Nonlinear seismic site response classification using K-means clustering algorithm: Case study of the September 6, 2018 Mw6.6 Hokkaido Iburi-Tobu earthquake, Japan

Abstract

Spectral ratio methods have been widely used in evaluation of nonlinear seismic site response. Nevertheless, it remains inefficient and subjective to identify stations with nonlinear site response according to empirical threshold values of spectral ratio nonlinear degree indicators. This study, which was the first to apply the machine learning clustering algorithm to address this problem, used the September 6, 2018 Mw6.6 Hokkaido Iburi-Tobu earthquake (Japan) as an example. First, we calculated the surface/borehole and horizontal/vertical spectral ratios using strong ground motion data recorded by KiK-net vertical array and K-NET stations, respectively. The degree of nonlinear site response (DNL) and percentage of nonlinear site response (PNL) were computed using the difference between the strong motion of the mainshock and weak aftershocks as the reference for linear site response. Then, the K-means clustering algorithm was incorporated in the identification of nonlinear site response using the DNL, PNL, strength of ground motion (PGA) and site condition (VS20 or VS30) as explanatory variables. After careful multicollinearity diagnosis and confirmation of the optimum clustering number, we successfully classified the stations into two clusters with nonlinear and linear site responses. Overall, the clustering results were found in good agreement with the classification results based on empirical thresholds of several nonlinear indicators. For the stations identified with nonlinear site response, the reduction of amplification and frequency shift could be observed from the spectral ratio curves regarding the ground motions in the mainshock and the reference weak aftershocks, demonstrating typical nonlinearity response characteristics. Furthermore, a comprehensive indicator of nonlinear site response occurrence probability (NLscore) was obtained from a linear weighted combination of the normalized variables (PGA, VS30/VS20, DNL and PNL). The NLscore ranking of the top several stations was found consistent with the clustering identification results, irrespective of the choice of combination scheme. It was demonstrated that the performance of clustering algorithm in this application was satisfactory and that the identification results were convincing and robust. This work provides an enlightening example of using state-of-art machine learning technique to solve the traditional earthquake engineering problems.