Structural damage identification based on unsupervised feature-extraction via Variational Auto-encoder

Abstract Structural health monitoring (SHM) is a practical tool for assessing the safety and system performance of existing structures. And structural damage identification has become the core of a SHM system. However, how to extract damage-sensitive features from structural response has become a challenging problem. Thus deep learning methods have attracted increasing attention for its ability to effectively extract high-level abstract features form raw data. This paper presents a damage detection method based on Variational Auto-encoder (VAE), one of the most important generative models in unsupervised deep learning. In this paper, VAE is used to process responses of the structure, which reduces the high-dimensional data to low-dimensional feature space, and then restores the original data from the low-dimensional representations. This structure forces the VAE to learn the essential features hidden behind the complex data. Taking advantage of this characteristic, we apply the VAE to damage identification task of a bridge under moving vehicle. The results of both numerical simulation and experiment are proved that the proposed method can accurately identify the structural damage/s. This method directly analyzes the measured responses of the structure without the structural element model and baseline data. It is a baseline-free data driven method, which is suitable for real engineering application in SHM.