一种全并行神经网络参数辨识方法

A New Parallel Identification Method UsingNeural Networks

针对在线参数辨识计算工作量大，造成难以实时给出参数估值的问题，利用Ｈｏｐｆｉｅｌｄ型网络的快速优化计算能力，通过对Ｈｏｐｆｉｅｌｄ网络改进，推出了一种全并行递推神经网络参数辨识方法，使计算量较传统的参数辨识方法大大减小。同时由于神经网络的互连作用，增强了辨识的鲁棒性，为实时给出参数估值提供了可靠的保障。

It is necessary for a control engineer to apply adaptive control to automatic pilot of a navigator to catch up with its development. The parameters of controlled system must be identified online to control the navigator adaptively in many cases, but existing serial identification method can not meet the realtime requirement of online identification. A new parallel method for on[CD*2]line parameter identification by neural networks is proposed in this paper to realize the realtime identification. The Hopfield Networks is modified by replacing its sigmoidal function with multilinear function, and then the link matrix and bias of the modified networks are set according to eq.(9) to guarantee the stability of the networks and to guarantee that the only one equilibrium of the networks be the least squares solution of the parameter identification. In order to meet the requirement of online identification, the recurrence formula, eq.(12), is given. The simulation results of an underwater vehicle system with the new method show the validity of the method through comparison with the results obtained with traditional least squares identification method.