Análise e priorização de alarmes industriais utilizando word embeddings e técnicas de aprendizado de máquina

• Ano de defesa: 2024
• Autor(a) principal: Medeiros, Isaac Emmanuel Azevedo de
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal da Paraíba; Brasil; Engenharia Elétrica; Programa de Pós-Graduação em Engenharia Elétrica; UFPB
• Palavras-chave em Português: Alarmes industriais; Incorporação de palavras; Gerenciamento de alarmes; Industrial Alarms; Alarm Management; Clustering; Classification; Natural Language Processing; Word embedding; CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA
• Link de acesso: https://repositorio.ufpb.br/jspui/handle/123456789/33691

The analysis of industrial alarm content is of utmost importance for the detection and prevention of failures in operational processes. Alarms function as an alert system, signaling the operations team about abnormal conditions and potential failures in real-time. However, the excessive generation of records by these systems can hinder the identification and effective response to critical situations. Therefore, it is essential to develop efficient alarm management, aiming to prioritize and intelligently group alarms. Additionally, by conducting a thorough analysis of industrial alarm data, it is possible to gain a deeper understanding of operational conditions, recognize recurring patterns, identify trends, and take proactive measures to prevent failures. In this study, event and alarm data were collected and used from the SCADA (Supervisory Control and Data Acquisition) system of a thermoelectric plant located in the state of Paraíba. An exploratory analysis was conducted to understand the operational impacts caused by the volume of alarms, and for these alarms and their respective clusters, patterns involving temporal sequences were sought, which may suggest causality and assist in determining root causes for specific records. Natural language processing (NLP) techniques were used in the preprocessing of alarm texts to generalize information, eliminating equipment identifiers and elements with low semantic relevance. The BERT (Bidirectional Encoder Representations from Transformers) language model was used for the numerical representation of the text, and clustering and classification techniques were applied for the efficient grouping of alarms. By clustering alarm messages using the K-means algorithm, and with the obtained clusters, the Support Vector Classifier (SVM) algorithm with a linear kernel was applied, achieving an accuracy greater than 99% on the test dataset. It was thus possible to label a new sample with considerable efficiency. The use of BERT to transform alarm messages into embeddings, as well as the text preprocessing, directly contributed to the results obtained. The approach taken in this work not only improves alarm management but also contributes to a safer and more efficient operational environment, which is essential for the sustainability and productivity of the industry.