A Data-Driven Approach To Process Alarm Threshold Optimization

Currently, an intractable problem of industrial process alarm systemsis acknowledged as numerous false or nuisance alarms which possiblyresulted from unreasonable process alarm thresholds. In the presence ofhigher alarm thresholds assigned, some critical alarms could be missed;on the contrary, there would be more nuisance alarms in the case of loweralarm thresholds. In order to improve the performance of process alarmsystems, it is imperative to optimize the assignments of process alarmthresholds.This paper initially conducts an overview of international anddomestic researches on alarm management system optimization andalarm threshold assignment methods. Motivated by challenges of thelimitations of traditional methods, a novel data-driven thresholdassignment approach from the perspective of optimization is proposed tosolve the problems of both false and missed alarms.Initially, non-parameter statistics kernel density estimation methodsare employed to estimate process alarm states based on historical data. A process alarm thresholds optimization problem is formulated, whichinvolves an objective function minimizing probabilities of false andmissed alarms and an enabling numerical solver.Case studies on TE process and an industrial DMF recovery processare carried out respectively. Comparison between the proposed andtraditional methods in terms of numbers of false and missed alarmsdemonstrate the benefits of the contribution. It reveals that the proposedapproach can help effectively reduce the numbers of both false andmissed alarms, save labor and ensure process security and reliability.As an application, an experimental system concerning process alarmthresholds optimization is developed based on GUI in MATLAB.