Accurate prediction of residual service life (RUL) is challenging due to large sample sizes and high-dimensional condition monitoring data. To address this, a two-stage selective deep neural network ensemble method was proposed for complex equipment like aviation turbofan engines. It aimed to improve prediction accuracy and convergence efficiency. The method involved generating a diverse candidate set in the first stage, considering multiple methods and eliminating internal coupling relationships using a heterogeneous neural network structure, multi time scale design, and algorithm parameter randomization. In the second stage, a genetic algorithm pruned redundant models, removing poor-performing learners and obtaining optimal candidate subsets. Prediction results were derived by averaging the integrated outputs, providing stable support for maintenance decision-making. Empirical results show that this ensemble method significantly enhances RUL prediction accuracy compared to individual models.