A carregar...
Publicação
MANUTENÇÃO PREDITIVA. ANÁLISE DE DADOS E APLICAÇÃO DE MODELOS PREDITIVOS A FALHAS DE EQUIPAMENTO INDUSTRIAL
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 10.66 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
A Internet of Things contribuiu para a Manutenção Preditiva permitindo às empresas do
setor da energia identificar potenciais problemas nos dispositivos de produção antes da
ocorrência da falha. É expectável o crescente interesse do setor da indústria nesta área,
uma vez que uma paragem na produção pode causar grandes prejuízos.
Neste projeto desafiado pela Galp, serão testados métodos de Machine Learning ba-
seados, com o intuito de comparação, nas três abordagens de aprendizagem existentes:
Supervisionada, Semissupervisionada e não Supervisionada para a deteção precoce de
falhas em duas bombas presentes na refinaria de Sines da Galp , utilizando as medições
existentes atualmente sensorizadas por dispositivos de controlo incorporados. A inte-
gração das técnicas de manutenção preditiva com a sensorização dos dados permitirá à
fábrica evitar a substituição desnecessária de equipamentos, poupar custos e melhorar a
segurança, disponibilidade e eficiência dos processos. Para tal, realizou-se uma análise
da literatura existente, considerando dois aspetos proeminentes: pré-processamento dos
dados monitorizados e a deteção precoce de anomalias/falhas.
Este trabalho apresenta um quadro para a seleção e análise de modelos de previsão
de falhas, que se destina a permitir o desenvolvimento eficiente de um conjunto de mode-
los explorados, sendo eles o Local Outlier Factor (LOF), One-Class Support Vector Machine
(OCSVM) e Random Forest (RF), para dois conjuntos de dados correspondentes aos dois
equipamentos (offline) com dados de Janeiro de 2014 a Setembro de 2021 (momento da
extração), com o objetivo de a Galp colocar estes modelos em produção com dados online,
para uma deteção de anomalias em tempo real.
O principal desafio cingiu-se na combinação de dois fatores: um pré-processamento
dos dados eficiente de modo que estes ganhassem valor e robustez antes de serem ingeri-
dos pelos modelos, e o de conseguir implementar os modelos com a respetiva otimização
de parâmetros com o intuito de encontrar o melhor ajuste possível, de modo a torná-los
capazes de detetar desvios do estado normal de funcionamento e verificar que eventos ou
falhas podem ser detetados.
The Internet of Things has contributed to Predictive Maintenance by allowing companies in the energy sector to identify potential problems in production devices before failure occurs. It is expected the growing interest of the industry sector in this area, since a production stoppage can cause great losses. In this project challenged by Galp, Ma- chine Learning methods based on the three existing learning approaches will be tested for comparison purposes: Supervised, Semi-supervised and Non-Supervised for the early detection of failures in two pumps present in Galp’s Sines refinery , using the existing measurements currently sensed by embedded control devices. The integration of predic- tive maintenance techniques with data sensing will allow the plant to avoid unnecessary equipment replacement, save costs and improve process safety, availability and efficiency. To this end, a review of existing literature was conducted, considering two prominent aspects: pre-processing of monitored data and the early detection of anomalies/faults. This paper presents a framework for fault prediction model selection and analysis, which is intended to enable the efficient development of a set of explored models, being the Local Outlier Factor (LOF), One-Class Support Vector Machine (OCSVM) and Random Forest (RF), for two data sets corresponding to the two equipments (offline) with data from January 2014 to September 2021 (time of extraction), with the objective of Galp putting these models in production with online data, for a real-time anomaly detection. The main challenge was the combination of two factors: efficient pre-processing of the data so that it gained value and robustness before being ingested by the models, and being able to implement the models with the respective optimization of parameters in order to find the best possible fit, so as to make them able to detect deviations from the normal operating state and check which events or faults can be detected.
The Internet of Things has contributed to Predictive Maintenance by allowing companies in the energy sector to identify potential problems in production devices before failure occurs. It is expected the growing interest of the industry sector in this area, since a production stoppage can cause great losses. In this project challenged by Galp, Ma- chine Learning methods based on the three existing learning approaches will be tested for comparison purposes: Supervised, Semi-supervised and Non-Supervised for the early detection of failures in two pumps present in Galp’s Sines refinery , using the existing measurements currently sensed by embedded control devices. The integration of predic- tive maintenance techniques with data sensing will allow the plant to avoid unnecessary equipment replacement, save costs and improve process safety, availability and efficiency. To this end, a review of existing literature was conducted, considering two prominent aspects: pre-processing of monitored data and the early detection of anomalies/faults. This paper presents a framework for fault prediction model selection and analysis, which is intended to enable the efficient development of a set of explored models, being the Local Outlier Factor (LOF), One-Class Support Vector Machine (OCSVM) and Random Forest (RF), for two data sets corresponding to the two equipments (offline) with data from January 2014 to September 2021 (time of extraction), with the objective of Galp putting these models in production with online data, for a real-time anomaly detection. The main challenge was the combination of two factors: efficient pre-processing of the data so that it gained value and robustness before being ingested by the models, and being able to implement the models with the respective optimization of parameters in order to find the best possible fit, so as to make them able to detect deviations from the normal operating state and check which events or faults can be detected.
Descrição
Palavras-chave
Manutenção Preditiva Machine Learning Deteção precoce de anomalias/falhas