A structure-preserving diffusion-based zero-shot learning framework for multimodal magnetic flux leakage signal analysis

Addressing the core challenges of weak defect signatures and difficult unknown defect identification in magnetic flux leakage (MFL) inspection of large-bore pipelines, this study proposes an intelligent detection method that integrates a Zero-Shot Structure-Preserving Diffusion Model (ZSSPDM) and cross-modal attention fusion. A Structure-Preserving Diffusion Model (SPDM) is designed to explicitly preserve defect edges and geometric structures during denoising via triple constraints—gradient consistency loss, morphological similarity loss, and frequency-domain regularization—enhancing the signal-to-noise ratio (SNR) of original MFL signals from 12.3 dB to 24.1 dB for high-quality feature input. A gated multi-head cross-modal attention network is constructed, taking MFL signals as queries to dynamically integrate ultrasonic testing (UT) and infrared (IR) features, mitigating inter-modal redundancy and conflicts while achieving a macro F1-score of 0.93 on known defect classes, outperforming early and late fusion strategies. A zero-shot recognizer based on visual-semantic dual-stream embedding is developed, establishing a semantic attribute space (geometry, depth level, causal type, directionality) and leveraging contrastive learning to enable knowledge transfer between known and unknown classes. On a test set containing four unseen defect categories, the method achieves a Zero-Shot Learning (ZSL) Accuracy of 0.84 and an H-Mean of 0.88, surpassing mainstream models such as GLEE and TransMIL. Cross-material and cross-pipeline tests demonstrate an average ZSL Accuracy of 0.81, confirming strong generalization and engineering applicability. This work provides a high-precision, robust solution for intelligent pipeline inspection, with significant advantages in signal enhancement, multimodal fusion, and zero-shot generalization.