Based On UAV And Infrared Image Processing Design Of Intelligent Inspection Algorithm For Heat Pipe Network

The thermal pipe network is the lifeblood of urban heating system.The normal operation of the thermal pipe network system guarantees the quality of urban heating.At present,if there is a fault in the heat network,it is often the workers carrying the leakage equipment to check along the heat pipeline,in order to determine the location of the leakage fault.Traditional inspection is time-consuming and laborious,and will increase the heat loss time.Compared with manual inspection,automatic inspection of UAV and infrared thermal imager can not only save human resources and reduce operation and maintenance costs,but also greatly improve the accuracy of leakage fault identification.In this thesis,the leakage problem of thermal pipe network was studied.Firstly,infrared heat source image data set was collected and made,and an improved Unet semantic segmentation algorithm was used to identify the categories,positions and contours of relevant heat sources in infrared images.Secondly,each hot well instance is obtained by clustering the classified hot well pixels.Again,the gray threshold adaptive watershed algorithm designed in this thesis is used to obtain the high temperature area.Finally,the area comparison determines whether the current inspection line has leakage fault.Firstly,the infrared image data set required for semantic segmentation network training is collected and annotated.The UAV flew along the set route to acquire infrared image data of the thermal pipe network in the form of taking photos and videos.The scope of inspection covers the streets,residential areas and the thermal pipelines laid in Chang qing Highway within the heating supply jurisdiction of Chang re Group.The pre-processed data were labeled,and the main heat sources in the figure were divided into four types of labels: hot well,street lamp,pedestrian and vehicle.Secondly,three mainstream semantic segmentation networks,Unet,PSPnet and Deeplabv3+,are selected to clarify their structure and characteristics,and a comparative experiment is conducted.After considering the total loss after training,m Io U(mean intersection over union),m PA(mean pixel accuracy)and other indexes,Unet was found to be the most suitable semantic segmentation network model for this subject.Thirdly,further improvement was made on the basis of Unet network model,the specific steps include: using transfer learning to train the network;The network structure was adjusted and VGG16 was used as the main trunk feature extraction network.A new learning rate adjustment strategy is designed as Adam optimizer plus cosine quasi-annealing algorithm.Adjust the input image size and add Dice loss term and Focal Loss term on the basis of Unet original loss function;The Convolutional Block Attention Module CBAM(Convolutional Block Attention Module)is introduced.Finally,the improved model CVGG-Unet was used for experiments.The results show that the average intersection ratio and average pixel accuracy of the improved model are significantly improved,and the image segmentation performance is further enhanced.Finally,different fault detection methods are designed for pipelines and well chambers in thermal pipe network.For the hot well chamber leakage fault,firstly,a single hot well target is separated by clustering algorithm,which is convenient for troubleshooting faults one by one.Secondly,the threshold adaptive watershed algorithm is designed to extract the high temperature area.Finally,the area is compared to determine whether there is a thermal leakage fault according to the ratio.For thermal pipeline leakage,an adaptive high temperature area alarm threshold is set according to the flight height of the UAV,and then the threshold adaptive watershed algorithm is applied to segment the area of the thermal diffusion area.If the area of the thermal diffusion area is greater than or equal to the alarm threshold,the leakage fault is judged in the inspection scene.