Long Short-Term Memory Neural Networks for Predicting Dynamic Response of Structures of High Complexities

Abstract

Abstract This paper presents an initial investigation on the feasibility of modeling structural dynamics of complex structures using the Long Short-Time Memory (LSTM) deep learning neural networks, and predicting the structures’ vibration responses due to random excitation. LSTM networks are applied to the responses of various simulated systems subjected to random excitation loads, including mass-spring-damper systems with linear or nonlinear Duffing springs, a cantilever beam, and a tapered, cambered wing structure. Given a known force input, the dynamic response of the system is simulated in Matlab or ANSYS, which is used to train the LSTM model. In the case of mass-spring-damper and beam systems, the excellent agreement between the test and LSTM-predicted responses demonstrates the potential of the LSTM method for predicting vibration responses. In the case of cambered wing structure, the LSTM shows difficulties in dealing with responses consisting of multiple modes. Parametric studies are also performed to discover possible means for performance improvement. The studied learning parameters include the number of hidden units, the number of LSTM layers, and the size of train data. It is found that all these parameters have significant impact on the model accuracy. While it is always beneficial to have as much and there could be an optimal setting for the number of is not determined by the studies, they provide valuable directions to.