Enhancing damage detection in truss bridges through structural stiffness reduction using 1DCNN, BiLSTM, and data augmentation techniques

• Published in: Structures (Oxford) 2024-10, Vol.68, p.107035, Article 107035
• Main Authors: Tran-Ngoc, Hoa, Nguyen-Huu, Quyet, Nguyen-Chi, Thanh, Bui-Tien, Thanh
• Format: Article
• Language: English
• Subjects: 1DCNN; Augmentation techniques; BiLSTM; Structural stiffness reduction; Truss bridge
• ISSN: 2352-0124; 2352-0124
• DOI: 10.1016/j.istruc.2024.107035

Time-series data plays an important role in bridge health monitoring (BHM), enabling early damage detection and ensuring the safety of the bridge. In recent years, researchers have employed several methods based on one-dimensional convolutional neural networks (1DCNN), recurrent neural networks (RNN), and long short-term memory (LSTM) to analyze time-series data from structural health monitoring (SHM) systems. However, these methods have limitations in capturing the bidirectional temporal dependencies and spatial correlations across the variables. This present study proposes a practical approach that integrates 1DCNN with bidirectional LSTM (BiLSTM), empowered with augmentation (AUG) techniques, for damage detection problems through time-series data analysis. 1DCNN is employed to automatically extract spatial features from the input data, whereas BiLSTM captures bidirectional time dependency as well as cross-variable relationships. Moreover, advanced augmentation techniques such as window cropping, flipping, and noise injection are used to enrich the data before training. The proposed 1DCNN-BiLSTM-AUG model significantly improves performance for identifying structural stiffness reduction in large-scale truss bridges. The accuracy reaches 0.984, and the loss is 0.067, which proves that it surpasses other models, including 1DCNN, LSTM, and 1DCNN-LSTM. The obtained results demonstrate that this approach is a promising tool for SHM.