Domain-Consistent Transfer Learning in Lightweight ۱D CNNs for Vibration-Based Damage Detection

Structural damage detection in civil infrastructure remains a major challenge, especially when relying on vibration signals captured under varying operational conditions. This study addresses the problem of damage detection in cable-stayed bridges using a lightweight one-dimensional convolutional neural network (۱D CNN) enhanced through domain-consistent transfer learning. A novel framework is proposed in which a compact ۱D CNN is pretrained on healthy-state vibration data from one cable-stayed bridge and subsequently fine-tuned on labeled damage-state data from another structurally similar bridge. The method leverages the temporal characteristics of acceleration signals and the consistency of domain-specific features across bridges to facilitate effective knowledge transfer. The designed ۱D CNN model consists of a minimal number of convolutional layers to ensure low computational cost, making it suitable for deployment in resource-limited structural health monitoring (SHM) systems. To validate the approach, a two-stage training process was implemented: the model was first pretrained using normal operational data from the Manitoba Bridge, and then fine-tuned using labeled damage scenarios from the Z۲۴ Bridge. Experimental results demonstrate that the proposed method achieves high classification accuracy while significantly outperforming a baseline model trained from scratch. Furthermore, confusion matrix analysis reveals the model's ability in binary classification of healthy and damaged states with notable reliability. The findings suggest that incorporating domain consistency in transfer learning can improve generalization performance in SHM applications where labeled damage data is scarce. The proposed framework offers a scalable and efficient solution for vibration-based damage detection across structurally analogous bridges, paving the way for more intelligent and adaptable monitoring strategies in real-world infrastructures.

کلیدواژه ها:

Transfer Learning ، SHM ، Vibration based Damage Detection ، Lightweight ۱D CNNs