- Research Article
2
- 10.1155/stc/6840921
A Semisupervised Transfer Learning Method for Structural Damage Identification and Its Application to a Cable‐Stayed Bridge
- Jan 01, 2025
- Structural Control and Health Monitoring
- Naiwei Lu + 4 more +4
Engineering structures usually have rare damage scenarios, which provide insufficient labels to identify the structural damage. Therefore, the effectiveness and accuracy of traditional data‐driven methods with limited training samples is an urgent task. This study develops a semisupervised transfer learning (TL) method for structural damage identification. The effectiveness is validated through experiments on a scale cable‐stayed bridge specimen. Initially, a convolutional neural network (CNN) was trained based on a numerical dataset simulated by a finite element model, where a source domain is modeled with strong recognition performance and generalizability. Subsequently, the network structure and hyperparameters in the source domain were transferred to the corresponding positions in the experimental training model to create a pretrained model in the target domain; this pretrained model was updated by using site‐specific measured data from the engineering structure in the target domain. Finally, a dynamic threshold self‐training algorithm was employed to further optimize the model: in the initial stage, a high‐confidence threshold of 90% was set to filter reliable pseudolabels, and the threshold decreases to 80% gradually, under the condition that the model cannot provide predictions with sufficiently high confidence. Experimental study on a scale cable‐stayed bridge specimen was conducted to demonstrate the effectiveness of the proposed method. Four networks were established based on (1) experimental samples; (2) both labeled and unlabeled experimental samples with the self‐training algorithm; (3) TL from the finite element model and experimental samples; and (4) applying self‐training with unlabeled samples to the target model derived from TL. The results indicate that the damage identification accuracy of the aforementioned models is 73.6%, 79.4%, 84.6%, and 89.8%, respectively. The TL improves the reliability of generating pseudolabels by utilizing the self‐training process and unlabeled data and decreases errors in pseudolabels. Both finite element simulation data and practical unlabeled sample data were successfully combined by the TL and self‐training semisupervised learning method for damage identification in cable‐stayed bridges effectively.
Read more