Objective The magnetic flux leakage technique for in-line inspection of pipelines is widely used for the online detection of pipeline defects, due to its advantages including being couplant-free and easy to automate. Subsequent accurate prediction of defect sizes and evaluation of applicability using inspection data are essential for making informed repair decisions.
Methods A deep learning target detection model based on Paddle Paddle-You Only Look Once Evolved (PP-YOLOE) is proposed, developed from collected magnetic flux leakage data and the characteristics of defect data. Color images converted from triaxial magnetic flux leakage data are input into the model for training in target detection, facilitating rapid defect positioning and data extraction, and creating accurate and reliable datasets for quantifying defect sizes. Additionally, a Multi-Task Learning (MTL) model is introduced, taking circumferential, axial, and radial magnetic flux leakage data of pipeline defects as input. This model outputs data regarding the length, width, and depth of defects in parallel, leveraging the characteristics of magnetic flux leakage data to achieve size evaluation and prediction of defects.
Results Comparing the data from pulling experiments revealed that the combination of PP-YOLOE and MTL improved defect identification efficiency and the prediction accuracy of defect sizes. The recall rate and accuracy rate for defect target detection reached 0.87 and 0.94, respectively. Compared with the application of Stochastic Gradient Descent (SGD), the Nesterov-accelerated Adaptive Moment Estimation (NADAM) algorithm achieved an increase in quantification accuracy for corrosion defects by 9% in length and 4% in depth.
Conclusion The proposed method is applicable for analyzing massive pipeline magnetic flux leakage testing data and improving the identification efficiency. It provides reliable foundational data support for assessing residual strength and predicting the remaining life of pipelines with defects.
DownLoad: