Artificial intelligence-based damage localization method for building structures using correlation of measured structural responses

Buildings deteriorate and suffer from unexpected severe loads, such as earthquakes and typhoons. Damages occurring in these buildings due to such loads can be disastrous. In the field of structural health monitoring, there have been many efforts to prevent disasters by identifying damage in structures. To deal with large quantities of data in responses measured from the structures for damage detection, various techniques based on artificial intelligence have been developed. The most recent methods employ data for damaged states that are usually obtained from numerical analysis as well as data corresponding to future behavioural predictions. Thus, supervised learning-based methods using future or labelled data cannot be applied to real-world situations. In this study, an unsupervised damage identification method is proposed using a convolutional neural network (CNN) trained exclusively with healthy state data. The discrepancy between healthy state data and output data from the CNN with the damaged state response is displayed as damage indicators. The distribution of the indicators is quantitatively analysed using a correlation coefficient (CC). The storey with the lowest correlation, or the storey where the data with a minimum CC value is extracted, is identified as the damage location. The validity of the presented method is examined by numerical and experimental studies.

Automated summary

Buildings can deteriorate and suffer severe damage from unexpected loads such as earthquakes and typhoons. This study proposes an unsupervised damage identification method using a convolutional neural network trained exclusively with healthy state data. The discrepancy between healthy state data and output data from the CNN is used as damage indicators, and a correlation coefficient is utilized to quantitatively analyze the distribution of the indicators. By identifying the storey with the lowest correlation, the damage location can be determined. The validity of the method is examined through numerical and experimental studies.