Mantenimiento Predictivo, Machine Learning para la detección automatizada de fallos
Loading...
Official URL
Full text at PDC
Publication date
2021
Authors
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citation
Abstract
Este Trabajo de Fin de Grado ha consistido en la resolución de un problema de mantenimiento predictivo a partir de un conjunto de datos reales proporcionados por la marca de camiones Scania. El objetivo que se persigue es minimizar una cierta función de coste que pondera de forma desigual los errores cometidos en la predicción de la aparición de fallos en los camiones según estos sean falsos positivos o falsos negativos. Algunas de las particularidades del conjunto de datos disponibles es que están etiquetados pero no balanceados y, además, existe en ellos una gran cantidad de errores de medición. Por este motivo, se han aplicado sobre los datos diversas técnicas de filtrado e imputación. A continuación, se han efectuado predicciones mediante distintos algoritmos de machine learning. Se ha escogido de entre ellos Random Forest por su rendimiento superior y se ha afinado para alcanzar el mejor resultado posible, tomando en todo momento como métrica la función de coste anteriormente mencionada. Finalmente, los resultados obtenidos han sido comparados con aquellos alcanzados sobre el mismo problema por parte de varios investigadores.
This Final Thesis has consisted in the solution of a predictive maintenance problem using a real-world dataset supplied by Scania, a lorry manufacturer company. The goal to pursue is to minimize a certain cost function that weighs differently prediction mistakes depending on whether they are false positives of false negatives. Some of the peculiarities of the available dataset are that data are classified but unbalanced and there is a considerable amount of measurement errors. For this reason, several filtering and imputing techniques have been applied over the data. Then, predictions have been made with different machine learning algorithms. Amongst them, Random Forest has been selected because of its superior performance. It has been refined to obtain the best possible result taking as metric the aforementioned cost function. Finally, our results have been compared to those achieved on the same problem by several researchers.
This Final Thesis has consisted in the solution of a predictive maintenance problem using a real-world dataset supplied by Scania, a lorry manufacturer company. The goal to pursue is to minimize a certain cost function that weighs differently prediction mistakes depending on whether they are false positives of false negatives. Some of the peculiarities of the available dataset are that data are classified but unbalanced and there is a considerable amount of measurement errors. For this reason, several filtering and imputing techniques have been applied over the data. Then, predictions have been made with different machine learning algorithms. Amongst them, Random Forest has been selected because of its superior performance. It has been refined to obtain the best possible result taking as metric the aforementioned cost function. Finally, our results have been compared to those achieved on the same problem by several researchers.