一种数据驱动的工业报警自适应阈值预测方法  

Data-Driven Approach to Process Alarm Management System Adaptive Threshold Prediction

作　　者：王佳 郝鑫[1] 徐砚 WANG Jia;HAO Xin;XU Yan(China Software Testing Center,Beijing 100048,China;China Cyber Security,Chengdu Sichuan 610000,China)

机构地区：[1]中国软件评测中心,北京100048 [2]中国电子科技网络信息安全有限公司,四川成都610000

出　　处：《通信技术》2020年第1期52-58,共7页Communications Technology

基　　金：国家重点研发计划资助（No.2018YFB0904900,No.2018YFB0904904）~~

摘　　要：化工过程常常工作在多个稳定状态。传统的报警阈值只是针对单个模式而设定,当变量从一个稳定状态到另一个稳定状态时,会产生误报警和漏报警,于是提出了一种报警阈值自适应预测方法。首先,通过历史数据得到各个阶段的带宽系数和贝叶斯估计的样本信息。其次,为了更新模型参数,在过渡过程采用基于蒙特卡罗方法的贝叶斯参数参数估计方法,利用后验分布函数的均值和方差,并在稳定过程采用递推迭代公式更新均值和方差。针对整个过程得到自适应的报警阈值,以此减小产生误报警和漏报警的数量。最后,通过一个工业实例数据验证了方法的有效性。Chemical processes operating in multi-steady states can usually undergo process transitions.However,the traditional alarm threshold are always assuming a single mode of process operations and hence result in false alarms and missed alarms.In this paper,firstly,the bandwidth coefficient of each mode and the Bayesian estimation of sample information are obtained from the historical data.Then,in order to update the model coefficient,the Bayesian method based on Monte Carlo method algorithm is discussed for parameter estimation of single variable linear regression model in the transitions;the recursive iteration formula is proposed in the stable process to update the mean and variance.Compared with the posterior distribution of mean and variance,variance sensitivity based-adaptive alarm thresholds(Tadp)are created to fix false and missing alarm problems.Finally,The potential of this method is demonstrated in an industrial dimethylformamide(DMF)recovery plants.

关 键 词：工业控制系统 报警管理 贝叶斯估计 自适应阈值 

分 类 号：TP277[自动化与计算机技术—检测技术与自动化装置]