Uma estratégia evolutiva para detecção e diagnóstico de falhas em sistemas dinâmicos

Abstract

Fault Detection and Diagnosis (FDD) in dynamic systems is an important problem in process engineering that has attracted much attention in recent years, Early detection and diagnosis of process failures in its initial stage can help avoid abnormal event progression and reduce productivity loss. This work proposes an approach for FDD based on the Participatory Clustering Algorithm; an evolutionary classifier that allows the system to loam how to classify the faults as they occur. It is able to detect new operation modes, and in each iteration a new group may be created, an existing group may have its parameters changed, or two redundant groups unified. Moreover, more than one group can be used to describe an operation mode. The end result is the proposition of Modyied Particqmtorjy Clustering Algorithm (MPCA), which innovates in the use of compatibility and alert indices, as well as in their calculation procedures. The MPCA introduces the concepts of candidate groups and consolidated groups, suggesting a new approach to the outlier problems. In order to analyze the algorithm performance, the MPCA is tested for FDD of two different dynamical systems: an induction motor (using simulated data), and a process on wire rod mill (using actual operation data). Finally, the results are discussed and the use of the MPCA in other FDD cases is evaluated.