Uma avaliação crítica sobre técnicas baseadas em PCA para detecção de falhas em processos da indústria química

Abstract

The aim of this dissertation was to study, develop and implement chemical process monitoring systems. In order to understand the limitations of Data Driven modeling techniques, classical techniques based on Principal Component Analysis (PCA) and related procedures were adopted as a starting point. It was sought to build a critical view of the classical methods of process monitoring. Data Driven models are quite limited, since previously learned models do not fit necessarily well new operating regions, whose data were not available during the modeling phase. In addition, it is possible to conclude that the techniques based on data currently available, such as PCA and its variants, do not have the capacity to actually model the dynamic behavior of a process. This was the main motivation for the study of Recurrence Plots. Based on the concept of recurrence, a new technique was developed to monitor processes with multiple operational points. A new control chart was proposed to monitor processes with multiple operation points, based on the Frobenius Norm. Many other aspects related to process monitoring, such as variable selection, preprocessing and removal of spurious values, were not addressed in this work, but are of fundamental importance for industrial application. Finally, it is important to note that, although the proposed monitoring methodology has been inspired by the concept of recurrence, it, in fact, is not able to reconstruct the phase space of the process.