Fault Detection Of Complex Industrial Processes Based On Data-driven Methods

In the past few years, the scales of the modern industrial processes are increased continuously. The production systems have become more and more integrated and complex with the developments of automatic control techniques. On the one hand, the interactions among different manufacturing plants make the faults more destructive for the process. To reduce the impacts of the process faults such as the decline of the product quality and the increase of the product costs, the effective process monitoring becomes one of the most urgent problems in modern industrial processes. It is important to improve the security of the industrial processes and increase the system reliability. However, the complexity of the processes makes the monitoring models difficult to be built by the traditional methods. On the other hand, with the filedbus tenology and the distributed control systems widely used, a large number of the historical operating data can be measured and stored automatically by the data acquisition and storage system system. In this context, data-driven monitoring methods have been intensively researched and arisen as an active research area in process monitoring.However, most of the traditional data-based fault detection methods often contain some assumptions such as linear, Gaussian and unimodal processes. All of assumptions limit the monitoring performance of the monitoring methods when applied in practical processes.To deal with the practical monitoring challenges, some improvements of the conventional multivariate statistical process monitoring methods and several novel monitoring strategies are propsed in this dissertation, which are summarized as follows:(1) For nonlinear process monitoring, a fault detection method called statistics kernel principal component analysis (SKPCA) is developed. The original data space is projected into a statistics space by the calculation of the the higher-order statistics of the data set. Then, KPCA is conducted in the statistics space to extract some dominant principal components. Since the higher-order statistics are involved in the statistics space, SKPCA can extract more meaningful knowledge and obtain a better monitoring results than KPCA(2) For multimode process monitoting, a novel multimode fault detection approach named distance space statistics analysis (DSSA) is proposed. Based on the analysis of the distribution characteristics of the multimode data, the distance relationships of the normal samples and the fault samples are compared. Then, the expressions of the samples in the original data space are obtained to build the distance space. Finally, the statistics analysis method is employed to extract the useful informations in the higher-order statistics. The proposed DSSA method use only one single model for multimode process monitoring.(3) In order to improve the monitoring performance for multimode process, a novel local neighborhood standardization (LNS) strategy is proposed as a data preprocessing method to address the challenges caused by the multimode characteristic of operating data. After a thorough analysis of LNS, the proposed data preprocessing method is integrated with principal component analysis method to provide a complete fault detection method in multimode processes. By using LNS, the multimode characteristic of operating data can be erased in the data processing phase while the correlations of variables are maintained. The contribution-based fault identification method can be directly applied in the LNS-PCA method. The new data preprocessing method makes the PCA method display a robust performance regardless of the multimode characteristic in process data.(4) The process complexity makes the within-mode data usually follow an uncertain combination of Gaussian and non-Gaussian distribution. A novel Mahalanobis distance-based local outlier factor (MDLOF) method is proposed to deal with the complex distribution of the process data. The local structure of variables is taken into account by employing Mahalanobis distances, and the local density of the surrounding neighborhoods is also considered. The density-based local outlier factor is employed as a monitoring statistic. The degree of the oulierness is calculated for each sample. Different from the traditional distance-based monitoring methods, the information in the local neighborhood is utilized in MDLOF. The density-based monitoring index makes MDLOF effective regardless of the data distribution.(5) To develop an efficient monitoring method with the multimodality and the within-mode uncertainty of data distribution, a novel neighborhood standardized local outlier factor (NSLOF) method is proposed by integrating the LNS and local outlier factor. A new normalized Euclidean distance based on the local neighborhood standardization strategy is employed during the calculation of the monitoring index. The utilization of the normalized Euclidean distance makes the contribution-based fault identification method available for the NSLOF method. Meanwhile, compared to the MDLOF, the monitoring performance is improved and the algorithm complexity is reduced.Several monitoring methods are compared with the proposed methods, and the numerical example and the Tennessee Eastman process are applied to illustrate the efficient of the proposed methods. Finally, based on the conclusion of the thesis, some future research directions are discussed.