针对单一深度学习网络对涡扇发动机退化特征提取不足、超参数选择困难的问题，提出一种改进一维卷积神经网络（1Dimensional Convolutional Neural Network，1DCNN）和长短时记忆网络（Long ShortTerm Memory，LSTM）的涡扇发动机剩余寿命预测方法。首先，利用相关性、单调性和离散性一系列评价指标对涡扇发动机的多维传感器特征参数进行评价和选择，将综合评价指标高的优选特征参数作为1DCNN的原始输入特征；然后，通过改进激活函数和Dropout函数来提升1DCNN的特征提取能力，构建表征发动机退化趋势的一维复合健康指标；最后，利用贝叶斯优化（Bayesian Optimization，BO）的LSTM挖掘一维复合健康指标的时间特征，并实现剩余寿命预测。为验证此方法的预测效果，采用美国国家航空航天局提供的涡扇发动机退化数据集进行剩余寿命预测，实验的均方根误差为14.040 2，评分函数值为314.607 8。结果表明：相比于单一深度学习方法和传统机器学习方法，该方法不仅能获得较高的剩余寿命预测精度，还能有效解决深度学习模型超参数选择困难的问题。

Aiming at the problems of insufficient extraction of degradation features and difficult selection of hyperparameters in single deep learning network for turbofan engine, an improved 1dimensional convolutional neural network (1DCNN) and long shortterm memory network (LSTM) was proposed to predict the remaining life of turbofan engine. Firstly, a series of evaluation indexes including correlation, monotonicity and discreteness were used to evaluate and select the characteristic parameters of the multidimensional sensor of turbofan engine. The optimal characteristic parameters with high comprehensive evaluation indexes were taken as the original input characteristics of 1DCNN; then, the feature extraction capability of 1DCNN was improved by improving the activation function and Dropout function, and a onedimensional composite health index was constructed to characterize the degradation trend of the engine; finally, LSTM based on Bayesian optimization (BO) was used to mine the time characteristics of onedimensional composite health indexes and predict the remaining life. In order to verify the prediction effect of this method, the residual life prediction was carried out by using the turbofan engine degradation dataset provided by NASA. The root mean square error of the experiment was 14.040 2, and the score function value was 314.607 8. The results show that compared with the single deep learning method and the traditional machine learning method, the proposed method can not only achieve higher prediction accuracy of residual life, but also effectively solve the problem of difficult selection of hyperparameters in the deep learning model.

V263.5