针对目前燃气轮机基于数据驱动的故障诊断技术诊断精度有待提升的问题，建立某型号燃气轮机的热力学模型并植入故障特征构造训练样本，在此基础上训练一种基于注意力机制的卷积神经网络与长短期记忆网络结合的神经网络模型。卷积层和注意力机制模块提取燃气轮机多维度的故障特征，长短期网络层进行时序动态故障参数处理。研究表明：相比于典型卷积神经网络，这种神经网络模型不仅能够识别多种故障的动态特征，对于各类故障的诊断能力均可达到93%以上，且加入注意力机制模块后对于不同的故障类型诊断准确率最高提升约3%。

Aiming at the problem that the diagnostic accuracy of current gas turbine fault diagnosis technology based on data drive need to be improved,establishes the thermodynamic model of a certain type of gas turbine and implants the fault features construction training sample.On this basis,a neural network model including convolutional neural network and long and short term memory network based on attention mechanism is trained.The convolution layer and attention mechanism multidimensional fault features of gas turbine,and then the long and short term network layer processer the timeseries dynamic fault parameters.The results show that compared with convolution neural networks,this attentionmechanismbased neural network model,which can recognize the dynamic characteristics of faults,can achieve more than 93% accuracy for all kinds of faults.After adding the attention mechanism module,the accuracy of different fault types can be improved by up to about 3%,which provides a new idea for fault diagnosis of gas turbine.

TK478

国家自然科学基金（51906138， 51876116）；基础科研重点项目（JCKY2019204B009）；专项基础研究（2017-I-0011-0012,2017-I-0002-0002）