Abstract
Anomaly detection of train wheels helps railway operators to find wheel defects and save cost by enabling condition-based maintenance. Existing approaches focus on applications on freight trains and use supervised data-driven methods which need substantial volume of fault data. This is often unavailable in passenger railways that normally provide reliable services. In this regard, this paper reports an unsupervised data-driven approach for anomaly detection of passenger train wheels. Particularly, short-time Fourier transform (STFT) is used to extract time–frequency features from the vibration signal collected with a pair of fiber Bragg grating (FBG) sensors during normal operating hours. Then, four common unsupervised learning algorithms, namely, non-negative matrix factorization (NMF), one-class support vector machine (OC-SVM), multilayer perceptron autoencoder (MLP-AE), and convolutional neural network autoencoder (CNN-AE), are used to derive five health indexes for monitoring the health condition of the train wheels. The proposed workflow is tested against a dataset complied with respect to the wheel turning record. Our results show that the health indexes obtained from the proposed workflow lead to improved correlation with the condition of train wheels compared with existing health metric. In addition, comparison of the four learning algorithms indicates that NMF and MLP-AE outperformed OC-SVM and CNN-AE.
Original language | English |
---|---|
Article number | 106037 |
Journal | Engineering Applications of Artificial Intelligence |
Volume | 122 |
DOIs | |
Publication status | Published - Jun 2023 |
Keywords
- Anomaly detection
- Autoencoders
- Fiber Bragg grating sensor
- Non-negative matrix factorization
- Short-time Fourier transform
- Unsupervised learning
- Wheel defects
ASJC Scopus subject areas
- Control and Systems Engineering
- Artificial Intelligence
- Electrical and Electronic Engineering