Automated Leakage Detection Of Crude Oil Transmission Pipes Using Multi-Feature Data Fusion

Crude oil,as “the fuel of industry,” is an important mean in ensuring economic development and national security.At present,China has become the largest crude oil importer in the world.The safety of crude oil transportation has become increasingly crucial,which has attracted widespread attention from the world.As a distribution method over long distances,pipeline networks have many advantages: large capacity and rate,low cost,stable and safe transportation process.The use of pipelines has become the main crude oil transportation method in China.Inevitably,crude oil may leak due to natural reasons or man-made damage during pipeline transportation,thereby threatening the safety of pipeline transportation and the environment.Specifically,the leakage in crude oil transportation pipelines will not only cause major economic loss,but also will threaten the environment and human health.Therefore,the leakage detection for crude oil transmission pipelines is of great significance to the transportation industry.Traditional detection methods are rely heavily on the manual inspection to determine the leakage level through the workers' experience.However,visual judgment is subjective,and this method also consumes extensive human resources and cost.Furthermore,time delay and undetected critical leakage with manual inspection can pose serious problems.Therefore,proposing an accurate and automated method for leakage detection of crude oil transportation pipelines is of great significance to the crude oil industry.The main objective of the present project and the thesis is to design an automated leakage level assessment method for crude oil transportation pipeline s.A leakage detection method using automatically acquired red-green-blue(RGB)images and thermal images,and subsequent image processing,feature extraction,feature fusion,with the use of neural networks for the detection and establishing the leakage level is proposed.Using an experimental setup it is verified that the developed methodology improves the accuracy of the leakage classification in a crude oil transportation pipeline and greatly reduces labor costs.Because there is no public dataset for leakage data,for the verification of the developed methodology,we designed a simulation experiment platform to obtain RGB images and thermal images under different leakage levels(large leak,moderate leak,minor leak,no leak).The developed experiment platform is composed of a fire hydrant and a pipeline with welded joints and threaded joints.The experiment platform has three kinds of joints: flange joints,welded joints and threaded joints.In order to simulate the heating process in crude oil before transportation,we use a water bath heating device to heat the oil to a specified temperature.We also placed a heating belt inside the pipe,and heated the outer surface of the pipe as well.We have carried out a secondary development with shooting software for the original thermal images to realize synchronous acquisition.We have used a UR5 robotic arm as an adjustable camera support structure,and adopted the trial teaching mode of the robotic arm to acquire leakage images from different angles.Due to the phenomenon of stress concentration,pipeline leakage often occurs at the joints of the pipes.To simulate this condition of leakage,we poured oil with different volumes and temperatures onto the pipe joints,and acquired images simultaneously.Then we carried out a series of preprocessing operation on the RGB images(since the resolution of thermal images is low,thermal images are not sensitive to light),including: image denoising,removing illumination influence,image enhancement,image cropping and stitching.The image preprocessing method are performed in MATLAB to verify the effectiveness of the proposed preprocessing method.In order to realize the automated leakage level assessment o f crude oil transmission pipes,we introduced three classification methods based on fuzzy neural network(FNN)and convolutional neural network(CNN).In the FNN,we first convert the RGB color model of RGB image and thermal images into HSV color model,and then extract the color moment features from the two types of images as the input to the FNN.The FNN output is the leakage level of the input sample.The accuracy of the FNN method was found to be only about 67%.Compared with FNN,convolutional neural networks are more widely used in the field of image classification.We proposed two feature fusion methods using CNN : end feature fusion and feature map cross fusion.Both fusion methods in CNN can realize the fusion of the information obtained from RGB images and thermal images.Two network paths have been designed for the feature fusion method,RGB-Network and Thermal-Network.For the end feature fusion method,we have designed a fusion layer after the Softmax classifiers of the two network paths.The results from the two classifiers are added to form the output of the fusion layer(the output is a 4-dimensional feature vector).In the feature map cross fusion method,Thermal-Network provides the pooled feature maps to the RGB-Network,to achieve the fusion of feature maps.The structure and the parameter selection of the two network paths are elaborated in C hapter 3.In order to improve the classification accuracy of the CNN,we used two types of classifiers in the cross feature map fusion method: Softmax classifier and Support Vector Machine(SVM)classifier,and we also introduced two common attention mechanisms into CNN: “Squeeze and Excitation Block ”(SE)and “Convolutional Block Attention Module”(CBAM),which enable the network to pay more attention to important features.In order to verify the network structure that has been developed in the present work,we built the mentioned convolutional neural network based on the Pytorch environment,and divided the dataset into a training dataset,a validation dataset and a testing dataset,in the ratio 6: 2: 2.Finally,through experimental verification,the accuracy of the end feature fusion method was established.An accuracy level of over96% has been achieved in this manner.In the cross feature map fusion method,the accuracy of the CNN+Softmax(with CBAM)method reached about 98%,and the the accuracy of the CNN+SVM(with CBAM)method was about 96%.Experimental results have verified that the two proposed CNN-based feature fusion methods have clear advantages over FNN and the classic convolutional neural networks.The dual-channel network architecture is found to solve the problem that a single image lacks enough features,and the feature fusion method has greatly improved the classification accuracy.In summary,the experimental results have verified the effectiveness and feasibility of the crude oil pipeline leakage classification method proposed in this thesis.