Multi-modality Hierarchical Attention Networks for Defect Identification in Pipeline MFL Detection

Abstract

Magnetic flux leakage (MFL) testing is widely used for acquiring MFL signals to detect pipeline defects, and data-driven approaches have been effectively investigated for MFL defect identification. However, with the increasing complexity of pipeline defects, current methods are constrained by the incomplete information from single modal data, which fails to meet detection requirements. Moreover, the incorporation of multimodal MFL data results in feature redundancy. Therefore, the Multi-Modality Hierarchical Attention Networks (MMHAN) are proposed for defect identification. Firstly, stacked residual blocks with Cross-Level Attention Module (CLAM) and multiscale 1D-CNNs with Multiscale Attention Module (MAM) are utilized to extract multiscale defect features. Secondly, the Multi-Modality Feature Enhancement Attention Module (MMFEAM) is developed to enhance critical defect features by leveraging correlations among multimodal features. Lastly, the Multi-Modality Feature Fusion Attention Module (MMFFAM) is designed to dynamically integrate multimodal features deeply, utilizing the consistency and complementarity of multimodal information. Extensive experiments were conducted on multimodal pipeline datasets to assess the proposed MMHAN. The experimental results demonstrate that MMHAN achieves a higher identification accuracy, validating its exceptional performance.