Image Data-Driven-Deep Learning in Geosystems: Exploratory Investigation into the Stability of Retaining Walls

SPONSOR: NATIONAL SCIENCE FOUNDATION

PI:  Zhen Liu

 

Project Period:  09.01.17 – 08.31.19

The project is proposed to introduce the most recent breakthroughs in computer vision attributed to deep learning to address a rarely discussed yet urgent issue in most engineering disciplines: how to analyze the explosively increasing image data including images and videos, which can hardly be analyzed with traditional methods? Therefore, the proposed work is very time-sensitive. The core concepts enabling the breakthroughs in image recognition and adopted in AlphaGo, i.e., deep convolutional neural nets, will be used to explore the possibility to accurately assessing the safety of retaining walls with image data. The analysis will be validated against traditional methods including limit analysis and numerical simulation. The research is hypothesis-driven and rationally built on our preliminary study, which clearly supports the hypothesis and shows the high potential of the proposed study. In addition to further proving the power of the proposed concept, this collaborative effort between geotechnical engineering and computer science will understand the data and deep learning in geotech analysis, relate image patterns to physical mechanisms, and investigate two key potential issues associated with the methodology.

The proposed work will be organized into three tasks. Task 1 is to understand how the geotechnical data and analysis results can be related to the input and output of deep learning using convolutional nets. This is a key to the successful application of deep learning in geotech. In Task 2, eight factors in three categories, i.e., geology (soil property), topography (geometry) and boundary conditions, will be evaluated regarding their influence on the safety factors. The relationships between these factors, their roles and significance in deep learning, and correlation between the patterns identified by the CNN and the physical mechanisms will be compared and revealed to connect the deep learning to the existing geotech knowledge pool. In Task 3, two additional key aspects representing the unknowns and possible weaknesses of the machine learning method will be analyzed: model robustness and extrapolation. Testing data prepared in ways different from the input data, i.e., appearance and ranges, will be used to assess and improve these two aspects.

Zhen Liu
Zhen Liu
Shiyan Hu
Shiyan Hu