一种有效提高控制图异常模式识别效果的方法  

On Improving the Performance of Trend Pattern Recognition in Control Chart

作　　者：吴德会[1] 虞耀君[1] 

机构地区：[1]九江学院电子工程系

出　　处：《机械科学与技术》2007年第11期1475-1478,共4页Mechanical Science and Technology for Aerospace Engineering

基　　金：国家自然科学基金项目(70272032)资助

摘　　要：提出了一种基于主元分析(PCA)的控制图特征提取方法,先用常规控制图提取生产过程数据集,再将其高维特征进行线性组合并向低维空间投影,从而降低了分类器的输入维数,提高特征的敏感性。用一个含有6种趋势的20维特征数据集进行测试,通过PCA预处理后,特征被降到了3维并保留了88%的分类信息;再用BP分类器对特征提取前后的数据集进行识别,结果优于新型RSFM网络进行直接识别的效果。实验结果表明了本文方法的可行性和有效性。A method for feature extraction from control chart based on principal component analysis（PCA） was proposed. Firstly, universal control chart was used to extract data sets of trend pattern. Secondly, through linearly associating the trend data sets in higher dimensional feature space and projecting them to lower dimensional feature space, we can reduce the number of recognizer input dimension and increase their sensitivity to pattern. A 20-dimensional simulated data set, including six patterns,was used to test. Through PCA, the data set was reduced to 3 dimensions and contained 88% classification-messages. Meanwhile, the abnormal pattern recognizer to the extracted characters was designed based on a BP（back-propagation） artificial neural network. Compared with RSFM（regional supervised feature mapping） methods which the original feature were delivered to as input data directly, the proposed method can obtain quicker result and higher recognition rate. Experimental results show that this feature extraction method is feasible and effective.

关 键 词：主元分析 控制图 特征提取 模式识别 

分 类 号：O235[理学—运筹学与控制论]