Fusing State-Space and Data-Driven Strategies for Computational Shock Response Prediction

This paper proposes a state-space-based approach for computational simulation and prediction of shock response for both time histories and shock response spectra. The method operates by developing a nominal model for the system through state-space identification and then modeling the resulting residual through an artificial neural network. The model for the residual is then folded back into the nominal system via Kalman filtering, allowing for forward computational simulation without any measurement information. The proposed identification method is applied to a three-degree-of-freedom system and a high-fidelity finite element model built in Abaqus. The proposed method provides reasonable predictions for time histories of excitations not seen during training or identification and produces useful predictions of the associated shock response spectra.