In April 2015, the number of operating High Speed Trains (HSTs) in the world has reached 3603. An efficient, effective and very reliable braking system is evidently very critical for trains running at a speed around 300 km/h. Failure of a highly reliable braking system is a rare event and, consequently, informative recorded data on fault conditions are scarce. This renders the fault detection problem a classification problem with highly unbalanced data. In this paper, a Support Vector Machine (SVM) framework, including feature selection, feature vector selection, model construction and decision boundary optimization, is proposed for tackling this problem. Feature vector selection can largely reduce the data size and, thus, the computational burden. The constructed model is a modified version of the least square SVM, in which a higher cost is assigned to the error of classification of faulty conditions than the error of classification of normal conditions. The proposed framework is successfully validated on a number of public unbalanced datasets. Then, it is applied for the fault detection of braking systems in HST: in comparison with several SVM approaches for unbalanced datasets, the proposed framework gives better results.
A SVM framework for fault detection of the braking system in a high speed train
Zio, Enrico
2017-01-01
Abstract
In April 2015, the number of operating High Speed Trains (HSTs) in the world has reached 3603. An efficient, effective and very reliable braking system is evidently very critical for trains running at a speed around 300 km/h. Failure of a highly reliable braking system is a rare event and, consequently, informative recorded data on fault conditions are scarce. This renders the fault detection problem a classification problem with highly unbalanced data. In this paper, a Support Vector Machine (SVM) framework, including feature selection, feature vector selection, model construction and decision boundary optimization, is proposed for tackling this problem. Feature vector selection can largely reduce the data size and, thus, the computational burden. The constructed model is a modified version of the least square SVM, in which a higher cost is assigned to the error of classification of faulty conditions than the error of classification of normal conditions. The proposed framework is successfully validated on a number of public unbalanced datasets. Then, it is applied for the fault detection of braking systems in HST: in comparison with several SVM approaches for unbalanced datasets, the proposed framework gives better results.File | Dimensione | Formato | |
---|---|---|---|
alstrom with FS FVS and FVC - final.pdf
accesso aperto
Descrizione: articolo
:
Post-Print (DRAFT o Author’s Accepted Manuscript-AAM)
Dimensione
740.27 kB
Formato
Adobe PDF
|
740.27 kB | Adobe PDF | Visualizza/Apri |
alstrom with FS FUS and FVC.pdf
Accesso riservato
:
Publisher’s version
Dimensione
532.85 kB
Formato
Adobe PDF
|
532.85 kB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.