Fault Prediction of On-Board Train Control Equipment Using a CGAN-Enhanced XGBoost Method with Unbalanced Samples

Abstract

On-board train control equipment is an important component of the Train Control System (TCS) of railway trains. In order to guarantee the safe and efficient operation of the railway system, Predictive Maintenance (PdM) is significantly required. The operation data of the on-board equipment allow us to build fault prediction models using a data-driven approach. However, the problem of unbalanced fault samples makes it difficult to achieve the expected modeling performance. In this paper, a Conditional Generative Adversarial Network (CGAN) is adopted to solve the unbalancing problem by generating synthetic samples corresponding to specific fault labels that belong to the minority classes. With this basis, a CGAN-enhanced eXtreme Gradient Boosting (XGBoost) solution is presented for training the fault prediction models. From the pre-processing to the field data, artificial fault samples are generated and integrated into the training sample sets, and the XGBoost models can be derived with multiple decision trees. Both the feature importance sequence list and the knowledge graph are derived to describe the characteristics obtained by the models. Filed data sets from practical operation are utilized to validate the proposed solution. By comparison with conventional machine learning algorithms, it can be found that higher accuracy, precision, recall, and F1 scores, which are up to 99.76%, can be achieved by the proposed solution. By involving the CGAN strategy, the maximum enhancement to the F1 score with the XGBoost approach reaches 6.13%. The advantages of the proposed solution show great potential in implementing equipment health management and intelligent condition-based maintenance.

Fuzzy Few-Shot Learning for Predictive Maintenance and Cost-Benefit Analysis in Semiconductor Manufacturing: A Case Study on Photomask Haze

“…Many existing PdM models are based on machine learning or deep learning methods [21], which generally require a large amount of failure data for training aimed at estimating the RUL or predicting impending anomalies [22]. Data collection is hindered by a severe imbalance between normal data (abundant) and abnormal data (scarce), which can render conventional machine learning or deep learning methods ineffective [23]. Several researchers have proposed the pre-training of deep learning models or the use of transfer learning to reduce data size requirements [24]; however, engineers have found that the inference process lacks interpretability.…”

Prediction of Maternity Recovery Rate of Group Long-Term Disability Insurance Using XGBoost

“…This combination allows XGBoost to achieve excellent model performance and high-speed processing. Liu et al (2023) utilized XGBoost to enhance the predicted outcomes developed from conventional machine learning algorithms and resulted in an increased F1-score of 6.13%. On imbalanced dataset such as personal credit evaluation, XGBoost performed better than the other tree-based models and logistic regression (Li et al, 2020).…”