Abstract:

hrough a large number of physico-mechanical tests and microstructure analysis of soft soils in Nansha area, Guangzhou, China, 40 groups of physico-mechanical indices and microstructure parameters are obtained. Using data analysis ability of grey-relation analysis method and nonlinear mapping ability of artificial neural networks, a model for the relationship between physico-mechanical indices and microstructure parameters of soft soil is established based on grey-relation analysis and radial basis function (RBF) neural networks. In this model, grey-relation analysis is applied to preprocess data and extract key components as the input of the neural networks. RBF algorithm can fully utilize the information contained in the training data，adaptively choose the centers of radial basis functions, the widths and the weights of networks; therefore, the problems of determining node number of hidden layer and centers，slow learning speed and weaken generalization ability of traditional RBF neural networks when the input data are generous and complex are solved. Model A and model B show that this method can reduce the structure of neural networks, and raise efficiency of training and accuracy of prediction, and provide an efficient way to quantitatively study about relationship between physico-mechanical properties and microstructure parameters of soft soils.

Key words: soft soil, physico-mechanical index, microstructure parameter, grey-relation analysis, radial basis function neural networks