基于动态集成算法的航空发动机气路参数预测

Aeroengine gas path parameter prediction based on dynamic ensemble algorithm

针对单一学习机对航空发动机气路参数预测困难的问题,提出了基于动态加权核密度估计(DWKDE)组合方法的集成预测算法,该组合方法选择测试样本的近邻样本,通过评估学习机在近邻样本的局部性能动态确定各学习机的权值,并基于该权值利用加权核密度估计实现数据序列的集成预测。该组合方法不易受离群值和样本不对称分布的影响,将该组合方法用于AdaBoost.RT和AdaBoost.R2算法,获得了改进后的集成学习算法。实验证明:相比于神经网络和原始集成学习算法,改进后的集成学习算法较好地提高了航空发动机气路参数序列的预测精度,方均根误差(RMSE)指标至少可降低27%。

For the problem that single global modeling methods cannot get satisfactory aeroengine gas path parameter prediction results,ensemble algorithms with a combination method called dynamic weighted kernel density estimation（DWKDE）were proposed.Neighboring samples of the test samples were chosen.The weights of the base learners were dynamically calculated by evaluating the base learners＇ local performance in the neighboring samples.Based on the calculated weights,the integrated prediction of the time series was realized by the weighted kernel density estimation.The proposed combination method was insensitive to outliers and deviations from normality.Applying this combination method to AdaBoost.RT and AdaBoost.R2,experiments were conducted on gas path parameters of aeroengine,verifying that the proposed algorithms can achieve higher prediction accuracy than the single neural network and the traditional ensemble learning algorithms,for example,theroot mean square error（RMSE）can be reduced by at least 27%.