GNSS denoising and anomaly detection method based on self-supervised learning

Abstract

In the context of Global Navigation Satellite System (GNSS) positioning, issues such as observation noise and outliers caused by factors like multipath effects, ionospheric disturbances, and receiver malfunctions are prevalent. Traditional methods mostly rely on filtering or statistical models, making it difficult to achieve robust processing without accurately labeled data. This paper proposes a self - supervised denoising and anomaly detection algorithm based on the joint optimization of contrastive learning and reconstruction loss. Specifically, a network structure composed of a spatio - temporal encoder and a masked sequence predictor is designed. First, contrastive learning is utilized to learn representations that can distinguish normal fluctuations from abnormal disturbances in the unlabeled original observation sequences. Then, by randomly masking part of the input sequences and training the decoder to reconstruct clean signals, the model implicitly models the underlying laws of GNSS data. On this basis, an anomaly scoring function is constructed by combining contrastive loss and reconstruction error to achieve end - to - end learning for noise suppression and anomaly detection. Experiments on real - world GNSS datasets show that, compared with baseline methods such as wavelet denoising, Kalman filtering, and isolation forest, this method significantly improves the signal - to - noise ratio while maintaining positioning accuracy and effectively reduces the false alarm and miss - alarm rates of anomalies, which verifies its applicability and robustness in complex urban environments.