Year: 2023
Type: dissertation
Abstract: Machine learning has been increasingly used in the prediction of laboratory earthquakes, yet traditional tree-based models face distinct limitations. These models are significantly reliant on the careful selection of features from Acoustic Emission (AE) signals, which limits their applicability across varied datasets and experimental settings. Additionally, the intrinsic simplicity of tree-based models restricts transfer learning capabilities, making it challenging to adapt them to data from different apparatuses. To mitigate these constraints, this study introduces a deep learning model that uses Temporal Convolutional Networks (TCN).The TCN model offers a few advantages over previously used Long short-term memory (LSTM) networks. It effectively captures extended dependencies from sequence inputs and allows for faster computation through parallelism. An optimal sequence length is also proposed to encompass most stick-slip events, thereby enhancing the model's performance. A significant contribution of this research is the model's adaptability, achieved through finetuning. This allows for generalization across diverse experiments and settings, underscoring the potential of transfer learning in this field. The findings highlight the efficacy of integrating deep learning with transfer learning strategies for predicting laboratory earthquakes. Moreover, the TCN model’s adaptability suggests its utility for broader applications, providing a foundation for more advanced and reliable seismic prediction systems.
Source: TSpace (University of Toronto)
Author: Wang, Chonglang
Institution: University of Toronto
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7133057454