Variable universe fuzzy inference decided neural network feed-forward compensation for pid control of motor position

Abstract

To improve the compensation accuracy of neural network decided by traditional fuzzy inference engine (FIE), a variable universe fuzzy inference (VUFI) decided neural network feed-forward compensation method for PID control of motor position is proposed. The method contains four control blocks: fuzzy inference (FI) block, variable universe (VU) block, neural network (NN) control block, and basic control block. The FI block adaptively decides the feed-forward control quantity of NN based on the error and the error change rate of control system to keep the dynamic performance of NN at the beginning of the control or the transient jump signal. The VU block constructs the contraction expansion (CE) factors from the absolute change quantity of NN weights and adjusts the universes of membership functions based on the learning situation of NN in real time so that high-precision inference can be achieved. The NN control block consists of a neural network identifier (NNI) and a neural network controller (NNC) that has the same structure with the NNI, and the NNI dynamically transfers network parameters to the NNC and outputs feed-forward compensation control quantity after learning the dynamic inverse model of DC servo motor online. The basic control block adopts a PID controller, which provides the learning samples for NN online. Experimental results proved that the proposed method can improve the inference accuracy of FIE without sacrificing steady-state accuracy, significantly reduce overshoot, and have better stability and dynamic performance.