基于K-均值聚类与粒子群核极限学习机的推力估计器设计

Thrust Estimator Design Based on K-MeansClustering and Particle Swarm OptimizationKernel Extreme Learning Machine

关键词:  航空发动机  推力估计器  K-均值聚类  粒子群核极限学习机  直接推力控制

Key words:  Aero-engine  Thrust estimator  K-means clustering  Particle swarm optimization kernel extreme learning machine  Direct thrust control

摘要:

鉴于航空发动机直接推力控制与健康管理需要高精度及高实时性的推力估计器，提出了一种基于K-均值聚类与粒子群优化的核极限学习机推力估计方法。采用K-均值聚类对全工况范围内的测量数据进行聚类，在每一个子类中，通过核极限学习机建立推力估计器，采用粒子群算法对核极限学习机的核参数和惩罚系数进行优化，利用了核极限学习机稳定性好、非线性拟合能力强的特点，实现了对发动机推力的估计。经涡扇发动机台架试车数据训练与测试表明，本推力估计方法平均预测时间为0.27ms，实时性满足机载在线状态评估和直接推力控制需求，且在估计精度上较现有方法存在一定优势。

Abstract:

In order to achieve direct thrust control and health management of aero-engines， the thrust estimator with high precision and high real-time is needed. A design method for thrust estimator was proposed， which was based on K-means clustering and particle swarm optimization (PSO) kernel extreme learning machine (KELM). Firstly， the K-means method was used to cluster the measured data in the whole behavior range. Then the thrust estimator model was established by KELM in each sub-class， and PSO was utilized to search the best kernel parameter and penalty coefficient. The stability and non-linear fitting ability of KELM were fully utilized to estimate engine thrust. Finally， the training and testing results of turbofan engine bench test data show that the average forecast time of thrust estimation method is 0.27ms， which meets the requirements of direct thrust control and airborne on-line state assessment， and it is more accurate than the existing methods.