嵌入维数自适应最小二乘支持向量机

状态时间序列预测方法

Condition Time Series Prediction Using Least Squares Support Vector Machine

with Adaptive Embedding Dimension



针对航空发动机状态时间序列预测中嵌入维数难于有效选取的问题, 提出一种基于嵌入维数自适应

最小二乘支持向量机( L SSVM ) 的预测方法。该方法将嵌入维数作为影响状态时间序列预测精度的重要参

数, 以交叉验证误差为评价准则, 利用粒子群优化( P SO ) 进化搜索LSSV M 预测模型的最优超参数与嵌入维

数, 同时通过矩阵变换原理提高交叉验证过程的计算效率, 并最终建立优化后的L SSVM 预测模型。航空发

动机排气温度( EGT ) 预测实例表明, 该方法可自适应选取适用于状态时间序列预测的最优嵌入维数且预测

精度高, 适用于航空发动机状态时间序列预测。- T o deal wit h the difficulty of selecting an appro pr iate embedding dimension for aeroeng ine co ndition

time series predictio n,　a metho d based o n least squar es　suppo rt　vecto r machine ( L SSVM )　with ada ptive em


bedding dimension is pro po sed.　I n the method,　the embedding dimensio n is　identified as　a parameter that af


fects the accuracy o f the aer oengine condition time series predictio n　 par ticle sw arm o ptimizat ion ( P SO)　is ap


plied to optimize the hyperpar ameter s and embedding dimension of the L SSV M pr edict ion model　 cro ss
v alida


tion is applied to evaluate the perfo rmance o f the L SSVM predictio n mo del　 and matr ix tr ansfo rm is applied to

the L SSVM　pr ediction model tr aining to accelerate the cross
validation evaluation pro cess.　Ex periments on an

aeroengine ex haust g as t emperatur e ( EGT )　predictio n demonst rates that the metho d is hig hly effective in em


bedding dimension selection.　In compar ison w ith co nv