Pei Sun

For some complex industrial processes that are nonlinear in nature, first-principle models are usually difficult or even impossible to obtain if a high accuracy is required. Data-driven modeling methods provide an alternative for these cases, as they are found to work well within a narrow operating regime. Common data-driven approaches include Artificial Neural Network and Support Vector Machine. In practical applications, however, the success of identifying accurate models from process data depends heavily on the availability of recorded data that provide rich information about the process. It is often the case that heuristic experience is used in designing and input sequence that excites the process with the aim of recording data that encapsulate such rich information. My research focuses to design experimental procedure that allows exciting the process with the aim of recording process data that contain a maximum information content. A second objective of my research is to then utilize the developed approaches to model fault conditions in complex operating systems with as few samples as possible. The benefits of this research is to detect fault conditions and to describe the fault conditions mechanistically by comparing an initial, or fault-free model, with the identified model describing the impact of the detected fault conditions. In addition to that, as the adapted model (describing the impact of a fault) more accurately reflects the current fault condition, it is possible to decide whether a plant shutdown is required or whether the process can continue to operate. Finally, if the process can continue to operate, the adapted model also allows to retune regulatory, supervisory as well as model-based controller to minimize the impact of a fault condition upon the process operation, particularly with respect to product quality.