@MASTERSTHESIS\{IMM2012-06457,
    author       = "J. Lozano-Angulo",
    title        = "Detection and One Class Classification of Transient Events in Train Track Noise",
    year         = "2012",
    school       = "Technical University of Denmark, {DTU} Informatics, {E-}mail: reception@imm.dtu.dk",
    address      = "Asmussens Alle, Building 305, {DK-}2800 Kgs. Lyngby, Denmark",
    type         = "",
    note         = "{DTU} supervisor: Jan Larsen, jl@imm.dtu.dk, {DTU} Informatics, and Shashank Chauhan from Brüel and Kj{\ae}r Sound and Vibration Measurement A/S. As co-supervisors, Finn T. Agerkvist from the Electrical Engineering department (Acoustic Technology group) at {DTU} and Karim Haddad from Brüel and Kj{\ae}r Sound and Vibration Measurement A/S.",
    url          = "http://www.imm.dtu.dk/English.aspx",
    abstract     = "The thesis is about detection and one class classification of transient events in train track noise. Two different detection approaches have been designed to
locate impulsive noise events in train track noise data. They make use of a selected set of features to perform the detection of these events. These approaches are novelty detection based approaches and simple threshold based approaches.
The novelty detection approaches take advantage of the abundance of train track noise, containing no transient events, to create a model of normality of the system. To perform detection, they compare any incoming data to the data model by assessing if the incoming data belongs or not to it.
The simple threshold based approaches apply a threshold to a specific set of feature values extracted from incoming data. Where abnormal high feature values indicate the presence of transient events.
Three different data sets have been extracted from a long duration train track noise measurement to create data models and to test and analyse the different proposed detection techniques. The performance of the detectors is studied from two different points of view. The first one is related to the {ROC} curve produced by the detectors using a training data set. The second one is related to the consistency of detection results in different data sets."
}