Research On Pipeline State Recognition And Defect Quantification Technology Based On Hierarchical Feature Mining And Bayesian Decision

Pipeline is an critical means of transportation for oil,natural gas and other important resources.It is necessary to regularly inspect and maintain it to reduce safety accidents.Eddy current nondestructive testing is a commonly used pipeline inspection technology,which has the advantages of high efficiency,accuracy and high trafficability.After the pipeline is detected by the internal detector system,analyzing the detection data to identify,locate and quantify the defects is an important step in the detection task.Due to the complexity of the real pipeline,the value and shape of the detection signal generated by various pipe structures are similar to the defect signal,which greatly interferes with the defect identification.Mileage wheel is prone to "slip" causing distance measurement errors,which is cumulative error.The amplitude and phase of defect signal are directly affected by many factors,which makes the depth and amplitude and phase have a multi-dimensional nonlinear relationship.This thesis focuses on the identification,location and quantification of real pipeline defects.The main research contents are as follows:1.For defect identification,a feature mining framework based on kurtosis adaptive stepwise residual is proposed.By mining corresponding features of various interference signals,the interference signals are identified and eliminated step by step to reduce false detection.In the process of defect identification,the kurtosis of the residual signal is used as the loss function of the framework to guide the algorithm to adjust hyperparameters and solve the problems of missed detection and the need for manual intervention in the algorithm’s inability to adapt to fully automatic identification.Through a large number of experiments,the algorithm proposed in this thesis achieved a defect identification rate of99% for simulated pipelines and 92% for real pipelines.2.For defect location,a two-level location method is proposed.Firstly,a primary location is carried out using pipeline construction drawings and pipe structure as reference points.Then,a secondary accurate location is obtained using the relative distance between the mileage wheel and the reference point,and the heavy hammer wheel.Through excavation and defect verification of field pipelines,this method achieves a location error within 10 cm for defects,while the error of a single mileage wheel location can be up to several tens of meters,which proves that the method effectively reduces the cumulative error caused by a single mileage wheel location.3.For defect quantification,a Monte Carlo approach is proposed based on multiple simulated pipeline defect amplitude-phase distributions to model the real defect amplitude-phase distribution.Combining the K-nearest neighbor method,the defect depth is determined.Utilizing the prior information of the amplitude-phase distribution of simulated pipeline defects,the Bayesian decision-based minimum error rate idea is employed to correct and fuse the quantification results of different pipe diameters,effectively solving this highly nonlinear problem from the perspective of statistical probability.The root mean square error of the depth prediction value for over 100 defects in multiple real pipelines using the proposed algorithm in this thesis is 0.92 mm,and the relative error is between 1.2% and 25.3%.This thesis extensively validates the proposed algorithm on multiple simulated and real pipelines,and compared with typical algorithms,it achieves significant improvements in defect identification,location,and quantification,with stronger robustness.