Imbalanced data augmentation for pipeline fault diagnosis: A multi-generator switching adversarial network

Shang, Rou; Dong, Hongli; Wang, Chuang; Chen, Shuangqing; Sun, Tong; Guan, Chuang

doi:10.1016/j.conengprac.2023.105839

Cited by 10 publications

(1 citation statement)

References 32 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…While several methods have been proposed to enhance time series data 12,13 , these methods are more suitable for data with no apparent phase, such as vibration signals, and may not be ideal for turnout data. In literature 14 , proposes a novel multi-generator switching adversarial network (MGSAN) to augment the pipeline data under different health states, especially faulty ones, but GANs showed limited effectiveness in data augmentation for one-dimensional signals. WeiTing Yang et al proposes an interpretable unsupervised machine learning model based on Bayesian networks (BN) as a base model to support process monitoring schemes 15 .…”

Section: Introductionmentioning

confidence: 99%

A fault diagnosis method for railway turnout systems based on Convolution Transformer through data augmentation

Fu,

Yang,

et al. 2024

Preprint

View full text Add to dashboard Cite

In this work, we tackle the challenge of improving fault diagnosis in railway turnout systems (RTS), a critical component of railway infrastructure. Addressing limitations in existing fault-diagnosis techniques, especially their efficacy with small training datasets, we focus on meticulous identification of RTS failure modes. The limited availability of fault samples and the issue of dataset imbalance further compound this challenge. To this end, we propose an innovative fault-diagnosis approach leveraging the Convolution Transformer (ConvTran) combined with advanced data augmentation techniques. This method is uniquely designed to ensure precise fault detection in situations characterized by restricted and skewed data. Existing augmentation techniques often inadequately capitalize on the specific features of time series data. To counteract this, our study proposes an innovative pseudo-synthetic data generation algorithm, utilizing feature combination to effectively overcome the challenges of sample imbalance and small sample sizes. The effectiveness of our model is rigorously tested and validated through comprehensive experiments using an actual dataset from a railway station monitoring system in China. The experimental outcomes decisively demonstrate that our proposed data augmentation method significantly improves the precision of turnout fault diagnosis, especially in contexts with limited sample sizes and imbalanced data. Impressively, our approach achieves a 92% accuracy rate in fault classification on small datasets, which further escalates to 99.7% following data augmentation. This exceptional performance not only outstrips that of other benchmark models but also underscores the superior efficiency and practical applicability of our proposed methodology in the field of RTS fault diagnosis.

show abstract