Structural Damage Condition of Buildings with a Sparse Number of Sensors Using Machine Learning: Case Study

Abstract

The optimum location of signal sensors in actual buildings to determine the structural damage condition using machine learning is discussed in this study. The target buildings are a local government office and a Fire Station in Japan, with two acceleration sensors located on the ground and the roof level of the buildings. An additional sensor location is considered in this study. The structural damage condition is evaluated by machine learning (ML) methods from the sensor signals for five cases of single and multiple sensor locations. The maximum story drift is used as an identifier of the structural damage condition. Seven ML methods are developed, and their accuracy is compared. Several intensity measures (IM) obtained from each sensor signal are used as input features for the ML models, and the prediction importance level of each IM is evaluated in order to establish its usefulness. Finally, the results are compared to the methodology using wavelet power spectrum and convolutional neural network to predict the damage condition of buildings.