Predicting structural dynamic responses using improved long short-term memory neural network

Accurately predicting the response of a structure is of great significance for health monitoring and safety assessment of structures under dynamic loads. Therefore, this paper proposes the use of an improved Long Short-Term Memory (LSTM) neural network for simulating and predicting the dynamic responses. The key point is to establish a robust LSTM network that learns dynamics from limited training data and predicts the structure's response to natural excitation. Additionally, this paper directly utilizes the structural response data as input to the neural network. Through the training of the improved LSTM, it is not necessary to use excitation data to train the model, and the response data at a specific time point in the future can be accurately predicted, which solves the problem that the full-size structure excitation cannot be measured and the structure cannot be updated based on the dynamic responses. The trained LSTM network can comprehensively learn actual structural dynamic behaviors. The proposed method bypasses complex structural modeling using finite element models, where important modeling assumptions may be ignored to cause errors. The proposed method was verified using simulated data from a four-story shear building. The results demonstrate that the trained LSTM network can accurately reproduce the responses and predict the responses which are not included in the training data.