Graph reduction techniques for exergy-based FDI on the Tennessee Eastman process

Abstract

As industrial processes, which form the backbone of the industrialized world, continue to become larger and more complex, control systems require fault diagnostic schemes that can maintain plant safety and product quality, even during fault conditions. As a result, graph- based exergy fault diagnostic schemes have become increasingly popular tools, especially in the petrochemical industry. However, when these graph-based schemes are applied to com- plex processes, the implementation becomes more complex and requires the deployment of more sensors as well as additional computational resources for the control system. Therefore, this study focuses on evaluating the concept of graph reduction by proposing several graph reduction techniques and assessing their efficacy at reducing complexity while preserving the performance of the fault diagnostic schemes. To determine the effect that the reduction tech- niques would have on the fault detection and isolation (FDI) methods, it is first necessary to determine the performance of these FDI methods prior to any graph reduction and use it as control data. The distance parameter, eigendecomposition, and residual-based FDI methods are used in this study. The attributed graph data used in this study is generated from the Ten- nessee Eastman process (TEP). Five graph reduction techniques are proposed based on three theoretical concepts, which rely on understanding the process used by FDI methods. These three concepts are concerned with finding redundant attributes and removing them from the graph. These reduction techniques are evaluated with an experimental process whereby the extent to which the technique reduces graph attributes (reduction interval) is increased, and the performance of the FDI methods using this reduced version of the graph data is recorded. Graph reduction is a viable concept when at least one reduction technique can reduce graph attributes while maintaining the level of FDI performance achieved prior to reducing any attributes. To validate this study, it is shown that graph reduction is a general solution by applying these five reduction techniques to the attributed graph data of the gas-to-liquids process (GTLP) and evaluating their effect on FDI performance. The three FDI methods are applied to the graph data of the GTLP to generate a set of control data. The reduction techniques are assessed with the same experimental process, which reduces more attributes from the GTLP graph data and measures FDI performance after each reduction increment. Since at least one reduction technique could reduce the attributed graph data of both the TEP and GTLP, while maintaining a similar level of FDI performance for at least one FDI method, graph reduction is considered a general solution, and the reduction techniques have been validated. This study clearly shows that it is possible to reduce the attributed graph data of a process and maintain or even improve upon, in some instances, the level of FDI performance achieved before reducing attributes. It will, therefore, contribute to mitigating the adverse effects resulting from applying graph-based FDI methods to large and complex processes.