Heavy-duty trucks are important links in the logistic chains of transport. Critical components in trucks include fuel injectors in which inaccuracies can lead to severe financial damage and higher emissions. Intelligent and efficient ways to detect such scenarios are thus of high importance. This thesis applies machine learning algorithms to measured or estimated engine data, focused on gas exchange signals, to detect inaccuracies in fueling quantities. The fueling inaccuracies considered were of low deviations from the nominal curve, with magnitudes not covered by the currently used fueling diagnostics. The data used for the models was generated from Scania test cell engines where different setups of injectors were deliberately set to over- or underfuel. Seven different machine learning models were used on the data and evaluated on how well they could detect deviations from nominal fueling. The tests were mainly done with a pure data-driven approach but also improved through different data selection techniques and using domain knowledge. An investigation to connect the findings within the thesis to real customer data was initiated in order to make the results useful for e.g. predictive maintenance. The complications connected to why this was not ultimately achieved were discussed.
| Identifer | oai:union.ndltd.org:UPSALLA1/oai:DiVA.org:uu-447180 |
| Date | January 2021 |
| Creators | Dufva, Johannes, Lindgren, Andreas |
| Publisher | Uppsala universitet, Avdelningen för systemteknik |
| Source Sets | DiVA Archive at Upsalla University |
| Language | English |
| Detected Language | English |
| Type | Student thesis, info:eu-repo/semantics/bachelorThesis, text |
| Format | application/pdf |
| Rights | info:eu-repo/semantics/openAccess |
| Relation | UPTEC F, 1401-5757 ; 21039 |
Page generated in 0.002 seconds