Detection of Anomalies in Operation of Continuous Plants and Systems

Abstract

Anomaly detection during real time operation of continuous plants and systems is a key activity of the operator that aims at determining the health status of the system. For such detection the information from real time multidimensional data streams supplied by multiple sensors has to be analyzed and com-pared for similarity analysis. This paper presents a methodology and algorithm for developing a computer monitoring system for anomaly detection in online mode. The proposed method includes creating Data Cloud models based on portions of data samplings with equal lengths and estimating the Similarity Level between them as a bounded value within the interval [0, 1]. Data cloud models use the concept of mesh of grid cells that capture the local density of the data points around the cells. The Similarity analysis is based on estimating the difference between the local densities of a pair of data clouds. The whole methodology is explained and illustrated in the paper by using numerous examples with real data. The algorithm and the monitoring system use the moving window concept for continuous analysis of the similarity level in online mode. The final section of the paper shows experimental results for anomaly detection based on using real operation data (temperatures) from a Petrochemical plant.