Abstract: To solve the problem of insufficient fault samples for gas turbine fault diagnosis,a novel fault diagnosis method under small sample based on transfer learning with Deep Auto-Encoder (DAE) was investigated.A large number of normal samples were used to train a DAE,and a feature learning model was established.Then,the feature learning model was transferred in feature learning task with limited fault data to learn the deep features of fault samples.The learned deep feature was classified by Support Vector Machine (SVM).To make the deep features learned by DAE more representative,the feature learning ability of a single AE under different neuron number in hidden layer was evaluated,which was then used for optimizing the neuron number of hidden layer in DAE.The validity of the developed fault diagnosis method was verified by the actual flight data of CFM56-7B series engines.

Key words: deep auto-encoder, gas turbine, fault diagnosis, support vector machine, small sample, transfer learning

摘要： 为解决民航发动机故障诊断面临的故障样本不足的问题,提出一种基于深度自动编码器(DAE)迁移学习的小样本故障诊断方法。该方法首先利用大量的正常样本对DAE进行训练,建立发动机状态特征提取模型;然后将该特征提取模型迁移到具有少量数据的发动机故障样本中,并对这些故障样本进行特征提取;最后利用支持向量机(SVM)实现小样本分类。为了使DAE能够学习到更具有代表性的深度特征,利用重构误差评估不同隐藏层神经元节点数下的单个自动编码器(AE)特征提取能力,进而通过单个AE特征提取能力对DAE隐藏层的神经元节点数进行优化。以某航空公司的CFM56-7B系列发动机的实际飞行历史数据验证了所提故障诊断方法的有效性。

关键词: 深度自编码器, 民航发动机, 故障诊断, 支持向量机, 小样本, 迁移学习