A comparison of machine learning techniques for LNG pumps fault prediction in regasification plants
| Autor: | Eduardo Candón, Márquez A. Crespo, Fuente A. De la, J. Gómez, Jean Serra |
|---|---|
| Přispěvatelé: | Universidad de Sevilla. Departamento de Organización Industrial y Gestión de Empresas I |
| Rok vydání: | 2020 |
| Předmět: |
0209 industrial biotechnology
Boosting (machine learning) Mean squared error Computer science Decision tree 02 engineering and technology Machine learning computer.software_genre Predictive maintenance Failure Analysis Machine Learning 020901 industrial engineering & automation Approximation error 0202 electrical engineering electronic engineering information engineering Asset and maintenance management Regasification Corrective maintenance Sensors business.industry 020208 electrical & electronic engineering Predictive Maintenance Condition monitoring Control and Systems Engineering Signal Analysis Artificial intelligence business Condition Monitoring computer Liquefied natural gas |
| Zdroj: | idUS. Depósito de Investigación de la Universidad de Sevilla instname |
| ISSN: | 2405-8963 |
| Popis: | Cuenta con otro editor: IFAC-PapersOnLine Se incluye en el Volumen 53, nº 3 Article number 168746 We present a comparative study on the most popular machine learning methods applied to the challenging problem of Liquefied Natural Gas pumps fault prediction in regasification plants. The proposed solution tries to address the problem of pump failure during operation, this failure makes the pump unavailable, with a high cost of corrective maintenance. It must be taken into account that the condition monitoring may be insufficient because they are cryogenic and inaccessible equipment once the tanks have been started up. The use of machine learning techniques allows us to anticipate the response time by detecting anomalies in the operation, and to be able to do the maintenance before the failure occurs. In our experiments, we predict the power consumption based on the parameters captured in real time during operation. For the composition of the dataset, data was collected between 2007 and 2017, resulting in a dataset of over 15,000 lines for training and validation. First, all models were applied and evaluated on a dataset collected from a real case study. In the second phase, the performance improvement offered by boosting was studied. In order to determine the most efficient parameter combinations we compare Root Mean Squared Error, Absolute Error, Relative Error, Squared Error, Correlation, Training Time and Scoring Time. Our results demonstrate clear superiority of the boosted versions of the models against the plain (non-boosted) versions. The fastest scoring and total time was the Decision Tree and the best overall was Gradient Boosted Trees. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|