Monitoramento de processos industriais sob o efeito de concept drift utilizando análise externa e aprendizado ativo

Abstract

Appropriate industrial processes monitoring is of paramount importance to detect, predict and diagnose faults, in addition to ensuring that the products have quality and are within the expected specifications. Thus, it is important to detect disturbances in the process, equipment malfunction or other harmful events as soon as possible, so that afterward, it is possible to identify the cause of the unwanted behavior and take corrective actions. There are several methods in the literature that deal with monitoring. Among them, the use of external analysis techniques and multivariate methods is widespread. However, there are process changes to which these methods are not robust enough to adapt and monitor, generating false alarms in normal situations. Thus, this work proposes a system of joining PCA (Principal Component Analysis) and external analysis techniques with an active learning method through Concept Drift detections to generate a monitoring system that adjusts to changes in the behavior of an industrial plant process. The method was applied to a real problem of an industrial fan and the results presented showed that the model was adapted to the process changes without storing excessive data for training the model and with a lower number of false alarms in relation to a model without updating.