Predicting Multi-Mode Dynamic Responses of Structures Using Long Short-Term Memory Neural Networks

Abstract

Abstract This paper reports the progress of a research project of using Long Short-Time Memory (LSTM) deep learning neural networks for modeling structural dynamics of complex structures and predicting the structures’ vibration responses due to random excitation. In a prior study, the feasibility of the approach was investigated. The LSTM networks were applied to the responses of various simulated systems subjected to random excitation loads. While the networks demonstrated promising capabilities of modeling and predicting responses of a single dynamic mode, they exhibited difficulties in modeling responses of multiple modes accurately. To resolve the multi-mode issue, this paper presents an improved approach that converts a multi-mode problem into a set of single-mode problems. The multi-mode responses are separated into individual modal response components. A set of mode-specific sub-LSTM networks are obtained, with each trained by using the original input data and a particular modal response component. These sub-LSTM networks are then combined to yield an equivalent LSTM network to model and predict the multi-mode responses. The proposed multi-mode approach is applied to a tapered, cambered wing structure numerically modeled with finite elements and subjected to a random force excitation. The multi-mode approach shows significant improvement over the original direct approach and yields an excellent match between the actual and predicted responses.