Time-Inferred Autoencoder for Construction and Prediction of Spatiotemporal Characteristics from Dynamic Systems Using Optical Data

Dynamic spatiotemporal (ST) characteristics of a structure, such as natural frequency, frequency spectra, and operational deflection shapes, are used to assess and monitor a system’s condition. Recent developments in machine learning and computer vision approaches have created new paradigms to utilize ST characteristics and extract complex dynamics. Commonly used machine learning models, such as traditional autoencoders, lack the ability to learn complex phenomena in the latent space and extract important structural dynamic parameters. Hence, this study proposes a novel model based on a time-inferred autoencoder (TIA) to learn the ST characteristics of a structure of interest. To validate the proposed approach, experiments were conducted by collecting high-speed videos of multi-degree of freedom systems, which were used to train the TIA model and understand the underlying dynamics of the targeted system. Following the training stage, the TIA model was able to reconstruct the full line-of-sight optical data that can be used for identifying resonant frequencies. The robustness of the TIA approach and its capability to adapt to changes in the dynamics of the inspected structure was evaluated as a function of different levels of structural damage. If further developed, TIA can be used as a structural health monitoring tool to learn the dynamics of a system from videos without having to utilize contact-based sensors.