Deep learning-driven multimodal early warning analysis for intelligent security in coal mine camps

Coal mine safety is crucial for both life and production.However, traditional monitoring relies on a single sensor, resulting in a high rate of missed alarms in complex underground environments.To address multiple challenges such as changes in light, dust interference, etc. this study proposes a deep learning early warning system that integrates video, infrared, and vibration data.Through cross-modal feature fusion and multi-task learning, it achieves collaborative perception of abnormal human behaviours and equipment failures.Experimental results show that the system achieves an area under the curve of 0.982 for abnormal behaviour detection on public datasets, which is approximately 7% higher than that of a single visual model; the accuracy of fire warning reaches 96.7%, and the false alarm rate is reduced by 5.3%.This method provides a highly reliable and scalable technical path for intelligent safety monitoring in coal mines around the clock.