基于邻域保持嵌入-支持向量数据描述的过程监控算法及其应用

Support-vector-data-description Process Monitoring Algorithm Based on Neighborhood Preserving Embedding and Its Application

针对非线性、多模态间歇过程的故障检测问题,提出一种基于邻域保持嵌入的支持向量数据描述(support vector data description based on neighborhood preserving embedding, NPE-SVDD)故障检测策略.首先,利用NPE算法将原始数据降维到特征空间.接下来,在特征空间建立SVDD模型,计算超球体的球心O和半径R.对于测试样本,计算其到球心的距离D,对比D与R的大小确定样本状态.检测样本状态后,应用距离贡献图法进行故障变量定位分析. NPE算法可以保留原始数据的局部信息;并通过结合SVDD分类规则代替原始NPE算法的T2和SPE统计量,消除了数据服从高斯分布的限制,提高了故障检测率.利用数值模拟过程和半导体蚀刻过程仿真,将实验结果与主元分析(principal component analysis, PCA)、 NPE、 SVDD等方法进行对比分析,验证了NPE-SVDD方法的有效性.

To detect faults in nonlinear and multimode batch processes,we propose a fault detection strategy that uses support vector data description based on neighborhood preserving embedding( NPE-SVDD). First,we project the raw data into the feature space using the NPE algorithm. Then,we build the SVDD model in the feature space and calculate the center of the sphere O and the radius R of the hyper sphere. For a test sample,we calculate the distance D from it to the center of the sphere,and compare the magnitudes of D and R to determine the sample state.After determining the sample state,we use the distance contribution mapping method to determine the location of the fault variable. The NPE algorithm can preserve the local information of the raw data. NPE-SVDD eliminates the limitations associated with data that obey a Gaussian distribution and improves the fault detection rate by replacing the T2 and SPE statistics of the original NPE algorithm with SVDD. We use numerical simulation and semiconductor-etching-process simulation to compare the experimental results with those obtained by principal component analysis,NPE,SVDD and other methods to verify the effectiveness of the NPE-SVDD method.