A Pipeline Defect Size Prediction Method Integrating Simulation-Driven, Data Augmentation, and Physics-Guided Approaches

Abstract

Magnetic Flux Leakage (MFL) testing is a vital technique for the quantitative assessment of internal defect dimensions in oil and gas pipelines. However, existing inversion methods face dual challenges: traditional physical forward computational models are excessively time-consuming, while purely data-driven deep learning models (i.e., “black-box” models) typically lack physical interpretability and generalization capability. To address these issues, this paper proposes a novel pipeline defect size prediction framework that integrates simulation-driven modeling, data augmentation, and physics-guided mechanisms. First, a defect dataset containing dual-channel MFL signals (axial and radial components) is constructed via finite element simulation. Subsequently, a Conditional Variational Autoencoder (CVAE) is employed to augment the dataset, thereby overcoming the bottleneck of scarce high-fidelity defect samples. Finally, a dual-channel Convolutional Neural Network (CNN) is constructed, into which a Physical Proxy Model (PPM) is introduced. The PPM embeds the “dimension-signal” physical mapping mechanism into the network as a loss constraint, enforcing the model to strictly adhere to underlying physical laws during prediction. Experimental results demonstrate that the proposed method achieves superior performance in predicting defect length, width, and depth, yielding an overall Mean Absolute Error (MAE) of 0.7525 mm and a Root Mean Square Error (RMSE) of 1.2128 mm. Compared to the benchmark LSTM model, the MAE is significantly reduced by 14.15%. The proposed framework realizes an effective fusion of data-driven and physics-driven mechanisms, offering a novel paradigm for defect prediction in complex scenarios where complete physical equations are unavailable.