Damage Classification Method for a Wooden Building Based on Hysteresis Loop Using CNN

Abstract

The shape of the restoring force-story drift angle relationship curve (hysteresis loop) derived from the seismic response data of a building changes depending on the structural health of the building. In this study, we utilized this characteristic for structural health monitoring and proposed a method to assess the risk of collapse due to future earthquake motions based on hysteresis loop images using a convolutional neural network (CNN). We investigated the impact of axis range and resolution settings in the proposed method, comparing the accuracy and recall of the unsafe model in each case. By adjusting the lower limit for the maximum input acceleration of the seismic wave, we achieved a recall of approximately 90% for the unsafe model. Therefore, it can be concluded that the limit of the maximum input acceleration is one of the most influential parameters in damage classification using CNN.