基于标准距离k近邻的多模态过程故障检测策略

Fault detection strategy of standard-distance-based k nearest neighbor rule in multimode processes

工业产品的生产经常需要在不同模态间切换,多模态过程数据具有多中心和方差差异大等特点.针对多模态过程数据的特征,通过构造标准距离,提出了基于标准距离k近邻的故障检测策略(SD–kNN).首先在标准距离度量下计算样本与其前k近邻的距离;其次将近邻距离的平方和的均值作为样本的统计量D^2;最后,根据D^2的分布确定检测方法的控制限,当新样本的D^2大于控制限时,判定其为故障,否则为正常.标准距离使不同模态中样本间的近邻距离能够在同一尺度下度量,使得SD–kNN的D^2能够准确反映样本间的相似程度.进行了数值模拟过程和青霉素发酵过程故障检测实验. SD–kNN方法检测出了数值模拟过程的全部故障和青霉素过程95%以上的故障,相对于PCA, kPCA, FD–kNN等方法具有更高的故障检测率. SD–kNN继承了FD–kNN对一般多模态过程的故障检测能力,还能够对方差差异显著的多模态过程进行故障检测.

The production of industrial products often switches between different modes, and the multi-mode process data has the characteristics of multi center and large difference of variances. According to the characteristics, a standard distance was constructed, and a fault detection strategy based on standard distance k nearest neighbor rule(SD–k NN)was proposed. Firstly, calculated the k nearest neighborhood distances between samples in the standard distance metric;secondly, the mean of square sum of the neighborhood distances was taking as the sample’s statistic D^2;finally, according to its distribution, the control limit of the detection method was determined. When D^2 of a new sample is greater than the control limit, it was judged as fault;otherwise it was normal. Since the standard distance enables that the neighborhood distances of samples in different modes are measured at the same scale, the statistic D^2 of SD–kNN can accurately reflect the similarity between samples. Fault detection experiments in numerical simulation process and penicillin fermentation process were carried out. The SD–kNN detected all faults in a numerical simulation process and more than 95% faults in penicillin fermentation process, and it had higher fault detection rate than PCA, kPCA, FD–kNN and so on. SD–kNN inherits fault detection ability of FD–kNN in the general multimode process, and detects fault in the multimode process with obviously different variances.