PSI - Issue 64

Ricardo Perera et al. / Procedia Structural Integrity 64 (2024) 1369–1375 Author name / Structural Integrity Procedia 00 (2019) 000–000

1370

2

1. Introduction Fiber-Reinforced-Polymers (FRP) has increased more and more its use in recent decades to strengthen concrete structures, both by external bonding of FRP sheets (Externally Bonded Reinforcement technique - EBR) (Dong et al. 2013) or by inserting a FRP sheet or bar in the concrete cover (Near Surface Mounted technique – NSM) (Al-Saadi et al. 2019), because of its excellent mechanical properties and durability (Siddika et al. 2019). However, brittle failure modes are typical of this type of reinforcement (Ortiz et al. 2023). Hence, the early identification and localization of damages in these structures are critical for their safe and reliable application. In recent years, Autoencoders (AEs) have been extensively studied and applied in the field of Artificial Intelligence (AI) as demonstrated by the large number of related works which have emerged. AEs are deep learning architectures capable of reconstructing data instances from their feature vectors. They work on all sorts of data and have been widely used for image data in dimension reduction (Dasan et al. 2021) and noise reduction (Ahmed et al. 2021), where they can learn to extract features such as edges, shapes, and textures from the input image. Furthermore, convolutional autoencoders, which combine the autoencoder and the convolutional neural network, are specially required for modeling image data because of their ability to better retain the connected information between the pixels of an image (Choi et al. 2021, Fettah et al. 2024). In the same way, due to their strong capacity for generalization, the application of autoencoders in vibration-based SHM has increased over the last few years. Studies performed in the past have also demonstrated the good performance accuracy for damage detection of convolutional autoencoders (Lee et al. 2021, Yuan et al. 2021, Yessoufou et al. 2023). In this work, a deep learning approach using convolutional autoencoders for exploiting the raw electromechanical impedance (EMI) signatures is implemented to automatically detect anomalies in an unsupervised manner for FRP strengthened concrete structures. To validate the effectiveness of the method, an experimental test campaign was performed. A concrete specimen strengthened with FRP and instrumented with PZT transducers in different location was subjected to different loading stages which provided different levels of damage. 2. Convolutional Autoencoders Autoencoders are deep learning models able to learn the feature representation of the input data in an unsupervised way. Autoencoders are made up of three main components, namely, an encoder, a bottleneck and a decoder (Fig. 1). With the encoder, a feature representation of the input data is achieved through their dimensionality reduction. On the other hand. the goal of the decoder is to transform the feature representation into the same dimension as the input data to achieve reconstruction of the input data. The encoder and decoder weights are typically initialized randomly and are trained to minimize a reconstruction loss, which measures the difference between the input and the reconstructed output. By using a training dataset and starting from initialized randomly encoder and decoder weights, the network parameter learning is completed by updating the weights in the autoencoder through error back-propagation and gradient descent algorithms. The loss between the input data and the reconstructed output, usually measured by the root mean square, must be minimized to ensure the effectiveness of the autoencoder.