基于选择性深度神经网络集成的涡扇发动机剩余寿命预测

大样本高维度状态监测数据对剩余使用寿命（RUL）精准预测有着技术性挑战，为了提高以航空涡轮风扇发动机为代表的复杂装备的预测精度和收敛效率，提出一种两阶段的选择性深度神经网络集成方法. 第1阶段为多方法联合扰动下的候选集生成方案，通过采用异质神经网络结构、多时间尺度设计和算法参数随机化消除模型内部耦合关系，强化候选深度神经网络集多样性；第2阶段利用遗传算法集成修剪冗余模型，有效剔除性能不佳的冗余学习器，以获取多样化最优候选子集，并按平均集成输出预测结果. 与个体模型的数据实验对比表明，所提方法通过同步增强集成模型准确性和多样性，提升了近20%的RUL预测精度，可为运维决策提供有力支撑.

Remaining useful life estimation of turbofan engine based on selective ensemble of deep neural networks

The large sample and high-dimensional condition monitoring data has technical challenges for accurate prediction of remaining useful life (RUL). A two-stage selective deep neural network ensemble method based on high-dimensional data was proposed to improve the prediction accuracy and convergence efficiency of complex equipments represented by aviation turbofan engine. The first stage was the candidate set generation scheme under the joint disturbance of multiple methods. By using heterogeneous neural network structure, multi time scale design and algorithm parameter randomization to eliminate the internal coupling relationship of the model, the diversity of candidate deep neural network sets was enhanced. In the second stage, the genetic algorithm was used to integrate pruning redundant models. The method effectively the eliminated redundant learners with poor performance to obtain the diversified optimal candidate subsets and output the prediction results according to the average integration. The final result was estimated by average. The experimental comparison with the data of the individual model shows that the proposed method enhances the accuracy and diversity of the ensemble model and improves the RUL prediction result by nearly 20%, which provides stable support for maintenance decision.