Base On Clustering Radial Basis Function Neural Networks And Application Research To Constitutive Model Of Geotechnical

Constitutive relationship of geotechnical materials is the base of design and calculation of geotechnical engineering. Mechanical properties of geotechnical materials are complex, being nonlinear, dilation rate, and anisotropy etc. Simultaneously, their constitutive model could be influenced by stress state and stress path. Traditional constitutive models have been set up mostly based on experiment data according to different theoretical hypothesis. Traditional calculating modes would produce two primary problems: firstly, models must simple for easily modeling and obtaining parameters, which leads to the demand of reflecting complex engineering problems hard to be satisfied; secondly, models must contain a mass of parameters for accurately reflecting the constitutive relationship of geotechnical materials, and these parameters may be determined just by introducing a lot of hypothesis according to traditional calculating mode, which brings much limitation of the application of models. Due to the limitation of distilling information of traditional calculating modes, it is difficult to considerate the both. The numerical method of ANN (Artificial Neural Network) pioneer a new approach to research the problem.ANN model can directly extract information from experiment data without any hypothesis. Compared with other artificial neural network, RBF (Radial Basis Function) neural network can approach any nonlinear function and deal with the intrinsic law of system that is hard to resolve, and they also possess rapid study velocity. Therefore, it is a perfect choice to research the constitutive model of geotechnical materials by RBF neural network. However, the application exists certain limitation, that is, neurons of association layer will be superabundant when the number of experimental data is large (namely big samples), which results in over-long training time. Meanwhile, the complexity of restricting relationship among data leads to the fall of predicting precision of model, which is hard to satisfy the needs of engineering.The topology structure and function characteristic of RBF neural network have been researched in the paper, and they are modified a little in allusion to the possible problem of RBF network in the case of big samples. An improved RBF neural network, namely, Base on Clustering Radial Basis Function Neuron Network (BC-RBFNN for short) has been proposed. Meanwhile, the BC-RBFNN has been analyzed and demonstrated both in theory and experiment, which shows that the complexity of network model is reduced and the predicting precision is also improved. Afterwards, the BC-RBFNN has been applied to the research of constitutive model of geotechnical materials. Combined with the numerical modeling method of geotechnical materials'constitutive relationship, the elasto-plastic constitutive model based on BC-RBFNN has been set up on the basis of the triaxial compression test data for moderate sandy clay. Finally, the predicting results and experiment values of BC-RBFNN model, the predicting values of RBF network model and the predicting values of BP network model have been compared and analyzed, which validates the feasibility, validity and superiority of BC-RBFNN model when applied to practical engineering.