Deep Convolutional Neural Network-Based Method for Strength Parameter Prediction of Jointed Rock Mass Using Drilling Logging Data

Abstract

Field evaluation of the strength properties of jointed rock masses is a challenging task in geotechnical engineering. Typically, laboratory tests using small jointed specimens have difficulty determining the strength parameters of jointed rock masses due to the scale dependence of discontinuities and because the tests are expensive and time-consuming. Fast and continuous estimation of the unconfined compressive strength σcm of a jointed rock mass directly using drilling via a deep convolutional neural network (CNN) is a novel and practical field investigation method. This paper presents a CNN framework that includes (1) obtaining a training dataset; (2) determining the unconfined compressive strength σcm via a rock mass quality rating (RMQR) system; (3) training the CNN model; and (4) validating the results using tunnel engineering calculations. A comparison of the CNN predictive results with the true values suggests that the CNN makes good predictions across a wide range of unconfined compressive strengths σc of intact rock, especially for high RQD values. Due to the joint orientation, the unconfined compressive strength σcm of a jointed rock mass cannot be reliably determined using the σcm/σc ∼ RQD relation. By incorporating the physical variables of RQD and σc, which are known to affect the unconfined compressive strength σcm of a jointed rock mass, into the CNN, the proposed CNN model can provide better predictions than the regular CNN model. All the results predicted by the physics-informed CNN are within the accepted error range of 10%. This method is applied to the excavation of the Huangshan Tunnel in the Hanjiang-to-Weihe River Project of China and is verified as reliable via comparative studies with previous works. Thus, the proposed method represents fast and efficient prediction of the strength of jointed rock masses in rock engineering.