Research And Application Of Fire Detection Algorithm For Ammonia Chemical Plants Based On Improved YOLOv3

In recent years,the occurrence of fires in chemical companies has increased,posing significant dangers to people’s production and daily life.Therefore,the prevention and detection of fires in hazardous chemical products are crucial for both the economic development of the country and the safety of the population.Various fire warning methods have emerged,and the use of convolutional neural networks(CNNs)has become an important tool in image classification,target recognition,and tracking.Therefore,applying convolutional neural network algorithms to fire detection helps to improve its target recognition and anti-interference capability in complex environments.This paper aims to explore the application of deep learning target detection algorithms to detect and prevent fires in hazardous chemical products,specifically in synthetic ammonia chemical plants.The goal is to develop accurate and efficient real-time fire detection techniques,providing theoretical support for their promotion and application in practice.The main research content and innovative contributions are as follows:Firstly,lightweight model design: The structure of the CNN-based target recognition algorithm is complex and computationally intensive,making it unsuitable for real-time detection on devices with limited resources.To address this issue,this paper introduces a lightweight Mobile Netv3 network as a replacement for the original YOLOv3 backbone network Darknet53,significantly reducing the amount of model computation.Secondly,optimization of small target detection: the detection of smoke and flame in the complex environment of a fire scene is subject to false alarms,false positives,and poor accuracy,affecting the detection of small targets.In this paper,the fire detection problem is transformed into a coordinate regression problem,and the K-means++algorithm is used for re-clustering.Methods such as focusing and pyramid pooling are introduced to improve the accuracy of the algorithm and identify smoke and flames more effectively during fires.Thirdly,optimization of border regression accuracy: the accuracy of the border regression in the target detection algorithm significantly impacts the performance of the final detection algorithm.This paper addresses the limitations of IOU and GIOU loss functions in non-overlapping situations and parallel and perpendicular border positions by using the DIOU loss function.This function directly optimizes the center distance between two frames by adjusting their distance,greatly improving the accuracy and convergence speed of the border regression.Fourthly,development of fire detection system for synthetic ammonia chemical plants.Based on the improved fire detection algorithm proposed in this paper,a fire detection system for ammonia chemical plants is designed and produced using Py Qt5.The system includes a login function,fire data loading function,fire detection identification display function,camera control part,data management part,camera parameter setting module,and fire alarm function.The system was tested for functionality and the results showed that the system can achieve effective detection of fire targets.Finally,comparison tests are set up based on the homemade fire dataset VOC＿Img2023.The experimental results show that the improved algorithm performs well in real-time and accuracy,especially in the detection of small targets in complex target contexts.This paper designs a fire detection system based on Py Qt5 that can be effectively applied to ammonia chemical plants to accomplish fire detection tasks.