An unsupervised method based on convolutional variational auto-encoder and anomaly detection algorithms for light rail squat localization

•An automatic light rail squat localization approach via axle box accelerations.•A convolutional variational auto-encoder for unsupervised feature extraction.•Anomaly detection algorithms for automatic light rail squat detection.•Effects of various factors on detection performance of the proposed me...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Construction & building materials 2021-12, Vol.313, p.125563, Article 125563
Hauptverfasser: Yuan, Zhandong, Zhu, Shengyang, Chang, Chao, Yuan, Xuancheng, Zhang, Qinglai, Zhai, Wanming
Format: Artikel
Sprache:eng
Schlagworte:
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:•An automatic light rail squat localization approach via axle box accelerations.•A convolutional variational auto-encoder for unsupervised feature extraction.•Anomaly detection algorithms for automatic light rail squat detection.•Effects of various factors on detection performance of the proposed method. As a typical short-wave rolling contact fatigue defect of rail surface, squats accelerate the degradation of the railway infrastructure. Early detection of rail squats contributes greatly to the maintenance of the railway system. This article proposes an automatic detection algorithm for light rail squat localization via the vehicle axle box acceleration signal. A convolutional variational auto-encoder, named CVAE, is tailored for feature extraction in an unsupervised manner. Then, two anomaly detection methods, namely elliptic envelope and one-class support vector machine (OCSVM) are employed as alternatives for rail squat detection. The validity and reliability of the proposed approach are verified by vehicle-track coupled dynamics simulation, where squats are considered as an irregularity excitation. And an in-lab experiment is conducted to further validate the effectiveness of the method. Comparative studies are conducted to investigate the influence of squat depth, feature extraction method, noise level and vehicle speed on the algorithm. The results demonstrate the elliptic envelope method is a better option for squat detection which can be explained by the consistency of the property of features extracted by CVAE and input data requirements of the elliptic envelope method. Furthermore, compared to the convolutional auto-encoder, CVAE is a better unsupervised feature extraction method because features extracted by CVAE are more sensitive to light squats. And it’s verified that the CVAE-elliptic envelope method is robust over signal noise and variability of vehicle speed.
ISSN:0950-0618
1879-0526
DOI:10.1016/j.conbuildmat.2021.125563