In this paper, an intelligent fire identification and response system based on high-precision infrared thermal imaging is proposed. Build a "cloud-edge-end" collaborative architecture, deploy lightweight algorithms at the edge to achieve second-level response, and conduct big data fusion analysis in the cloud. Core algorithmic innovations include: (1) designing a multi-feature fusion deep learning model that concurrently extracts temperature flow, spatial gradient flow, and temporal differential flow features, while incorporating a Convolutional Block Attention Module (CBAM) to enhance perception of critical regions; (2) the spatio-temporal context verification mechanism is introduced, and the spatial semantic constraints are combined with the three-dimensional model of the facility, and the sliding window time sequence criterion is used to filter out the instantaneous interference. The experiment builds 12,850 video clip data sets based on the real scene of natural gas treatment plant. The results show that the recall rate of the proposed method is 98.8%, the accuracy rate is 96.5%, the F1-Score is 0.976, the average false alarm rate is reduced to 0.7 times/day, and the median alarm delay is only 1.5s The fire detection has successfully leapt from "minute level" to "second level", which is a digital transformation for the safety management of oil and gas facilities.
Chen et al. (Sun,) studied this question.