为了实现船用燃气轮机剩余使用寿命的预测，对燃气轮机健康监测参数进行斯皮尔曼（Spearman）相关关系分析，采用平均影响值（Mean Impact Value，MIV）进一步分析监测参数对性能退化的敏感性，筛选出敏感特征；对得到的燃气轮机特征参数进行预处理，以消除外界环境的影响；研究了一维卷积神经网络（One Dimension Convolutional Neural Networks，1DCNN），挖掘滑窗特征参数与运行时间的映射关系，实现燃气轮机剩余使用寿命预测。基于美国国家航天局发布的航空发动机退化数据集，验证了SMIV1DCNN剩余使用寿命预测方法的有效性；开展了船用燃气轮机性能退化剩余使用寿命预测仿真试验。仿真试验结果表明，该方法不受燃气轮机初始状态影响，剩余使用寿命预测绝对误差56.10、平均绝对百分误差107.87、均方误差70.95，预测性能优于BP神经网络、LSTM神经网络与GRU神经网络。

In order to predict the remaining useful life of marine gas turbine, the Spearman correlation analysis of gas turbine health monitoring parameters was carried out, the mean impact value (MIV) was used to further analyze the sensitivity of monitoring parameters to performance degradation, and the sensitive features were extracted. The obtained feature parameters of gas turbine were preprocessed to eliminate the influence of external environment.One dimensional convolutional neural network (1DCNN) was studied to learn the mapping relationship between the feature parameters of sliding window and service time, so as to accurately predict the remaining useful life (RUL) of gas turbine. The feasibility and effectivity of the SMIV1DCNN remaining useful life prediction method was verified by using the aeroengine degradation dataset released by NASA.The simulation testing of marine gas turbine performance degradation remaining useful life prediction was carried out. The results show that the proposed method is not affected by the initial state of the gas turbine,and the mean absolute error,the average absolute percentage error and the root mean squared error of RUL prediction results are 56.10,107.87 and 70.95 respectively, which are better than BP neural network, LSTM neural network and GRU neural network.

TK478

国家科技重大专项（2017-Ⅰ-0007-0008)