Abstract
Accurate prediction of the rock mass parameters is of great significance in the safety and efficiency of tunnel construction by tunnel boring machine (TBM). In this fields, machine learning method has become the main stream to build the mapping between rock mass parameters and TBM driving data and achieves good results. However, tunneling data have hundreds of features and high acquiring frequency, which results in the structural difference between tunneling data and rock mass data. Current solution aiming at these differences rely on human experience and may lose data information, or increase the model complexity and time costs. This paper proposed a pre-treated method suitable for TBM tunneling data based on principal component analysis (PCA), with can reduce the feature number of tunneling data, and simplify its structure to match the rock mass data, with low data information lost. In addition, the proposed method can match multiple machine learning method, such as back propagation neuron network (BPNN), support vector regression (SVR), and random forest (RF), the corresponding integrity prediction models of the uniaxial compressive strength and joint frequency achieve acceptable accuracy. The method is verified by totally 1272 filed samples from five different projects. The predicted mean absolute percentage errors (MAPE) of different targets and different projects are controlled within 15%, which is acceptable in actual projects. In addition, compared with model involved full tunneling variable as input, the proposed pre-treated method can reduce the time cost from 3-5 minutes to 3-6 seconds, which improves the calculation efficiency.