عنوان مقاله [English]
One of the most important parameters in back analysis is analysis of measured deformations of excavated tunnel. Numerical methods as a conventional method are used for back analysis. An artificial neural network that has been trained with sufficient number of examples made with numerical methods can be used instead of them. These networks are accurate enough and faster than numerical methods and an operator can utilize it without knowledge of rock mechanics or numerical methods.
In this research, a multilayer artificial neural network has been presented which is able to predict deformations around a tunnel after excavation. Input parameters of this artificial network are deformation modulus, Poisson ratio, tensile strength, cohesion, friction angle, initial vertical stress and horizontal to vertical stress ratio.
A tunnel model with 183 cases was executed with ''FLAC'' code and results (deformations) as a database was used for training and testing of the neural network. By training and testing of different neural networks, optimized values for layers and nods and architecture were found. According to obtained results, a neural network was selected. This network was able to predict deformation in roof and sides wall of tunnel accurately without having any knowledge about rock behavior. Relative strength of effect (R.S.E.) factor which is a mathematical relation in neural networks is presented that shows relative influence of parameter i as input on parameter k as output. With study of R.S.E., it was found that each parameter has a special influence on the deformations around the tunnel and some parameters has small influence in all conditions. So in analysis or back analysis of underground structures more attention should be paid to parameters that have more influence on results. By considering the effect of each parameter in numerical analysis, amount of effort for determination of the parameter and priority in geotechnical exploration can be fined.
Keywords: Relative strength of effect, Tunnel, Artificial intelligence, Artificial neural network, Deformation.