Wave Propagation Modeling via Neural Networks for Emulating a Wave Response Signal

Abstract

AbstractWave propagation in structures is generally computed by numerical methods such as finite element method, spectral element method, etc. In these numerical methods, spatio-temporal discretization of the partial differential equations is performed using a fine mesh which leads to high computation cost but precise results. A trade-off between accuracy and computational cost can be achieved by adopting deep learning-based approaches. This research demonstrates an alternative deep learning-based approach for predictive modeling of wave propagation signals within damaged structural elements. Our goal is to evaluate the wave propagation spatio-temporal solution matrix for a given crack depth and crack location within the structural element. In this framework, deep-learning-based surrogate modeling is proposed by utilizing a deep convolutional autoencoder (DCAE) to learn the wavefield representation and project it to a compressed domain called latent space. This latent space works as labels for a feed-forward neural network (FFNN) followed by DCAE. This process eliminates the need to solve the system’s governing equations each time, leading to significant savings in computational costs and making the method excellent for issues that require repeated model computations. In DCAE architecture, we integrated the squeeze-and-excitation (SE) block which works as a channel-wise attention mechanism and enhanced the performance of the model. The DCAE with SE block achieved the very good reconstruction accuracy. This deep learning-based wave propagation predictive model can be a valuable resource for generating data for a given crack depth and location, which can be used for inverse formulations and various structural health monitoring (SHM) application.KeywordsWave propagationDeep-learningSurrogate modelingConvolution neural networksAutoencoder