Vibration-based damage detection of rail fastener clip using convolutional neural network: Experiment and simulation. (January 2021)

• Record Type: Journal Article
• Authors: Yuan, Zhandong; Zhu, Shengyang; Yuan, Xuancheng; Zhai, Wanming
• Is Part Of: Engineering failure analysis. Volume 119(2021)
• Publication Date: 2021-01
• Subjects: Rail fastener clip -- Damage detection -- Convolutional neural network -- Vehicle-track coupled dynamics; System failures (Engineering) -- Periodicals; Fracture mechanics -- Periodicals; Reliability (Engineering) -- Periodicals; Pannes -- Périodiques; Rupture, Mécanique de la -- Périodiques; Fiabilité -- Périodiques; Fracture mechanics; Reliability (Engineering); System failures (Engineering); Periodicals; Electronic journals; 620.112
• DOI: 10.1016/j.engfailanal.2020.104906
• Languages: English
• ISSNs: 1350-6307

Abstract

Highlights: A one dimensional CNN is designed to learn optimal damage-sensitive features and identify the health condition of rail fastener clips automatically. Repeated impact tests are conducted on the track system under different health conditions of fastener clips in laboratory. Parametric studies are performed on the number of convolution blocks, location of sensor and robustness to noise level. A modified vehicle-track coupled dynamics model with a detailed fastening dynamics model is established to generate numerical datasets. Abstract: With the rapid development of rail transportation, health monitoring of railway track structure becomes a challenging problem. In this work, a novel and efficient approach is proposed to carry out damage detection of fastener clips using one dimensional convolutional neural network (CNN). A one dimensional CNN is designed to learn optimal damage-sensitive features from the raw acceleration response and identify the health condition of rail fastener clips automatically. Two case studies are implemented experimentally and numerically to validate its feasibility. First, repeated impact tests are conducted on the track system under different health conditions of fastener clips in laboratory. The time-domain data recorded by accelerometers on the rail are employed for the CNN training and evaluation. Parametric studies are performed on the number of convolution blocks, location of sensor and robustness to noise level. It is found that the CNN achieves a high detecting accuracy and good robustness. Furthermore, in order to collect rail response induced by the passing train under variational clip health condition, a modified vehicle-track coupled dynamics model is established to generate numerical datasets of the rail vertical acceleration under different damage scenarios of the fastener clips. Thereafter, the CNN is trained and evaluated on the numerical datasets, showing a high detection accuracy. Finally, the t-distribution stochastic neighbor embedding (t-SNE) technique is applied to manifest the super feature extraction capability of CNN.