摘要
航空发动机参数具有高维时序性,参数特征能够用于表征发动机剩余寿命。从起飞到着陆全周期内,不同机型所采集和存储的发动机参数量和数据规模差异明显。为了解决参数维数不同导致的特征提取细粒度不一致,从而影响发动机寿命预测精度的问题,提出一种用重参数化结构改进的vision transformer(VIT)模型。建立多维数域映射算法,将发动机参数源域数据集转换为彩色图像数据集,从数据源端改善了泛化性。改进VIT模型的多头注意力卷积结构,引入重参数化结构及全连接层,将源域数据的时序性转换为图像数据的空间特性,提高了模型寿命预测精度。在公开数据集(CMAPSS)上实验表明,寿命预测方均根误差(RMSE)范围为[10.83,14.68],预测精度至少提高了4.3%。此外,该方法在公开数据集(N-CMAPSS)测试,RMSE为2.07,进一步验证了模型泛化性能。
Time-series aero-engine parameters are originated from different data sources.Its feature might be relevant with the remaining useful life prediction of the aero-engine.During the whole periods from the airplane’s taking off to landing,different types of aircrafts have particular engine parameters and data scale.In order to improve the remaining useful life prediction precision and apply it into different dimension engine datasets,the new prediction model was constructed based on the revised VIT(vision transformer)with the re-parameterization.The algorithm of datasets mapping into the RGB image was given,and the model generalization got better.Re-parameterization was combined with multi-dimension attention of VIT,and the precision was improved after the transferring between the time series datasets and space feature.The result of the experiment in datasets(CMAPSS)of the datasets showed that the value of root mean square error(RMSE)lied in[10.83,14.68].The prediction model and algorithm were better than the traditional method with the smaller RMSE of 4.3%decrease in the least.And also,it could be applicable in another datasets(N-CMAPSS)with the RMSE value of 2.07.
作者
郭晓静
郭佳豪
徐琛
GUO Xiaojing;GUO Jiahao;XU Chen(National Demonstration Center for Experimental Engineering Techniques Training Education,Civil Aviation University of China,Tianjin 300300,China;College of Electronic Information and Automation,Civil Aviation University of China,Tianjin 300300,China)
出处
《航空动力学报》
2026年第3期207-217,共11页
Journal of Aerospace Power
关键词
航空发动机剩余寿命
数域映射
重参数化结构
多头注意力
图像特征提取
aero-engine remaining useful life
data field mapping
re-parameterized structure
multi-dimension attention
image feature extraction
