Multi-source data fusion-based structural strain prediction of long-span arch bridges via CNN-GRU-Attention network

Résumé fourni par la source

Abstract Accurate prediction of bridge strain is crucial for ensuring the structural safety of bridges and developing scientifically sound maintenance strategies. However, bridge strain is influenced by the coupled effects of multiple factors, including external loads, temperature, and wind speed, exhibiting strong nonlinearity, time-dependence, and sensitivity to sudden changes; traditional prediction methods struggle to effectively capture these complex patterns. To this end, this paper proposes a convolutional neural network (CNN) gated recurrent unit (GRU)-Attention bridge strain prediction model that integrates CNN, GRU, and attention mechanisms. In particular, CNN is used to extract local spatial features from multi-source monitoring data (such as temperature, cable tension, and wind speed); GRU addresses the issue of long-term dependencies in time series data; compared to long short-term mesolimory (LSTM) networks, it has fewer parameters and is more efficient to train. The Attention mechanism enhances the response to critical events such as peak vehicle loads and sudden temperature changes. A health monitoring system for an in-service steel-concrete composite rib arch bridge was used to collect multi-source monitoring data for model validation, and the results were compared with those of the CNN-LSTM-Attention and CNN-BiLSTM-Attention models. The results show that the CNN-GRU-Attention model exhibits a better fit between its predicted curves and the actual strain curves, accurately capturing strain peaks and sudden changes. The strain predictions for the main girder and steel arch structures have mean absolute error of 7.0848 and 7.4447, respectively, which demonstrates superior performance compared to the benchmark models. In addition, the correlation analysis between multi-source monitoring data and bridge strain confirmed the critical role of data fusion in improving prediction accuracy.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.