神经网络Kalman滤波算法及多组分光度分析应用

Kalman Filtering Neural Networks Algorithm and Its Application to Multicomponent Spectrophotometric Analysis

前馈神经网络ＮＮ误差反向传播算法（ＢＰ）收敛速度较慢且常陷入局部极优值等，针对此种缺陷提出了一种基于扩展Ｋａｌｍａｎ滤波的快速学习新算法（ＥＦ）。与ＢＰ相比，ＥＦ法不仅学习效率高收敛速度快，数值稳定性好，而且所需学习次数少，调节参数少，由非线性系统建模与辨识的模拟结果表明，ＥＦ是提高网络收敛速度改善神经学习性能的一种有效方法，谈谈用于多组分光谱分析，结果良好。

In view of the shortcomings of the backpropagation(BP)algorithm,such as slow convergence and frequent local-optimization deadlock ,we propose a new quick-learning neural network algorithm based on extended Kalman filtering (EF).As compared with the BP algorithm,EF method has a higher learning efficiency, faster convergence, better stability ,less learning cycle and smaller hidden-neuron number. Simulation based on nonlinear-system modeling and recognition indicates,that EF is an efficient way to improve convergent rate and to promote learning capability of the neural networks.It has been satisfactorily applied to multicomponent analysis of pharmaceutical preparations.