Safe learning-scheduled allocation for spacecraft with variable-speed control moment gyroscopes

Agile spacecraft using variable-speed control moment gyroscopes (VSCMGs) require online allocation choices among torque tracking, singularity management, wheel-speed margins, gimbal motion, command smoothness, and degraded-actuator usage. This paper develops a safety-constrained steering architecture in which supervised learning is used only for bounded scheduling of these secondary allocation objectives. An extreme learning machine maps online risk indicators to positive objective multipliers, bounded wheel-speed and gimbal-angle bias targets, and a bounded robust-compensation authority. The physical actuator command is not produced by the learner; it is selected by a torque-priority finite-preview allocator that projects candidate wheel-acceleration and gimbal-rate commands onto actuator limits and retains an unscheduled torque-residual penalty. A metric-triggered transient layer activates admissible internal-motion profiles when conditioning, singular-value, wheel-margin, or actuator-health indicators show elevated risk. The final trained setting is evaluated through deterministic stress tests and a 500-case Monte Carlo campaign including near-singular configurations, tracking, inertia uncertainty, bounded disturbances, and partial actuator-effectiveness loss. The campaign achieves 478 successful trials, corresponding to a 95.6\% success rate with a 95\% Wilson confidence interval of 93.4--97.1\%. A stage-wise ablation against profile-disabled, fixed-weight, heuristic-labeler, null-motion, singularity-robust damped-inverse, and weighted-inverse alternatives shows that the learning scheduler improves the allocation tradeoff without replacing the analytical attitude-control and actuator-feasibility layers.