Probabilistic Approaches Based Process Data Modeling And Fault Detection

To ensure producing high value-added products and operation safety, process monitoring and control methods have played a crucial role in modern industries. As the rapid development of the computer technology and the application of the advanced control systems, a huge number of process data are collected and recorded, which reveals the process operation status, product qualities and even economic cost. For deeply mining the useful information from the massive amounts of process data, the data driven technologies, especially multivariate statistical process monitoring (MSPM) has become a hot topic in recent years and has been applied in real industries. The traditional MSPM is mostly based on the assumption that it is used in the linear systems which are run in a single stable operation condition. However, the processes are actually operated in a more complex environment, which leads to process data also complex. In practice, the high-dimensional industrial data often contains cross-correlations, auto-correlations, nonlinear relationship, time-varying and mutilmode behaviors and constrained relationships between process and quality data. In this thesis, several novel models and methods for complex process fault detection are proposed based on probabilistic approaches. In each section, one of the complex characteristics of process data is discussed as follows.(l)An auto-regressive factor analysis (ARFA) model is proposed to capture both dynamic and static relationships in data simultaneously. In ARFA, the dynamic relationship is described using an AR model, which represents auto-correlations in data. Besides, the dynamic factors are restricted in a low-dimensional subspace, which indicates that the cross-correlated observations are explained by few dynamic factors. The proposed method is a rather general dynamic model which can improve the performance of modeling and process monitoring, especially for high-order dynamic systems. The corresponding statistics based on the residuals of the dynamic factors and observations can detect dynamic process faults correctly without smearing from previous measurements.(2)A probabilistic latent variables regression (PLVR) is proposed by including both process data and quality information for monitoring performance improvement. In PLVR, the effective latent variables, which are more sensitive to the variations of quality variables by monitoring the input subspace, can be determined through the covariance with information from the process and the quality variables. In the probabilistic framework, the traditional statistics are evaluated and new statistics for PLVR is also presented for easy tracking of the operating process. It can also monitor the occurrence of the upsets of product qualities.(3) A semi-supervised probabilistic latent variable regressionmodel (SSPLVR) is proposed to deal with the unequal sample sizes of the process and the quality variables. In SSPLVR, the model is trained based on both labeled and unlabeled data samples. When the process variables are incorporated into the model, most of them do not have corresponding quality variables, but the SSPLVR model is still helpful for improving the accuracy of the regression model. Since the problem with only part of the labeled input data and a large number of the unlabeled data is often met in the conventional continuous and batch processes of the chemical plants. Thus, two new monitoring strategies based on SSPLVR for continuous and batch processes are developed respectively.(4)An adaptive monitoring scheme based on the recursive Gaussian process (RGP) model is designed. In chemical batch processes with slow responses and a long duration, it is time-consuming and expensive to obtain sufficient normal data for statistical analysis. With the persistent accumulation of the newly evolving data, the modelling becomes adequate gradually and the subsequent batches will change slightly owing to the slow time-varying behavior.To efficiently make use of the small number of initial data, a Gaussian process model and the corresponding SPE statistic are constructed at first. With the accumulation of the newly evolving data, an updating strategy based on SPE statistic is introduced to judge if the new data should be added so that the model can be enhanced. Then a recursive GP model is used for implementation of updating, in which way the proposed monitoring scheme is effective for adaptive modelling and online monitoring.