基于改进LSTM的航空发动机气路参数预测方法  

Aeroengine Gas Path Parameter Prediction Based on Improved LSTM

作　　者：马帅 吴亚锋[1] 郑华[1] 缑林峰[1] MA Shuai;WU Yafeng;ZHENG Hua;GOU Linfeng(School of Power and Energy,Northwestern Polytechnical University,Xi'an 710072,China)

机构地区：[1]西北工业大学动力与能源学院,陕西西安710072

出　　处：《测控技术》2024年第2期1-10,共10页Measurement & Control Technology

基　　金：国家科技重大专项(2017-V-0011-0062)。

摘　　要：以航空燃气涡轮发动机气路故障诊断为导向,提出了一种用于发动机气路参数预测的特征注意力增强型长短时记忆网络(Feature Attention Enhanced Long Short-Term Memory Network, FAE-LSTM)。FAE-LSTM是具有编码-解码结构的动态网络,首先通过编码器中的特征注意力单元对工况序列进行动态特征提取,然后通过特征拼接层融合编码器输出序列、工况序列和历史性能参数,最后通过解码器实现最终的参数预测。FAE-LSTM基于发动机飞行过程数据建立发动机在健康状态下的动态模型,从而作为参数预测模型应用于基于残差的故障诊断系统中。针对网络的预测性能和应用方式进行了仿真分析,结果表明,相比于其他常用多变量时间序列预测模型,FAE-LSTM的长期预测误差最低减少24.5%;相比于使用串-并联结构,故障检测系统使用并联结构的FAE-LSTM网络能够获得更精确的检测结果。Oriented by the gas-path fault diagnosis,a novel parameter prediction approach based on feature attention enhanced long short-term memory network(FAE-LSTM)is proposed for gas turbine engines.FAE-LSTM is a dynamic network with an encoding-decoding structure.Firstly,the encoder extracts the dynamic features of the exogenous input sequence,and then a feature fusion process is applied through a concatenate layer by concatenating the encoder output,the exogenous input sequence and historical predicted parameters.Finally,the final prediction results are achieved by the decoder.The FAE-LSTM models the engine in a healthy state based on the engine flight process data,which is used as a parameter prediction model in a residual-based fault diagnosis system.The analysis and research are carried out on prediction performance and the application mode.The results show that,compared with other commonly used multivariate time series prediction models,the long-term prediction error of FAE-LSTM is reduced by at least 24.5%,and compared with the series-parallel structure,the FAE-LSTM network with parallel structure can obtain more accurate detection results.

关 键 词：航空发动机 性能参数预测 故障诊断 特征注意力机制 LSTM网络 

分 类 号：V263.6[航空宇航科学与技术—航空宇航制造工程]