神经网络扩展滤波算法及其多元光谱分辨应用

Extended Kalman Filtering Tranined Neural Networks and its Applications to Multicomponent Spectral Resolution

前馈神经网络(NN)误差反向传播算法(BP)应用较广,但收敛较慢且易陷入局部极优,针对这一不足,本文提出了一种基于扩展滤波的快速学习新颖算法(EF).与BP相比较,该法不仅具有学习效率高,收敛速度快,所需学习次数少,数值稳定性好等优点,而且所需调节参数少.由非线性系统建模与辨识的模拟结果表明,EF是一种有效的神经学习新算法.该法用于多元光谱校正与分辨,获得良好结果.

Backpropagation (BP), one of the most useful algorithms to train neural networks (NN), has some deficiencies and/or inadequacies such as low convergence and local optima. A novel learning method for training NN,extended Kalman filtering(EF), has been develped with rapid convergence speed, few iteration cycles and small hidden neurons. This EFNN method was used for multicomponent spectral resolution and simultaneous determination of composite pharmaceutical APC preparations and mixed aromatic species samples with satisfactory results.