Seismometer Fault Detection Based on Unsupervised Machine Learning Using Time Series Features

Seismometers are critical instruments in earthquake monitoring networks, and their reliable performance ensures the accuracy of earthquake data. While several networks have implemented methods to monitor system health, these methods are often broad and require further analysis to pinpoint specific problems. Due to their high sensitivity, seismometers are prone to failures, and early detection of malfunctions is essential. However, seismometer fault detection remains an underexplored research area despite its importance. To address this gap, we adopt a novelty detection approach within an industrial predictive maintenance (PdM) framework, demonstrating effective fault detection. Three unsupervised machine learning (ML) models—Isolation Forest (IF), One-Class Support Vector Machine (OCSVM), and Local Outlier Factor (LOF)—are implemented. Optimal parameter is achieved to enhance each model's performance by setting the contamination parameter in IF to 0.005, nu in OCSVM to 0.05, and enabling the "true" option for LOF novelties. Suppose performance was significantly worse among the three models. LOF had an impressive accuracy of 0.98, closely followed by OCSVM. These results provide a preliminary step toward developing an effective PdM strategy. More broadly, this approach can be applied across various domains related to anomaly detection, offering a robust alternative to traditional labeling techniques for large datasets.