Combining functional modeling and reasoning with on-line event analytics

Research output: Book/ReportPh.D. thesis

34 Downloads (Pure)

Abstract

Modern industrial plants rely heavily on automation to increase energy and economic efficiency, however, human operators are tasked with the supervision of the plant and have to intervene when abnormal situations occur. To determine whether and how to mitigate an abnormal situation, control rooms provide a large number of sensor and status information to the operators and alarms are intended to draw operators attention to situations that require immediate action. In reality, however, the amount of information provided to operators can actually reduce operators' focus and distract them from maintaining awareness of the situation. The most severe situations that overload operators with information are alarm floods, in which a large number of events is presented to the operator requiring immediate attention. As coping strategy operators tend to resort to treating symptoms as quickly as possible to reduce the number of alarms, when rationally analysing the whole set of occurring events and treating underlying root causes would be more effective. Identifying these root causes from the large amount of data available and providing operators with the context of how occurring alarms can reduce the risk of overloading operators and ensure safe and efficient plant operation.
This work contributes to an operator support approach based on a functional representation of the process flow and operating goals in Mulitlevel Flow Modelling (MFM).
MFM is an established modelling methodology for operations and process knowledge that has been applied for process design and diagnosis in a variety of industrial contexts.
The qualitative diagnosis in MFM facilitates the identification of all possible fault scenarios.
Two major aspects are considered in this work: maintaining correct causal analysis based on all occurring alarms from the plant and ensuring that the MFM model used for diagnosis fits the current plant situation.

Toward the correct on-line analysis an improved causal reasoning system leveraging the connection of occurring alarms has been developed and shown to increase efficiency and reasoning speed.
Approaches to ranking the identified root cause candidates have been defined to provide meaningful information to operators.
By providing a ranking of the root causes, operators' focus can be directed toward distinguishing the most likely root causes and determining an efficient mitigation strategy.

The configuration and operating goals of a plant are frequently changed in accordance with pre-defined operating procedures.
To ensure that the correct model is used for the causal reasoning, methods to link these operating procedures to MFM have been investigated.
In an industrial setting operators tend to adapt operating procedures, based on their experience and out of execution efficiency and convenience.
To account for these adaptations, a structured representation of operating procedures and method for validating the representation against operations logs have been proposed.

All proposed methods have been demonstrated on case studies with industrial relevance for the chemical or petrochemical industry.
Combining the qualitative modelling and reasoning in MFM with the analysis of alarms and events from the control system in real time facilitates the contextual situation assessment necessary to support operators.
With ongoing research into machine-learning based alarm generation and MFM based counter action planning, the proposed methods provide the core functionality for establishing a comprehensive operator support system, which can relieve operators from the repetitive task of filtering out relevant events and provide assistance for efficient mitigation of abnormal situations.
Original languageEnglish
PublisherTechnical University of Denmark
Number of pages164
Publication statusPublished - 2020

Fingerprint

Dive into the research topics of 'Combining functional modeling and reasoning with on-line event analytics'. Together they form a unique fingerprint.

Cite this