Contribuição ao uso de inteligência artificial para detecção e diagnóstico de falhas em máquinas rotativas

• Ano de defesa: 2024
• Autor(a) principal: Gonçalves, Daniel Ferreira
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Uberlândia; Brasil; Programa de Pós-graduação em Engenharia Mecânica
• Palavras-chave em Português: Máquinas rotativas; Mancais hidrodinâmicos; Monitoramento e diagnóstico de falhas; Inteligência artificial explicável; Rotating machines; Hydrodynamic bearings; Fault monitoring and diagnosis; Explainable artificial intelligence; CNPQ::ENGENHARIAS::ENGENHARIA MECANICA::MECANICA DOS SOLIDOS; Engenharia mecânica; ODS::ODS 9. Indústria, Inovação e infraestrutura - Construir infraestrutura resiliente, promover a industrialização inclusiva e sustentável, e fomentar a inovação.
• Link de acesso: https://repositorio.ufu.br/handle/123456789/42114; https://doi.org/10.14393/ufu.te.2024.293

Rotary systems are widely employed in modern industry. The implementation of methodologies for proactive maintenance, which reduces unscheduled downtime and increases operational efficiency, is on the rise in the industry. However, the spread of Artificial Intelligence (AI) based monitoring and fault diagnosis systems still faces challenges in both academia and industry. Some failures, such as imbalance, misalignment, and cracks, have similar symptoms, making it difficult to diagnose accurately. The absence of labeled historical data and the scarcity of explainable models that are more transparent and understandable to end users also make the use of this type of system unfeasible in the industry. In view of this, the present work provides a methodology for an intelligent monitoring and diagnosis system based on the combination of explainable AI models and vibration analysis techniques. The proposed methodology is a low-cost solution for the industry, since it does not require dedicated sensors or high-performance hardware for diagnosis and presents possibilities of integration with online services. Data from numerical models of rotors under various operating conditions and experimental tests performed in a rotating system supported by hydrodynamic bearings were used. The analysis of experimental data provided valuable information about the real behavior of the machine and enabled the implementation and adjustment of reliable mathematical models and, thus, a more assertive diagnosis. Techniques for trait extraction in the time, frequency, and time-frequency domains were explored. The possibility of reducing the data by selecting the best features or reducing dimensionality is also discussed. Learning models ranging from an ensemble of clusterizers to traditional methods, such as the k Nearest Neighbor (k-NN) and the Support Vector Machines (SVM), were employed in the recognition of failures. Subsequently, explainability tools were applied to better understand the predictions obtained. The high success rates of the models, combined with their interpretability, make the proposed methodology a promising tool for monitoring and diagnosing failures in rotating systems in industrial environments.