Abstract: Aero-engine thrust is unmeasurable，and computing the thrust based on component model is inaccurate and time consuming. The aero-engine thrust estimate method based on clustering by fast search and find of density peaks(CFSFDP)and particle swarm optimization extreme learning machine(PSO-ELM)is proposed. Firstly，the CFSFDP method was utilized to cluster the test bench data in the whole behavior range，and then，a sub-estimator was designed in each cluster using PSO-ELM. In the process of designing the sub-estimator with PSO-ELM，the particle swarm optimization algorithm was utilized to search the best hidden nodes number of extreme learning machine for obtaining the best topological structure. Finally，the training and testing results show that the maximum mean relative error is 3.06‰ which is better than the RBF network with 7.25‰ and BP network with 14.84‰. It can satisfy the demand of thrust control and onboard real time state assessment. And the method can be used for estimate other immeasurable parameters.

Key words: Aero-engine；Thrust estimate；Clustering by fast search and find of density peaks；Particle swarm optimization extreme learning machine；Direct thrust control

摘要： 针对航空发动机推力不可测，部件级模型求解推力精度不高、实时性差的问题，提出了基于快速寻找密度极点聚类与粒子群极限学习机的航空发动机推力估计方法。首先利用基于快速寻找密度极点的聚类算法对全工况范围内的台架试车数据聚类，然后在每一个子类中，用粒子群极限学习机设计了子推力估计器。在子类推力估计过程中，为使网络拓扑结构最优，用粒子群算法寻找极限学习机的最优隐层神经元数目的方法。训练与测试表明，推力估计测试相对误差最大值为3.06‰，优于传统的RBF(7.25‰)与BP(14.84‰)神经网络方法，能够满足直接推力控制与机载在线实时状态评估的需求，且可将方法扩展到其他不可测参数的估计。

关键词: 航空发动机；推力估计；快速寻找密度极点聚类；粒子群极限学习机；直接推力控制