Unsupervised Structural Health Monitoring of Large-Scale Structures Using Frequency Spectrum Including Time Information and Deep Learning

Guidelines for seismic performance evaluation of existing structures have been published by FEMA and the European Commission. Additionally, these organizations have presented frameworks for strengthening the resilience of society and industry. These frameworks require countermeasures for early recovery after natural disasters. For rapid post-earthquake recovery, it is necessary to assess the structural integrity and implement repairs. To achieve this, the utilization of structural health monitoring is essential. In structural damage assessment, loads and response displacements are important physical quantities, but these are difficult to obtain directly from structures. Therefore, in structural health monitoring, structural integrity is evaluated using measured accelerations. However, the superposition of uncertainties related to seismic motion and structures can sometimes complicate the setting of thresholds for structural integrity evaluation. To detect structural anomalies caused by earthquakes rapidly and with high accuracy, structural health monitoring methods incorporating deep learning have begun to be researched. Additionally, by extracting features from measured data using deep learning, prior setting of evaluation thresholds becomes unnecessary. Generally, training deep learning models requires large amounts of training data. However, since it is difficult to acquire anomaly data from existing structures, seismic response analysis is utilized. Typically, to represent real phenomena, analytical models require many degrees of freedom. In frequency domain response analysis, the frequency response function is defined as the ratio of the Fourier spectra of the structural seismic response to the seismic motion. By modeling the frequency response function, the seismic response of structures can be calculated. In this paper, we develop a method to calculate structural seismic responses by modeling the frequency response function with an AR model. This eliminates the need to set parameters such as degrees of freedom, mass, and stiffness of the analytical model. Furthermore, we propose a structural health monitoring method that uses this approach for training deep learning models. Verification of this method using experimental data from a full-scale steel building demonstrated that anomaly detection was possible from limited observational data only.