Advancing Rockfall Hazard Assessment through Data-Driven Modelling Based on Laboratory-Scale Experiments
Loading...
Authors
Ghahramanieisalou, Milad
Issue Date
2025
Type
Dissertation
Language
en_US
Keywords
Coefficient of Restitution (COR) , Multi-Impact Analysis , Non-Spherical Rocks , Rockfall Dynamics
Alternative Title
Abstract
Understanding the dynamics of rockfalls is critical for predicting hazards and mitigating risks in both natural and engineered environments. This research investigates the relationship between rock shape, release angle, and rockfall behavior, combining experimental analysis with machine learning (ML) modeling to address the complexities of rockfall mechanics. Laboratory experiments were conducted to analyze the behavior of three distinct rock shapes—ellipsoidal, octahedral, and spherical—dropped from a pendulum arm at two specific release angles (15° and 30°) onto a horizontal concrete surface. Motion data captured from dual camera angles were analyzed to evaluate key parameters, including the coefficient of restitution (COR), translational and angular velocity, runout distance, and trajectory dispersion. The results highlighted significant variations in behavior based on shape, with non-spherical rocks, particularly ellipsoidal and octahedral forms, exhibiting erratic trajectories, greater lateral dispersion, and wider variability in COR compared to spherical counterparts. These findings underscore the critical influence of shape and impact conditions on rockfall dynamics, emphasizing the need for nuanced hazard prediction models.
To extend these insights, ML techniques were employed to predict rockfall parameters using the experimental data. Three models—K-Nearest Neighbors (KNN), perceptron, and deep neural networks (DNNs)—were evaluated for their ability to handle the nonlinear and irregular patterns inherent in rockfall behavior. While perceptron models struggled with the complexity of the data and DNNs faced challenges with overfitting and interpretability, KNN emerged as the most effective approach. By leveraging localized, instance-based predictions, KNN demonstrated robust accuracy and adaptability, effectively modeling the dynamics across diverse shapes and release angles.
Furthermore, the study compares the predictive performance of the KNN model with RocFall, a physics-based software widely used for rockfall simulations. This comparison provides insights into the strengths and limitations of both approaches, highlighting the potential for ML to complement traditional physics-based models by offering data-driven adaptability and improved predictive capabilities for specific rockfall scenarios.
The research presents a comprehensive framework for understanding and predicting rockfall behavior, combining experimental insights with an interpretable ML approach. The findings improve the predictive capabilities of hazard models and provide practical tools for safety management in mining and civil engineering. By addressing the challenges of complex physical interactions and high-dimensional datasets, the approach enhances risk assessment and mitigation strategies.