With the rapid growth of global air transport, traditional time-based maintenance models relying on flight hours have struggled to adapt to complex operational environments and the nonlinear degradation characteristics of equipment. Meanwhile, frequent cyberattacks during data transmission pose severe challenges to aviation maintenance safety. To address these issues, this paper proposes ReForGe, a secure predictive maintenance model based on an ensemble learning framework. The model integrates recursive feature elimination with decision path analysis to accurately identify core features from multi-source heterogeneous sensor data. Additionally, it embeds a lightweight security verification mechanism to effectively mitigate risks of malicious tampering and data theft during transmission. Furthermore, by introducing an expected-cost-driven dynamic weight adjustment mechanism, ReForGe adaptively balances operational risks between over-maintenance and fault omission, significantly improving the accuracy of remaining useful life prediction and the reliability of maintenance decisions. Experimental results demonstrate that ReForGe achieves a fault prediction accuracy of 93%, with its F1-score surpassing conventional models by 4.49%. It also exhibits exceptional robustness and cost-effectiveness under complex operational conditions.