Research On Early Warning Method Of Plant Fire Based On Deep Learning

Traditional fire alarm sensors are widely used in various scenarios due to their low price and easy installation.However,the complexity of the industrial plant environment makes it difficult for ordinary sensors to function.At the same time,the detection and early warning method based on fire image processing usually needs to manually select the characteristics of flame or smoke for analysis,which has the disadvantages of difficult feature extraction and poor real-time performance.It can be seen that the current traditional fire detection and early warning methods do not have high reliability in the complex environment of the plant building.Therefore,this paper conducts an in-depth study on the fire early warning method in the plant environment based on deep learning,which mainly includes the following three parts:(1)Improvement of model accuracy based on YOLOv5 s algorithm.First,the model structure is adjusted,and the channel adjustment and matching CA attention mechanism is introduced before the SPPF module of the model Backbone layer to reduce the influence of complex backgrounds on the target;second,based on the idea of weighted bidirectional structure,Bi FPN＿Add multi-scale is proposed Feature fusion method,which assigns matching weights to features with different contributions to improve the fusion effect of multi-scale features of the model.Third,in response to the situation where similar target objects are easily missed when they are close to each other in complex factory backgrounds,the Soft＿NMS algorithm is used to process the boundary boxes in the final stage of the model.By combining the above three improvement methods,the CBS+YOLOv5s model is proposed and experimentally evaluated.The experimental results show that the above three independent improvement methods have certain improvement effects on the algorithm accuracy.The m AP value of the CBS+YOLOv5s fire detection and warning model with comprehensive improvement is increased by 1.91% compared with the original YOLOv5 model,and the Precision and Recall values are also increased by 1.21% and 0.09%,respectively.This indicates that the CBS+YOLOv5s algorithm proposed in this paper can more accurately detect and label the location of the warning fire occurrence.(2)Selection of lightweight solutions.Mobile Net V3 and Ghost Net,two lightweight networks,improve the original YOLOv5 s model,and evaluate the effectiveness of lightweight solutions.Experimental comparisons on the same dataset show that although both networks significantly reduce model computational complexity,the performance of the Ghost Net network is superior.Compared to the original YOLOv5 s,the Ghost Net solution improves the FPS value by 33.63% and reduces the computational complexity by 48.72%,indicating the feasibility of the Ghost Net solution.(3)Propose CBSG+YOLOv5s plant fire detection and early warning algorithm.In order to improve the portability of the CBS+YOLOv5s algorithm,the Ghost Net network is used to carry out lightweight processing on the CBS+YOLOv5s algorithm to achieve a relative balance between the model detection accuracy and the calculation amount.The experimental results show that compared with the CBS+YOLOv5s algorithm with a large amount of calculation,the CBSG+YOLOv5s algorithm proposed in Chapter 4 of this paper can reduce the calculation amount of the model by 50% in the case of less m AP loss,and the weight reduction effect is very significant.The portability of the model is greatly enhanced.In summary,regarding for fire detection and early warning in plant environment,the two methods proposed in this paper have shown improvements in different optimization directions compared with the original YOLOv5 s algorithm,and can better achieve the fire detection and warning in different plant environment.