PSI - Issue 64

Lim Boon Xuan et al. / Procedia Structural Integrity 64 (2024) 791–798 Lim et al./ Structural Integrity Procedia 00 (2019) 000–000

792

2

1. Introduction A continuous girder bridge is a widely used statically indeterminate bridge structure, which offers certain inherent benefits compared to simple-supported girder bridges. From a structural perspective, the deflection induced by the live loads at the midspan of a continuous bridge is smaller than that observed in a simply supported bridge of equivalent span length (Hossain et al., 2014). Apart from the live loads, temperature effect results in additional internal forces, which induce deflection in the form of daily, seasonal, and annual periodic trend (Zhu and Meng., 2017). However, the problem of cracking in the bridge will besides contribute to the deflection (Liu and Xu., 2012). Therefore, it is worthwhile to model the temperature effect on deflection under normal condition for the purpose of future condition assessment and damage detection. Despite an abundance of studies on the temperature effect on bridge deflection, most have pertained to some large-span bridges, such as cable-stay bridge and suspension bridge (Zhou and Sun et al., 2019). So far, there is lack of research on multi-span continuous girder bridge, the most common type of bridge. In practice, the time-lag effect between temperature and temperature-induced deflection is the main challenge in precisely characterizing and modeling their relation. Due to the robust non-linear fitting performance of deep learning technology, the non-linear relation characteristic caused by time-lag effect can be computed. Among various deep learning techniques, the long short-term memory network (LSTM), first introduced by Hochreiter and Schmidhuber in 1997, outperforms autoregressive model (AR) in the prediction of long-term time series due to its ability to capture long-term dependencies in sequential data (Gunarto et al., 2023). In recent years, researchers have dedicated to apply LSTM network to bridge structural health monitoring (SHM) for condition assessment. Notably, several studies have investigated the mapping relationship between temperature and temperature-induced deflection or temperature-induced strain using LSTM, but these mostly pertained to cable-stayed and suspension bridges (Wang et al., 2022; Yue et al., 2022a; Yue et al., 2022b; Sun et al., 2021; Yue et al., 2021). The conclusion of these prior studies demonstrated that the LSTM network is capable of learning the complex patterns and non-linear relation between temperature and temperature-induced deflection. As mentioned above, the deflection of a bridge is not composed solely by temperature but several factors. To eliminate random variables during LSTM network training, the temperature-induced deflection should be extracted from raw deflection to isolate the temperature’s effect. While prior studies have used hourly or minutely averaging method to extract temperature-induced deflection (Yue et al., 2021; Sun et al., 2019; Yang et al., 2018), these methods may eliminate valuable information from the deflection data. Instead, this paper adopts an adaptive method capable of decomposing signals into different frequency components and suitable for non-linear and non-stationary signals, known as empirical mode decomposition (EMD) (Huang et al., 1998; Xie et al., 2022). EMD, first introduced by Huang and his colleagues in 1998, divides signals into intrinsic mode functions (IMFs) to extract valuable information, making it beneficial for noise reduction and feature extraction in various fields (Wang et al., 2022). EMD has been discussed in bridge SHM previously by researchers for decomposing response signals into IMFs to extract informative features for structural condition assessment (Yu and Ren., 2005; He et al., 2011; Shao et al., 2019). Notably, prior research has shown that EMD can extract quasi-static features such as temperature-induced strains from high frequency live load (Chang and Kim et al., 2012; Li et al., 2016). In summary, EMD is adopted in this paper to extract temperature-induced deflection from raw deflection. Given the aforementioned context, a case study utilizing deep learning technique to analyze the temperature induced deflection of a multi-span continuous box girder bridges was conducted, aided by the big data retrieved from the structural health monitoring (SHM) system of the bridge. 2. Methodology 2.1. Empirical Mode Decomposition (EMD) Empirical mode decomposition (EMD) was introduced based on the assumptions of existing different simple intrinsic mode of oscillation in non-stationary and non-linear time series signals. This technique first identifies the intrinsic oscillatory mode of a raw signal empirically. Afterward, a process called sifting, in which most oscillations without zero crossings between their extrema, known as riding waves, can be effectively removed, thus smoothing