Vision-based bridge deflection monitoring using a chain-camera system and CNN-BiLSTM-attention network with cross-scenario data training strategy

Bridge deflection is a key indicator for evaluating structural performance. Although vision-based monitoring enables non-contact and multi-point measurements, it is still affected by limited measurement range, noise, temporal drift, and insufficient labeled field data. To address these challenges, this study proposes a vision-based bridge deflection monitoring framework that integrates a chain-camera system, a CNN-BiLSTM-Attention network, and cross-scenario data training strategy. The chain-camera system enables long-distance, multi-point synchronous measurements through displacement relay, overcoming the range limitations of conventional vision-based approaches. A hybrid deep learning model combining convolutional neural networks, bidirectional long short-term memory, and a multi-head attention mechanism is developed to suppress noise and correct temporal drift in displacement signals. Furthermore, a cross-scenario data training strategy is introduced to leverage high-quality laboratory data to enhance the generalization capability of the real-bridge model. Laboratory experiments and field tests on an in-service bridge validate the effectiveness of the proposed framework. Results show that, in laboratory experiments, the proposed model improves R 2 by 61.72% on average and reduces MAE and RMSE by 24.46% and 24.89%, respectively, compared with baseline models. In field tests, the absolute magnitude of the negative R 2 value is reduced by 75.74%, while MAE and RMSE are reduced by 28.59% and 39.37%, respectively. These results demonstrate the effectiveness of the proposed framework for bridge deflection monitoring under both laboratory and field conditions.