Multi-source domain dynamic adversarial adaptation network for thin-walled metal structures damage detection using guided wave under compound perturbations.

Ultrasonic guided wave is recognized as a promising technique for damage detection of thin-walled metal structures. However, such structures are chronically subjected to dynamic environments, where the induced compound perturbations such as temperature variation and structural deformation will distort signal characteristics and reduce detection accuracy. To address this challenge, a multi-source domain dynamic adversarial adaptation network (MS-DDAAN) is proposed to enhance damage detection performance under compound perturbations. Firstly, single-frequency signals are obtained through a series of processes, including broadband chirp excitation, frequency deconvolution, filtering, and normalization. Meanwhile, signals acquired under different operating conditions are categorized into multiple source domains and a target domain. Afterward, a dual-layer feature extraction architecture is designed, integrating dynamic convolution and self-attention mechanism to adapt to signal distribution shifts and extract domain invariant damage features. Additionally, an adversarial adaptation network, which contains global domain discriminator and local domain discriminator, is introduced to achieve distribution alignment and feature alignment across different domains. Finally, a label classifier is employed to map damage features to state labels, and a dynamic harmonic factor is introduced to balance label alignment with domain alignment during the training process. The proposed method is validated on thin-walled metal plate under 12 composite conditions of temperature variation and structural deformation. Experimental results demonstrate that MS-DDAAN achieves an accuracy over 95% in damage identification across all test conditions and its performance is superior to the compared other transfer learning methods.