基于CNN-Seq2Seq的航空发动机喘振诊断模型的研究  

Research on Aero-Engine Surge Diagnosis Model Based on CNN-Seq2Seq

作　　者：姚艳玲 袁化成[1] 陆超 唐晓澜 黄爱华[4] YAO Yan-ling;YUAN Hua-cheng;LU Chao;TANG Xiao-lan;HUANG Ai-hua(School of Energy and Power,Nanjing University of Aeronautics and Astronautics,Nanjing 210016,China;AECC Sichuan Gas Turbine Establishment,Mianyang 621000,China;School of Computer Science,Southwest Petroleum University,Chengdu 610500,China;Chengdu Aeronautic Polytechnic,Chengdu 610000,China)

机构地区：[1]南京航空航天大学能源与动力学院,江苏南京210016 [2]中国航发四川燃气涡轮研究院,四川绵阳621000 [3]西南石油大学计算机科学学院,四川成都610500 [4]成都航空职业技术学院,四川成都610000

出　　处：《测控技术》2022年第5期45-50,共6页Measurement & Control Technology

基　　金：四川省科技计划重点研发项目(2019YFG0424);四川省教育厅科研计划(18ZA0030)。

摘　　要：喘振是航空发动机在运行过程中一种危害性极大的气动失稳现象,严重时会导致飞行事故,引起人员和财产的巨大损失,因此寻求一种快速准确的发动机喘振故障的诊断方法对于保障航空发动机飞行安全具有重要的现实意义和价值。当前针对航空发动机喘振故障诊断的模型存在诊断时间长、诊断准确率不高的特点。为了解决这些问题,在序列到序列(Seq2Seq)模型的基础上,使用卷积神经网络(CNN)代替Seq2Seq中编码器-解码器结构上的长短时记忆网络(LSTM)构建得到基于CNN-Seq2Seq的航空发动机喘振诊断模型。该模型利用CNN局部连接以及权值共享的特点实现并行计算,从而大幅提高计算效率并降低诊断耗时。使用某型发动机上的多传感器试验数据对模型进行实验,实验结果表明该模型在显著降低模型大小和计算量的同时,在精确率、召回率及F1 Score上仍然获得了94.3%、92.1%、93.2%的表现,同时单次预测耗时仅需2 ms。Surge is a very harmful phenomenon of aerodynamic instability in the process of aero-engine operation,which will lead to flight accidents and huge loss of personnel and property.Therefore,it is of great practical significance and value to seek a rapid and accurate diagnosis method for engine surge fault to ensure the flight safety of aero-engine.At present,the model for aero-engine surge fault diagnosis has the characteristics of long diagnosis time and low diagnosis accuracy.In order to solve these problems,based on the sequence to sequence(Seq2 Seq)model,convolutional neural network(CNN)to replace long short term memory network(LSTM)in Seq2 Seq is used to construct the aero-engine surge diagnosis model based on CNN-Seq2 Seq,so as to realize parallel computing,thus greatly improving the computing efficiency and reducing the diagnosis time.The model is tested with multi-sensor test data of an engine.The experimental results show that the model can significantly reduce the size of the model and the amount of calculation,while still achieve 94.3%,92.1%and 93.2%performance in accuracy,recall and F1 score.At the same time,a single prediction time only takes 2 ms.

关 键 词：发动机喘振 故障诊断 序列到序列 卷积神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]