Improved Principal Component Analysis Approaches for Chemical Processes Monitoring

Author:Lou Zhi Jiang

Supervisor:wang you qing

Database:Doctor

Degree Year:2017

Download:74

Pages:123

Size:10579K

Keyword:

Chemical processes usually work in extreme conditions such as high temperature and high pressure.Therefore,even a slight abnormal changes may cause the collapse of the whole production process,resulting in the interruption of production process,leakage and diffusion of poison gas,or even explosion accidents.Chemical process faults not only results in serious economic losses,but also causes serious pollution to the surrounding environment,and even threaten the safety of the workers.Therefore,early detection of process faults can effectively improve production efficiency,enhance the safety of equipment operation,and protect the personal security.One of the most common Multivariate statistical process control(MSPC)methods is principal component analysis(PCA).The main idea of PCA is to reduce the data dimensions by projecting correlated variables onto a smaller set of new variables that are uncorrelated and retain most of the original variance.The traditional PCA is a Gaussian,static,and unimode algorithm and hence it cannot handle the non-Gaussian,dynamic,and multimode features in chemical processes.To address these issues,this paper proposed several strategies for PCA,which are summarized as follows:(1)This paper proposed a preliminary-summation-based PCA(PS-PCA)to handle the non-Gaussian feature in chemical process.The main idea of PS-PCA is preliminary summation the process data to make the variable distribution close to Gaussian distribution,and then monitoring these summations by using traditional PCA.Another function of preliminary summation is accumulating the fault information in process data and hence PS-PCA is more sensitive to abnormal changes in process,so PS-PCA can achieve much higher fault detection rate than the traditional PCA.However,preliminary summation also can accumulate the information of the outliers and causes large false alarm rate,which is called as "summation infection".To eliminate the impact of "summation infection",a robust preliminary summation(RPS)method is proposed,of which the main idea is distinguishing outlier and the faulty data by the consecutive detection results and then removing the influence of outliers.In addition,the RPS method can use the outlier to amplify the information of the fault,and hence some faults can be detected more easily.(2)To address the dynamic problem in chemical processes,this paper put forward a new dynamic structure to present the dynamic property,and proposes the two-step PCA(PS-PCA).In this new dynamic structure,the process data is divided into two parts:the dynamic component and the innovation component.The dynamic component represents the component can be predicted by using historical data;the innovation component represents the independent disturbance introduced to the process each time.The innovation component is time uncorrected and has fixed expectation and variance,so it can be normalized and monitored as in traditional PCA.Then a two-step PCA(TS-PCA)is put forward to monitor these two components:identifying the dynamic structure and then use the dynamic structure to estimate the innovation component,then monitoring the innovation component by using conventional PCA.In addition,a new algorithm based on least squares algorithm is proposed to identify the dynamic structure,which can be used in both steady state and unsteady state.TS-PCA does not restricts that the dynamic processes should be in steady state,and it can monitor process in both steady state and unsteady state.(3)This paper proposed a novel combination of hidden semi-Markov model(HSMM)and PCA,termed as HSMM-PCA,to address the multimode problem.HSMM is very suitable for describing multimode process:the hidden states in HSMM can be used to describe the different operation modes,the state transition probability matrix can be used to describe the operation order,and state duration probability distribution matrix can be used to describe the duration of each operation mode.As a result,this paper combines HSMM with PCA to address the multimode problem,where HSMM divides and identifies the process modes,and then PCA is adopted for monitoring the data in each operation mode.Compared with the traditional multimode approaches,HSMM-PCA can detect the mode disorder fault and achieves much better fault detection performance.