The utilization of reusable launch vehicles has emerged as a prominent research focus owing to their significantly lower launch costs. Vertical sub-stage recovery technology is one of the most successful rocket recovery technologies at present， which requires that the recovery section needs to control the attitude of the rocket to keep it perpendicular to the ground. However， in the terminal deceleration phase of recovery， controlling the rocket relies solely on engine gimballing angles to obtain control moments， thus encountering constraints related to moment saturation. Moreover， deviations in thrust estimation contribute to errors in control moments. In this paper， we introduce an intelligent model predictive control algorithm to achieve attitude control during the terminal deceleration phase. Building upon traditional model predictive control， it employs neural networks to approximate the optimal value function， effectively reducing the computational load of model predictive control. Furthermore， in consideration of potential engine faults， along with constraints related to saturation and thrust estimation errors， a fault-tolerant optimal control allocation algorithm based on quadratic programming is devised. Simulation results substantiate the effectiveness of the proposed methodology.