Transformer model combining cross-attention and self-attention guided by damage index for pipeline damage localization based on helical guided waves
Hong Zhang, Guoxiang Wang, Feiyu Teng, Shanshan Lv, Lei Zhang, Faye Zhang, Mingshun Jiang
Abstract
The structural health monitoring technology based on ultrasonic guided waves (UGW) has broad application prospects in pipeline structural damage detection. This paper proposes a Transformer model based on damage index (DI) guidance and fusion of cross- attention and self-attention (DCAS-Transformer). Firstly, based on the principle of damage and guided waves, the DI representing the correlation between healthy signals and damaged signals is calculated and used as a weight to guide subsequent model training. This step enhances the correlation between signals and space. Secondly, based on cross-attention mechanism, a channel attention module is constructed to jointly input DI and signals into the module, achieving feature fusion and further emphasizing information related to damage. Finally, in the transformer module, based on its unique spatial attention mechanism, deep damage information extraction is achieved and damage localization is completed. The results show that the average localization error is 9.6 mm and the relative error is 1.92 %, and the proposed method performs well under different sensor layouts and noise levels. Compared with other deep learning methods, the proposed method has more stable performance and better generalization.