Context-based recognition of process states using neural networks

Patterns whose interpretation varies across contexts are common in many engineering domains. The resulting one-to-many mapping between patterns and their classes cannot be adequately handled by traditional pattern recognition approaches. This important class of pattern recognition problems although common has not received any attention in chemical engineering domain. In this paper, we show that identification of the state of chemical or biological processes is context dependent. Two types of features are important for context-based pattern recognition-primary features, which determine the class of a pattern, and contextual features, which cannot themselves predict the class, but can improve the effectiveness of the primary features. Process measurements can be used as primary features for identifying the current process state, and the previous process state provides the context in which the primary features have to be interpreted. We also propose a dynamic neural network architecture for context-based operating state identification. Three variations of the architecture, each using a different approach to identify change of context, are described. These are illustrated using two case studies for operating state identification-the startup of a simulated fluidized catalytic cracking unit and operation of a lab-scale fermentation process. A comparison with traditional neural networks reveals that the performance of the proposed context-based pattern recognition architecture is superior in all cases.