Cardoso, António João MarquesFarinha, José Manuel TorresRodrigues, João Carlos Antunes2023-12-122023-12-122023-12http://hdl.handle.net/10400.6/13828The maintenance of physical assets is increasingly assuming a leading role in the success of companies, whether industrial or services. The pressure of budgets, combined with a strict maintenance policy, gives companies competitive advantages in an increasingly demanding market. This PhD thesis emerged with the aim of solving some Predictive Maintenance problems. The research also aims to respond to gaps identified in the state of the art as the prediction and classification of the long-term state of equipment. The present work describes novel contributions to the state of the art of lubricating oils. The results show that it is possible to create good models using Artificial Neural Networks (ANN) to classify oils considering all variables. Models can even possibly rank lubricants with a small error. Using Principal Component Analysis (PCA), the relevance of each variable for oil analysis was determined, thus providing a better insight into the importance of each parameter under analysis. The results also show that a neural model does not need to use all variables. Principal Component Analysis also allowed the creation of an algorithm that calculates the percentage of degradation of a lubricating oil, from the manufacturer's standard references so, this algorithm works for any industrial lubricant. It is noteworthy that lubricant classifiers (PCA, RNA and Human Experts) were compared with each other, having converged in more than 90%, which confirms the reliability of the classification. The developed algorithms can support industries, in a general way, since they provide information that is easy to interpret, and helps them to make decisions about the most appropriate time to replace oil in the assets. Through exhaustive research of the state of the art in prediction and industrial forecast, it was concluded that, to date, there is no published model for predicting failures with such a long-time span, which demonstrates the innovation and contribution of the present research for science and for the competitiveness of the industry. It should be noted that the developed algorithms have already been tested and applied, showing in general a prediction error below 10%. The current project has a short-term prediction model and another long-term prediction model, both using neural networks. The long-term forecasting model can predict asset failures 90 days in advance, which allows industries to make scheduled stops on their assets, thus avoiding losses resulting from unscheduled stops. Adequate feature input vectors in Artificial Neural Networks using sliding windows along time series greatly improved the training, leading to the conclusion that overlapping windows allow the network to learn in less iterations. Larger windows make it easier to capture peak values, but the optimal window size needs to be determined experimentally. Regarding short-term forecasting, it has been shown that data resampling can make the forecasting process faster, as it considerably reduces the input data set in the network. An algorithm was also developed to determine the expected equipment state through classification of the predicted sensor values. This way, the algorithm will be able to classify the probable state of the assets in the future in normal operation, alert or malfunction. This PhD thesis is very important for the industrial area, especially in the areas of maintenance, safety, quality, sustainability and efficiency, as it will maximize the availability of assets, contributing to the success of the Quality Management Systems and Maintenance Management Systems. Such positive increments in several sectors will have as main consequence the reduction of costs, increase of equipment availability and improvement of quality, what will become a competitive differentiator for the industries, because they will be able to approach the market with more competitive prices and quality. Part of the work was published in scientific articles and presented at several congresses and received the distinction of best presentation award in TEPEN 2021 & IncoME-VI congress in China. It also received the 2nd Young Engineer Innovation Award- PIJE 2021, from Ordem dos Engenheiros, Portugal.A manutenção e a gestão de ativos têm um papel preponderante no sucesso de qualquer indústria. O desenvolvimento da sensorização e o aumento da capacidade de armazenamento e processamento de dados, aliados à Inteligência Artificial, vieram permitir uma melhoria significativa nas técnicas de manutenção e gestão do ciclo de vida dos ativos, contribuindo para uma maior disponibilidade e eficiência dos mesmos com menores custos de manutenção. O projeto doutoral proposto descreve modelos de manutenção preditiva baseados em Inteligência Artificial. A fiabilidade e o desempenho dos motores Diesel dependem bastante da qualidade e condição dos seus óleos lubrificantes. A presente tese descreve modelos para classificar automaticamente a condição do óleo, utilizando Redes Neuronais Artificiais e Análise de Componentes Principais. Os resultados dos modelos classificadores de lubrificantes foram comparados com as classificações de peritos humanos. A comparação mostra que os modelos de classificação desenvolvidos são credíveis. A presente tese de doutoramento apresenta modelos de previsão de valores de sensores a curto, médio e longo prazo, ambos usando redes neuronais. O modelo de previsão de longo prazo é capaz de prever o valor de sensores até 90 dias de antecedência. Usaram-se métodos de aprendizagem supervisionados e não supervisionados para criar modelos de classificação do estado de uma máquina industrial. O principal objetivo era determinar quando o ativo se encontrava no seu estado de funcionamento normal ou fora desta zona, correndo assim o risco de falha. Os resultados mostraram que é possível classificar e prever o estado de máquinas industriais utilizando redes neuronais artificiais. Os modelos propostos apoiam a monitorização e manutenção de ativos, sendo que as principais implicações são a melhoria da disponibilidade operacional, incremento de qualidade, menor impacto ambiental, mais segurança e racionalização de custos.engManutenção PreditivaDisponibilidadeAtivos FísicosRedes Neuronais ArtificiaisPrevisãoAnálise de Componentes PrincipaisInteligência ArtificialAnálise de DadosPrevisãoClassificaçãoPredictive MaintenanceArtificial Neural Networksdoctoral thesis101718330