Seismic Anomaly Detection and Instrument Health Forecasting with Deep Learning

Jiun-Ting Lin, Ana C. Aguiar, Qingkai Kong, Amanda C. Price, & Steve C. Myers

Submitted September 7, 2025, SCEC Contribution #14734, 2025 SCEC Annual Meeting Poster #TBD

Reliable seismic analysis relies on rigorous data quality assessment; however, traditional approaches often fail to catch uncommon or station-specific data anomalies. This is critical because erroneous data can be subtle and embedded within the apparent “ground truth” that affects the accuracy of subsequent analyses. Furthermore, identifying when a station is unhealthy or nearing failure enables operators to diagnose issues and perform maintenance earlier to prevent a complete instrument failure and data loss. Here, we introduce an unsupervised deep autoencoder approach that detects seismic data anomalies without prior knowledge about data quality. This new tool helps to identify anomalous data and long-term trends, potentially indicate impending instrument failure. We test the model on the U.S. International Monitoring System and successfully detect data anomalies on a monthly scale. Comparing to existing models in a manually selected testing dataset, our model achieves 88% accuracy, comparable to a supervised ML model (90%), and significantly better than the standard quality assessment package (78%). When applied to previously unseen stations containing novel anomaly types, our model outperforms both baselines, demonstrating its transferability. This shows the potential for application to other networks or data centers such as Global Seismographic Network, EarthScope Data Services, and California Earthquake Data Center. This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344. LLNL-ABS-2009486.

Key Words
Seismic Anomaly Detection, Failure Detection, Predictive Maintenance

Citation
Lin, J., Aguiar, A. C., Kong, Q., Price, A. C., & Myers, S. C. (2025, 09). Seismic Anomaly Detection and Instrument Health Forecasting with Deep Learning. Poster Presentation at 2025 SCEC Annual Meeting.

Related Projects & Working Groups
Research Computing (RC)