CORRECT-Net: A Multimodal Vibration–Current Fusion Network for Coal–Rock Cutting State Recognition in Shearers

Coal–rock cutting state recognition is essential for adaptive cutting and intelligent speed regulation in shearers. To address the limited representational capability of individual signals, the confusion between adjacent gangue-bearing cutting conditions, and the domain discrepancy between simulation and experimental data, a vibration–current multimodal fusion method based on CORRECT-Net is proposed. First, an EDEM–RecurDyn–MATLAB/Simulink co-simulation system was developed to generate cutting records for four coal–rock states. After screening for physical equivalence and label conflicts, 158 valid records were retained and grouped into 150 physical-condition groups, which were partitioned at the group level into training, validation, and test sets. Subsequently, SincNet was employed to extract frequency-band-constrained features, a Transformer was used to model long-range temporal dependencies, and a residual importance-guided GATv2 module was introduced to perform cross-modal fusion of vibration-impact and current-load features. On 2500 test windows, CORRECT-Net achieved an accuracy of 96.20% ± 0.11%, a macro-F1 score of 95.14% ± 0.21%, and a hazardous-condition miss rate of 0.58% ± 0.13%. Compared with the multimodal 1D-CNN, TCN, and Bi-LSTM models, CORRECT-Net improved the accuracy by 8.80, 4.00, and 2.08 percentage points, respectively. In the progressive ablation study, the accuracy increased from 87.40% ± 0.26% to 96.20% ± 0.11%, while the macro-F1 score increased from 84.57% ± 0.34% to 95.14% ± 0.21%. Under Gaussian noise with a standard deviation of 0.05, the model retained an accuracy of 92.76% ± 0.24%. When the vibration and current modalities were separately unavailable, the corresponding accuracies were 86.56% ± 0.37% and 92.44% ± 0.25%, respectively. A five-fold simulation-to-experiment transfer evaluation was further conducted at the independent-run level using five experimental records per class. Without adaptation using experimental samples, the model achieved an accuracy of 91.33% ± 5.19%. When 20% and 50% of the experimental windows were used for adaptation, the accuracy increased to 96.33% ± 0.75% and 98.67% ± 1.39%, respectively. These results demonstrate that CORRECT-Net effectively integrates mechanical vibration responses and motor-load information and, under the present simulation and experimental conditions, achieves high recognition accuracy, a low hazardous-condition miss rate, and effective adaptability to the experimental domain.