Impact diagnosis in stiffened structural panels using a deep learning approach

Abstract

Low-velocity impact on a structure emanates an elastic wave that propagates through the structure carrying a wealth of information about the impact event. This propagating wave can be visualized through a series of images (time-frames in the context of computer-vision) in the time–space domain collectively referred to as the wavefield. An approach for the autonomous analysis of these wavefields is presented in this article for the purpose of impact diagnosis, that is, identifying the impact location and reconstructing the impact force time-history. The high spatio-temporal dimensionality of the wavefield mandates the use of deep neural networks for analysis; however, unlike the traditional object detection problem in computer-vision, the nature of the impact diagnosis problem requires the capturing of context from the wavefield evolution. This necessitates learning across multiple time-frames of the wavefield simultaneously rather than focusing independently on each frame. While scanning simultaneously across multiple time-frames provides indispensable information about the wave propagation phenomenon in terms of its interactions with geometric features, boundaries, and so on, it mandates the use of deep learning models that can analyze this complex phenomenon in both spatial and temporal domains. A unified CNN-RNN network architecture is employed in this article to address this issue of spatio-temporal information extraction. The proposed approach is verified using simulated wavefields obtained from the finite element analysis of a five-bay stiffened aluminum panel. In order to demonstrate the generalization capabilities of the model, simulated wavefields corresponding to highly idealized impact scenarios are used for training, whereas for testing, the ones corresponding to more realistic impacts are used. It is shown that by incorporating the physics-based concept of time-reversal in the recurrent part of the network, better network performance can be achieved. The potential extension of the proposed methodology to an end-to-end vision-based impact monitoring system is also discussed at the end.