A Conditional Dual-Branch Cross-Fusion Network with Cross-Modal Attention for Vibro-Acoustic Engine Fault Diagnosis

Reliable diagnosis of diesel-engine thermal and mechanical faults remains difficult when labelled samples are scarce, operating conditions vary, and measurements are strongly contaminated by noise. A conditional dual-branch cross-fusion network with cross-modal attention (CDCFN) is proposed for joint analysis of vibration and acoustic signals. Modality-specific convolutional modules extract fault-related features. A multimodal cross-attention mechanism applies temporal attention separately to the two modalities, concatenates the attended features for joint channel attention, and then separates the reweighted streams. A cross-modal conditional-decision module forms two directional decisions by using one modality to weight the other. CDCFN was evaluated on thermal- and mechanical-fault datasets using 50 training and 100 test sample pairs per class. Across the tested load conditions, the best mean accuracies were 98.23 ± 0.68% for CDCFN_V/A on the thermal-fault dataset and 97.99 ± 0.63% for CDCFN_A/V on the mechanical-fault dataset; the corresponding weighted F1 scores were 98.22% and 98.05%. Under matched Gaussian-noise training and testing at −12 dB, CDCFN_A/V achieved across-load mean accuracies of 94.89 ± 0.34% and 74.71 ± 1.45%, respectively. LayerCAM indicated attention to intake-valve seating, exhaust-valve seating, piston reversal, and combustion responses. These results support directional cross-modal attention for the tested small-sample, variable-load, and matched-noise settings.