Contrastive Learning with Prototype-Based Negative Mixing for Satellite Telemetry Anomaly Detection

Telemetry data are the most important basis for ground operators to assess the status of satellites in orbit, and telemetry data-based anomaly detection has become a key tool to improve the reliability and safety of spacecrafts. Recent research on anomaly detection focuses on constructing a normal profile of telemetry data using deep learning methods. However, these methods cannot effectively capture the complex correlations between the various dimensions of telemetry data, and thus cannot accurately model the normal profile of telemetry data, resulting in poor anomaly detection performance. This paper presents CLPNM-AD, contrastive learning with prototype-based negative mixing for correlation anomaly detection. The CLPNM-AD framework first employs an augmentation process with random feature corruption to generate augmented samples. Following that, a consistency strategy is employed to capture the prototype of samples, and then prototype-based negative mixing contrastive learning is used to build a normal profile. Finally, a prototype-based anomaly score function is proposed for anomaly decision-making. Experimental results on public datasets and datasets from the actual scientific satellite mission show that CLPNM-AD outperforms the baseline methods, achieves up to 11.5% improvement based on the standard [Formula: see text] score and is more robust against noise.