Structural Response Prediction of a Flat Beam with Geometric Nonlinearities with Temporal Convolutional Networks

Abstract

Complex structural systems are ubiquitous in the aerospace industry. The structural response characterization of these complex systems and components often involves extensive nondestructive testing. When testing is not an option, analysts rely on numerical simulations to predict the structural response. These simulations are performed by solving the partial differential equations (PDE) of motion. In the presence of nonlinearities, solving the relevant PDEs becomes computationally intractable for realsystems. Furthermore, the physical parameters required to perform these simulations may be unknown or hard to obtain. To overcome these challenges, we propose to train a data-driven model that can be used in place of more expensive simulations. The focus of this paper will be the exploration of the Temporal Convolutional Network (TCN). The TCN has been demonstrated to be a powerful architecture in the field of speech recognition and speech-to-text generation but it has not been studied for the application to structural response. The TCN model is enhanced with physical constraints that act as regularizers by leveraging relevant domain-knowledge. The TCN performance is compared to the traditional autoregressive model which is a popular approach in this field. Error metrics are evaluated in the time and frequency domain to assess the performance of the methods considered.