Achieve the best performance during abnormal conditions

Abnormal conditions that are not resolved quickly become incidents with economic, safety and environmental consequences. This whitepaper, authored by President and CEO Samy Achour, sets out why conventional alarm systems and traditional decision support systems are not sufficient for abnormal condition management, and what an intelligent architecture has to do instead.

The paper grounds the problem in published industry data, citing Abnormal Situation Management Consortium® findings on incident frequency and cost in the US petrochemical industry, refinery major-incident averages, and insurance claim figures for equipment damage with the observation that many of these shutdowns were preventable.

Its central technical argument concerns alarms. Alarms are a limited set of conclusions on a few individual variables, with thresholds deliberately set far from normal operation to avoid nuisance alerts. As a result they do not detect deviations from optimum operations or from best practice, and they miss minor problems that only become major after their effects propagate, meaning that by the time a temperature or pressure alarm fires, significant time has passed since the initial failure. In effect, alarms look backwards in time. Alarm filtering compounds the issue by hiding information under grouped alarms, and alarm management systems focus on DCS/SCADA alarms without including higher-level business events.

Traditional decision support systems are examined next: rules-based engines built for a particular situation, often unable to span the enterprise, difficult to update as the plant changes, and requiring extensive customisation.

The proposed alternative is KnowledgeNet® (KNet), an action-oriented architecture that complements rather than replaces alarm management. The paper details its three-step decision-making process, KRules detects symptoms of abnormal conditions and transmits them to the KRCA module for root cause diagnosis, which then guides the operator to a fix through operator interface messages or the KWorkflow engine with priority-to-colour mapping on schematic displays so process health is visible at a glance.

A substantial section covers data validation in KRules: spike rejection, limit checks, noise filtering, zero standard deviation to detect stuck sensors, sudden noise shifts, exponential and moving-average filtering, variance-weighted estimation across redundant sensors, and alarm validation against independent measurements. A full keyword glossary closes the paper.