To overcome the difficulty of early detecting transient actuator faults in hypersonic vehicles—where physics-based diagnostics are time-consuming and of limited accuracy—this work develops a reliability-oriented fault-prediction and diagnosis algorithm for unmanned systems. First, a decision-tree/Random-Forest importance-ranking scheme is designed to handle the high-dimensional flight-state data; five key features are retained, reducing dimensionality by 62%. Next, a Long Short-Term Memory (LSTM) network performs multi-step forecasting on the reduced feature sequence, capturing latent fault-evolution trends. The predicted states are then fed to a Random-Forest classifier, which issues early warnings of bias, stuck-in-place and drift faults. As a proof-of-concept, 1,500 MATLAB simulations of hypersonic actuator operating and fault scenarios are conducted. The proposed pipeline shortens average diagnostic latency while boosting fault-identification accuracy to 92%. The resulting “feature selection–sequence prediction–intelligent diagnosis” strategy delivers speed, accuracy and foresight, providing a valuable technical reference for reliability assurance in hypersonic unmanned systems.