基于时变神经网络的迭代学习辨识算法

Iterative learning identification algorithm based on time- varying neural network

为了实现在有限时间区间上可重复运行的离散时变非线性系统辨识,给出基于时变神经网络的迭代学习辨识算法。对于每一个固定时刻,以该时刻的神经网络逼近该时刻系统输入输出间的映射关系,提出了在同一时刻沿迭代轴训练网络权值的带死区迭代学习最小二乘算法,为防止收敛速度下降过快,进一步提出了协方差阵可重调的改进算法。所提算法有较快的收敛速度,且时变神经网络对非线性时变系统的辨识精度也较高。

In order to achieve that the discrete time-varying nonlinear system identification can run repeateadly on the finite time interval,an iterative learning identification algorithm based on time-varying neural networks is given. First,for each fixed time,the neural network of the moment which approachs the mapping relationship between input and output of the system,the iterative learning least squares algorithm with dead-zone for the weights updating along the iteration axis is proposed. To prevent the convergence rate from falling too fast,the improved algorithm whose covariance matrix can be retuned is proposed further. The proposed algorithm guarantees that the estimation error converges to a bound point wisely over the entire time interval,and the neighborhood radius depends on the modeling accuracy of the neural network. Experimental results show that the proposed algorithm has a faster convergence rate,and identification accuracy of the time-varying neural networks for nonlinear time-varying system is higher.