Abstract: TBM tunneling and advance borehole drilling are analogous processes that generate substantial rock–machine interaction data for perceiving ahead-of-face geology. To address dynamic rock mass classification during TBM construction, we analyzed the intrinsic characteristics of TBM operational and drilling parameters. Operational parameters reflect rock–machine coupling responses with high dimensionality and strong correlations yet relatively high signal-to-noise ratios. Drilling parameters directly indicate forward rock drillability, offering geological foresight but suffering from high noise and poor stability. A data processing workflow extracted multi-dimensional core feature subsets. Machine learning confirmed the theoretical feasibility of using drilling parameters alone for rock classification, but current technological constraints limit its accuracy, highlighting the need for multi-source fusion. We therefore developed Random Forest and Neural Network-based fusion models. Results show fusion models significantly outperform single-source models. Notably, an optimized concise feature model achieves optimal balance among accuracy, efficiency, and foresight by leveraging both parameter types. Multi-source data fusion, particularly with refined feature selection, offers an effective solution with strong engineering applicability for advancing TBM intelligent perception.