Classification of Intact Rock Uniaxial Compressive Strength Using Measurement-While-Drilling Data from Tunnel Drilling Jumbos: Method Development and Field Validation on the Sichuan–Tibet Railway

Identifying rock conditions ahead of a tunnel face is essential for drilling-parameter optimization, construction safety, and support design. Measurement-while-drilling (MWD) data from drilling jumbos offer a continuous, low-cost basis for this task, but non-steady operational states introduce strong disturbances. Based on the XCMG XTD3133C drilling jumbo, this study develops an MWD-based method for classifying intact-rock uniaxial compressive strength (UCS) grade. Bench-scale and engineering field data are assembled into an annotated database linking UCS results to borehole positions, and the two datasets are modelled and evaluated separately. Non-steady drilling states are identified and removed using physical correlations among pressure, rate, and speed signals, and the penetration-rate signal is smoothed by a moving average. Long short-term memory (LSTM) and deep neural network (DNN) classifiers are compared across input-parameter combinations, with the borehole as the unit of data partitioning. LSTM-3 and DNN-3 reach accuracies of 0.957 and 0.983 on bench-scale data, and 0.917 and 0.928 on field data from a Sichuan–Tibet Railway section, with balanced accuracies of 0.914 and 0.927. Considering acquisition cost, model complexity, and deployment stability, DNN-3—using three pressures and the penetration rate—is selected as the preferred candidate for on-board deployment.