Enhancement of laser-generated ultrasonic wavefield in noisy experimental environments using latent assimilation

Abstract

This study explores the fusion of models and data in dynamic systems for structural health monitoring, using Kalman filter-based estimations. For linear Gaussian state-space models, the Kalman filter is a low-complexity optimal solution that merges observation data with state vectors derived from dynamic process modeling. However, Kalman filtering struggles with high-dimensional systems due to the large matrix inversion required in the algorithm. To overcome this limitation, we propose using reduced-order modeling (ROM) through latent assimilation (LA) to refine ultrasonic wavefields affected by environmental noise. Specifically, we employ a convolutional autoencoder (CAE)-based dimensionality reduction approach. The CAE encodes physical space into a latent space, allowing data assimilation (DA), performed via the Kalman filter, to be executed efficiently. To capture the evolution of latent state-space vectors, as opposed to the trend in the literature, we used the Gated Recurrent Unit (GRU) for its advantage in terms of computational efficiency and training speed, due to its simpler architecture and fewer parameters. The filtered latent vectors are then decoded back into physical space via the CAE. This method is demonstrated on a laser-ultrasonic wavefield collected in a noisy environment, with results aligning with the study’s hypotheses. The proposed technique is therefore applicable to ultrasonic datasets acquired in an inhospitable environment.