Vibration-based multiclass damage detection and localization using long short-term memory networks

This paper proposes a novel damage detection and localization method of civil structures using a windowed Long Short-Term Memory (LSTM) network. A sequence of windowed samples are extracted from acceleration responses in a novel data pre-processing pipeline, and an LSTM network is developed to classify the signals into multiple classes. Predicted classification of a signal by the LSTM network into one of the damage levels indicates the presence of damage. Furthermore, multiple structural responses obtained from the vibration sensors placed on a structure are provided as input to the LSTM model, and the resulting predicted class probabilities are used to identify the locations with a high probability of damage. The proposed method is validated on the experimental benchmark data of the Qatar University Grandstand Simulator (QUGS) for binary classification, as well as the Z24 bridge benchmark data for multiclass damage classification associated with different levels of pier settlement and the numbers of ruptured tendons. The results show that the proposed LSTM-based method performs on par with the one dimensional convolutional neural network (1D CNN) on the QUGS dataset and outperforms 1D CNN on the Z24 bridge dataset. The novelty of this paper lies in the use of recurrent neural network-based windowed LSTM for multiclass damage identification and localization using vibration response of the structure.