Motion-state recognition of satellite active components based on high-frequency photometric signals

Estimating the states of distant non-cooperative satellites from photometric signals is essential for space situational awareness. Conventional analyses operate on second-level or longer timescales, where rapid photometric variations induced by moving components are averaged out or aliased, leaving component motion states unidentifiable. We propose observing satellites with high-frequency photometric signals to perceive component motion states from reflected-light variations induced by component motion. We built a hardware-in-the-loop system comprising a satellite model, a photodetection chain, and synchronized acquisition, recording a large volume of signals at 1 kHz with component motion states and vibration frequencies. Time-frequency spectra reveal speed-dependent lines and harmonics during both changing and constant wheel speeds, abrupt amplitude changes at antenna and array start-stop, and persistent narrowband lines during steady payload rotation. To highlight the motion characteristics of the target, four physical descriptors, namely dominant-frequency trajectory, harmonic prominence, envelope standard deviation, and frequency stability, are constructed, and a time-frequency multi-task transformer network (TFMT-Former) model is proposed to exploit their correspondence with the time-frequency spectrum via branch-wise cross-attention, learnable task routing, and gated fusion. The model achieves 93.41% ± 0.22% Macro-F1 and 82.32% ± 0.41% subset accuracy, outperforming the best baseline by 9.31 and 11.27 percentage points. Ablation and descriptor-shuffling tests show gains stem from the descriptor-spectrum correspondence rather than increased input dimensionality. This extends photometry analysis from whole-satellite macroscopic states to active-component motion states, providing an experimental and methodological basis for fine-grained state perception of distant non-cooperative satellites.