Application of the Support Vector Machine Classifier in Earthquake Magnitude Estimation
Description:
The ability of an Earthquake Early Warning (EEW) algorithm to resolve reliably the event magnitude from a few stations is critical for an efficient EEW application. When applied to a single-station seismic record, traditional approaches based on tc and Pd methods exhibit somewhat larger errors than the machine-learning methods. We use the Support Vector Machine (SVM) method with a Gaussian kernel to predict magnitude from a single station triaxial record. Unlike other approaches, raw records are used without downsampling or converting them to a vector of parameters. Preliminary investigations demonstrated that SVM approach is superior to the Deep Neural Networks as it provides better accuracy both in the training and validation data sets. The subset of the KiK-net database generated by the NIED strong-motion seismograph network incorporates ~4,000 triaxial records of P waves over the magnitude range 5.5-7.6. This was used to train the SVM model and investigate the accuracy of the results by means of the normalized confusion matrix. Results demonstrated that the confusion matrix is almost diagonal, which suggests that the spectral norm of this matrix is a good measure of the accuracy of the solution. Several techniques were evaluated for the pre-processing of the P wave records to investigate their influence on the solution with the intention of providing a rigorous physics-based interpretation of the results.
The outcomes of this study will be implemented in the network based EEW system operated by the Greater Vancouver Water District. The system was designed, built, and commissioned in 2021/2022 as part of the pilot project “EEW and Strategic Response System”. The project deliverables included a robust EEW application, and a Structural Health Monitoring service for some of the critical infrastructures operated by Metro Vancouver. Currently, the system is being utilized by operations staff within the water system infrastructure.
Session: End-to-End Advancements in Earthquake Early Warning Systems -IV
Type: Oral
Date: 5/3/2024
Presentation Time: 11:15 AM (local time)
Presenting Author: Anton
Student Presenter: No
Invited Presentation:
Authors
Anton Zaicenco Presenting Author Corresponding Author anton.zaicenco@weir-jones.com Weir-Jones Engineering Consultants |
Iain Weir-Jones iainw@weir-jones.com Weir-Jones Engineering Consultants |
|
|
|
|
|
|
|
Application of the Support Vector Machine Classifier in Earthquake Magnitude Estimation
Category
End-to-End Advancements in Earthquake Early Warning Systems