simple, stable, reliable, easy to adjust, which makes it the best choice for people in industrial control. However, PID control is hard to get ideal with some large delay, uncertainty and multivariable-nonlinear strong coupling complex systems. The development of artificial neural network inspires researchers to put the essence of the neural network into the PID control.Professor Shu Huailin proposed an embedded neural network-PID neural network(PIDNN), contacting with PID law. Its hidden layers are closely related to the performance of the PID control law:differential, integral, proportional. The advantage of PIDNN is its initial weights can be set depend on the PID control law. But some control systems without PID control experience,have to set initial weights randomly. These random initial weights would make the PID neural network in trouble easily. To solve this problem, many researchers improved the PID neural network from all sides, and these improved algorithms were applied in the actual system successfully.In this paper, on the basis of these scholars’research, the mufti-variable PID neural network was improved from the aspects of output function. This paper proposed a new function model, to replace the original MPIDNN proportional threshold function, and then experimental example was simulated by MATLAB software. After the improvement, the network got better results in the convergence speed and approximation than the original one. The main content of this paper is:Firstly, research’s background and significance and the domestic and foreign research situation would be expounded. Then, the basic knowledge of PIDNN-the related concepts of neural network and BP neural network would be introduced. The principle of PID control, PID controller based on BP network structure and defects were also described in the paper.Second PID neural network structure, the structure and algorithm of PID neural network (single control system-SPIDNN and multivariable control system-MPIDNN) were showed with emphasis. And then the stability theorem of PID neural network and the key role of its initial weights were expounded.Finally from the aspect of output function of MPIDNN, a new output function model was proposed, to replace the original network output function-proportional threshold function. A 3 input and 3 output coupling system was simulated in this paper. The simulation shows the convergence speed of the improved MPIDNN algorithm was speed up significantly and the control error became better too.The content of this paper is summarized in the final phase of this paper. On the basis of the content and the research direction of this paper, the research prospect and development space of PIDNN in the further would be discussed too.