摘要
依据小波的非线性逼近能力和神经网络的自学习特性 ,提出一种适合高维输入的小波神经网络建模方法 ,这种小波网络结构类似多层感知器 ,不同的是隐层神经元的激励函数为小波函数。分别对 3种小波函数进行试验 ,利用多种优化算法训练神经网络 ,经比较 ,选择 B-样条函数为激励函数 ,利用L- M算法较为理想 ,成功解决了 32维输入的大型多辊热连轧机钢板材质量建模问题。经过 86 0 0组实测数据拟合和检验 ,测试结果表明 ,拟合命中率达 82 .3% ,测试命中率达 80 .5 % ,表明了该方法的有效性。
Based on the function approximation ability and the learning characteristic of neural network, a wavelet based neural network(WNN) is introduced to handle the high dimension input problem. The structure of the WNN is similar to that of multi layer perceptron, but the activation function of hidden nodes is replaced by a wavelet function. The WNN is taken as the steel plate quality model of large scale hot rolling mill and a 32 input modeling problem is resolved. Several different wavelet functions and different algorithms have been tested and compared. The authors choose B spline function as the activation function, Levenberg Marquardt algorithm is used to train the WNN Finally. After modeling and examination of 8600 samples of engineering data with noise, the best result is 80.5% error value between forecast value and real value satisfies the practice engineering demand. Simulation results demonstrate the effectiveness of the methodology.
出处
《系统工程》
CSCD
北大核心
2002年第5期55-58,共4页
Systems Engineering
基金
国家 8 6 3计划资助项目 (86 3- 5 1- 0 11)
西安交通大学自然科学基金资助项目 (0 90 0 v- 730 2 4 )
