A Deep Learning-Based Framework for Damage Detection With Time Series

Natural hazards have caused damages to buildings and infrastructures and economic losses in many countries. Immediate emergency response requires accurate damage detection. In recent studies, advances in deep learning approaches have put forward the development of data-driven damage detection. However, these proposed approaches can successfully classify damaged samples into different damage states without quantifying them. Therefore, we propose a deep learning-based framework for damage detection, which comprises signal processing and damage recognition modules. The signal processing module is introduced to denoise the noisy signals collected from sensors and process the denoised signals via fast Fourier transform. The damage recognition module is characterized by a novel network with the architecture of encoder-decoder-encoder. We train the network with only undamaged samples to capture the pattern in the undamaged state. In the testing stage, given an unknown damage state, we feed damaged samples into the trained network to quantify it by outputting the corresponding probability of damage. We conduct experiments to verify the validity and feasibility of the proposed framework with a dataset of a building subjected to excitations. Our approach can map samples from various damage states to a specific probability of damage, which increases as the intensity of excitation increases.