Pattern reconciliation: A new approach involving constrained clustering of time series

Abstract

Abstract In spite of the advances in strategies involving clustering and pattern recognition in time series, there are no approaches capable of directly associating the recognized patterns with the dynamic behavior of the process investigated. This paper presents a new approach involving pattern reconciliation in the clustering of time series. The method is based on Fuzzy C-Means and considers the process dynamics as a soft constraint in order to ensure the feasibility of the recognized patterns. The first case study comprises the diagnosis of abnormal operation of a non-isothermal Continuous Stirred Tank Reactor (CSTR), a benchmark system used for the assessment of Fault Detection and Diagnosis (FDD) techniques. The second comprises a real industrial scenario which involves the recognition of starting patterns in a gas turbine for fault detection purposes. The results show that the proposed method is able to recognize feasible patterns preserving the quality of clustering and classification.