Anomaly Detection in weekly SIRGAS GNSS Time Series: A TimesNet assessment in real data

Abstract

Time series anomaly detection is essential in a range of fields because it allows for the identification of atypical patterns that differ from the expected behavior. Time series data of GNSS coordinates of SIRGAS stations are valuable for research purposes, including the determination of tectonic drift (velocity) and strain rate deformation. Changes in the patterns of a GNSS time series at a station due to abnormal events such as earthquakes, equipment changes, or malfunctions can significantly impact the accuracy of velocity estimation. In this work, we assess the efficacy of a neural network called TimesNet for detecting anomalies in real GNSS time series data from 95 SIRGAS stations in Brazil and Ecuador. As TimesNet achieved state-of-the-art efficiency on many benchmark datasets, it was chosen as the primary model for this research. Hyperparameter optimization was employed for TimesNet using two stations: BRAZ, in a low seismic activity area in Brazil, and CHEC in a high seismic activity area in Ecuador. The results with both hyperparameters sets were compared against three established anomaly detection filters (Gaussian process, Kalman and Moving Average) and a LSTM neural network that were paired with three anomaly scorer methods (norm, K-means and difference). We tested two methods to predict labels from the anomaly scores: 1) Using a fixed percentile (Top 1%) threshold, and 2) Using the Top S scores, where S is the number of anomalies of each station. The TimesNet with hyperparameters optimized in the CHEC station obtained better results than BRAZ, both in Brazil and Ecuador stations, showing that the choice of hyperparameters may present an impact in the anomaly detection results. The best F1Top S and MSE in Brazil were obtained with the Kalman filter (F1Top S=57.60 | MSE = 0.0094), closely followed by TimesNet (F1Top S=55.67 | MSE = 0.0095). In Ecuador, the best F1Top S and MSE were obtained by TimesNet (F1Top S=63.64 | MSE = 0.0145), followed by the Kalman filter (F1Top S=59.09 | MSE = 0.0344). Therefore, the anomaly scores produced by TimesNet were less noisy than those produced by the other methods on Ecuador. To foster further research, the source code for this analysis is publicly available at https://github.com/mauriciodev/deep_anomaly_gnss.