Sensitivity clustering and ROC curve based alarm threshold optimization

Abstract In industrial practice, to reduce the variable alarm rate and ensure the safety and stability of device production, a variable alarm threshold is optimized by taking into account the receiver operating characteristic (ROC) curve that corresponds to sensitivity clustering, false alarm rate (FAR), and missed alarm rate (MAR). In this paper, the sensitivity value of the variable calculated and the grouping rule recommended by the engineering equipment and materials users association (EEMUA) are first used to cluster the variables into groups and to calculate the relevant weight ω1. In this approach, in addition to the original weights, ω1 and ω2 are the remaining weights, which correspond to the FAR and MAR, respectively. Later, the ROC functional relationship between ω1 and ω2 is obtained by the correlativity between the FAR and MAR. The optimized objective function with respect to the FAR, MAR, and original weights is then established, with the clustering weight ω1 and ω2 added to the original weights of the FAR and MAR, respectively. Eventually, an objective function is optimized to obtain the optimal alarm threshold using the particle swarm optimization (PSO) algorithm. The experimental results on the Tennessee Eastman (TE) industrial simulation data show that the proposed method can greatly reduce the FAR according to the variable sensitivity effect on the system, and it can decrease the number of alarms with a reduction rate of 37.8 % in comparison to the initial situation totally.