Prediction method of condition degradation for network-level bridges based on U-Net++ convolutional neural network

• The bridge feature selection is carried out by eta square coefficient and spearman coefficient; • The prediction of condition degradation for network-level bridges is realized by U-Net++ convolutional neural network ; • The influence of different data missing rate on the prediction effect of the model is analyzed; • Several machine learning models and improved U-Net architectures are selected for comparison to prove the superiority of the proposed method. The condition degradation prediction for network-level bridges is conducive to the decision-making of bridge maintenance. However, traditional prediction methods mainly belong to shallow neural networks and have difficulty in extracting the common degradation features of network-level bridges, thus obtaining a limited prediction accuracy. To this end, this study proposes a prediction method of condition degradation for network-level bridges based on U-Net++ convolutional neural network , in which the U-Net++ is employed to capture the common degradation characteristics of network-level bridges by the fusion of multi-scale feature. Firstly, a dataset of bridge condition is established based on the inspection data of 539 bridges in a typical city. A correlation analysis is performed to preliminarily reveal the relevance between the key features of the bridge and the bridge condition level. Then, the U-Net++ network model is utilized to establish a nonlinear mapping relationship between the key features of the bridge and the bridge condition level. By the established model, the predication of bridge condition level can be achieved. Several machine learning models and three U-Net-based model are employed to verify the advantages of the proposed method. The results show that the proposed model may be the optimal network architecture for bridge condition degradation prediction. It overcomes the shortcomings of traditional U-Net model with a large fluctuation in the prediction and obtains a prediction accuracy of over 80 % for both the primary components of the bridges and the whole bridge.