A satellite telemetry data anomaly detection method integrating wavelet packet transform and diffusion models (WPT-diffusion)

Satellite telemetry data are high-dimensional, non-stationary time series, yet existing anomaly detection methods are hindered by restrictive distributional assumptions and insufficient capacity to model complex normal behavior manifolds. In particular, prevailing deep learning approaches predominantly exploit time-domain information while neglecting multi-scale frequency characteristics, limiting sensitivity to subtle spectral anomalies. To address this, we propose wavelet packet transform–diffusion (WPT-diffusion), which embeds multi-scale time–frequency features as structural conditional guidance within a diffusion generative model. A learnable Morlet wavelet kernel and frequency-domain attention mechanism are introduced to adaptively enhance discriminative subband features, while temporal attention captures long-range dependencies to improve reconstruction fidelity. Experimental results on the ESA OPS-SAT real-world dataset indicate that WPT-diffusion achieves a precision of 0.90, a recall of 0.88, and an F1-score of 0.890, compared favorably against CATCH, USAD, TimesNet, TranAD and DDTAD. These findings support the effectiveness of the proposed method for real-world satellite mission applications.