Detecção de Falhas de Processos Industriais em Múltiplos Pontos de Operação via Análise Externa Linear e Não Linear

Abstract

The methods of multivariate statistical control of processes focused on fault detection have great potential to extract relevant information from the data generated by instruments and control systems of industrial plants, which are usually underutilized. However, traditional multivariate statistical control techniques should not be used for processes operating at multiple points of operation due to the inability to distinguish a failure from a normal change of operation. This limitation makes it difficult to use these techniques in real processes. In this context, this work presents studies and proposes five methods based on Nonlinear External Analysis and External Analysis with multiple linear models, to detect failures in industrial processes, which are naturally nonlinear and work at multiple points of operation. These methods will be applied to a literature benchmark simulator and to the monitoring of a real vibration process of a large process fan used in an iron ore pelletizing furnace. The results show that the proposed methods can distinguish failures from normal variations of operating points of industrial processes, keeping the level of false alarms in the specified value. Additionally, the results show that these methods have the potential to detect failures automatically in advance, allowing for corrective actions that may reduce or even avoid damages to equipment of a certain process, generating a potential for financial gains.