地基土压缩系数的RBF神经网络预测

Prediction on compressibility coefficient of soil based on RBF neural network

采用了输入层节点数为4、隐含层节点数为29、输出层节点数为1的RBF神经网络结构;RBF神经网络学习时,设置中心化方法为K-means聚类法,训练速率取0.15,加权种子数取2,Sigma参数取0.1,权重为0.2,最大迭代次数为16 000,误差均值控制为0.01.研究发现,训练RBF神经网络时,30组数据的土压缩系数的拟合值与实测值的相对误差为-2.540 0%～2.600 0%,有25组数据的相对误差为0,相对误差绝对值的平均值为0.185 14%;验证RBF神经网络时,土压缩系数的预测值与实测值的相对误差为-2.500 0%～12.000 0%,相对误差绝对值的平均值为5.669 27%.对岩土工程,一般误差小于25%是可以接受的,该地基土压缩系数RBF神经网络预测模型是可行的.

The neural network structure adopted is that,the node number of input layer of 4,the node number of hidden layer of 29,and the node number of output layer of 1 in the neural network structure are adopted.During the training of RBF neural network,the clustering method of K-means is used.The training speed of 0.15,the seed number of weighting of 2,the sigma parameter of 0.1,the weighting of 0.2,the maximum of iteration time of 16 000,and the average value of controlled of error 0.01 are adopted.It is found that the relative error of fitting value of compressibility coefficient compared with the observed value for 30 groups of independent variables in training RBF neural network model is between-2.540 0%～2.600 0%,and there are 25 groups of data whose relative error is 0,the absolute value of the relative error is 0.185 14%： In general,the error（25%） is feasible in geotechnical enginee-ring,so the prediction model of coefficient of compressibility with RBF neural network is feasible.