Enhancing predictive safety monitoring in longwall mining via interpretable multi-scale data fusion

The predictive performance of safety monitoring systems in longwall mining faces is frequently compromised by complex, non-linear environmental noise and dynamic extraction processes. While conventional time-series decomposition techniques can mitigate sensor noise to extract salient temporal patterns, they frequently underperform in capturing the complex multivariate interactions among distributed sensors, thereby limiting the effectiveness of the monitoring system. To alleviate this performance bottleneck, this study proposes an integrated forecasting framework that combines multi-scale data decomposition with a self-attention-based fusion mechanism. This hybrid approach dynamically synthesises diverse temporal scales and multivariate feature interactions, contributing to the system’s resilience against volatile data fluctuations. Rigorous evaluations using a canonical public dataset from the Upper Silesian coal basin reveal that, across a comprehensive suite of architectures (including Encoder-Decoder LSTM/GRU, standard Transformer, Informer, and Autoformer), models trained on the fused decomposed features generally exhibit enhanced forecasting robustness and sustained predictive stability across most temporal configurations. Furthermore, SHapley Additive exPlanations (SHAP) are integrated to provide feature-level transparency. The interpretability analysis demonstrates that the self-attention mechanism effectively assigns higher weights to underemphasised, distant sensor inputs critical for robust forecasting, compensating for the limitations of traditional decomposition methods in capturing complex multivariate dependencies. By facilitating both high predictive fidelity and transparent reasoning, this framework offers a robust analytical foundation for proactive hazard mitigation and advanced safety monitoring in underground mining.