一种引入动量项的小波神经网络软测量建模方法

Soft sensing based on wavelet neural networks with momentum

针对小波神经网络存在收敛速度缓慢且容易陷入局部极小的问题,提出一种参数修正改进算法。首先,输出层神经元函数采用双曲正切函数代替传统的Sigmoid函数;其次,通过在权值调整式中增加动量项的方法选择学习步长,以提高网络学习效率。利用所提改进方法建立催化裂化主分馏塔轻柴油凝点软测量模型,结果表明,改进小波神经网络模型的预测精度和泛化能力优于BP神经网络和小波神经网络模型,对生产有重要的指导意义。

As wavelet neural networks（WNN）algorithm usually has low convergence rate and easily falls into local minimum,an improved wavelet neural networks（IWNN）is proposed to modify the parameters.Firstly,the traditional Sigmoid function is replaced by hyperbolic tangent function for output layer.Secondly,the learning step is selected by adding momentum to the weight adjustment to improve learning efficiency.The proposed IWNN method is used to build soft sensing of light diesel pour point from fluidized catalytic unit（FCCU）main fractionator.Compared with the models of BP and WNN,the results obtained by the IWNN approach showed a better prediction accuracy and generalization capability.Moreover this modeling can be used to guide production efficiently.