基于时空近邻标准化和PCA的故障检测方法

Fault Detection Method Based on Time-Space Neighborhood Standardization and Principal Component Analysis

针对多阶段复杂工况数据非线性、动态性的特点,提出了基于时空近邻标准化和主元分析相结合的故障检测方法(TSNS-PCA).首先在时间层次上搜索样本的近邻集,其次寻找样本的时间近邻在空间层次上的近邻样本集,然后在标准化样本集上通过主元分析方法计算监控模型的T 2和SPE控制限,最后计算待检测样本的相应统计值,并且与控制限相比较,达到故障检测目的.TSNS方法不仅可以将复杂非线性、动态性、多阶段数据中心变换至坐标原点,并使新数据近似服从单一正态分布,还保持了离群点与正常点间的偏离程度.利用一个非线性动态数值模拟过程和青霉素仿真过程证明在动态多阶段过程监控中TSNS-PCA方法的效果明显优于PCA、KPCA、DPCA、WkNN、TSNS-LOF等方法.

With the development of industrial technology,industrial processes usually have multiple production stages.Therefore,the study of multi-stages fault detection methods is of great significance.Multi-stage complex working condition data has non-linear and dynamic characteristics.Aiming at the characteristics of the above data,a fault detection method based on the combination of time-space local nearest neighborhood standardization and principal component analysis(TSNS-PCA)is proposed.Firstly,the nearest neighbor set of the sample is searched at the time level,secondly,the nearest neighbor sample set of the time nearest neighbor of the sample is found at the space level,and then the T 2 and SPE control limits of the monitoring model are calculated on the standardized sample set by the principal component analysis method.Finally,the corresponding statistical value of the sample to be detected are calculated and compared with the control limits to achieve the purpose of fault detection.The TSNS method can transform the complex nonlinear,dynamic and multi-stage data to the origin of coordinates and make the new data approximately to a single normal distribution,while maintaining the degree of deviation between outliers and normal points.A non-linear dynamic numerical example and penicillin simulation process are uesd to prove that the effect of TSNS-PCA method is significantly better than that of PCA,LNS-PCA,DPCA,WkNN,TSNS-LOF and other methods in dynamic multi-stage process monitoring.