Anomaly Detection for Satellite Telemetry Parameters Based on Time-Frequency Feature Analysis and Adversarial Training

Satellite telemetry data directly reflects in-orbit operational status, making anomaly detection critical for ensuring satellite safety and reliability. This paper presents a novel anomaly detection approach integrating time-frequency feature analysis with adversarial training to address limitations of existing methods. A frequency-domain attention mechanism is designed to capture periodic patterns and subtle anomaly signatures in spectral domain. Synergistic time-frequency analysis establishes cross-domain feature relationships, enabling comprehensive perception of complex anomaly patterns by combining temporal evolution characteristics and spectral features. Built on a simplified Transformer architecture, the proposed model leverages adversarial training to enhance anomaly discrimination capability. Comparative experiments and ablation studies on three benchmark datasets and a scientific satellite thermal control telemetry dataset validate the model's superior performance. The method provides an effective technical solution for satellite telemetry anomaly detection, supporting reliable on-orbit satellite operations.