Feature-Based Neural Networks for Impact Localization in Composite Structures

Accurate impact localization is critical for the Structural Health Monitoring of composite materials, particularly in aerospace applications where early damage detection ensures safety and reduces maintenance costs. This study explores a neural network-based approach to localizing impacts on composite structures using engineered features extracted from piezoelectric sensor signals. Multiple MLP-based models have been trained with feature sets extracted from this signals using different supervised and unsupervised techniques. This systematic comparison highlights how the choice of feature selection strategy influences localization accuracy and model generalization. Notably, certain selection techniques yield models with stronger robustness when localizing impacts outside the original training grid. These findings suggest that careful feature selection, combined with domain-informed engineering, can enhance model performance and spatial generalization in Structural Health Monitoring applications involving composite structures.