DKit: A blackboard-based, distributed, multi-expert environment for abnormal situation management

Mylaraswamy Dinkar, Purdue University

Abstract

Abnormal situation management (ASM) involves timely detection, diagnosis and correction of abnormal process conditions. Industrial statistics estimate the economic impact due to abnormal situations to be about $10 billion per year in the petrochemical industries in the U.S. alone. Process fault diagnosis, which forms the first step in ASM, deals with detection and analysis of root causes of abnormal behavior. Most diagnostic methods studied in literature tend to be restricted in their scope of application leading to the inadequacy of a single diagnostic method in meeting all the requirements of a good diagnostic system. Designing a hybrid framework based on collective problem solving is the theme of this thesis. A blackboard-based, distributed diagnostic tool kit called DKit was developed in this thesis for online real time process fault diagnosis and abnormal situation management in general. First, a set of desirable features is identified for a good diagnostic system. Different diagnostic methods are compared based on the form of process knowledge used. The inadequacy of a single method to meet all the features is the motivation for designing collective problem solving-based strategies. Second, a blackboard-based framework (DKit) is proposed and developed as an attractive alternative to individual diagnostic methods. DKit is possibly the first concrete realization of integration concepts for large scale process fault diagnosis. Key components of DKit, namely, the diagnostic methods and a scheduler which coordinates the function of different diagnostic experts are discussed in detail. A hierachical design for the scheduler with model-based digraph diagnosis at the bottom is proposed in the current design of DKit. Third, a fluid catalytic cracking unit-based testbed (called CATSIM) is developed for comprehensive testing of the proposed framework. The utility of DKit is shown through simulation runs. A hybrid neural network which combines process model information with history data for better fault diagnosis is proposed as a general framework for enhancing diagnosis by using all available process knowledge.

Degree

Ph.D.

Advisors

Venkatasubramanian, Purdue University.

Subject Area

Chemical engineering|Artificial intelligence|Industrial engineering

Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server
.

Share

COinS