基于注意力机制循环神经网络的液体火箭发动机故障检测  

Rocket Engine Fault Detection with Attention based Recurrent Neural Networks

作　　者：张万旋 卢哲 张箭 薛薇[1] 张楠[1] ZHANG Wanxuan;LU Zhe;ZHANG Jian;XUE Wei;ZHANG Nan(Beijing Aerospace Propulsion Institute,Beijing,100076)

机构地区：[1]北京航天动力研究所,北京100076

出　　处：《导弹与航天运载技术（中英文）》2024年第2期25-31,共7页Missiles and Space Vehicles

摘　　要：针对液体火箭发动机主级段工作过程,采用多变量非线性时间序列分析理论,在两级注意力机制循环神经网络(Dual Stage Attention Based Recurrent Neural Networks,DA-RNN)的基础上,提出一种新型时序分析工具——卷积两级注意力机制循环神经网络(Convolutional Dual Stage Attention Based Recurrent Neural Networks,CDA-RNN),从而建立故障趋势预测模型。通过对预测残差进行自相关性分析并定义故障置信概率,提出了故障检测量化依据。利用发生微弱故障的热试车数据进行验证,结果表明,CDA-RNN模型对非稳态工作段微弱故障多参数检测具有良好鲁棒性,该方法十分有效,具有直接应用价值。Focusing on the main working phase of liquid rocket engine,with the aid of multivariate non-linear time series analysis,and based on Dual Stage Attention Based Recurrent Neural Networks(DA-RNN),a new time series analysis tool,Convolutional Dual Stage Attention Based Recurrent Neural Networks(CDA-RNN),is proposed,by which a fault trend prediction model is established.Compared with LSTM,DA-RNN,etc,this model shows higher prediction accuracy.Combined with autocorrelation analysis of the prediction residual,a quantitative basis of fault detection is proposed after introducing failure confidence probability.Using hot test data with weak fault to validate the model,result shows that the CDA-RNN model enables robust weak fault muti-parameter detection in unsteady working process.This strategy is so effective that it calls for direct engineering application.

关 键 词：多变量时间序列 注意力机制 循环神经网络 卷积神经网络 自相关性分析 

分 类 号：V231.1[航空宇航科学与技术—航空宇航推进理论与工程]