An adaptive perception-based identification method for disturbed EMAT defect echoes

Defect echoes acquired by electromagnetic acoustic transducers (EMATs) are often affected by time-delay drift, local waveform deformation, noise interference, and limited labelled data, making defect-type identification difficult. To address these issues, this paper proposes an adaptive perception-based identification method and develops an adaptive perception echo fusion network (APEF-Net). First, defect echo signals are acquired using a self-developed unidirectional focusing EMAT testing platform, and a dataset is constructed. Second, a deformation adaptive echo module (DAEM) is constructed to dynamically adapt to time-delay drift and local deformation in the echoes. A directional local-global fusion module (DLGF) is designed to fuse local waveform details with global echo contextual information, alleviate detail degradation and background interference in deep features. Finally, a source-domain pretraining and target-domain transfer fine-tuning strategy is combined to improve identification performance under limited labelled data conditions. APEF-Net achieves classification accuracies of 98.25% on the target dataset and 97.22% on an independent weld ultrasonic defect dataset, outperforming the comparison methods and demonstrating good applicability. SHapley Additive exPlanations (SHAP) analysis results show that the model decision basis is in good agreement with defect geometric features and their ultrasonic scattering patterns.