%Aigaion2 BibTeX export from Knowledge Engineering Publications
%Saturday 05 December 2020 04:20:29 AM

@MASTERSTHESIS{,
    author = {Nair, Sandesh},
     month = oct,
     title = {Predictive Maintenance in a Railway Scenario using One-Class Support Vector Machines},
      year = {2016},
    school = {TU Darmstadt},
  abstract = {In almost any commercial service domain, to be competitive the best and most reliable service must
be offered to attract customers. Commercial trains are no exception to this, with reliability being a
crucial issue. A locomotive failure is also not just restricted to a single train but the whole network is
directly affected. Besides, fixing the component failure later on has added cost overheads of its own. To
avoid all this a predictive maintenance approach must be adopted, so that problems can be identified
before hand, hence minimizing the risk of actual failures. The goal of this thesis is to predict component
failures, specifically power converter failure. For this, we train one-class svm models with multiple
parametrisations exclusively on non-failure cases. These models are then validated by testing on a
hold-out set to determine the accuracy. If good results are achieved then this approach can be further
enhanced to study other types of component failures, with the final aim being developing a prototype
which can be used on a real time basis. The experiments show encouraging results, with up to 25.5 \%
failures predicted correctly with keeping a non-failure prediction rate higher than 90 \%.}
}