Research On Fire Alarm System Based On Deep Learning

Fires pose a serious threat to human life and cause significant economic losses.To prevent fires,fire detection and alarm have always been a key issue in fire-fighting work.Traditional fire alarm systems rely on sensors that need to be activated at close range and require human intervention to confirm the presence of a fire.These systems are not suitable for critical environments and have limited detection distance,slow response times,and low accuracy.With the widespread application of deep learning in video surveillance,fire detection technology has evolved from traditional image processing methods to deep learning-based methods.However,there are still issues with low detection accuracy,high miss rates,poor real-time performance,and missing alarm information,and there is a lack of a complete and feasible automatic fire alarm system.In response to the low detection accuracy and high miss rates for small flame targets in fire detection models,a small flame target detection head was added to the original YOLOv5 network to focus on extracting features of small targets.A 4×4 detection head flame feature fusion structure was designed based on the ASFF mechanism.In addition,the YOLOv5 Conv module was improved with the SPD module to enhance detection accuracy for low-resolution and small images.Finally,the overall performance of the flame detector was improved using the SA attention mechanism and the WIo U loss function.Experimental results show that the performance of the YOLO-SSA model is better than that of the original YOLOv5 model and other mainstream object detection algorithms,achieving an average precision of 78.35%.To address issues such as missing alarm information and poor real-time performance in traditional fire alarm systems,a lightweight network model was designed for assisting multi-object extraction tasks at fire scenes.The on-site information of the fire was transformed into text and fed back to the firefighting system as auxiliary alarm information.The complexity of the model was greatly reduced by using lightweight Ghost modules and lightweight general upsampling factor CARAFE,and the model efficiency was improved by pruning the model.The proposed lightweight YOLO-GC model was validated on the COCO128 dataset and achieved an average precision of 88.63% for multi-object extraction tasks,with a 74% reduction in parameter volume and a 71%reduction in computational complexity compared to YOLOv5s.In response to the inadequate functionality of existing fire alarm systems,a desktop fire monitoring software,Android client application,and web page have been developed with features such as image detection,video file detection,and real-time video stream detection.In order to improve the efficiency of fire dispatching and simplify the dispatching process,a front-end and back-end separation system based on Nodejs and Vue was designed to improve firefighting tasks.This achieved an intelligent fire alarm system that combines intelligent detection,intelligent alarm,intelligent reception,and intelligent dispatch,completing a truly "smart firefighting" solution.