Research On Pipeline Defects Intelligent Identification Based On Convolutional Neural Network With Small Sample

Pipeline is widely used as an important role of transportation tool,and its safe operation and management is one of the key tasks.Especially,urban pipeline is routed to the residential areas closely,its integration would affect the quality of life and the safety of residents seriously.Unfortunately,the explosion of urban gas pipeline often causes casualties,the leakage of urban drainage pipeline will affect civilian quality of life,which not only causes a waste of resources,but also threatens to human life and health.Therefore,it is necessary the adopt pipeline internal inspection device to detect the pipeline defects regularly.In this paper,an intelligent pipeline defect identification technology based on Convolutional Neural Network(CNN)is proposed in conditions of small samples,small pipeline inner diameter(D<15cm)and low detection cost as well.This technology is not limited by the pipeline material and its usage environment,the internal video transmitted by the camera can be recognized automatically,and then identify the pipeline defect categories.In addition,this technology can mark the defect area in the pipeline video screen,which could improve the efficiency of the pipeline defects inspection,reduce the cost of human labor,and enhance the functions of pipeline safety operation management system greatly.First of all,this paper analyses the advantages and disadvantages of current pipeline detection methods in depth,builds a CNN model for intelligent identification of pipeline defects for urban small-diameter pipelines,and designs an overall implementation scheme.Moreover,the pipeline detection technology based on CNN is summarized,the CNN and deep learning framework are introduced,and the comparative analysis of the usually utilized target detection algorithms at the current stage is elaborated.These lay the foundation for the in-depth research of the contents of this paper.Secondly,the pipeline defect data platform is acquired and built,which is convenient for the development of pipeline defects detection technology.At the same time,in order to train the neural network model of pipeline defect identification efficiently,a training environment is set up.In this paper,the neural network model of pipeline defect intelligent recognition based on YOLOv3 target recognition algorithm is trained,and the model is verified and evaluated by pictures from pipeline defect data platform.Next,the CNN model of pipeline defects recognition is optimized for the characteristics of the target recognition task with a small number of pipeline defect data sets.In this research,Contrast Limited Adaptive Histogram Equalization(CLAHE)is equalized.The project completed the expansion of the data set by using the method of filling the image background.The paper also introduced the Relation-aware Global Attention(RGA)of correlation into the CNN of pipeline defect recognition,and optimized the loss function of the bounding box of the recognition model.Finally,the optimized CNN based intelligent identification model for small sample pipeline defects is evaluated and verified.A series of experiments are conducted to evaluate its detection performance and robustness.The pipeline defect identification accuracy of the optimized model is reached over 85%for small samples during the final verification experiment,which satisfy the accuracy criterions of pipeline industry in small sample defects identification.