航空发动机健康估计的神经网络修正卡尔曼滤波算法

Neural Network Corrected Kalman Filter Algorithm for Aero-Engine Health Parameters Estimation

关键词:  商用航空发动机  故障诊断  神经网络  卡尔曼滤波  粒子滤波

Key words:  Commercial aircraft engine  Fault diagnosis  Neural network  Kalman filter  Particle filter

摘要:

针对商用航空发动机与气路相关的传感器分布不均、且个数小于气路健康参数的个数、使用卡尔曼滤波算法估计全部气路健康参数时容易出现误判的特点，提出一种神经网络修正的卡尔曼滤波算法。该算法在每个采样周期内利用BP神经网络来修正个体的偏移方向，按粒子滤波算法计算每个个体的权值用以估计总体的均值和协方差，然后利用卡尔曼滤波算法更新所有个体，并将总体的均值作为当前时刻的估计结果。通过对商用航空发动机部件级模型在多个飞行状态点数字仿真模拟9种气路突变故障，由7个可测输出估计全部10个健康参数，该混合算法的估计误差相比BP神经网络与无迹卡尔曼滤波算法分别平均降低了34.6%与47.9%。

Abstract:

Aiming at the fact that Kalman filter is apt to misjudge the gas path health parameters of a commercial aircraft engine since available on-board sensors are unevenly distributed and the number of them is usually less than the number of health parameters， a Neural Network corrected Kalman Filter (NN-KF) algorithm is proposed. Individuals of the algorithm are corrected by Back Propagation-Neural Network (BP-NN) and their weights are calculated on the basis of Particle Filter (PF) during each sampling period in order to estimate mean and covariance， then KF is adopted to update individuals. The mean vector is regarded as estimated result every time step. Several types of abrupt gas path faults of a commercial aircraft engine at multiple fly points have been numerically simulated and all 10 health parameters are estimated based on 7 measured outputs. The simulation results suggest that the hybrid method compared with BP-NN and Unscented Kalman Filter (UKF) reduces estimated errors by averages of 34.6% and 47.9%， respectively.