← Back to Sessions

Denoising Score Matching for Online Change Point Detection

Description: 

Earthquake prediction is a major scientific challenge and controversial topic in earthquake science. It relies on identifying precursory signals occurring only before large damaging earthquakes. However, despite decades of research, no reliable precursory signals were found, and the current view is that individual earthquakes cannot be accurately predicted. Recently, there has been a renewed interest in studying earthquake precursory phenomenon, largely driven by improved ultra-dense near-field observations by seismic and geodetic recordings, availability of big data and related products such as high-resolution earthquake catalogs, and new developments in machine learning methods in earthquake science. These precursors often appear as change points in sequential data streams, posing challenges for detection due to the complexity of the data.
In this study, we propose offline and online versions of a denoising score-matching change-point detection algorithm, where the pre- and post-change data distributions are unknown and can be arbitrarily complex. Our key novelties are: (i) We estimate the score function of the data distribution as a component of the detection statistic, ensuring computational scalability and enabling strong modeling capacity through deep neural networks; (ii) We introduce a diffusion process to the data stream by injecting noise at appropriate scales, which helps increase the detection power of the statistics by enhancing score estimation in regions where pre- and post-change distributions overlap. We illustrate the advantages of our methods by theoretically deriving the upper bounds on the detection delays and empirically validating our methods' efficacy on both synthetic and real geophysical monitoring signals right before the 2014 M6.6 Jinggu earthquake in Yunnan, China. Our preliminary results show that the precursory signals exist, and our method can effectively extract them compared to traditional methods.

Session: Building and Decoding High-resolution Earthquake Catalogs With Statistical and Machine-learning Tools [Poster]

Type: Poster

Date: 4/15/2025

Presentation Time: 08:00 AM (local time)

Presenting Author: Wenbin

Student Presenter: Yes

Invited Presentation: 

Poster Number: 42

Authors