PSI - Issue 44

2

Valentina Giglioni et al. / Procedia Structural Integrity 44 (2023) 1948–1955 Valentina Giglioni et al. / Structural Integrity Procedia 00 (2022) 000–000

1949

Nomenclature ℎ ! i-th node at the hidden layer " j-th value of the input vector " j-th value of the output vector weight matrix of the encoder ′ weight matrix of the decoder # biases term in the encoder $ biases term in the decoder activation function of the encoder activation function of the decoder

reconstruction loss number of sensors # threshold to label short sequences as healthy or damaged $ threshold to evaluate the percentage of damaged short sequences within a macro-sequence 1. Main text Earthquake-induced effects in civil structures and infrastructures require a prompt and effective evaluation to preserve their integrity and serviceability and ensure safety (Zhang et al. (2018), Bao and Li (2021)). Indeed, bridges health conditions are expected to be rapidly investigated after a critical event by carrying out additional visual inspections, in order to prevent collapse as well as economic and human losses. Nevertheless, this approach may lead to important delays and unreliable results. For these reasons, strategies to enable real-time post-earthquake bridge assessment are particularly needed to support decision making and guide maintenance activities, especially in high seismicity zones like Italy (Liu et al. (2022), Lu et al. (2021), Todorov and Muntasir (2022)). In this framework, following the recent developments in computer science and Machine Learning (ML) fields (Worden and Manson (2007), Rafiei and Adeli (2018)), an autoencoder-based technique for damage detection of roadway bridges is proposed. The methodology lies in the field of unsupervised learning, which is deemed particularly suitable for addressing civil engineering Stuctural Health Monitoring (SHM) problems. This relies on the fact that unsupervised approaches require input-only data for training, without any labels indicating structural classes (Rastin et al. (2021), Jiang et al. (2021)). In this paper, a Multi-Layer Perceptron (MLP) autoencoder is independently trained for each sensor to correctly reconstruct raw acceleration sequences measured during operational conditions, wherein the bridge is assumed to be healthy. Same length testing sequences stemming from unknown scenarios are afterwards fed into the autoencoder to test its performance. To this aim, a specific index quantifying the difference between the original and the reconstructed input is selected as a feature for anomaly detection, enabling to classify newly acquired sequences as healthy/damaged. However, it should be highlighted that, whilst training is carried out by analyzing short sequences, damage detection is performed at the level of macro-sequences. This allows (i) to limit the computational costs, (ii) to avoid the risk of losing some dynamics information and (iii) to reduce the rate of false detection errors. The procedure is tested on the Z24 benchmark bridge (Peeters and De Roeck (2000)). First of all, a FE model is built and calibrated using previous knowledge about materials, geometry and structural dynamics properties (Masciotta et al. (2016)). Then, load time histories randomly generated from white noise spectrum are applied to the model in order to simulate vibrations during operational conditions. The number and intensity of such forces are determined to minimize the differences between the estimated and real measurements in terms of the root-mean-square (RMS) acceleration. Then, a realistic earthquake-induced damage scenario is introduced to the FE model by reducing the concrete’s elastic modulus of the supporting piers. As such, the acceleration responses simulated in healthy/damage conditions through several time-histories analysis are extracted from different nodes and used to train and test the autoencoder network built for each sensor. The implemented methodology is deemed effective in detecting specific