基于局部信息的LNS-PCA的多模态过程故障监测

Multimodal process fault monitoring of LNS-PCA based on local information

为了满足各种不同的企业生产需求,在实际化工过程中往往包括许多不同的运行模态。因为涉及内部各种反应的复杂性和微妙性,且各个环节相互联系,导致化工过程数据呈现高斯与非高斯数据相互混合的情况,传统多元统计监控(Multivariate Statistical Process Monitoring, MSPM)在多模态数据故障检测领域精确度较低。因此进行准确的复杂化工过程故障检测仍然是一大难题。针对这一类问题,本工作提出了一种新的基于局部信息的近邻标准化和主成分分析(Local Information Local Neighbor Standardization and Principal Component Analysis, LLNS-PCA)的方法建立高精确度的故障诊断模型。首先对样本利用高斯混合模型(Gaussian Mixture Model, GMM)方法分解成多个局部样本,应用每一个局部样本的平均值和方差进行近邻标准化,再使用主元分析进行故障监测。基于PCA监测模型,采用T2和SPE两种监测统计量对多模态过程进行监测。最后通过数值例子和青霉素生产过程验证其有效性。结果表明,相对于PCA, KPCA, LNS-PCA等方法,LLNS-PCA在多模态故障检测领域有更加迅速的反应率和更高的准确率,因此能够保障多模态生产过程的安全性和产品的高品质。

Led by market demand,the industrial process needs to switch to a variety of working modes,the industrial process detection system is becoming more and more complex,and data often presents the characteristics of the multi-mode complex distribution.It is of great significance to study multimode fault detection technology for ensuring the safe operation of industrial processes.The statistical process control method represented by principal component analysis(PCA)is a typical fault detection method based on the data drive.It is widely used to analyze whether there is a fault in the production process through the data collected by the system,which does not depend on prior knowledge and mathematical model.However,it requires that the data must conform to the Gaussian distribution,which cannot be satisfied in the multi-mode production process.To improve the performance of industrial process fault detection and eliminate the multi-modal and non-Gaussian characteristics of data,this work proposes a multi-mode process fault monitoring method based on local information LNS-PCA(LLNS-PCA).Firstly,the Gaussian mixture model(GMM)was used to divide the sample into several local samples.Secondly,for each sample data,the mean and variance of the local sample are standardized to make the data follow the Gaussian distribution.Finally,the data of each local sample were combined and PCA model was trained to obtain T2 statistics and SPE statistics for fault monitoring.The LLNS-PCA algorithm was validated with numerical examples and penicillin production data as training samples.Under the same conditions,PCA,KPCA,and LNS-PCA are used to detect anomalies.The results showed that the LNS-PCA based on local information proposed in this work has a better detection effect.In conclusion,LLNS-PCA was superior to PCA,KPCA,and LNS-PCA,which was worth promoting.