This paper proposes a deep learning-based active fault-tolerant control framework subject to safety-margin constraints for transatmospheric vehicles with actuator faults. The framework consists of a CNN–LSTM fault diagnosis module and a GRU-based control reconstruction module. The CNN–LSTM identifies control-surface jamming and actuator loss-of-effectiveness faults, and the actuator effectiveness safety-margin is evaluated as the difference between the remaining actuator effectiveness and a preset lower safety limit. A finite-time fault-tolerant controller generates supervised control data within the admissible operating range. The GRU-based control reconstruction module learns the nonlinear mapping from flight states, attitude errors, and fault diagnosis results to auxiliary control inputs, enabling online control reconstruction. Simulation results demonstrate that the proposed framework accurately identifies actuator faults and achieves effective attitude tracking and rapid post-fault recovery of stability under the tested safety-margin conditions.