Multivariate Statistical Monitoring Methods For Dynamic Process Data

The statistical monitoring methods studied in this dissertation are substantially data-driven and may be applied to process monitoring either on-line or off-line. Generally speaking, all of these methods may be confined to the field of MSPC(multivariate statistical process control) in process system engineering. The common process data are dynamic in nature, which can be caused by the complicated process mechanism , random errors or disturbances, the extensive adoption of feed-back control or the requirement of real-time sampling, et al. However, it is difficult to build an dynamic model for a large multivariate system , which result in the monitoring of the dynamic process data a very challenge problem. Fortunately, compared to conventional system identification, filtering or control , which usually are based on an relatively accurate dynamic model, process monitoring has few requirements of the model, but focuses on how to describe the data variation, which may be reflected by some statistic indices. Therefore, in this thesis we mainly study the model-free methods related to dynamic process data monitoring. Combined with the knowledge in multivariate statistical analysis, statistical quality control, dynamic description and time-frequency transform, and based on statistical and statistical learning theory, the aforesaid model free methods can be applied to monitor dynamic process data effectively without much model parameters to determine.The main contents of this thesis are as follows: .(1) Fault detection, fault variable identification and fault identification are challenging problems in MSPC. A integrated novel MSPC method is proposed by combining multivariate feature extraction with three SVM-based methods commonly used in one-class classifier design, key feature selection and multi-class classifier design, respectively. The fiirst aspect of this method is its ablity to calculate control limits of multiple statistics for fault detection simultaneously without conventional theoretical distribution assumptions. Secondly, the method determine the key variables for fault identification based on both their magnitude changes and their contributions to fault classification in residual space, improving the identification accuracy. In the third method; Thirdly, fault identification is implemented by taking advantage of the well-known properties of SVM-based multi-class classifier which avoids introducing specific discriminant criteria. Using principal component analysis (PCA) as feature extraction method, the foresaid SVM-based MSPC method is illustrated with application to a benchmark simulator Tennessee Eastman process and its effectiveness is verified.(2) The methods to modify control charts for dynamic multivariate data which don't following IID(Independent and Identical Distribution) assumption are studied. Firstly, the conventional methods either by adjusting control limits using non-parameter methods or by creating new statistic are introduced, then combining the advantages of EWMA(Exponentially Weighted Moving Average) control chart and MBB(Moving Block Bootstrap) control chart for monitoring auto-correlated data, an modified MBB method-eMBB(EWMA-MBB) is proposed.In this method, the latent variables are firstly extracted, then a new eMBB bootstrap statistic is defined to account for more extensive dynamics than conventional MBB. In a simulation example, for weakly dependent multivariate data, particularly in the case of small sample size, the foresaid eMBB and MBB method has advantage over the conventional PCA method according to their empirical ARL(Average Run Length) performance, and the empirical ARL of eMBB is closer to the theoretical value than that of MBB.(3) To overcome the difficulty to model the complex dynamic system, the conventional dynamic latent variable method is deeply explored to improve the statistical monitoring performance. Firstly, the properties of dynamic latent variable are confirmed to contain more dynamic information than conventional latent variable, but they own some of autocorrelation and cross-correlation. Thus, we suggest adopting the non-parametric methods to modify the control charts. For monitoring the residual space, the corresponding non-parameter methods are also recommended. In the second aspect, a method to choose the lagged variables and the time-lagged length is proposed, which taking process knowledge and empirical in-control ARL validation into account. The third feature in our research is proposing a new strategy to identify fault variable, which is based on cumulative sum of each variable's residual and an RFE(Recursive Feature Elimination) algorithm. The properties of the foresaid methods are verified through two typical simulations.(4) In the face of autocorrelation and multiscale of process data, the application of time-frequency transform to monitor multivariate dynamic process is investigated based on the basic framework of MSPCA(Multiscale Principal Component Analysis). Firstly, the superiorities brought out by discrete wavelet transform are pointed out, as well as the properties of the scale features extracted by MSPCA, and the reason why MSPCA can substitute the conventional spectral PCA is illustrated. Secondly, two fault detection methods aiming at detecting abrupt fault and the fault with stationary scale features respectively ,are presented, therefore reinforce the basic MSPCA; thirdly, to identify the fault with stationary scale features , a SVM-based classifier is proposed using PCA to extract corresponding scale feature. The properties and effectiveness of the above methods are illustrated by application to an standard simulation process of CSTR(Continuous Stirred Tank Reactor).(5) Integrate all kinds of methods explored in the thesis and merge them into an statistical monitoring system which includes other related methods, and investigate the key points that will be met in industrial application. Lastly, the method to design the monitoring system of an industrial fluidized reactor is presented, and an statistical model for monitoring the chunk in the reactor is constructed based on the real process data.