Application of Artificial Intelligence Techniques to Predict Peak Shear Strength Property of Fiber-Reinforced Soil

• Main Authors: Jie-Ying Lin, 林婕嫈
• Other Authors: Jui-Sheng Chou
• Format: Others
• Language: zh-TW
• Published: 2015
• Online Access: http://ndltd.ncl.edu.tw/handle/40443739510803129281

碩士 === 國立臺灣科技大學 === 營建工程系 === 103 === Fiber-reinforced materials exhibit high tensile strength, durability, and a light mass. Such materials, when used in geotechnical engineering, can improve soil strength and the overall stability of reinforced structures such as side slopes. In addition, fiber-reinforced materials can be employed for landscaping and facilitate integral ecology. Scholars have studied the mechanics of fiber-reinforced soil (FRS) by using triaxial and direct shear tests, proposing theories and empirical models for predicting the shear strength of reinforced soils. However, mixing fibers into soil can be regarded as a random distribution, demonstrating uncertainty. The shear strength parameters cannot be accurately predicted using theories or empirical prediction equations. Therefore, this study established an FRS characteristic database comprising relevant literature from 1983 to 2015. The parameters collected and analyzed in this study included soil parameters (i.e., size of test piece, friction angle of the soil, and soil cohesion), fiber parameters (i.e., aspect ratio of fiber, volume content, and interface friction between fiber and soil), and stress parameters (i.e., confining pressure and normal stress). After datafication, data mining technologies were employed to identify factors influencing shear strength and to predict the friction angle of FRS. The analysis techniques included (1) classification and regression methods, such as linear regression analysis, classification and regression tree analysis, a generalized linear model, and chi-squared automatic interaction detection; (2) machine learners, such as an artificial neural network and support vector machine/regression; and (3) meta ensemble models, such as Voting, Bagging, Stacking and Tiering. The results indicated that fiber content, fiber aspect ratio, friction angle of soil and stress parameter were the primary factors influencing the shear strength of FRS. Through subsequent model training,cross-validation and testing, the optimal model obtained was the Tiering SVM-(SVR/SVR) model. The correlation coefficient of the prediction values with the actual values recorded in the literature was 0.9, indicating a high correlation. The mean absolute percentage error was < 4%, root mean square error was < 2°, and mean absolute error was < 2°. The overall improvement in performance measures compared with that demonstrated using conventional theory or empirical equation was 9.31%-79.50%. This study contributed to the domain knowledge by proposing an effective artificial intelligence model for predicting the friction angle of FRS.