A knowledge-based system architecture for diagnosis and sensor validation in chemical process plants

A knowledge-based system architecture for malfunction diagnosis and sensor validation in chemical process plants is discussed. The sensor validation is performed during diagnosis to identify sensor failures and determine alternate data for failed sensors such that correct symptomatic data are provided for the diagnosis of malfunctioning process equipment. The architecture consists of three distinct problem-solving modules and effectively integrates qualitative and quantitative knowledge for sensor validation and diagnosis. The diagnostic module involves a malfunction hierarchy which represents malfunction hypotheses in its nodes. A systematic functional decomposition strategy is developed to help identify process subsystems in levels of detail ranging from general functional systems to specific equipment components. It facilitates the construction of a malfunction hierarchy from the process flowsheet. Functional decomposition also identifies control system (Category I) sensors which need to be represented in the malfunction hierarchy so that their failures are explicitly considered in the diagnosis as the possible causes of undesirable operating conditions. For non-control system (Category II) sensors, their failures are considered to ensure that correct symptomatic information is used for diagnosis and process monitoring. For both sensor categories, four different modes of sensor failure are considered, which include mechanical failures as well as biases. Recognizing the characteristics of chemical plants and the validation problem, a novel sensor validation strategy is developed. The strategy streamlines the validation by leveraging functional and operational knowledge about the sensors. The knowledge for sensor validation, including the past operational sensor reliability ranking and the methods for determining alternate values for the sensors, is organized in the validation module. The knowledge representation is flexible so that any applicable methods for validation, ranging from qualitative estimates to quantitative process models, can be incorporated. The validation algorithm is integrated into malfunction diagnosis such that the diagnosis specifies which sensors need to be validated. Finally, numerical process data are stored in the data abstraction module, along with the various parameters for qualitative data abstraction and independent data checks. The computational architecture is demonstrated with a working prototype system developed for a dynamically simulated chemical process.