Target Recognition Method Based On Faster-RCNN Pipeline Magnetic Flux Leakage Signal

With the rapid development of the world economy,the total use of oil and natural gas continues to rise.It is particularly important to transport oil and natural gas and other energy sources to actual industrial sites in a reasonable and safe manner.Pipeline transportation has many advantages such as being unaffected by climate and ground,sustainable work,and huge transportation volume,making pipeline transportation of oil and natural gas called "energy bloodline",which plays a crucial role in the national economy.Important role.Therefore,it is more and more important to ensure the safe transportation of pipelines and realize the regular inspection of pipeline status.At present,the most effective means for pipeline safety detection at home and abroad is the magnetic flux leakage internal detection.At the same time,the analysis of pipeline abnormal signals is particularly important in pipeline magnetic flux leakage detection.It is an important part,so it is very important to find a reasonable and efficient pipeline magnetic flux leakage signal anomaly detection algorithm.This thesis uses the experimental data collected by the detector in the pipeline for analysis and research,and mainly completes the following aspects:the use of Faster R-CNN detection algorithm in deep learning to achieve the initial detection of defects and multiple components;for the detection of small defects with insufficient accuracy In order to solve the problem,an adaptive data-to-color method and an improved Faster R-CNN algorithm were designed.For the problem of insufficient samples of magnetic flux leakage data,an anomaly signal detection algorithm based on the combination of SimCLR and Faster R-CNN was proposed.The specific research contents are as follows:First,use the Faster R-CNN detection algorithm to initially realize the detection of pipeline abnormal signals.First,a method based on median data to color map is proposed,and then the R-CNN algorithm and Faster R-CNN algorithm are used to detect defects,tees and branch pipes in the abnormal signal of the pipeline,respectively,through R-CNN and Faster R-The comparison of CNN detection in terms of speed and accuracy establishes the feasibility of Faster R-CNN in the detection of abnormal signals of magnetic flux leakage data.Second,an adaptive detection algorithm for pipeline magnetic flux leakage data based on Faster R-CNN.In view of the problem of small defects in the process of converting magnetic flux data to color maps due to the influence of surrounding signals,the small defect signals are submerged after the color map is converted,and the small defects are not detected due to the mismatch between the actual size of the small defects and the feature extraction anchor window.Multiresolution magnetic flux leakage data to color method,and improve the feature extraction anchor window in Faster R-CNN.Finally,through the adaptive data-to-color method and the improved Faster R-CNN algorithm,the detection accuracy of small defects in the magnetic flux leakage data anomaly signal is improved,thereby improving the overall detection accuracy of the pipeline magnetic flux leakage signal for components and defects.Third,based on SimCLR and Faster R-CNN to achieve pipeline abnormal signal detection in the case of insufficient samplesWhen using deep neural networks for training,it is always necessary to label a large number of samples,but the number of samples that can be labeled under magnetic flux leakage data is not large enough to train too deep neural networks.If you train with a small number of samples directly under a randomly initialized model,it is prone to under-fitting the model.If you use transfer learning to train on the basis of the pre-trained model generated in the ImageNet public dataset,over-fitting of the model is easy to occur,so this article It is proposed to use the combination of SimCLR and Faster R-CNN in comparative learning to realize the detection of abnormal signals of magnetic flux leakage data with few samples.There are not many data sets that can be marked in the magnetic flux leakage data.In this paper,the unsupervised training method is used to generate the pre-training model in the magnetic flux leakage data,and then the generated pre-training model is used to initialize the parameters of the Faster R-CNN feature extraction network.And in turn reduce the possibility of overfitting and underfitting of the training model,improve the model generalization ability,and thus improve the overall detection accuracy.This thesis proposes a deep learning detection method for pipeline magnetic flux leakage defects and components,which effectively improves the detection accuracy,speed and positioning accuracy,and has great application value in actual pipeline detection in reality.