PSI - Issue 64

Marco Pirrò et al. / Procedia Structural Integrity 64 (2024) 661–668 Author name / Structural Integrity Procedia 00 (2019) 000–000

662

2

1. Introduction

In the last years, research in Structural Health Monitoring (SHM) of bridges experienced increasing interest, mainly because bridges are continuously exposed to deterioration caused by increased service loads and climate changes; in addition, a large number of bridge infrastructures is approaching the end of design lifetime (Federal Highway Administration, 2008). A well-assessed OMA-based SHM approach aims to continuously collect the dynamic responses in order to analyze the time evolution of the modal characteristics (e.g., natural frequencies and mode shapes): in fact, since modal parameters are linked to the mass and stiffness properties of the structure, their variation along time can provide evidence of degrading processes (Peeters and De Roeck 2001, Farrar and Worden 2007, Magalhães et al. 2012). However, in real SHM applications, the traditional OMA-based approach might present some drawbacks, related, for example, to possible estimation error during the modal identification. Furthermore, modal parameters are sensitive not only to structural changes, but also to environmental and operational variability (EOV): in bridges’ applications, EOV is mainly due to temperature and vehicular traffic, affecting the evolution of modal characteristics (Peeters and De Roeck 2001, Magalhães et al. 2012, Borlenghi et al. 2023). Therefore, prior to any anomaly detection procedure, the EOV effects must be understood and minimized, since EOV could lead to false alarms during the monitoring activities. It should be noticed that Deep Learning (DL) algorithms (Goodfellow et al. 2016) are becoming more and more attractive also for SHM and increasing attention recently raised towards Artificial Intelligence (AI) framework (Worden and Manson 2007, Bao and Li 2021). DL architectures map a set of input data through internal layers of artificial neurons, with the aim of modeling the representative information contained in the input data. Among the DL architectures, the autoencoder (AE) network seems to be especially suitable to SHM applications, because it can learn the hidden characteristics of the input data (e.g. dynamic measurements) by reducing the dimensionality of the input set (Pathirage et al. 2018). More specifically, an AE model outputs an approximation of the input time series using only its lower-dimensional representation, which has been coded by layers of neurons (Goodfellow et al. 2016). Differently from OMA-based SHM, the AE model can implicitly account for the influence of the EOV in the time series. The SHM strategies using AE networks generally involves the training of a single AE model for each sensor of the monitoring system (Wang and Chan 2021, Finotti et al. 2021, Finotti et al. 2022, Giglioni et al. 2022). The present paper proposes a different training procedure, with just one AE network being trained by using simultaneously all the available sensors. Of course, since training is performed in an unsupervised way, the investigated structure is supposed to be in its intact conditions under normal EOV during the training period. When newly collected data are available, the trained AE model is used for data reconstruction purposes: if the structural condition and the EOV are similar to those of the training period, the reconstructed output will approximate well the original input data. Conversely, if the structure is experiencing any anomaly, the AE model would not reconstruct the input data with good approximation: in order to evaluate such reconstruction ability (conceivably linked to the integrity of the investigated structure), the Mean Absolute Error (MAE) between the input data (e.g. dynamic measurements) and the reconstructed output is computed. It has to be expected that the MAE values will increase as soon as the structural condition departs from the normal condition. The proposed algorithm is applied to data collected during the continuous monitoring of the KW51 railway bridge (Maes and Lombaert 2021). The KW51 bridge (Leuven, Belgium) has been monitored for 14 months before, during and after one retrofitting intervention. In this context, an AE model is trained using the accelerations measured after the structural retrofitting and under changing environmental (i.e., temperature and humidity) and operational (i.e., train passages) conditions. The trained AE turns out to recognize – through the MAE increase − the condition of the bridge before the retrofitting as different from the one accounted during training. 2. Theoretical background on autoencoder An AE network is used to reconstruct a set of input time series, by means of subsequent layers of artificial neurons. For sake of simplicity, this section deals with the case in which the AE is made by only three layers (Fig. 1) − encoder, hidden and decoder layer − but the mathematical derivation can be generalized to multiple layers (Pathirage et al. 2018). The encoder layer aims to learn a set of lower-dimensional features (called hidden