Deep Learning-based Methods for Structural Health Monitoring Data Improvement and Augmentation

Abstract

This thesis addresses data quality challenges in Structural Health Monitoring (SHM) using deep learning techniques. A Transformer-based generative adversarial network is developed to reconstruct missing signal. An unsupervised domain adaptation-based methodology is proposed to impute missing data. A segmentation method for detecting anomalous data is developed by employing denoising diffusion probabilistic models (DDPMs). Additionally, a generative model using DDPMs is proposed to synthesize realistic monitoring data, enhancing reconstruction accuracy and data augmentation in SHM applications.