A Multi-Task Learning Framework with Physics Embedded for Signal Reconstruction and State Prediction in Underground...

Accurate forecasting of the collective state of hydraulic support groups is paramount to ensuring operational safety and optimizing efficiency in longwall coal mining. Existing monitoring methods face four critical bottlenecks. Contact sensors commonly employed for position prediction are highly vulnerable to harsh underground conditions, resulting in unreliable measurements. Vision-based non-contact alternatives are often hindered by labor-intensive manual data annotation requirements. Furthermore, occlusion of underground cameras by coal dust frequently results in missing spatio-temporal data, yet existing methods cannot reconstruct such missing signals. Moreover, most forecasting models either overlook the intrinsic spatio-temporal attributes of hydraulic support group behavior or unnecessarily complicate spatial modeling. Specifically, these models fail to recognize that hydraulic supports exhibit negligible spatial indistinguishability, resulting in redundant model complexity. To address these challenges, this paper proposes an integrated approach combining: (1) an unsupervised vision-based localization module for hydraulic support positioning; (2) comprehensive benchmarking of spatio-temporal models; (3) a Multi-Task Learning Framework with Physics Embedded (MTLPE) for simultaneous signal reconstruction and position prediction; and (4) an iterative learning strategy. Extensive ablation experiments validated MTLPE's superiority across four operational scenarios. The model consistently outperformed all baseline methods, achieving an RMSE of 0.43 mm, compared with the best-performing baseline of 0.48 mm. This work provides comprehensive insights and innovative approaches for signal reconstruction and position prediction of hydraulic support groups and other multi-rigid-body systems.