Gated Differential Vector Attention Unit Network for Damage Localization by Ultrasonic-Guided Wave

Damage localization based on ultrasonic-guided waves (UGW) ensures the reliability and safety of engineering structures, including both composite materials and aluminum materials. Efficient and accurate damage localization requires a thorough analysis of the variations in guided wave amplitude at each moment. However, existing deep learning-based damage localization methods overlook the amplitude discrepancy information in the guided wave signals. Therefore, this paper proposes a Gated Differential Vector Attention Unit (GDVU) Network for damage localization. This method introduces a novel feature extraction unit called GDVU, which effectively extracts and learns the amplitude discrepancy information between two neighboring moments, generating differential vectors that adaptively guide the update of the network’s state. Building upon GDVU, we design Bidirectional Gated Differential Vector Attention Unit (BiGDVU), which captures amplitude discrepancy information in both directions, further enhancing the network’s ability to localize damage. Furthermore, we enhanced the existing multimodal input mode by combining the more damage-characterized scattered wave signals and energy density signals to feed the network. This allows for an effective mapping of the correlation between the damage zone and signal. Subsequent experimental results on the CFRP plate and aluminum specimen show that the GDVU-based network achieves superior damage localization accuracy and robustness compared to other existing methods.