Abstract: Underground coal-mine construction presents a complex environment in which direct mapping between downhole drilling-tool responses and lithology is difficult, limiting the accuracy of lithology identification. To address this challenge, we developed a polycrystalline diamond compact (PDC) bit-drill-rod rock-breaking simulation model in Abaqus and analyzed tool response characteristics across composite strata under different drilling conditions. With Rate of Penetration (ROP) held constant, increasing revolutions per minute (RPM) reduced bit torque and its fluctuation while leaving weight on bit (WOB) fluctuations relatively unchanged; with RPM held constant, increasing ROP raised both the mean WOB and torque. The tri-axial vibration acceleration of the drill rod increased with RPM, whereas with increasing ROP it showed a nonmonotonic trend (initial decrease followed by increase). All of these parameters display abrupt changes at lithologic interfaces, indicating that their joint use can serve as effective features for lithology recognition. On this basis, we propose a lithology recognition model—BiLSTM-Att-KAN, which integrates a Bidirectional Long Short-Term Memory Network (BiLSTM), an attention mechanism, and a Kolmogorov-Arnold Network (KAN). Replacing the conventional Multilayer Perceptron (MLP) classifier with KAN overcomes the nonlinear-mapping limitations imposed by fixed activation functions (e.g., Sigmoid/ Rectified Linear Unit (ReLU)) in MLPs and, in theory, enhances recognition of complex lithologies. Tests using both simulated and field drilling data show that BiLSTM-Att-KAN achieves 91.64% accuracy on the simulation dataset, outperforming traditional classifiers such as k-nearest neighbors and random forest by 7.00%～22.08%. On real borehole data, the model yields accuracies of 86.81% on the training-test split and 85.28% on an independent test set, validating its effectiveness and demonstrating its potential as a technical tool for lithology identification in coal-bearing strata.