Anomaly Detection in Real-World Seismic Time Series: Evaluation of Timer Model with COGNOS Framework

Time series anomaly detection is one of the core tasks in time series data analysis, aiming to identify abnormal events or behaviors from normal temporal data, and plays a critically important role across numerous domains. In seismic data analysis, the identification of anomaly intervals preceding earthquake occurrences holds significant research value for subsequent earthquake prediction. This study evaluates the effectiveness of a large time series model named Timer in detecting anomalies within seismic time series data collected by tiltmeters. As a multi-task time series model, Timer has achieved state-of-the-art performance on forecasting, imputation, and anomaly detection tasks across multiple benchmark datasets, and is therefore selected as the core model in this study. This study employs the Timer model combined with the Constrained Gaussian-Noise Optimization and Smoothing (COGNOS) method for anomaly detection in seismic time series data, and compares the results with those of the original Timer model. Additionally, several mainstream time series anomaly detection models are selected as baselines for comparison. To comprehensively validate the model’s performance, comparative experiments were conducted on both the publicly available SIRGAS GNSS dataset and a real-world seismic dataset collected by tiltmeters. The experimental results demonstrate that the Timer model exhibits outstanding performance in anomaly detection on seismic time series data, and achieves further improvement when integrated with the COGNOS method.