Abordagens de machine learning aplicadas à manutenção preditiva industrial para a detecção de falhas

• Ano de defesa: 2024
• Autor(a) principal: Bromberger, Dani Antonini
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Santa Maria; Brasil; Engenharia de Produção; UFSM; Programa de Pós-Graduação em Engenharia de Produção; Centro de Tecnologia
• Palavras-chave em Português: Manutenção preditiva; Inteligência artificial; Machine learning; Datasampling; SHAP values; Predictive maintenance; Artificial intelligence; CNPQ::ENGENHARIAS::ENGENHARIA DE PRODUCAO
• Link de acesso: http://repositorio.ufsm.br/handle/1/31838

Artificial intelligence (AI) approaches in predictive maintenance (PdM) are reshaping how industries manage equipment maintenance. With AI techniques such as machine learning (ML) and real-time data analysis, it is possible to accurately and efficiently identify equip-ment failures. However, despite a favorable scenario, the implementation of AI in PdM faces challenges due to the absence of reliable historical equipment data, class imbalance in da-tasets, and the lack of interpretability in model decisions. For that reason, this research aims to develop AI approaches for equipment PdM to enhance the reliability of historical data, ad-dress class imbalance, and improve model interpretability. To contextualize the problem, a literature review was conducted, and 33 articles were selected and analyzed. The analysis re-vealed that leading methodologies utilize a dataset with information on equipment operating history. Following dataset preprocessing, AI algorithms are employed to identify patterns and anomalies in the data. A case study was conducted using a real-world water pump dataset to validate the effectiveness of the AI approaches. The dataset included sensor readings and the operational history of the water pump. After an initial analysis through Exploratory Data Analysis (EDA), data preprocessing techniques, such as forward fill propagation, normalizer, LabelEncoder, Principal Components Analysis (PCA), and linear correlation matrix, were applied to enhance data reliability. Subsequently, three ML algorithms — Random Forest (RF), Support Vector Machine (SVM), and k-nearest Neighbors (k-NN) — were adopted for model training, and validated using k-fold Cross-Validation. Five datasampling techniques—Random OverSampling, Borderline SMOTE, TomekLinks, NearMiss, and Cluster Cen-troids—were implemented to mitigate class imbalance. Furthermore, hyperparameter tuning via Grid Search was applied to refine the model's learning process. The models' decisions were made more interpretable by adopting SHAP values to identify key features influencing failure prediction probability. The results demonstrated that the Random Forest model with Cluster Centroids (RF_CC) yielded superior performance in terms of recall and AUC ROC. Additionally, features such as sensor_04, sensor_35, and sensor_33 were found to be most influential in model decision-making. In conclusion, the research successfully developed and evaluated AI approaches for PdM, showcasing the potential to enhance equipment reliability and optimize maintenance strategies.