Abstract: The scientific prediction of the TBM penetration rate is of great significance to the selection of hydraulic tunnel construction methods, construction schedule and cost estimation. In view of the high nonlinearity, fuzziness and complexity of TBM excavation process, and in order to improve the prediction accuracy and computational efficiency, the partial least squares regression (PLSR) has been applied to extract the principal components of the influencing parameters. Then the deep neural network (DNN) is employed to train and forecast the TBM penetration rate. A prediction model of TBM penetration rate based on the coupled method of PLSR and DNN is proposed. Based on the measured data of the double-shield TBM construction of a water conveyance tunnel in the Lanzhou water source construction project, six impact parameters including the rock uniaxial compressive strength, rock uniaxial tensile strength, cutter head thrust, cutter head speed, rock mass integrity coefficient and rock Cerchar abrasiveness index are selected to verify the prediction reasonability of the model. The fitting and prediction accuracy of the different prediction methods are compared and analyzed. The research results show that the PLSR can effectively overcome the problem of multiple collinearity between the independent variables. The extracted principal components are trained as the input layer of the DNN, which simplifies the structure of the neural network. The PLSR-DNN coupled model effectively avoids the over-fitting and inadequate fitting problems. It has the characteristics of fast convergence, stable solution and high fitting accuracy. The average relative fitting error of the PLSR-DNN prediction model is 2.96%, and the average relative prediction error is 3.27%. The fitting accuracy and prediction accuracy of the PLSR-DNN prediction model is significantly higher than those of PLSR model alone, BP neural network model and SVR model, respectively.

Key words: tunnel boring machine, penetration rate, partial least squares regression, deep neural network, coupling prediction model