基于鲁棒测地线距离的复杂间歇过程故障检测  

Complex batch process fault detection based on robust geodesic distance

作　　者：郭小萍[1] 李婷[1] 李元[1] 

机构地区：[1]沈阳化工大学信息工程学院,沈阳110142

出　　处：《仪器仪表学报》2014年第12期2681-2689,共9页Chinese Journal of Scientific Instrument

基　　金：国家自然科学基金(61034006;60774070;61174119);辽宁省教育厅科学研究(L2013155);辽宁省博士启动基金(20131089)项目资助

摘　　要：针对多工况间歇过程受噪声等影响使数据呈现出离群点、非线性等特征,本文提出一种基于鲁棒测地线距离(RGD)的故障检测方法。首先利用局部离群概率(LOP)剔除离群点之后标准化,减轻噪声干扰,提高故障检测的鲁棒性;接着,采用多向主元分析(MPCA)进行特征提取,获得主元;然后,在主元空间求得赋权邻接矩阵,通过改进的迪科斯彻(IDijkstra)算法计算各批次间的测地线距离,表征数据间的实际最短距离,从而更好地体现批次数据间的局部近邻关系。最后,通过构造测地线距离α次方统计量Lo D进行过程故障检测,与欧氏距离平方和统计量相比可提高样本的有效性,减小边缘训练数据距离的偏离程度。在人工三臂螺旋数值仿真和半导体工业实例中进行应用,结果表明了所提算法的有效性。According to outliers and nonlinear characteristics the fact thatof the multimode batch process data shows outliers and nonlinear characteristics due to the influence of noise disturbance, fault detection based on Robust Geodesic Distance（RGD） is proposed in this article. To start, Local Outlier Probability （LOP） which could mitigate noise and improve the robustness of the fault detection by excluding outliers is utilized and for standardization afterward;. Then, using Multiway Principal Component Analysis （MPCA） is used for data extraction and getting PCs; Next, Get and the empowered adjacency matrix is gotter in the reduced space. Improved Dijkstra （IDijkstra） algorithm is proposed to calculate the Geodesic distance between each batch so as to characterize the actual shortest distance of the nonlinear data and reflect the local neighborhood relations between batch data;. Finally, statisticsStatistics based on power of Geodesic distance is structured for fault monitoring, which could improve the effectiveness of the sample and reduce the deviation of distance from the edge of the training data compared with Euclidean distance squared statistics is structured for fault monitoring. The application results in artificial three-arm spiral numerical simulation and semiconductor industries instances show the effectiveness of the proposed algorithm.

关 键 词：多工况 间歇过程监视 局部离群概率 测地线距离 改进迪科斯彻算法 

分 类 号：TP277[自动化与计算机技术—检测技术与自动化装置] TH165.3[自动化与计算机技术—控制科学与工程]