Study On Intelligent Recognition Of Pipeline Magnetic Flux Leakage Signal Based On Deep Learning

The main way of energy transportation such as Petroleum,Natural gas is pipeline transport.Pipeline leakage accidents occur frequently due to corrosion,abrasion and accidental injury,which not only cause important directly or indirectly economic losses,but also cause serious environmental pollution.Internal detection method of magnetic flux leakage(MFL)of pipeline is one of the most effective detection methods for pipeline transportation of petroleum and natural gas,so pipeline can be detected on a regular time by pipeline MFL internal detector,and potential flaws can be found and repaired in time by recognizing and analyzing MFL signals that is collected by internal measurement detector.Therefore,it plays an important role in long oil and gas pipeline.Currently,recognition and analysis of pipeline MFL data mostly adopt manual interpretation.However,the pipeline laying distance of petroleum and natural gas pipelines is typically coming with huge amount of MFL internal detection data in many types.Thus,there are some problems in identification and analysis,such as time-consuming,missed detection,false detection and so on.At present,automatic recognition methods for pipeline MFL signals are still on its exploring stage.Consequently,it is a hot research topic how to achieve the recognition and judgment of flaws by efficient,high-precision intelligent detection of MFL signals in order to issue pipeline flaw report excavation sheet.Deep learning theories can provide feature representation based on learning,which is widely used in many applications like image classification,detection,segmentation,etc.Currently,there are fewer researches on the recognition of pipeline magnetic flux leakage signals.Consequently,we study the intelligent recognition of pipeline magnetic flux leakage signals of girth weld,spiral weld and defection based on deep learning as the basic framework.There are four main parts in this paper which are listed in the following:For the problem of imaging pipeline magnetic flux leakage signal data,the curvilinear imaging and pseudo-color imaging are employed for its visualization.Firs of all,magnetic flux leakage signal detected are extracted from detector according to their formats and types determined in advance.Secondly,pipeline magnetic flux leakage signals are converted into curve image and pseudo-color image according to their features by pseudo-color conversion and tracing imaging.We found that pseudo-color image can improve the ability todifferentiate of image details and gain clear,natural images by experiment comparison.Color value can be considered as the third dimension feature of magnetic flux leakage signal in order to utilize color to reveal whole magnetic flux leakage signal data.Finally,nonlinear bilateral filter are used to smooth the images of pipeline magnetic flux leakage for the images after preprocessing.Pseudo-color images and curve images of pipeline magnetic flux leakage are collected to form the dataset and the corresponding image database as the network input of the subsequent classification and recognition experiments.For the problem of blurry edge of pipeline magnetic flux leakage images,we proposed an approach for the edge enhancement of pipeline magnetic flux leakage image based on multi-scale mathematical morphology and Laplacian.Multi-scale structure elements are constructed and combine the rectangular structure elements with circular structure elements for the enhancement of curve images.For the pseudo-color images of pipeline magnetic flux leakage,we fisrt use Laplacian to conduct the edge detection for each component of pseudo-color images to get the points on the edge,and then determine the edge points detected by the edge color constraint to remove the false edge points.Finally,the edge detection images of each component are overlaid with original image for the pseudo-color images after edge enhancement.The experimental shows that our approach can effectively enhance the edges of weld,defect images,detail the information of contour and improve the distinguishability of images.For the problem of automatic classification of pipeline magnetic flux leakage welding line image,we proposed a classification method of pipeline magnetic flux leakage welding line image based on convolution kernel sparse self-coding.Firstly,a deep learning platform base on Caffe is constructed,which is an optimization method based on Convolutional Neural Network(CNN)as classification network and model.In order to enhance the features extraction capacity of convolution kernel for pipeline magnetic flux leakage signal,convolution kernel is inputted into sparse self-coding network to proceeds optimal pre-training.Sparse autoencoder network is used to automatically learn the features of images,and then determine the information amount of image structure that each convolution carries by introducing image entropy similarity constraint rule for the equilibrium of image entropy of convolution kernel.Next,we remove the convolution kernels with similar weights by similarity determination condition to make the convolution kernels to have various learning capacity.The feature extraction capacity of convolution kernel is improvedby pretraining network and the difference between goal feature information and background feature information is enhanced to improve network classification ability.Finally,six types of pipeline magnetic flux leakage images as input are exploited to validate the CNN with convolution kernel optimized,which are girth weld,spiral weld,tee,flange,defect and non-defect.The experimental results show that the average accuracy of our approach can reach 94% for the classification of the six types of pipelines above,which is better than traditional convolution neural network model and has good feature extraction ability and generalization ability.For the problems of intelligent recognition of pseudo-color images of pipeline magnetic flux leakage,the recognition method of pseudo-color images of pipeline magnetic flux leakage based on DAR-SSD network is proposed in this paper.Firstly,we use Single Shot MultiBox Detector(SSD)as the basic target detection network model.In order to improve recognition accuracy of SSD algorithm with respect to small flaws in pipeline magnetic flux leakage signal,empty convolution and attention residuals module are added to SSD model.Empty convolution is utilized to expand perception of network model,and low-resolution and high-semantic information feature image is combined with high-resolution and low-semantic information feature image to improve the learning capacity of network for small target detail features.Then attention residuals module is used to reinforce the area interested by network to make the target detected to be obvious.The experimental results show that our algorithms can automatically recognize the location of girth weld,spiral weld,flaws of magnetic flux leakage data while accuracy increases to 97.33%,false positive is2.05%,leakage detection rate is 0.62%,the positioning error is less than 2 meters..Meanwhile,the algorithm proposed has better robustness and is obviously effective with respect to small flaw target detection.In conclusion,this paper focuses on the improvement of feature representation capacity of network from the perspective of model structure,including designing network structure,introducing sparse self-coding,empty convolution and algorithm optimized of attention residuals module.The experimental results show that the method improved can improve the recognition performance of network.