Topology-aware cross-modal subspace learning for vibration–acoustic rock hardness detection in integrated excavation–support–anchoring equipment

Paired vibration and acoustic signals share a drilling process but may exhibit different local feature geometries. This study proposes topology-aware cross-modal subspace learning (TACSL), which preserves shared granular memberships, constructs a regional graph with sample-proportion weights, and regularizes within-view covariance while retaining paired cross-view covariance. The evaluated first-acquisition archive contains 1785 feature pairs from three hardness classes, with one specimen per class. Historical frequency-coordinate descriptors are rescaled to the confirmed 10 kHz acquisition rate and two duplicate descriptors are removed. Across five segment-level splits, using 270 training and 25 test samples per class, TACSL achieves 85.87% ± 3.35% accuracy; five guarded ordered blocks yield 80.22% ± 11.29%. Most baseline contrasts and all partition-strategy contrasts show no significant differences after Holm correction under the disclosed search budgets. These results support recognition within the evaluated acquisition. Cross-acquisition transfer and specimen-independent generalization are not evaluated in the present experiments. The contribution is the synchronized regional representation and covariance-regularization formulation; incomplete waveform provenance, one specimen per class, and exploratory model development limit the interpretation of its performance.