A Generalized Zero-Shot Regression Framework for Cross-Domain Damage Identification in Composite Structures

TY - JOUR

T1 - A Generalized Zero-Shot Regression Framework for Cross-Domain Damage Identification in Composite Structures

AU - Chen, Gang

AU - Shao, Weihan

AU - Liu, Qijian

AU - Sun, Hu

AU - Huo, Shiyu

AU - Fu, Hailing

AU - Qing, Xinlin

N1 - Publisher Copyright: © 2005-2012 IEEE.

PY - 2026/7/1

Y1 - 2026/7/1

N2 - In Lamb wave-based structural health monitoring (SHM), traditional data-driven methods face two main challenges: domain shifts between different monitoring regions and scarcity of damage samples in the target domain, often resulting in zero-shot scenarios. Existing zero-shot learning approaches are confined to classification tasks in SHM and remain inadequate for addressing cross-domain regression challenges. To overcome this, a novel paradigm termed generalized zero-shot regression (GZSR) is proposed, aiming to achieve cross-domain damage localization in composite structures without requiring training samples from the target domain. In GZSR, semantic representation serves as the key bridge for generalizing from seen to unseen locations. To tackle the lack of damage semantics in the target domain, a semantic construction module is introduced, which encodes textual descriptions of damage from the source domain into semantic vectors using a pretrained language model, along with a semantic reconstructor to inversely reconstruct semantic representations from the extracted features. This design encourages the model to learn intrinsic relationships between damage features and physical parameters, thereby enabling the mapping of damage features from the target domain into a unified semantic space. Furthermore, a cross-modal generative adversarial module is proposed. It generates pseudofeatures conditioned on semantics, with consistency between features and semantics constrained by cosine similarity. This mechanism effectively supplies domain-invariant features to the regressor, thereby transforming GZSR into a supervised learning task. Experimental results on four zero-shot cross-domain damage identification tasks demonstrate that GZSR achieves lower localization errors for both seen and unseen damage locations compared to state-of-the-art models.

AB - In Lamb wave-based structural health monitoring (SHM), traditional data-driven methods face two main challenges: domain shifts between different monitoring regions and scarcity of damage samples in the target domain, often resulting in zero-shot scenarios. Existing zero-shot learning approaches are confined to classification tasks in SHM and remain inadequate for addressing cross-domain regression challenges. To overcome this, a novel paradigm termed generalized zero-shot regression (GZSR) is proposed, aiming to achieve cross-domain damage localization in composite structures without requiring training samples from the target domain. In GZSR, semantic representation serves as the key bridge for generalizing from seen to unseen locations. To tackle the lack of damage semantics in the target domain, a semantic construction module is introduced, which encodes textual descriptions of damage from the source domain into semantic vectors using a pretrained language model, along with a semantic reconstructor to inversely reconstruct semantic representations from the extracted features. This design encourages the model to learn intrinsic relationships between damage features and physical parameters, thereby enabling the mapping of damage features from the target domain into a unified semantic space. Furthermore, a cross-modal generative adversarial module is proposed. It generates pseudofeatures conditioned on semantics, with consistency between features and semantics constrained by cosine similarity. This mechanism effectively supplies domain-invariant features to the regressor, thereby transforming GZSR into a supervised learning task. Experimental results on four zero-shot cross-domain damage identification tasks demonstrate that GZSR achieves lower localization errors for both seen and unseen damage locations compared to state-of-the-art models.

KW - Data-driven

KW - Lamb wave

KW - generalized zero-shot learning (GZSL)

KW - machine learning

KW - structural health monitoring

UR - https://www.scopus.com/pages/publications/105035630910

U2 - 10.1109/TII.2026.3675653

DO - 10.1109/TII.2026.3675653

M3 - Article

AN - SCOPUS:105035630910

SN - 1551-3203

VL - 22

SP - 6164

EP - 6175

JO - IEEE Transactions on Industrial Informatics

JF - IEEE Transactions on Industrial Informatics

IS - 7

ER -