一种快速AdaBoost.RT集成算法时间序列预测研究  

Study on time series prediction of a fast AdaBoost.RT integration algorithm

作　　者：严智 张鹏[1] 谢川[1] 张钰林 李保军 Yan Zhi;Zhang Peng;Xie Chuan;Zhang Yulin;Li Baojun(Aeronautics Engineering College, Air Force Engineering University, Xi'an 710038, China)

机构地区：[1]空军工程大学航空工程学院

出　　处：《电子测量与仪器学报》2019年第6期82-88,共7页Journal of Electronic Measurement and Instrumentation

基　　金：装备预研共用技术公开项目(41402010102);陕西省自然科学基金(2017JQ6034)资助项目

摘　　要：传统AdaBoost.RT算法的训练样本容易向小值样本集中,难以避免加权错误率低而真实错误率高的弱学习机,且迭代训练的速度较慢。针对这一问题,首先重新设计了相对误差函数和样本权重的更新方式;然后通过减少迭代训练中的样本规模提出了基于权重的自适应样本剔除快速AdaBoost.RT算法;最后将AdaBoost.RT算法应用于航空发动机起动阶段状态趋势监控。实验结果表明,快速AdaBoost.RT算法预测误差均值减少了0. 128 4和0. 263 2,误差标准差减少了0. 022 3和1. 794 4,虚警次数减少了5次,训练速度提升了53%。实验表明,快速AdaBoost.RT算法能有效监控航空发动机起动阶段的状态趋势,具有预测误差小、训练速度快、虚警率低等优点,对提高装备维护效率具有一定的参考意义。The training samples of the traditional AdaBoost.RT algorithm are easy to concentrate on small value samples,it is difficult to avoid weak learning machines with low weighted error rate and high real error rate,and the iterative training is slow. Aiming at this problem,firstly,the relative error function and updating method of sample weight are redesigned. Then,the weight-based adaptive sample elimination fast AdaBoost.RT algorithm is proposed by reducing the sample size in iterative training. Finally,the AdaBoost. RT algorithm is applied to the state trend monitoring of the aeroengine during the start-up phase. The experimental results show that the mean of prediction error of the fast AdaBoost.RT algorithm is reduced by 0. 128 4 and 0. 263 2,the standard deviation of error is reduced by 0. 022 3 and 1. 794 4,the number of false alarms is reduced by 5 times,and the training speed is increased by 53%. Experiments show that the fast AdaBoost.RT algorithm can effectively monitor the state trend of aeroengines during the starting phase,which has the advantages of small prediction error,fast training speed and low false alarm rate. It has certain reference significance for improving the efficiency of equipment maintenance.

关 键 词：AdaBoost.RT 时间序列 自适应样本剔除 集成学习 航空发动机 趋势监控 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]