推进技术 ›› 2013, Vol. 34 ›› Issue (10): 1406-1413.

• 控 制 • 上一篇    下一篇

基于奇异值趋势分解和LS-SVR的航空发动机性能指数组合预测

李 冬1,李本威2,赵 凯1,张 赟2,宋 岩3   

  1. 海军航空工程学院 研究生管理大队, 山东 烟台 264001;海军航空工程学院 飞行器工程系, 山东 烟台 264001;海军航空工程学院 研究生管理大队, 山东 烟台 264001;海军航空工程学院 飞行器工程系, 山东 烟台 264001;海军航空工程学院 基础实验部, 山东 烟台 264001
  • 发布日期:2021-08-15
  • 作者简介:李 冬(1984—),男,博士生,研究领域为航空发动机性能衰退、评估与预测。 E-mail:happyli.dong@163.com

Combined Prediction of Aero-Engine Performance Index Based on Singular Value Trend Decomposition and LS-SVR

  1. Graduate Students’ Brigade, Naval Aeronautical and Astronautical University, Yantai 264001, China;Department of Aircraft Engineering, Naval Aeronautical and Astronautical University, Yantai 264001,China;Graduate Students’ Brigade, Naval Aeronautical and Astronautical University, Yantai 264001, China;Department of Aircraft Engineering, Naval Aeronautical and Astronautical University, Yantai 264001,China;Department of Basic Experiment, Naval Aeronautical and Astronautical University,Yantai 264001,China
  • Published:2021-08-15

摘要: 发动机性能指数是衡量其性能优劣的重要指标之一。针对发动机性能指数具有非线性、非平稳的特征,引入多层次多尺度的思想,在此基础上提出一种基于奇异值趋势分解的组合预测方法。利用奇异值趋势分解提取原始数据的趋势项和波动项;以改进粒子群算法分别获取趋势项和波动项在最小二乘支持向量回归模型中的最佳参数组合(嵌入维数、延迟时间、惩罚因子、核参数),并引入回归移动的思想,在此基础上利用最佳的最小二乘支持向量回归模型进行预测。预测结果表明预测精度明显增加,计算时间也相对减少。提前预测步长在5步之内时,精度变化不大;步长超过10步,精度下降很快。与不同预测方法比较,证明了方法的有效性。 

关键词: 航空发动机;奇异值分解;性能参数;组合预测;粒子群参数优化 

Abstract: Engine performance index is one of most important index which weights good or bad performance. Aiming at performance index having characteristic of nonlinearity and nonstationary,multi-level and multi-scaling were introduced and a combined predicting method based on SVD (Singular Value Decomposition) was presented. Trend and fluctuation were extracted using SVD. The best parameter combination of LS-SVR (Least Square-Support Vector Regression) model in trend and fluctuation (embed dimension, delay time, punish coefficient, kernel parameter) was generated using improved PSO (Particle Swarm Optimization). The regression and motion were introduced. They were predicted respectively using best LS-SVR on this basis. The results indicate that predicting precision increases apparently and the computational time decreases. Precision changes unobviously when predicting step is within 5. The precision decreases rapidly when step is beyond 10.Compared with different methods, it verifies effectiveness of this method. 

Key words: Aeroengine;Singular value decomposition; Performance parameter; Combined prediction; Parameter optimization of particle swarm optimization