University of Twente Student Theses

Login

Investigating the Potential of Machine Learning Methods to Predict Soil Variables for Dike’s Macro Stability Analysis

Yaghi, Mostafa (2024) Investigating the Potential of Machine Learning Methods to Predict Soil Variables for Dike’s Macro Stability Analysis.

[img] PDF
13MB
Abstract:This report examines the idea of predicting Dike’s stability analysis variables using machine learning and linear regression models. This is done to address the limitations of the current practice, which involves obtaining the variables through three tests: clay inspection, clay compaction, and triaxial tests. While the clay inspection is straightforward, the other two are expensive, time-consuming, and not available in the design phase of the project. Furthermore, those tests are unavailable in the design phase, forcing the designers to estimate the needed variables. These limitations underscore the need for a more efficient and accurate method for predicting these variables. On that point, before performing these tests, the current practice estimates the clay variables by knowing the clay and sand ratio of the clay. Then, it predicts the unit weight and friction angle by assuming weak, moderate, and strong consistency indices. Therefore, the thesis hypothesis explores predicting the dry unit weight and friction angle before performing the compaction and triaxial tests and explores the idea of having a more precise estimate of the variables following the same order as the current estimation process. This study involved gathering Dutch data from the Fugro database, which was then input into two machine learning models: the Neural Networks and Random Forest. These models efficiently processed the data frames and predicted the variables using a set of scenarios. Each scenario represents what an engineer would know about the clay sample in the design phase. Scenario one explored the idea of having the Atterberg limits; scenario two explored knowing the contents of the clay (clay, sand, and silt), while scenarios three and four explored knowing the Triangular classification (NEN 5014) of the clay and the plasticity diagram categories, respectively. Then, the research showed that using the Random Forest led to better predictions, with scenario one having the best prediction towards the other variables, especially the dry unit weight. This was then combined with estimated water content variables to calculate the unit weight. These water content variables were based on different consistency index measures (0.60, 0.75, and 0.85). Therefore, the study shows that it is possible to predict the unit weight with low uncertainty by knowing the plasticity diagram classifications and Atterberg limits, which formed a combined scenario to mimic the order of the current estimation method. Furthermore, the research could not use the triaxial data because of the poor quality and small amount of data, isolating the friction angle from being predicted as the unit weight. In conclusion, the research underscores the potential for significant improvement in the current practices. This is particularly evident in the difference in the unit weight values between the different clay classes, a factor not accounted for in the current estimated values. The research also suggests that the unit weight could be predicted accurately for the design phase, eliminating the need for the compaction and triaxial tests for more accurate estimation. Keywords: Neural Networks, Random Forests, Auto-encoders, Clay contents, Atterberg limits, Dry unit weight, Friction angle, Clay inspection, Clay compaction, Triaxial tests, stability analysis.
Item Type:Essay (Master)
Clients:
Fugro, Utrecht, Netherlands
Faculty:ET: Engineering Technology
Programme:Civil Engineering and Management MSc (60026)
Link to this item:https://purl.utwente.nl/essays/104197
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page