Detecção de eventos de obstrução em válvula no processo de lingotamento contínuo com uso de aprendizado de máquinas

• Ano de defesa: 2024
• Autor(a) principal: Diniz, Ana Paula Miranda
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Espírito Santo; BR; Doutorado em Engenharia Elétrica; Centro Tecnológico; UFES; Programa de Pós-Graduação em Engenharia Elétrica
• Palavras-chave em Português: Lingotamento Contínuo; Bocal da Válvula Submersa; Clogging; Aprendizado de Máquinas; Deep Learning; Continuous casting; Submerged Nozzle Valve; Machine Learning; Engenharia Elétrica
• Link de acesso: http://repositorio.ufes.br/handle/10/17722

The continuous casting process, used in the manufacture of steel plates, is currently the most economical and efficient way of production within the industry. Although continuous casting is a widely used process, some problems associated with the process have not yet been resolved, one of them being the obstruction of the Submerged Entry Nozzle (SEN), which controls the flow of steel between the tundish and the mold. This obstruction, also called clogging, not only impairs the quality of the product but also results in lower process yield, resulting in losses. Thus, clogging detection is of fundamental importance, because control actions can allow the system to operate for a longer time. In this work, methodologies based on Machine Learning and Deep Learning are presented and compared to detect the occurrences of clogging from historical data of process variables. In general, the performance of the classifiers achieved very promising results in real data with unbalanced classes. In particular, the method employing spatiotemporal analysis, using four process variables, obtained a remarkably superior performance when compared to the others, reaching averages of Precision and Recall, respectively, of 95.53% and 97.33%. To reduce the false positive and negative rates, a post-processing heuristic was implemented and applied to the model output, achieving a Precision and a Recall, respectively, of approximately 98% and 99%. To the best of our knowledge, these results have never been found in the literature. Although a detailed comparison is unfeasible due to the differences between the datasets and their inaccessibility, the modeling proposed here reached higher performance levels when compared to the results found in the literature to solve this industry’s problem. The high and unprecedented results obtained in this work, therefore, will contribute both to the improvement of the quality of the final product and to the reduction of costs associated with steel production, since the more effective classification of clogging occurrences can help operators in the corrective action planning