Proposta de framework para detecção de eventos em processos industriais e diagnóstico de causa raiz : abordagens baseadas em aprendizado on-line

Abstract

Due to the increased complexity of industrial plants and their high digitization, the classic monitoring technique based on threshold logic has lost use. That technique addresses the problem in a univariable way, which in general leads to a high rate of alarm generation. In place of the previous ones, techniques based on machine learning are being used, using approachs such as artificial neural networks, fuzzy systems and statistical methods. With them, it is possible to handle the detection of events in a multivariable way, from the perspective of classification of patterns, being the theme explored in the present work. In order to develop a tool that can be used in modern industrial plants, without a database labeled with events that have occurred, or not to use information a priori, an approach based on evolving systems was developed. In such systems, the knowledge base about the process behavior is built from incremental and online learning. That way, in addition to adapting parameters, it carries out the adapt the models’s structure. Based on this, this work proposes a portable and scalable framework for use in monitoring industrial processes. The proposed structure is based entirely on evolving methods, allowing any method to be incorporated, regardless of its approach. In addition, in order to assist the supervisor in understanding the behavior of the process through the evolving method, a method of root cause is used to justify behaviors of the process that were detected. To demonstrate its use, four methods proposed in the literature with different grounds are implemented and incorporated. Experiments are performed with a benchmark and with a real gas compressor’s seal system. The results obtained indicate that the framework is a promising tool for use in monitoring processes.