Impact Damage Estimation and Localization in Composite Sandwich Plates Using Deep Learning

We introduce a novel method for the simultaneous estimation and localization of impact damage in composite sandwich plates using branched neural network. The dataset consists of 25 specimens, with low-velocity impact damage varying in impact energy and impact location. Fast Fourier Transform (FFT) and Power Spectral Density (PSD) of each specimen is obtained experimentally and Principal Component Analysis (PCA) is used for feature extraction. FFT features demonstrated better clusters in two-dimensional t-distributed Stochastic Neighbor Embedding (t-SNE) feature plots, providing higher model validation accuracies when compared to PSD for all the tasks. Shallow neural networks outperformed machine learning approaches providing best validation accuracies of 97.29% and 98.52% for damage estimation and damage localization tasks respectively. Branched neural networks with varying architectures have been trained and validated for the multi-output task of simultaneous impact damage estimation and localization, which provided best validation accuracies of 96.82% and 95.96% for the respective sub-tasks.