Research And Implementation Of Smoke Fire Detection Algorithm For Flour Mills Based On Deep Learning

The flour industry is an indispensable and critical link in China’s wheat industry chain.Since there is a large a mount of dust in the air of the flour mill workshop,failure to detect early fires in a timely manner may lead to dust explosions,which are extremely dangerous.Currently,smoke fire detection in flour mills is mainly based on sensor technology,but the method has problems such as small detection range,low sensitivity,and susceptibility to interference from the surrounding environment,which leads to a high rate of false alarms and missed alarms.In order to resolve the difficulties of detecting flour mill fires,this paper implements smoke fire detection in flour mills based on deep learning technology.Firstly,the flour mill smoke and flame dataset is constructed,and the flour mill fire detection model based on SSA-YOLOXs is constructed through the comparison experiments of classical target detection networks,and different optimization attempts are made for feature extraction,feature fusion and boosting speed,etc.The optimized detection model can meet both detection accuracy and real-time performance.Finally,a flour mill smoke and flame detection system is developed based on the optimized detection model.The specific work is as follows:(1)Constructing a Flour Mill Smoke and Flame Dataset.To solve the problem of missing flour mill smoke and flame data sets,images were collected in three ways: by simulating flour mill fire scenes shot,downloading fire images of flour mill related scenes from the network,and adding fog to some of the data.A total of 7,000 flour mill smoke and flame images were collected.After that,the smoke and flames in these images are labeled using Label Img annotation software,and then the dataset is divided into training,validation,and test sets in a Pascal VOC format at a suitable scale.(2)A smoke and flame detection model based on SSA-YOLOXs for flour mills is proposed.To address the problem that small target smoke and flame are difficult to detect in the flour mill environment,the SE channel attention module is embedded in the residual structure of YOLOXs backbone network and used to enhance the feature representation capability.In view of the complex background of the flour mill scene and the existence of more objects similar to smoke and flame interfering with the detection,Swin-Transformer is introduced into the last layer of YOLOXs backbone network to enhance the feature extraction of global contextual information.Due to the variable scale of flour mill smoke and flame,the AFF module was used to replace the Concat operation in YOLOXs neck network to better fuse feature information from different size feature maps.To meet the requirement of real-time flour mill fire detection,the SPP module of YOLOXs algorithm is replaced with SPPF module to improve the detection speed of the model.The experimental results show that compared with the original YOLOXs,the average accuracy of SSA-YOLOXs detection model is increased by 2.4% to 88.75%,and the detection efficiency is 32 frames/second,which fulfills the demands of accuracy and timeliness in flour mill smoke and flame detection tasks,and has better robustness in the complex scenarios of flour mills.(3)SSA-YOLOXs smoke and flame detection algorithm is deployed to the web side to build a flour mill fire detection system for the actual needs of flour mill fire detection.The system can detect and alert smoke and flame in images and videos.The system mainly consists of smoke and flame detection,fire warning,abnormal frame interception,abnormal frame management and other functional modules.The experimental results verify that the flour mill fire detection system built in this paper has a good detection effect.