(595e) A Novel Approach to Develop Real-Time Leading Indicators to Predict Incidents in Chemical Plants by Utilizing near-Misses

A recent
study by Pariyani et al. (in press) introduced a new approach for identifying
near-misses in chemical plants.  Abnormal events, which occur
when plant variables depart from and return to their normal operating ranges,
are recognized as near-misses, as they are precursors to incidents.  In this paper,
a new technique involving near-miss utilization and management is proposed to
identify escalations in the probabilities of the occurrence of the incidents,
particularly shutdowns, and to flag alert signals ? permitting operators to be forewarned
of major incidents likely to occur in the near future.  Also, the alerts detect
the onset and/or presence of inherent faults, or special-causes, likely to lead
to most-critical abnormal events and trips.  As an example, consider a chemical
process in normal operation with few variables out of their normal operating
ranges.  Shortly after a disturbance, a high frequency of its variables move
out of their normal operating ranges, creating a flood of alarms.  For such a
dynamic process, machine learning techniques (involving support vector
machines) are used to: (a) track patterns of absolute increases in
abnormal events in real-time by groups of variables, and (b) detect gradual
shifts in various performance indicators introduced by Pariyani et al. (in
press), and flag alerts to forewarn the operators of potential undesirable
events several hours before they occur.  These techniques have been targeted to
minimize the number of false-positives (alerts having lead-times greater than
24-48 hours) and false-negatives (undetected incidents) ? providing a more
proactive warning system, leading to reduced risk levels. 

Also, when creating an alert, as new abnormal
events are recorded, ASPEN DYNAMICS simulations are used to identify the
special-causes leading to the abnormal events underlying the alert.  Then, a
novel moving-horizon predictive method is introduced to predict the lead-times prior
to incidents.  The results and conclusions are presented for a case study
involving an industrial-scale air-separation plant. 

Reference

1.   Pariyani A.,
Seider W. D., Oktem U. G., and Soroush M., ?Incident investigation and dynamic analysis
of large alarm databases in chemical plants: A fluidized-catalytic-cracking-unit
case study,? Ind. Chem. Eng. Res., in press.