Research and Application on the Relationship between Tunnel Drilling Parameters and Rock Mass Properties Based on Field Experiments and Deep Learning

Abstract Conventional geological prediction methods, constrained by insufficient sampling density, fail to capture rock properties at each excavation cycle, leading to suboptimal performance of static blasting designs under variable rock conditions and compromising rock fragmentation efficiency and construction productivity. To overcome these limitations, this study establishes a relationship between drilling parameters and rock properties through field experiments and deep learning, enabling real-time inversion of rock strength and rock mass integrity, as well as dynamic adjustment of blast design. Core samples of 47 m from 10 excavation faces, representing 86 boreholes, were collected and tested for mechanical properties. For data homology, intelligent jumbos drilled within 2 m of sampling points while recording parameter curves. The dynamic time warping (DTW) algorithm resolved signal asynchrony between sensing and valve control parameters, increasing their correlation coefficients from 0.43–0.83 to 0.54–0.91. Based on the distinct responses of drilling parameters to rock strength and integrity, an long short-term memory (LSTM) network was developed for strength identification and a one-dimensional convolutional neural network (1D-CNN) for integrity recognition. The study found that 4 m·min −1 serves as an operational threshold for the on-site penetration rate: below this value, the energy required to maintain a given rate correlates positively with rock strength; above it, the hammer energy is underutilized for rock fragmentation, leading to reduced drilling quality and increased risk of rod sticking and breakage. As rock integrity deteriorates, drilling parameters exhibit a trend of decreasing mean values and increasing variance. Both developed models demonstrate robust overall performance. The DTW-LSTM/1D-CNN hybrid architecture not only effectively addresses the random and asynchronous issues of the drilling parameter curves in the field but also achieves synchronous and efficient prediction of rock strength and rock integrity through a dual-branch deep learning network. About 84.4% of strength predictions achieve errors below 10 MPa, and the accuracy for integrity recognition reaches 86.0%. It has good applicability and generalization for tunnel engineering with similar slate and rock properties and structural conditions. Based on the prediction results, rock mass grades can be classified, and hole count and spacing can be optimized, thereby improving rock fragmentation and slag removal efficiency and providing key technical support for intelligent tunnel blasting.