<p>Accurate real-time classification of tunnel surrounding rock is essential for construction safety and process control. This study develops a deep learning framework to infer rock-mass grades from measurement-while-drilling (MWD) data recorded by an intelligent drilling jumbo. Drilling sequences from 4,891 segments in two tunnels are first cleaned using Kalman filtering and interquartile range rules, then normalised on a per-borehole basis. A sliding-window strategy constructs fixed-length sequences, and a feed-speed variance (FSV) feature, computed as a short-window median absolute deviation, is introduced to characterise local fluctuations induced by stratigraphic transitions. A hybrid Convolutional Neural Network–Long Short-Term Memory (CNN–LSTM) model is designed to jointly extract spatial patterns and temporal dependencies. On the held-out test set, the proposed model attains an overall accuracy of 97.3%, with grade-wise accuracies of 97.3%, 98.2% and 96.8% for surrounding rock classes II, III and IV, respectively. Compared with eleven representative machine-learning and deep-learning baselines, including Temporal Convolutional Network (TCN), Long Short-Term Memory–Fully Convolutional Network (LSTM–FCN) and a transformer-based model, the CNN–LSTM improves accuracy by about 6–7% points under optimised settings. An ablation study further shows that the dynamic feed-speed variance feature provides a consistent improvement in classification performance when it is kept temporally aligned with the drilling parameters. The proposed workflow offers a practical basis for real-time geological perception in mechanised tunnelling.</p>