Research on a Multimodal Joint Quantification Algorithm for Pipeline Defects Based on Flux Leakage Testing

Magnetic Flux Leakage (MFL) pipeline defect detection faces numerous challenges, with traditional manual analysis being inefficient and highly subjective, while deep learning methods struggle to balance prediction accuracy across three features. This paper proposes an innovative defect quantification framework that establishes a multimodal joint prediction mechanism, comprehensively integrating magnetic flux and image data to effectively enhance defect detection accuracy. In magnetic flux quantification, the framework designs a hybrid optimizer combining genetic algorithms and ant colony optimization. By employing a weighted objective function, it balances prediction accuracy of different output features and introduces diversified feature engineering techniques to enhance model generalizability and numerical stability. In image quantization, an improved HCFNet combined with Linear Deformable Convolution innovatively proposes a Parallelized Patch‐Aware Attention module, improving small object detection precision. The introduction of the Softplus activation function and Huber Loss function further improves training stability and model convergence speed. Experimental results demonstrate that the proposed method significantly outperforms current mainstream detection models in defect quantification accuracy and feature quantification capabilities. © 2025 Institute of Electrical Engineers of Japan and Wiley Periodicals LLC.