Study on structural health monitoring system using transfer learning

Abstract

Japan's geographic location is in a seismically active area, occurring large earthquakes in the future, disaster prevention has required to realize a sustainable society. In addition to earthquake preparedness, measures for existing infrastructure are needed to realize a sustainable society. Due to the aging of infrastructure and the shortage of inspection personnel, more efficient structural health monitoring methods are required. In addition, structural health monitoring systems face the problem of difficulty in setting threshold values for comparison with acquired data due to, for example, variations in the stiffness and yield displacement of the structure. To solve those issues, structural health monitoring methods utilizing deep learning have been studied in Japan and overseas. In image identification tasks using deep learning used in those methods, it is desirable to extract more information about the estimation target to improve identification accuracy. However, most previous studies have used only image data generated by response data as input to CNNs. This means that the image data do not include the correction information between the input and response data. Therefore, the authors attempted that the input image data, to input the CNNs, incorporates a feature value not only for the seismic response but also for input ground motions to improve the accuracy of structural health judgment by deep learning. The analytical study was conducted to verify our proposed method. This paper describes that our generating image method improves the damage detection accuracy significantly.