Development of advanced fault diagnosis techniques for complex industrial processes

Modern industrial processes are systems with a high degree of complexity. These systems comprise of a large number of components functioning in harmony to produce high quality products. In practice, the operating states of these components are monitored in real-time to determine whether there are abnormalities in process operation. There are a number of challenges associated with monitoring a large number of process components, for instance, high monitoring cost and flooding of false alarms. To address these problems, a multivariate statistical process monitoring (MSPM) framework has been developed in recent years. The MSPM performs multivariate statistical analysis on real-time process data to generate two monitoring statistics capable of identifying abnormalities in all aspects of process operations. Many researchers have proposed a variety of techniques within the framework of MSPM. This thesis advances these developments by proposing several novel extensions in the areas of features extraction, robust online fault diagnosis and multivariate dynamic risk assessment. The first major contribution of this work is the use of Copula, a method for modelling complete dependence structure between random variables, for non-Gaussian feature extraction. The Copula method is then combined with Spearman's rank correlation coefficient for nonlinear feature extraction. Due to the use of the Spearman's correlation coefficient, the proposed technique is also robust to data contamination. Another type of technique based on Nonlinear Gaussian Belief Network is also proposed for robust feature extraction from noisy data with nonlinear variations. The second major contribution is the development of a powerful visualization tool for real-time process monitoring. This visualization tool is derived from the well-known nonlinear feature extraction algorithm, the Self-organizing Map. A direct visualization of the real-time operating state of processes is presented on a 2D map. A number of operating regions have also been identified on the 2D map, allowing for a more refined process monitoring. The third main contribution of this thesis is the integration of the process monitoring techniques into the operational risk assessment framework. The process monitoring statistics are transformed to indicate the real-time probability of faulty conditions. In the meantime, the possible process losses due to likely fault condition are also estimated. The probability of fault and possible process losses are then combined to determine the operational risk of process. Based on the risk level, the most effective remedial measures can be easily determined.