A fault detection and diagnosis tool for chemical processes based on hybrid-dynamic Bayesian belief network

点击翻译按钮获取中文摘要

The race to manufacture products of the highest quality at lowest cost has made process industries very complex over the years. Complex operations are more vulnerable to process upsets and equipment breakdown. Complete reliance on human operators is an overburden on the operators and puts the plants at risk. The difficulty arises due to the broad scope of faults and the size of the process plants. For example, in a large process plant there may be as many as 1500 process variables recorded every second. Furthermore, the efficiency of the operator depends on her/his knowledge, experience,and mental and physical state at a given time. Many empirical results have shown that human intuitive judgment and decision making can be far from optimal, and it deteriorates even further with complexity and stress. Industrial statistics show 70% of industrial accidents are due to human error. In the fault detection and diagnosis literature many quantitative and qualitative fault detection and diagnosis methods have been proposed. However, these detection and diagnosis methods have failed to make an impact in the industry. Most of these methods are able to detect faults early but diagnosis of the root cause of fault is poor. These methods lack transparency in the diagnosis, do not incorporate process knowledge, lacks flexibility to allow for operator’s input, and do not offer mitigation actions. As a result, although many expert systems are available, in reality, abnormal events are solely managed by process operators. Our goal is to develop a comprehensive decision-support system that will supplement the human cognitive decision-making process by assimilating information from various quantitative detection and diagnostic tools, accessing relevant process knowledge, and aiding the process of structuring the decision. The backbone of the proposed decision support system will be based on hybrid dynamic Bayesian belief network (hybrid-DBBN). The hybrid-DBBN is a flexible structure that allows integrating discrete and continuous information from a process. The structure of the network captures the known process interactions; conditional probabilities of the network are calculated using historical data. The network will be updated dynamically, based on the detection and diagnostic information from various univariate and multivariate fault detection and diagnosis tools. It will assimilate quantitative diagnostic information with process knowledge, and operator input to derive the most reasonable explanation for the fault. Finally, if there are many different alternative mitigating actions, it will use a risk based criteria to select the most appropriate operator action (s).The proposed research will advance the theory of Bayesian belief network to adequately represent chemical processes, such as, cyclic loops in a process, change of causality with time, and assimilation of discrete and continuous information. Detailed methodologies will be developed for constructing hybrid-DBBN for process fault diagnosis. Methodologies will be validated using extensive simulation studies and laboratory scale process equipment, and industrial case studies. The proposed research will make original contributions to the theory and application of DBBN. This new unified approach will give clear and concise diagnostic information and suggest appropriate action to the operator and thus make process operations safer. This has the potential to make significant economic impact; research suggests that the petrochemical industry in the US alone loses 30 billion dollars annually. In the long run, this will give the Canadian process industry a competitive edge in the business and save lives.