Abstract: In this paper, a smoldering fire classification and prediction method based on the fusion of infrared images and environmental information with a convolutional neural network (CNN) is proposed to address the issues of low detection efficiency and accuracy that are common in currently used smoldering fire detection methods. Firstly, environmental information and infrared images of experimental samples were obtained in real—time, and the rate of area change, the rate of perimeter change, circularity and target movement characteristics were used as the characteristic parameters for the discrimination basis of smoldering fire detection; Secondly, the importance of the four image feature parameters was evaluated using a Random Forest algorithm (RF), and weight allocation was assigned based on their respective importance. Finally, image feature fusion was performed based on the assigned weights, and the fused information was related to environmental information as input parameters for training and testing of the CNN. The results indicate a significant improvement in the early detection efficiency and overall detection accuracy of smoldering fires using the proposed method. The recall rate of smoldering fire detection increases by 65% during the 15 seconds before the occurrence of smoldering fire, while the detection accuracy for the entire detection period is improved by 6.25%. The research provide a new approach for early warning of smoldering fires.

Key words: smoldering fire detection, infrared image, convolutional neural network, random forest algorithm