Multivariate Alarm Systems for Time-Varying Processes Using Bayesian Filters With Applications to Electrical Pumps

Alarm systems are critically important for safety and efficiency of industrial plants. However, many alarm variables in contemporary alarm systems are generated in a way being isolated from related process variables, resulting in false and missing alarms. This paper is motivated by abnormality detection for condensate-water electrical pumps in thermal power plants and proposes a method to design multivariate alarm systems for time-varying processes. A novel feature to distinguish normal and abnormal conditions is observed on the variation rates of multiple linear regression model parameters. A model estimator based on Bayesian filters is formulated to track the variations of model parameters in normal conditions, and not to do so in abnormal conditions so that absolute cumulative modeling errors are large enough to raise alarms. The effectiveness of the proposed method is validated by industrial case studies.