Early warning of underground conveyor belt breakage based on improved RT-DETR

Early warning of underground conveyor belt breakage faces challenges including the poor interference resistance of single-modality signals, difficulties in capturing early microdamage features, and high redundancy in multi-source heterogeneous data fusion. Existing multimodal fusion methods remain inadequate in jointly utilizing video, audio, operating condition, and environmental information during belt breakage evolution. To address these challenges, a Top-k cross-modal feature enhancement module was introduced into RT-DETR to construct a Top-k cross-modal feature-enhanced RT-DETR model. Temporal alignment at the data level and dimensional unification at the feature level were performed on the four types of information, establishing cross-modal associations anchored to visual features and addressing the asynchrony and heterogeneity of multi-source data. A Top-k sparse attention mechanism was introduced to dynamically select auxiliary features closely associated with breakage evolution during cross-modal interactions, effectively suppressing contamination of fused features by noise in complex underground environments. The enhanced multiscale visual features were fed into the RT-DETR detection framework to enable progressive state recognition from microdamage to precursors of breakage. A complete early-warning decision pipeline was then constructed, incorporating state classification, quantitative risk scoring, four-level warning threshold assessment, and temporal sliding-window voting for noise suppression. Experimental results showed that the model achieved an early-warning accuracy of 90.7% and an F₁ score of 0.893 on the test set, significantly outperforming comparison models such as the standard Transformer and YOLOv8. Compared with the original DETR framework, accuracy improved by 4.4%, inference time decreased by 40%, and the number of parameters decreased by approximately 20%. Accuracy remained above 88% under high dust concentrations and sharp temperature and humidity fluctuations. The model significantly improves inference efficiency while maintaining high detection accuracy and retains good early-warning stability under extreme operating conditions, providing effective technical support for intelligent operation and maintenance of belt conveyors under complex operating conditions.