http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
- 中文 을 입력하시려면 zhongwen을 입력하시고 space를누르시면됩니다.
- 北京 을 입력하시려면 beijing을 입력하시고 space를 누르시면 됩니다.
RISS 인기검색어
- 검색키워드
- 저자 : Elipse, Carlo Alvarado (검색결과 1 건)
-
Elipse, Carlo Alvarado 세종대학교 대학원 2022 국내박사
In the recent years, numerous occurrences of pavement subsidence were observed around Seoul City area that raised concerns of the city government. Underground inspection of Seoul city roadways using ground penetrating radar (GPR) was conducted to discover the cause of the pavement subsidence and it was found to be underground cavity. Typically, to locate the underground cavity from the GPR survey data, an expert manually inspect the location of hyperbola signals that represents cavities in the whole GPR survey, however, this method is very time-consuming. To solve this problem, Tran (2020) developed a RetinaNet-based deep learning neural network that automatically detect and identify underground cavities from the GPR survey data. Although the Tran’s developed method successfully detects cavity locations, there are numerous misclassified locations that are not cavity but has almost the same features with the cavity locations, called non-cavity locations. Therefore, there’s a need for an additional process that can further eliminate non-cavity locations from the candidates while preserving the real cavity locations. In this study, a deep learning neural network, specifically an artificial neural network (ANN), that further classifies cavity and non-cavity locations based on geometric and numerical parameters from the GPR survey data was proposed. To determine the significant parameters to be used as input for the training database, parametric investigation was conducted in a collection 282 cavity and 572 non-cavity locations, summing up to 854 locations. It was found that there are 2 geometric and 2 numerical parameters that has a strong relationship in the classification of the locations. For the geometric parameters, the ratio of the height of the hyperbola signal from the B-scan and D-scan, called the H Ratio, and the ratio of the width of the hyperbola signal from the B-scan and D-scan, called the W Ratio, were considered. Meanwhile, there were also two parameters found in the investigation of numerical data from the GPR survey: polarity ratio standard deviation (PRS) and the pseudo-dielectric constant (PDC). PRS is defined as the standard deviation of the ratio of positive peak signal and the negative peak signal of the points located at the upper half of the hyperbola while PDC represents the effective dielectric constant of the underground object and the medium above it. A training database was formed including the four parameters identified from the parametric investigation and the classification of each location. The total accuracy of the trained model on the whole training database yielded to 95.20%. As a validation, the trained model was applied to 74 locations consist of 37 cavity locations and 37 non-cavity locations. All of these locations were classified as cavity by the RetinaNet-based deep learning neural network and the non-cavity locations selected has the same hyperbola signal appearance to that of the cavity locations. As a result, 35 out of the 37 cavity locations were correctly classified while 28 out of the 37 non-cavity locations were correctly classified as well, yielding to a total accuracy of 85.14%. It was observed that 75.68% of misclassified non-cavity locations were correctly classified as non-cavity and were discarded from the final candidate locations. Furthermore, additional validation was conducted on 30-km GPR survey data of Seoul city. The RetinaNet-based deep learning neural network detected a total of 358 cavity candidates which includes 36 true cavity locations. After applying the proposed model, the number of cavity candidates was reduced to 131 locations including 35 true cavity locations, resulting to an accuracy of 97.22% for cavity detection. Moreover, the over prediction reduced from nine times to four times the number of true cavity locations. Therefore, the proposed model proved its effectivity in eliminating misclassified non-cavity locations while preserving the real cavity locations.
연관 검색어 추천
이 검색어로 많이 본 자료
활용도 높은 자료
