Recent machine learning (ML) algorithms have resulted in new paradigms for extracting spatio-temporal characteristics (STCs), such as frequency spectra of critical infrastructures and performing structural health monitoring. However, the accuracy of the STCs extracted using ML is affected by any noise in the data used to train and test the ML algorithms. While noise reduction methods have been successfully proposed, they are application-specific, and none consider the dynamics of the targeted system. Hence, a novel framework named time-inferred autoencoder (TIA) is proposed. The TIA is based on a long, short-term memory (LSTM) neural network to learn the dynamics of the system and a maximum correntropy loss function for noise removal. The robustness of the TIA is validated by collecting a video of an undamaged beam for training the framework and learning the structure’s STCs. Later, the capability of the trained TIA to adapt to changes in the system’s dynamics and reconstruct the STCs is validated by recording noise-corrupted videos of a beam in three different damaged configurations. Results of laboratory tests showed that the TIA reconstructs the natural frequencies of the structure with an error of less than 1%. If further developed, the proposed framework can be used as a structural dynamics tool given its robustness and capability to adapt to changes and noise in the system.
|