Generative Adversarial Networks for Data Augmentation in Structural Health Monitoring

The present study contributes to the development of an automated system that leverages Deep Convolutional Neural Network-based approaches for the monitoring of prestressing wire breakage in concrete structures subjected to severe aging factors, such as bridges. Advanced methodologies for data acquisition and signal processing within the framework of Structural Health Monitoring (SHM) are explored, focusing on data augmentation techniques to address the critical issue of limited data availability. The work involved conducting controlled destructive tests on two prestressed concrete bridges in L'Aquila, Italy, prior to their planned demolition. A combination of accelerometers and acoustic emission sensors was utilized to capture vibration data during the controlled breakage of prestressing wires. This approach provided essential real-world data, which is crucial for comprehensive analysis. The acquired elastic wave were transformed into time-frequency representations via the Short-Time Fourier Transform (STFT), employing various window sizes to find an optimal balance between time and frequency resolution. The resulting spectrograms, normalized for consistency, served as the primary input for training Generative Adversarial Networks (GANs), which were utilized to address the data scarcity. This study compares different GAN architectures, including the Deep Convolutional GAN (DCGAN), Wasserstein GAN (WGAN), and Least Squares GAN (LSGAN). Among these, the LSGAN showed superior performance, producing stable and high-quality augmented STFT images. The generation of synthetic datasets plays a central role in enhancing deep learning algorithms to identify structural anomalies, thereby improving predictive maintenance and critical degradation detection capabilities. Through these methodologies, the thesis contributes new perspectives into the application of deep learning for SHM, emphasizing the importance of data augmentation to support more effective and reliable infrastructure monitoring. The findings indicate the robustness of the GANs for augementation of STFT images to enhance SHM, which can lead to an increment in the resilience, safety and longevity of critical civil infrastructure, especially bridges.