MODEL-FREE DATA DRIVEN REINFORCEMENT LEARNING FOR SOIL CONSOLIDOMETER
The increasing demand for infrastructure due to industrialization and urbanization necessitates engineering and infrastructural solutions to address soil stability issues, particularly soil consolidation, which can directly impact infrastructure durability. Research on soil consolidation has prog...
Saved in:
| Main Author: | |
|---|---|
| Format: | Theses |
| Language: | Indonesia |
| Online Access: | https://digilib.itb.ac.id/gdl/view/86851 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Institut Teknologi Bandung |
| Language: | Indonesia |
| Summary: | The increasing demand for infrastructure due to industrialization and urbanization
necessitates engineering and infrastructural solutions to address soil stability
issues, particularly soil consolidation, which can directly impact infrastructure
durability. Research on soil consolidation has progressed from Terzaghi's theory
(1925) to the adoption of the Constant Rate of Strain (CRS) method, which is more
efficient than the Incremental Load (IL) method. However, the development of
modern consolidometer instrumentation still faces challenges in control accuracy
and testing time. However, mathematical modeling of soil consolidation behavior
presents significant challenges due to its complex, non-linear, and oscillatory
nature, which cannot be captured by traditional methods. These challenges,
combined with uncertainties in soil properties, stratification, and environmental
factors, complicate the prediction of soil consolidation and settlement. This
research develops a model-free consolidometer using Q-Learning control, based
on the Constant Rate of Strain (CRS) method, integrated with Arduino and
LabVIEW for real-time data acquisition. This study utilizes Q-Learning control to
regulate the soil consolidation Process, comparing to traditional methods: On/Off
control and the Incremental Load method. In particular, Q-Learning was further
divided into two types: Hybrid Q-Learning and Online Q-Learning. The results
demonstrate that Online Q-Learning outperformed the other methods in terms of
steady-State error (ESS), transient response, and system stability. Online Q-
Learning achieved the lowest ESS for both Pressure and distance (1.94% and
0.40%, respectively), the fastest rise time for deformation (8160 seconds), and
minimal overshoot (0.94% for Pressure and 0% for deformation). |
|---|