A Digital Twin-Driven Dual-Stage Adversarial Transfer Learning Method for Lamb Wave-Based Structural Damage Localization Under Limited Sensing Data

Abstract

Structural health monitoring (SHM) based on Lamb waves relies on sensors to acquire structural response signals. However, sensor data acquisition is severely constrained under complex damage conditions. Digital twins (DTs) can enhance damage monitoring capabilities in Lamb wave SHM by integrating simulation and experimental sensor data. Nevertheless, performance remains limited by discrepancies in signal distribution between digital and physical domains, as well as cross-domain optimization conflicts. This study proposes a digital twin-driven dual-stage adversarial and transfer learning method with multi-objective optimization (DT-DSATMO) for Lamb wave-based structural damage localization under limited sensing conditions. Firstly, a strategy for hierarchical feature enhancement and conditional generation incorporating physical prior knowledge is introduced to construct distribution-consistent feature representations in the digital domain. Secondly, it achieves adaptive alignment between the two domains via a lightweight domain adversarial transfer network, improving cross-domain feature transferability. Furthermore, a Pareto frontier-based multi-objective optimization strategy is employed to balance damage localization accuracy, cross-domain robustness, and feature consistency. The proposed method is experimentally validated on a representative aircraft wing-box panel equipped with four lead zirconate titanate (PZT) sensors. The case study results show that it substantially enhances damage localization accuracy and cross-domain generalization under limited sensing data.

Keywords: damage localization; digital twin; dual-stage adversarial and transfer learning; limited sensing data; multi-objective optimization.

Figure 1

Schematic diagram of the holistic approach.

Figure 2

Stage 1: physics-informed feature enhancement in the adversarial transfer framework.

Figure 3

Stage 2: Lightweight BiGRU-Attention adversarial transfer network in the adversarial transfer framework.

Figure 4

Schematic of the multi-objective optimization mechanism.

Figure 5

(a) Experimental platform and (b) five-cycle tone-burst excitation.

Figure 6

Experimental training datasets collected for multiple damage locations ((a–d) represent four datasets).

Figure 7

Configuration and validation of the PML. (a) Schematic configuration of the PML applied around the plate boundary, and (b) comparison of simulated sensor signals with and without PML implementation.

Figure 8

Finite element model of PZT actuation–sensing electromechanical coupling.

Figure 9

Propagation cloud maps of Lamb waves in the structure at different time instants. (a) t = 1.98 × 10⁻⁵ s, (b) 5.61 × 10⁻⁵ s, and (c) 9.24 × 10⁻⁵ s.

Figure 10

Comparison of simulated and experimental guided wave signals in the healthy state.

Figure 11

Schematic diagram of the damaged structure simulation. (a) FE meshing in the damaged state. (b) Top view of damaged structure. (c) Cross-sectional view of damaged structure.

Figure 12

Comparison of simulated and experimental guided wave signals for a damage size of 20 mm.

Figure 13

Sample distribution of the simulated training dataset.

Figure 14

T-distributed stochastic neighbor embedding (t-SNE) visualization of cross-domain feature alignment at different training stages.

Figure 15

Evolution of RBF-MMD between digital and physical domains during training.

Figure 16

Damage localization results of DT-DSATMO under six cross-domain scenarios. (a) Scenario S1, (b) Scenario S2, (c) Scenario S3, (d) Scenario S4, (e) Scenario S5, and (f) Scenario S6.

Figure 17

Analysis of the proposed DT-DSATMO method: (a) feature distribution comparison; (b) trend evaluation.

Figure 18

Imaging results of DAS for damage at different locations. (a) D1, (b) D2, and (c) D3.

Figure 19

Imaging results of EPI for damage at different locations. (a) D1, (b) D2, and (c) D3.

Figure 20

The localization error results for each model configuration.

Figure 21

Variation in MAPE and MRE with the number of physical domain samples.

Figure 22

Localization errors for different unseen damage sizes.