2021 Volume 2 Issue J2 Pages 598-608
This paper presents a case study of the validation of the deep learning system for determining the degree of liquefaction based only on the strong-motion records. The authors assessed the damage degree of liquefaction via the deep learning system with 140 strong-motion records in the Kanto region during the 2011 off the Pacific coast of Tohoku Earthquake, and compared the results with the field survey reports where the liquefaction damages were confirmed. The damage degree of liquefaction was defined as "DDL", and the excess pore water pressure ratio (EPWR) in the ground was used as the correct answer label for deep learning, which was divided into five levels from 0 to 4. As a result, DDL = 0 to 2 (low EPWR) were found in 87 locations, and DDL = 3 or 4 (high EPWR) were found in 53 locations. The results were compared with the liquefaction hysteresis map of the Tohoku earthquake. About 8% of the sites with DDL=3 or higher corresponded to the observed liquefaction sites, and about 91% of the sites with DDL=2 or lower corresponded to the non-liquefied sites.
