An Optical Temporal and Spatial Vibration-Based Damage Detection Using Convolutional Neural Networks and Long Short-Term Memory

Structural dynamics provide critical information for structural health monitoring (SHM), such as changes in the modal behavior which indicate damages. However, for complex systems with noisy operational environments, many factors may influence the estimation of natural frequencies and other modal domain SHM features, such as varying mass distribution of bridges and the variation due to fluctuating temperature and unideal boundary conditions. For this reason, allying the mode shapes with the natural frequencies to forecast damages would pose a more robust solution. Among the techniques existent to perform damage detection, data-driven models, such as machine learning algorithms, are becoming widely used currently. For mode shape extraction, convolutional neural networks (CNN) have been applied to imagery data, allowing to extract full-field mode shapes with a denser spatial resolution (quasi-full field) of the structure if compared to traditional hardware. Combining CNN with long short-term memory (LSTM) network will associate the temporal dependency of the frames with its features which will be more specific for SHM decision-makings. In addition, for the circumstances with low vibration amplitude and subpixel image resolution, applying phase-based motion estimation (PME) and phase-based motion magnification (PMM) allows to extract the natural frequencies with subtle motion magnified at the resonances aiding to emphasize the dynamic features desired. As the training of the deep learning model, a lab-scale truss structure was adopted with different load conditions in order to obtain the required data, and the performance is cross-validated.