RBF composite parameter model for structural surface roughness with factor analysis
摘要
The characterization of the structural surface roughness is the groundwork for predicting the peak shear strength. The single parameter cannot fully reflect the characteristics of structural surface morphology. The index system composed of some single characterization parameters is a nonlinear system with much overlapping information. The factor analysis is conducted to reduce the dimension and strip the overlapping information by canonical normalization. At the same time, the standardized index system is transformed from nonlinear into linear by the RBF neural network structure. In practice, 6 statistical parameters reflecting the structural surface roughness are selected, the inverse calculation of JRC is established, 76 training samples and 37 group test samples are built. The multi-index composite parameters are established considering the characteristics of embossment of joints, such as height, angle and contact degree. In the mean time, the number of neurons in the hidden layer is fixed to improve the calculation speed. The prediction accuracy of the model is verified by the sample data and the direct shear tests of rock joints. The relative error and determination coefficient are calculated through the measured data to evaluate the performance. Finally, the applicability of the factor analysis and possible error analysis are discussed.