摘要
主元分析方法通常采用累积贡献率(CPV)法来确定主元个数,而CPV法具有一定的主观性。本文提出一种基于阈值法来自适应实时确定主元个数的方法,有效克服传统累积贡献率法的缺点;在阈值法的基础上,提出一种对数据样本及协方差矩阵均加入遗忘因子的移动窗递推PCA(MWRPCA)过程监测方法。结合TE过程仿真,该方法能够提高监测精度,降低计算时间。将MWRPCA和RPCA、改进RPCA和MWPCA方法相比较,通过对比分析各主元分析方法TE过程的仿真结果,MWRPCA方法可以运用最少的计算时间达到最高的平均CPV值,同时有效的检测出故障,更利于在线实时监控,仿真结果验证了所提出方法的有效性。
Principal component analysis methods usually use Cumulative Percent Variance（CPV） method to obtain the number of principal components,however CPV method is subjective to some extent.A new threshold-based method to adaptively obtain the number of principal components is proposed in the paper,which can eliminate the disadvantages of traditional CPV method.At the same time,a Moving Window Recursive Principal Component Analysis（MWRPCA）,which puts the factors into data samples and covariance matrix,is put forward to monitor the process in industry.From the simulation of Tennessee Eastman（TE） process,and the comparison results with RPCA,Modified RPCA and MWPCA, MWRPCA can achieve the highest monitoring accuracy with the least computation time,and we can detect the faults effectively.So the simulation results reveal that MWRPCA is more suitable for online monitoring.The effectiveness of new MWRPCA method is proved.
出处
《计算机与应用化学》
CAS
CSCD
北大核心
2011年第8期1022-1026,共5页
Computers and Applied Chemistry
基金
中央高校基本科研业务费(ZZ1136
ZY1111)
