Document Type
Conference Proceeding
Publication Date
9-2020
Abstract
In this paper, deep learning techniques are applied to predict a building’s structural response to strong ground motions. Data from sensors near and inside structures measure accelerations during strong ground motions. The Building Research Institute (BRI) ANX building, an eight-story structure, has experienced major earthquakes since 1998. Sensors in the building provide and accumulate big data of historic events. Changes of the natural frequency of the ANX building from the big data is initially quantified. The time-series data of the historic events can be used to predict future response to future events using deep learning models rapidly. Although previous literature attempted similar time-series predictions, the building’s multiple floors’ accelerations were not predicted from base/ground acceleration. Accurate deep learning models could predict the relationship between ground/base motion and several floors’ motions. Four different machine learning algorithms are implemented to predict multiple floor responses from ground accelerations. The supervised AI algorithms used are: Deep Neural Networks (DNN), Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), and Levenberg–Marquardt Recurrent Neural Networks (LM-RNN). The neuron structures for all types of neural networks are explained. Big data management and filtering is necessary for importing accurate meaningful features to neural networks. A new proposed algorithm is presented for all learning models and training is performed with a new selection of hyperparameters. Firstly, the supervised learning techniques are trained with a window of N-past features. Subsequently, the models are deployed for inference on seismic events which the models have not been trained on. Parametric studies of windowing, down-sampling, batching, network structure, number of training epochs, and dropout rates are compared. The different AI techniques’ accuracies are compared. Although there are slight changes in the natural frequency of the BRI ANX building over time, the LSTM neural networks could still predict acceleration response of recent events accurately. Fourier Spectrum comparisons between the sensors’ true acceleration signal and the neural network predicted acceleration signals are used as validation. Inference of trained models can be used to continuously predict multiple floor responses due to earthquakes.
Recommended Citation
Torky, Ahmed A.; Ohno, Susumu Prof.; and Kashima, Toshide Prof., "DEEP LEARNING TECHNIQUES FOR STRUCTURAL RESPONSE PREDICTION DURING STRONG GROUND MOTIONS" (2020). Civil Engineering. 110.
https://buescholar.bue.edu.eg/civil_eng/110
Comments
The authors would like to acknowledge the support provided by the Building Research Institute (Japan) for providing strong motion records of the BRI annex building. This study was supported by JSPS KAKENHI Grant number 19K22002.