Deep continual learning for predicting blast-induced overbreak in tunnel construction / He Biao
Tunnel construction, a critical component of modern infrastructure development, faces the persistent challenge of blast-induced overbreak. Overbreak, the excessive removal of rock mass beyond the planned tunnel profile, poses significant safety risks, increases costs, and causes project delays. T...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Published: |
2024
|
| Subjects: | |
| Online Access: | http://studentsrepo.um.edu.my/15626/1/Biao_He.pdf http://studentsrepo.um.edu.my/15626/2/He_Biao.pdf http://studentsrepo.um.edu.my/15626/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Tunnel construction, a critical component of modern infrastructure development, faces
the persistent challenge of blast-induced overbreak. Overbreak, the excessive removal of
rock mass beyond the planned tunnel profile, poses significant safety risks, increases costs,
and causes project delays. Traditional methods have been developed to predict overbreak.
These predictions use either empirical-, statistical-, or numerical-based models. However,
traditional methods for predicting overbreak are often inadequate because they simplify
the dynamic and complex nature of rock blasting. The development of an advanced overbreak
prediction model is, therefore, becoming essential.
This thesis addresses the limitations of existing overbreak prediction methods by introducing
a novel data-driven approach based on deep continual learning. The primary
objectives are to develop a more accurate and adaptable predictive model and to integrate
this model into the operational workflow of tunnel blasting. The developed model is expected
to possess the ability of continual learning, which is particularly advantageous in
dynamic environments like tunnel blasting.
To achieve this, this thesis adopts a three-pronged methodological approach. Firstly,
the Conditional Tabular Generative Adversarial Networks (CTGAN) model is used to
augment the real-world overbreak dataset. This aims to ensure a comprehensive representation
of various overbreak scenarios. Secondly, a self-attention multi-layer perceptron
(MLP) model, integrated with two continual learning strategies (Elastic Weight Consolidation
(EWC) and Memory Replay (MR)), is developed and trained on this augmented
overbreak dataset. This step enables the overbreak prediction model to possess the ability to continuously learn real-world scenarios and adapt to the dynamic environment of tunnel
blasting. Third, the overbreak prediction model is further integrated with metaheuristic
algorithms, aiming to identify the optimal blasting parameters that can minimize overbreak.
By adjusting the blasting parameters accordingly, the model can offer a dynamic
and responsive approach to overbreak management.
The findings of this thesis are significant. First, the CTGAN model effectively enriches
the original overbreak dataset by capturing the complex nature of real-world overbreak
scenarios. It can be utilized to create a comprehensive overbreak dataset that possesses
high representativeness and diversity. Second, the self-attention MLP model, empowered
by EWC and MR, demonstrates superior adaptability and accuracy in predicting overbreak.
Its ability of continual learning is highly applicable to actual tunnel blasting cases.
Third, the integration of metaheuristic algorithms further ascertains the optimal blasting
parameters for overbreak minimization under specific rock sections.
The achievements of this thesis indicate a substantial step forward in the application
of deep continual learning in tunnel blasting. This thesis offers a promising solution to
the longstanding challenge of blast-induced overbreak.
|
|---|