Joint extraction of space target micro-motion parameters at low SNR based on consistency variational autoencoder and cascaded transformer

Abstract

Micro-Doppler technology has significant applications in the recognition of space non-cooperative targets. However, in practical radar detection, background noise caused by low signal-to-noise ratios and the nonlinear coupling among various micro-motion parameters limit the feature extraction accuracy and engineering application of existing deep learning methods. To address the computational burden caused by high-dimensional noisy time-frequency representations and the physical multi-solution problem, this paper proposes a joint extraction architecture for space target micro-motion parameters (including target type, precession frequency, attitude angle, and spin frequency) based on a consistency-constrained Variational Autoencoder and a cascaded Transformer. First, a front-end network based on consistency constraints is designed to achieve label-free blind denoising via self-supervised latent-space consistency alignment and feature dimensionality reduction. Second, a temporal frame slicing cross-domain bridging mechanism is proposed to preserve the physical topological integrity while reducing the computational cost. Finally, a cascaded guidance network based on category context injection is constructed to narrow the solution space of micro-parameters, resolving the ambiguity in parameter estimation. Experiments show that at a low signal-to-noise ratio of 10 dB, the proposed method achieves a target classification accuracy of 100% and reduces the mean squared error of parameter regression to 0.0067. Furthermore, the extraction accuracy for precession and spin frequencies remains above 90% across the full spatial observation angle ( 0 ∘ ∼ 90 ∘ ). Compared with mainstream architectures, this method achieves an effective balance among noise robustness, parameter estimation accuracy, and lightweight deployment for spaceborne edge computing.