Researches On Performance Monitoring And Fault Diagnosis For Process Industry Based On Data-driven Technique

Performance monitoring system is one of the key elements of CIPS(Computer Integrated Process System) in process industry. In order to improve product quality and economic benefit, the process conditions should be closely monitored and faults should be timely detected. On the other hand, with wide application of DCS(Distributed Control System) and computer technology in process industry, large volumes of data are sampled and collected. However, these data which contain information about process conditions are not well exploited, so that there exists "data rich, information poor" in process industry. It is one of the most active research area in the field of process control that how to transform these collected data into valuable information, and mine deep-level information about process operation to improve the performance of process monitoring, which also is the focus of this thesis.This thesis involves around principal component analysis(PCA). In view of the characteristics of different industrial processes, some improvements of traditional PCA have been made at different degree, and some new monitoring algorithms are also proposed. All the strategies don't require the mathematical model, but are based on data-driven technique. Considering the fact that continuous process and batch process are the two important production modes in process industry, and each of them has its respective characteristic, our works are divided into two parts, those are, monitoring of continuous processes and of batch processes.The main contribution of this thesis is as follows,1 Multivariate kernel-density estimation method is used to calculate the distribution of data and assess the impact of parametric uncertainty on the monitoring performance. The simulation result of a continuous stirred tank reactor (CSTR) shows that the data would deviate from normal distribution under parametric uncertainty, and different parameters have distinct effect on data distribution at the same uncertain degree. Compared with traditional PCA-based methods which assumes that the data are independent and follow normal distribution, this kernel-density based method doesn't require any assumption on data distribution.2 An integrated framework for process monitoring and fault diagnosis is presented, which combines independent component analysis (ICA) for feature extraction and support vector machine (SVM) for identification of different fault source. ICA as a new signal processing technique is used to determine the projection coefficient matrix which represents the features characterizing the current operating condition. Multiple well-trained support vector machines use the projection coefficient matrix as their inputs to identify the fault. TheICA-based method without assumption that data follow normal distribution, can easily calculate the joint probability density function of independent components(ICs), because each of ICs statistically independent. Furthermore, these ICs span the characteristic space of normal operating condition (NOC), and the changes of coefficients matrix are monitored which are obtained by projection of process variables onto the space of NOC. The monitoring performance of the proposed method and that of the PCA-based method are compared with application to the Tennessee Eastman process(TE process). The results show the superiority of the proposed ICA-based method over the PCA-based method, and less detection delay can be obtained. In view of insufficient number of fault samples in process industry, support vector machines(SVM) with the capacity of sufficient learning from a limited training set are used to identify the projection coefficient matrix of TE process, and satisfying results are also obtained.3 A method based on distance between principal component subspace is proposed to identify different operating condition. This method clarifies the essence of PCA from the view of space, and every different subspace represents different operational mode and process performance. Based on that, distance bet...