427680 Early Warning System Design: The Combination of Expert Knowledge and Historical Data

Tuesday, November 10, 2015: 12:45 PM
Salon F (Salt Lake Marriott Downtown at City Creek)
Hangzhou Wang, Engineering and Applied Science, Memorial University of Newfoundland, St. John's, NF, Canada, Faisal Khan, Faculty of Engineering and Applied Sciences, Memorial University of Newfoundland, St. John's, NF, Canada and Salim Ahmed, Faculty of Engineering and Applied Science, Memorial University of Newfoundland, St. John's, NF, Canada

Early warning system design: the combination of expert knowledge and historical data

Hangzhou Wang, Faisal Khan*, Salim Ahmed

Safety and Risk Engineering Group, Faculty of Engineering and Applied Science,

Memorial University of Newfoundland, St. John’s, NL, A1B 3X5, Canada

Since process industries deal with hazardous materials in daily operations, it is important to monitor the state of a process in real time to identify any vulnerable condition before it leads to a more severe event. Alarm systems are introduced to monitor these key variables states and give early warnings before severe event occur. Despite efforts to improve alarm systems, alarm flooding remains a significant problem in the process industries. Besides, these warnings only indicate these monitored variables deviations states, it is difficult to find the root cause for these warnings, or to predict what will happen with these changes. To solve this problem, an early warning system design methodology, combining expert knowledge and historical data, is proposed. The method comprises three steps as below:

1. Step 1: Identify events and scenarios in the process

1) Identify events by HAZOP analysis

2) Allocate variable to these events

3) Identify corresponded scenario and links with scenario  

2. Step 2: Develop Bayesian Network based early warning system

1) Learn structure of monitored variables Bayesian Network from historical data

2) Learn parameters for monitored variables Bayesian Network with the given data set

3) Appending expert knowledge of events nodes to monitored variables Bayesian Network to construct the early warning system

3. Step 3: Implement the early warning system

1) Gather evidences of monitored variables states updating from real time variable state

2) Analyze the root causes of the current warning

3) Predict the occurrences of these events in real time

4) Take desired actions according to these occurrences probability

In this proposed methodology, with these continuously observed state changes of variables as evidence, the Bayesian Network forward inference method is used to update these events’ probabilities. The Bayesian Network backward diagnosis method is used to find the root causes of the state change quickly. Once the probability of a certain event exceeds the acceptable upper threshold, an event warning message will be generated and conveyed to operators, together with the root causes if available. Compared to a conventional DCS alarm record, more information can be exploited from the same sequence of variables states’ change with this proposed method; also more meaningful event warning messages are conveyed to operators, and desired action can be applied by operators in time to reduce the occurrence probability of this unfavorable event.

Keywords: Bayesian Networks, early warning system, structure learning, expert knowledge, historical data


* Corresponding author. Tel.: +1 709 864 8939

E-mail address: fikhan@mun.ca (Faisal Khan)

Extended Abstract: File Not Uploaded
See more of this Session: Data Analysis and Big Data in Chemical Engineering
See more of this Group/Topical: Computing and Systems Technology Division