Deep learning approaches for impact identification on composite structures under environmental and operational variabilities: a comparative study

This study presents a comprehensive evaluation of deep learning approaches for impact identification in composite structures under environmental and operational variabilities (EOVs). Five representative architectures—Convolutional Neural Networks (CNNs), Temporal Convolutional Networks (TCNs), Recurrent Neural Networks (RNNs), Graph Neural Networks (GNNs), and Transformers (XFMRs)—are compared across two key tasks: impact localisation (predicting spatial coordinates) and impact force reconstruction (estimating time-varying force histories). Particular emphasis is placed on model robustness when testing conditions deviate from those used in training, including temperature changes and impact mass variation. Additionally, the effects of critical data acquisition parameters—such as sampling frequency, signal window length, and sensor density—on model performance and generalisability are systematically investigated. Experimental validation is conducted using controlled impact tests on composite panels, providing insight into the strengths and limitations of each model architecture in realistic structural health monitoring scenarios.