Identification for non-linear systems based on particle swarm optimization and recurrent neural network [ultrasonic motor control applications]

A PSO-based learning algorithm for dynamic recursive Elman neural networks is proposed, which performs the evolution of network constitution, weights, initial inputs of context nodes and self-feedback coefficient together. A dynamic identification algorithm for nonlinear systems is constructed, which integrates the proposed algorithm with two identification strategies to perform the speed identification for ultrasonic motors. One of the strategies is that only the initial state and structure of the Elman network is optimized by the proposed learning algorithm and then the network is trained by a pure gradient descent-based learning algorithm, and the other is that the Elman network is trained on line by the proposed algorithm in the whole identification process. Numerical results show that the proposed algorithms not only realize the fully automatic optimization design for the dynamic recursive neural network, but also improve the precision of convergence for model identification, which provides an effective approach for nonlinear dynamic system identification.