Condition level deteriorations modeling of RC beam bridges with U-Net convolutional neural networks

Abstract

Reinforced concrete (RC) beam bridges have suffered structural deterioration due to loads, environmental conditions, etc. Regular visual inspections of bridges effectively monitor the structural condition level and provide a vast amount of condition-related data for years. This study proposes a deep learning-based condition level deterioration modeling method with a U-Net model to improve the prediction accuracy of future structural conditions. The proposed method is supported by the data gathered from the years of regional bridge inspection reports. Before training the model, the regional condition-related features regarding the influence of bridge ages and the superstructure types are investigated, and the correlations between selected features and structural conditions are also revealed. The acquired inspection database validated the high prediction accuracy and classification performance of each bridge's main part and system with the proposed deterioration modeling method. Its robustness is tested under a variety of data missing rate scenarios. The optimum model architecture and its effectiveness are also validated through comparative studies. This study provides a novel method to predict the future structural condition with inspection data and could serve as a reference for more reasonable utilization of the bridge condition deterioration model in structural condition assessment and management.