Sensor Fault Detection in Power Plants

Abstract

This paper presents a sensor fault detection and diagnosis approach for industrial combustion processes. Clustering algorithms are applied to the measurements of controllable process variables involved in single-input-single-output feedback control loops. Current data points from the process are compared with the clusters to identify sensor faults. Once the measurements of controllable process variables are obtained, a decision-tree algorithm monitors response process variables based on the controllable and noncontrollable process variables as predictors (inputs). Test data and training data residuals generated by the decision-tree algorithm are analyzed with statistical process control limits to identify sensor faults. The proposed approach handles data from temporal processes by periodic updates of the knowledge base. An industrial boiler combustion process is used to test the ideas presented in this paper.