Early-stage thermal damage identification in concrete using attention-enhanced coda wave learning

This is the authors' abstract. We don't add key points for this paper.

Identifying early-stage thermal damage in concrete remains challenging due to the strong scattering characteristics of ultrasonic waves in heterogeneous media. Although coda wave (CW) interferometry (CWI) is highly sensitivity, its signals can be affected by non-uniform material distributions and variations in transducer coupling, which makes conventional CWI-based indices unstable. To address this issue, this study proposes a dual-branch one-dimensional convolutional neural network with a convolutional block attention module (1D-CNN-CBAM) to identify thermal damage from ultrasonic CW. The time-domain and frequency-domain signals are processed in parallel and then fused to form complementary feature representations. Concrete specimens subjected to controlled thermal exposure below 300°C are tested to evaluate the proposed framework. The results indicate that the proposed 1D-CNN-CBAM model achieves a mean identification accuracy of 97.10% on the test set. In addition, the model shows good generalization, reaching 95.83% accuracy on independently fabricated specimens. Overall, these findings demonstrate that the integration of attention-enhanced deep learning with CW signals provides an accurate and robust approach for non-destructive evaluation of thermal damage in complex concrete structures.