用于大工业过程建模的新型小波神经网络结构

New architecture of wavelet neural network for modeling large-scale industrial processes

提出一种新的小波神经网络结构 ,旨在解决输入变量比较多、变量分先后次序起作用的一类问题。该网络结构类似于多层前向神经网络 ,不同的是将一部分输入节点移至隐层 ,输入变量不是由同一层输入 ,而是根据变量起作用的前后次序分别在网络的不同层输入 ,从而使网络的规模减小 ;同时 ,隐层神经元的激励函数是一维小波函数 ,避免了多元小波函数带来的维数灾难问题。因此 ,该神经网络是处理高维问题的有效工具 ,尤其适用于包含多道加工工序的大工业过程的建模。将该神经网络用于热连轧产品质量建模 ,并经过了实测数据拟合与检验。试验结果表明 ,提出的小波神经网络结构是可行的 ,而且有很好的应用前景。

A new architecture based on wavelets and neural networks is presented aiming to handle a class of problems in which the input variables are very large and do not occur simultaneously. The architecture is similar to that of a feed-forward neural network, except that here some input nodes are moved to hidden layers. The input variables are not input in one layer, but in different layers according to their action sequences; the activation function of the hidden nodes is replaced by an one-dimension wavelet function,which can avoid the curse of dimensionality brought about by multi-wavelet functions. Thus the proposed neural network is a powerful tool for solving high-dimension problems, especially for modeling large-scale industrial processes with several work procedures. The developed methodology is tested by modeling the product quality of hot rolling mill with engineering data. Experimental results show that the proposed method is suitable for such problems and has good prospects.