基於全球資料庫的地區性土壤液化潛能預測

Abstract

在大地工程領域中，土壤液化是相當重要的議題，工程師在設計時必須評估土壤液化潛能，因此有許多的學者提出預測模型來預測土壤液化潛能，然而這些預測模型都將全球的液化數據視為同一群體進行分析，但土壤液化應該受各區域以及土壤性質所影響，不應將全球的液化數據視為同一群體進行分析。近年來大數據分析和機器學習在各領域的應用上皆有一定的發展，因此本研究希望透過大數據分析建立出可以準確地預測土壤液化潛能之模型，且此模型需保有各區域的特性。 首先，會先回顧前人有關土壤液化的文獻，了解預測土壤液化潛能的各種方法，以及通常都運用哪些土壤參數進行分析，本研究會收集較具代表性的土壤參數，包括總應力與有效應力比值(σ_v⁄(σ_v^' ))、正規化N值((N_1)_60)、正規化q_c值(q_c1N)、土壤類型因子(I_c)、正規化V_s 值(V_s1)和細粒料含量(FC)，並代入前人的預測模型檢驗資料的正確性。 接著利用改良式層級貝氏模型(RHBM)進行分析，過程中會使用Johnson分布系統將參數轉換至標準常態空間，接著搭配吉普森取樣及貝氏分析中的共軛條件來學習參數間的相關性，且同時學習資料庫中各區域的行為特性，並將資料庫中的空缺資料填補，搭配目標區域中有限的已知資訊來推估未知資料的分布，推估出資料後即可運用此資料進行土壤液化機率的計算。

In geotechnical engineering, soil liquefaction is very important problem. Engineers must evaluate soil liquefaction potential when designing. Therefore, many scholars have proposed prediction models to predict soil liquefaction potential. However, these prediction models all regard global liquefaction data as the same group for analysis, but soil liquefaction should be affected by various regions and soil properties, liquefaction data should not be analyzed as the same group. In recent years, the application of big data analysis and machine learning in various fields has developed to a certain extent. Therefore, this study hopes to establish a model that can accurately predict the potential of soil liquefaction through big data analysis, and this model have to maintain the characteristics of each region. First of all, we will review the previous literature on soil liquefaction, understand various methods for predicting the potential of soil liquefaction, and what soil parameters are usually used for analysis. This study will collect more representative soil parameters, including the ratio of total stress to effective stress (σ_v⁄(σ_v^' )), normalized N value ((N_1)_60), Normalization q_c value q_c1N, soil type factor (I_c), normalized V_s value (V_s1) and fines content (FC) were substituted into previous prediction models to verify the accuracy of the data. Then the revised hierarchical bayesian model (RHBM) is used for analysis. In the process, the Johnson distribution system will be used to convert the parameters to the standard normal distribution, and then the gibbs sampling and conjugate conditions in bayesian analysis will be used to learn the correlation between the parameters. At the same time, the behavior characteristics of each region in the database will be learned, and the unknown data will be estimated with the limited known information in the target region, the data can be used to calculate the probability of soil liquefaction.