Pattern Recognition using Multivariable Time Series for Fault Detection in a Thermoeletric Unit

摘要

This paper presents a methodology for recognition of operating patterns of a gas turbine in a thermoelectric power plant (Brazilian Oil Company). Patterns related to the normal starts (without failure) and starts with failure (trip) were recognized. The process data were obtained from the plant information management system (PIMS) and techniques of data mining suitable for multivariable time series were adopted with emphasis on similarity metrics, linear scan and clustering, among others. The recognized patterns represent important and useful results to support the development of dynamic system for the monitoring and predicting the probability of failure in the equipment.