M²TRANS: A Real-Time Edge-Based Multimodal Multi-Task Transformer for Explainable Hazard Detection and Predictive Risk Analytics in Underground Mining Systems

Underground mining remains one of the most hazardous industrial environments due to low visibility, equipment congestion, unstable geotechnical conditions, toxic gas exposure, dust, heat, and the continuous interaction between human operators and heavy machinery. Conventional safety monitoring systems are often siloed, reactive, and unable to jointly reason across heterogeneous sensor streams in real time. This paper proposes M²Trans, a real-time edge-based multimodal multi-task transformer for explainable hazard detection and predictive risk analytics in underground mining systems. The model integrates video, thermal imagery, acoustic signatures, gas concentration measurements, vibration signals, environmental telemetry, and worker-location data into a unified transformer-based architecture deployed on edge computing devices for low-latency inference. M²Trans jointly performs three safety-critical tasks: hazard detection, hazard severity classification, and short-horizon predictive risk forecasting. To improve trust and operational adoption, the framework includes an explainability layer combining cross-modal attention visualization, feature attribution, and rule-grounded textual rationale generation for mine safety supervisors. A lightweight edge optimization pipeline based on model pruning, quantization, and asynchronous sensor fusion enables near-real-time performance under constrained underground connectivity. The paper presents the conceptual architecture, mathematical formulation, deployment workflow, and evaluation protocol of M²Trans, and reports illustrative experimental results on a multimodal underground mining benchmark assembled from synchronized simulated and field-inspired sensor streams. The framework demonstrates the potential of transformer-based multimodal intelligence to shift mine safety from delayed incident response to proactive and explainable risk prevention.