Generating Phone-quality Records to Train Machine Learning Models for Smartphone-based Earthquake Early Warning

Zengwei Zheng, Lifei Shi, Sha Zhao, Jianmin Hou, Lin Sun, Dong Lin, Yi Fang, Jie Liu, Shijian Li, Gang PanResearch 2020 · DOI

Abstract

Earthquake Early Warning (EEW) systems detect earthquakes and send early warnings to areas likely to be affected, playing a significant role in reducing earthquake damage. Smartphone-based EEW has emerged as a promising interdisciplinary approach, utilizing built-in MEMS accelerometers. However, collecting sufficient phone-recorded seismic data for training machine learning models remains a challenge. This paper proposes TMAP (Transformation Method from Accelerograph to Phone-quality records), which transforms high-fidelity seismic network data into realistic smartphone-quality records by modeling electronic noise and physical sliding. Experiments demonstrate that TMAP significantly outperforms existing methods, and machine learning models trained on TMAP-generated data show superior performance in earthquake intensity prediction.

Introduction

Earthquake Early Warning (EEW) systems provide critical seconds of notice before destructive seismic waves arrive. While traditional systems rely on expensive, sparse networks of professional-grade accelerographs, the widespread distribution of smartphones offers a potential "crowdsourced" alternative. However, the primary bottleneck for smartphone-based EEW is data scarcity. Machine learning models require massive datasets to distinguish between seismic events and everyday noise. Since major earthquakes are rare, there is a lack of real-world smartphone recordings of high-magnitude events. While researchers use shake tables to simulate quakes, these experiments are limited in scale. Conversely, decades of high-quality data from professional networks exist, but this data is "too clean"—it does not account for the low-precision sensors and physical "slipping" behavior of a phone sitting on a desk.

Methodology

The researchers developed TMAP to bridge the gap between professional seismic data and smartphone reality. TMAP incorporates two distinct models:

  1. The FIR (Finite Impulse Response) Model: This handles electronic differences, accounting for the lower bit-precision and higher noise floors of smartphone sensors compared to professional equipment.
  2. The Physical Sliding Model: This simulates how a phone slides on a surface during violent shaking. Based on static and dynamic friction coefficients, it models the mechanical interaction between the phone and its environment.
Workflow of TMAP training and its application in the SEEW system.

To validate TMAP, a benchmark dataset was created using 46 real earthquake records from Japan (KiK-Net and K-Net) and China. A shake table experiment involved 48 different smartphones (covering brands like Huawei, iPhone, Samsung, and Xiaomi) placed alongside professional accelerographs. The study tested both "Fixed" (taped down) and "Free" (loose) scenarios to evaluate sensor noise and sliding models respectively, resulting in a dataset of 7,728 smartphone records.

Results and Discussion

TMAP successfully replicates the "jump" phenomenon—sudden, sharp acceleration shifts observed when a phone slides. Unlike previous methods (such as TMK) that focused primarily on noise, TMAP accurately models the phase mismatch caused by physical sliding.

The effectiveness of TMAP was tested by training four machine learning models (SVR, Bayesian Ridge, Linear Regression, and Random Forest) for earthquake intensity prediction. TMAP records maintained high similarity to real phone data even during intense shaking. Notably, the Absolute Mean Error (AME) for a Bayesian Ridge model dropped from 0.370 (using older transformation methods) to just 0.094 when trained on TMAP data. These results suggest that the physical interaction between the phone and its environment is a critical signal component often overlooked by simple noise-addition methods.

Conclusion

TMAP provides a robust framework for generating synthetic smartphone-quality datasets from professional seismic records, addressing the critical data scarcity issue in smartphone-based EEW. While currently optimized for phones on flat, horizontal surfaces, future research will expand the model to include more complex scenarios, such as phones in pockets or hands. The discovery of the "sliding jump" characteristic also opens new avenues for developing more sensitive earthquake detection algorithms that can better distinguish quakes from everyday human activity.

Summary updated

Abstract

Abstract Earthquake Early Warning (EEW) system detects earthquakes and sends an early warning to areas likely to be affected, which plays a significant role in reducing earthquake damage. In recent years, as with the widespread distribution of smartphones, as well as their powerful computing ability and advanced built-in sensors, a new interdisciplinary research method of smartphone-based earthquake early warning has emerged. Smartphones-based earthquake early warning system applies signal processing techniques and machine learning algorithms to the sensor data recorded by smartphones for better monitoring earthquakes. But it is challenging to collect abundant phone-recorded seismic data for training related machine learning models and selecting appropriate features for these models. One alternative way to solve this problem is to transform the data recorded by seismic networks into phone-quality data. In this paper, we propose such a transformation method by learning the differences between the data recorded by seismic networks and smartphones, in two scenarios: phone fixed and free located on tables, respectively. By doing this, we can easily generate abundant phone-quality earthquake data to train machine learning models used in EEW systems. We evaluate our transformation method by conducting various experiments, and our method performs much better than existing methods. Furthermore, we set up a case study where we use the transformed records to train machine learning models for earthquake intensity prediction. The results show that the model trained by using our transformed data produces superior performance, suggesting that our transformation method is useful for smartphone-based earthquake early warning.
BibTeX
@inproceedings{zheng2020generating,
  author = {Zengwei Zheng and Lifei Shi and Sha Zhao and Jianmin Hou and Lin Sun and Dong Lin and Yi Fang and Jie Liu and Shijian Li and Gang Pan},
  title = {{Generating Phone-quality Records to Train Machine Learning Models for Smartphone-based Earthquake Early Warning}},
  booktitle = {Research},
  year = {2020},
  doi = {10.21203/rs.3.rs-89489/v1},
}
Figure