Research On Fire Detection Of Wind Turbine Engine Room Based On Image Recognition

Speeding up the development of wind power is one of the major measures to implement the deployment of the CPC Central Committee and The State Council,support the realization of the target task of "carbon peak and carbon neutrality" as scheduled,and promote the clean and low-carbon energy transformation.While vigorously promoting the construction of wind power,safety accidents also occur frequently,among which the fire in the wind power engine room accounts for a certain proportion in the statistics of safety accidents.If it cannot be discovered and solved in time,There will be significant economic losses.In view of the above problems,this paper mainly studies the fire detection algorithm of wind power engine room based on image recognition,and mainly completes the following research content.Firstly,the fire controllability analysis and image preprocessing in the wind power engine room are completed.For the purpose of research,the data set is made for the initial fire in the simulated wind turbine engine room.The original fire image may have some problems such as blur or image missing.The smoke and flame image of the training set is preprocessed.The means used include color space conversion,coordinate transformation,gray transformation,gamma transformation and image filtering,to realize the image convenient detection.Then,the flame smoke detection method and experimental results in the wind turbine engine room are analyzed.This paper discusses the traditional fire detection algorithm and deep learning algorithm based on image processing,introduces the detection process and principle of several algorithms such as RCNN,SSD and YOLO,and analyzes the advantages and disadvantages of these deep learning algorithms in the actual fire detection process according to the comparison of their performance indicators on Pascal VOC2007 fire data set.The YOLO algorithm was determined to be more in line with the performance requirements of the required speed and accuracy,and the YOLOv5s detection model was used to discuss the training results of flame smoke data set,and the evaluation indexes in the training process were analyzed.Finally,the performance of the YOLOv5s model was improved by adding the 3-D attention-mechanism module SimAM and the lightweight network FasterNet.In order to further improve the speed of flame smoke detection and the performance of flame and smoke feature extraction,we proposed to add attention module and improve the network structure to further improve the speed.Finally,we proposed a YOLOv5s model combining the 3-D attention mechanism module SimAM and lightweight network FasterNet.Based on the early fire data set in the simulated wind turbine engine room,various performance of four models,YOLOv5s,SimAM,C3-Faster and SimAM+C3-Faster,were analyzed and compared through experimental comparison,so as to prove the improvement of the expected function.The experimental results show that the addition of the 3-D attention mechanism SimAM can improve the accurate locking of the detection target in the actual detection process,but improve the model reasoning speed,frames per second and flexibility.Although there is a small loss in the loss function,the FPS value of the YOLOv5s model combining the 3-D attention mechanism SimAM and the lightweight network FasterNet increases by 0.3141 frames/s and the mAP₀.5 value increases by 2.384%compared with the original YOLOv5s model.At the same time,the detection accuracy of flame and smoke is higher,and the error detection and leakage detection are reduced,which is more in line with the experimental expectation.