Intelligent Prediction Model For The Penetration Rate Of Shield Machines In The Subway Construction And It’s Application

In modern city,subway is an efficient vehicle which greatly relieves the pressure of ground transportation.With increasing the number of subway lines,the geological conditions encountered during the tunnel excavation becomes complex.However,shield machines are sensitive to geological conditions.Therefore,effective prediction for the performance of shield machines under specific geological conditions is critical to the choice of construction method and schedule,project budget.Because the interaction between the subway shield machine and the geological conditions is very complex,it is difficult to fully reveal its correlation from the perspective of theory.In the present study,machine learning research methods were employed to establish the intelligent model for predicting the penetration rate of shield machines,including BP neural network,hybrid algorithm neural network,support vector regression(SVR)and Multiple kernel learning support vector regression(MKL-SVR).As a result,the penetration rate of subway shield machines was predicted accurately.The main contents of this paper are as follows:(1)According to the geological data and operational data collected from Beitian North Road Station to Beier Road Station of Shenzhen Metro Line 10,10 main factors that influences penetration rate of shield machines were obtained,including shield machine parameters,soil layer parameters and lone stone parameters.In detail,shield machine parameters include earth pressure,total thrust,torque acting on cutter head and cutter speed.The soil layer parameters includes cohesion,internal friction angle and compression modulus.The parameters of the boulder contain the proportion of solitary stone,compressive strength and rock quality designation(RQD).Subsequently,After processing these parameters,sample sets for machine learning were obtained.(2)The structure and algorithm of the Back Propagation neural network are studied.The training set and the test set are used to train and test the model.After that,the number of layers of the BP neural network model,the number of hidden layer neurons,the expected error,the activation function,and the training function are determined.Based on BP neural network model,a prediction model for penetration rate was established.(3)Considering the slow convergence rate of BP algorithm and the problem of local minimum value,imperialist competition algorithm(ICA),particle swarm optimization(PSO)and genetic algorithm(GA)are used to optimize artificial neural network(ANN).Consequently,a hybrid algorithm neural network model was established.In the search space,ICA,PSO and GA were employed to search the global minimum value.Then the optimal results were obtained in ANN.By training and testing the hybrid algorithm neural network model,a prediction model for penetration rate was established on the basis of the hybrid algorithm neural network model.Compared with the BP neural network model,more accurate results can be obtained from the prediction model.(4)The principle of SVR machine is studied.Simultaneously,the characteristics of each kernel function are discussed.Since the kernel function and kernel parameters directly determine the performance of the model,the radial core with strong compatibility was selected as the kernel function in the model.Then,grid search algorithm(GSA),PSO and GA were used to optimize the penalty and kernel parameters.As a result,a new prediction model for penetration rate was established based on SVR.(5)Considering the complexity of SVR machine and the limitations of dealing with heterogeneous problems when selecting kernel function and kernel parameters,multiple kernel functions were combined by using weighted linear combination.Consequently,a new prediction model for penetration rate was established based on MKL-SVR.The prediction model reduces the burden of selecting kernel functions and kernel parameters and improves the generalization ability of the model.