Abstract: Traditional aero engine components fault diagnosis methods based on single hidden layer neural network are limited in test accuracy due to their shallow network structure. While deep learning methods，such as deep belief network (DBN)，have the problems as time-consuming and difficult to train parameters. In order to solve the problems of existing data-driven aero engine components fault diagnosis methods and improve test accuracy at the same time shorten training time，we proposed a new deep kernel extreme learning machine (DK-ELM) which combines kernel method with multilayer extreme learning machine. It first abstracts the original input data layer by layer gradually. Then，the abstracted feature was mapped to higher dimensional space through kernel function and is classified. These measures help to improve the algorithm in its testing accuracy and generalization performance. And it has more fast training speed than deep learning methods. We have compared the proposed method with other deep learning methods and ELM based methods on some benchmark classification data sets and engine components fault data set. Results show that the DK-ELM based method is effective，reliable and easy to implement，thus it proves to be a good and functional tool for engine components fault diagnosis.

Key words: Gas turbine fan engine；Component；Defect diagnosis；Deep neural network；Extreme learning machine；Kernel method

摘要： 运用传统单隐层的神经网络进行航空发动机部件故障诊断识别受其浅层结构影响，精度不高，而用深度置信网络(Deep belief network，DBN)等深度学习方法则存在耗时、参数训练复杂的问题。为解决现有的基于数据驱动的发动机部件故障诊断方法的不足，提高诊断精度，缩短训练时间，将核方法和多层极限学习机(Multilayer extreme learning machine，M-ELM)相结合，提出一种深度核极限学习机(Deep kernel extreme learning machine，DK-ELM)。算法首先利用深度网络结构对输入数据进行逐层的特征提取，抽象得到的特征通过核函数实现高维空间映射分类。这些措施有利于提高算法的分类精度和泛化性能，在训练速度上较深度学习也有明显的提高。将该算法与深度学习和其他极限学习机算法进行综合比较研究，结果表明:基于DK-ELM的诊断方法有效、可靠，便于实现，为航空发动机部件故障诊断提供一个更为优秀实用的工具。

关键词: 涡扇发动机；部件；故障诊断；深度神经网络；极限学习机；核方法