Resumen:
The storage of a large amount of historical data in production processes has contributed to the
development of techniques related to data mining (DM) and the extraction of useful
knowledge about processes (Knowledge Discovery in Data bases, KDD). Although there are
many studies related to Fault Detection and Diagnosis (FDD), few of them are based on
grouping and pattern recognition in time series, especially in multivariate series. In addition,
there are no work related to the recognition of patterns in time series that consider the process
model as a constraint. This study proposes a new method for the recognition of patterns in uni
and multivariate time series, based on the Fuzzy C-Means (FCM) algorithm, which directly
considers the process dynamics in the clustering problem in order to guarantee the viability of
the standards recognized. The proposed method is applied in two case studies, both related to
clustering and recognition of patterns of abnormal operation (failures) and normal operation.
The first case study is a Continuous Stirred Tank Reactor (CSTR), a well-known reference
process used to evaluate control strategies and techniques for FDD. The second application
involved a real industrial scenario comprising a commercial scale gas turbine located at the
Rômulo Almeida thermoelectric plant (UTE), an integral part of the Companhia Brasileira de
Petróleo park. The results show that the FCM algorithm and a typical metric of similarity
between time series, based on the Principal Component Analysis (PCA), do not guarantee the
recognition of patterns consistent with the process dynamics, even if good results are obtained
classification and grouping. On the other hand, the results obtained from the reconciliation
approaches proposed in this study show the obtaining of consistent and reconciled patterns
with the dynamic reality of the process, without prejudice to the quality of the results of
grouping and classification.
