The intelligent control of coal mines faces challenges of dynamic response lag and multi-source data fragmentation, and traditional models are difficult to capture transient anomalies underground and collaboratively analyze multimodal data. This article proposes a Multi Task Adaptive Architecture (MTA-Transformer) based on Transformer, which uses cross modal feature fusion and shared encoders to unify the modeling of equipment vibration, gas concentration, and other data, achieving multi-scale dynamic perception of the mining environment and solving the problems of dynamic monitoring and risk early warning of the mining environment. Experiments have shown that in bearing fault detection tasks, the accuracy of this model reaches 93.5%, with a false alarm rate (FAR) of 2.0%, and the response time is within 5ms, which is a significant improvement compared to traditional models; The NRMSE for predicting gas concentration is 7.83%, the coverage probability of the prediction interval (PICP) is 91.7%, and the advance warning time can reach 6 hours. MTA Transformer provides a feasible technological paradigm for the intelligent construction of mines.