Research On Classification Method Of Pipeline Defect Fault Based On Mid-level Features

With the rapid development of China’s economy,the demand for oil and natural gas in industrial production has risen sharply,and the dependence on oil pipelines has become stronger.In the field of pipeline fault diagnosis,even small faults lead to expensive maintenance costs.In order to reduce pipeline failures and economic losses caused by faults,condition monitoring and fault diagnosis of subsea pipelines is increasingly important in reducing downtime and maintenance costs.At present,there are problems that small anomalies are difficult to find and many types of complex anomalies are difficult to identify in the field of submarine pipeline anomaly identification.In order to solve these problems,this thesis proposes a classification method for pipeline defect faults based on middle-level features,which mainly completes the following aspects:Firstly,aiming at the problem of low accuracy when using the existing defect fault classification method to classify small defects and the parts of defect,the middle layer feature extraction method based on BOW(Bag of Words)model is proposed.The model transforms the underlying features of the image into middle-level semantic features.Since the magnetic flux leakage data collected by the detector in the sea tube contains noise,the magnetic flux leakage data is preprocessed,including signal amplification,interpolation and filtering.Then,the magnetic flux leakage data is circled,and the data in the circle frame is mapped between 0 and 255 to form an abnormal data sample set.Secondly,aiming at the problem that the BOW model ignores the spatial position information between image feature points,a middle-level feature extraction method based on SPM(Spatial Pyramid Matching)model is proposed.The method can extract the spatial information between the feature points in the extracted middle layer feature,further enhance the expression power of the image content,thereby improving the accuracy of the abnormal classification.In the process of forming visual words,due to the problem that the amount of SIFT(Scale Invariant Feature Transform)feature vector to be clustered is large,resulting in memory overflow and low efficiency,a method of firstly clustering SIFT feature vectors in one image and recycling other images is proposed.The method can effectively solve the problem of insufficient memory and low clustering efficiency.Thirdly,aiming at the problem that the classification accuracy of the radial basis kernel function in SVM(Support Vector Machine)is constrained,a method of customizing the histogram cross kernel to be the kernel function of SVM and classifying the fault image samples of the pipeline is proposed.The kernel function can further enhance the expressive power of image semantics and increase the discrimination between different images,thereby improving the classification effect.The performance of fault classification method based on middle-level features is affected by multiple parameters.Simulation experiments are carried out under different combinations of parameters,so that the best classification accuracy can be obtained,and the variation law of classification accuracy under the influence of different parameters is studied.