Fire Detection System In Photovoltaic Power Plant Based On Video Monitoring

With the vigorous development of the photovoltaic industry,operation and maintenance of photovoltaic power stations have become increasingly important.Fire is one of the common problems of photovoltaic power plants due to some factors such as component damage and hot spots.Therefore,research on fire detection systems for photovoltaic power plants based on video surveillance is imperative.This thesis takes fire video images as research objects,and studies several different fire detection methods to reduce the rate of missed detection and false detection of fire detection.The thesis first discusses the research background and significance of this topic,summarizes the fire detection methods from three aspects of traditional image segmentation,traditional machine learning and deep learning,and analyzes the limitations of existing methods.Then,for these limitations and photovoltaic power plant scenarios,three different fire detection methods are proposed.Finally,the hardware and software parts of the fire detection system are designed and implemented,and several fire detection algorithms are integrated into the software system.For video image preprocessing,I improve the traditional image repairment algorithm Criminisi to improve the accuracy and completeness of the repairment.Then,for the noise appearing in the video image,three filters of mean filtering,median filtering and Gaussian filtering are discussed separately.The experiment proves the effectiveness of median filtering on salt and pepper noise.Finally,in order to improve the resolution of video images and improve the sharpness accordingly,an image enhancement algorithm based on fusion interpolation and spatial domain is proposed.For the fire detection method based on traditional image segmentation,the motion foreground is segmented first,and the Vi Be algorithm is mainly used,and Vi Be is improved in two aspects: removing dynamic shadows,and complementing the lack of motion areas.Then,the smoke and flame are further segmented according to the color and diffusion characteristics of the smoke,and the color and flicker characteristics of the flame,respectively,so as to realize the detection of fire.Finally,the area of the flame is measured by the method of pixel statistics to measure the severity of the fire.For the fire detection method based on dynamic and static feature fusion,the static features of smoke are first extracted,including color feature,HOG feature and static texture feature.Based on the traditional LBP operator,a LBP operator combined with Laplace transform is proposed to reduce the dimension of the static texture feature vector.Then,to make full use of the timeline information of moving smoke,a three-dimensional LBP operator based on the space vector is proposed inspired by the VLBP operator.Compared with the two-dimensional LBP operator,the extracted texture features are more abundant.Aiming at a variety of feature dimensions,a local linear embedding(LLE)algorithm is used to achieve feature dimensionality reduction.These features are fused,and SVM is used to classify them.Aiming at the combined detection method of smoke and flame based on deep learning,based on investigating the application of deep learning in video image,target detection and recognition,especially fire detection,a method of combined detection of smoke and flame based on improved Faster RCNN is proposed.First of all,a data set for fire detection in photovoltaic power plant scenarios is collected and produced,which is divided into three categories: smoke,flame and background.Then,the deep learning method is used to suppress the complex interference problems in the photovoltaic power plant scene,such as the effect of smog.Finally,the original Faster RCNN structure is improved to be suitable for smoke and flame detection,and to improve detection rate.