Logo image
Back
Subway door fault prediction employing stacking ensemble learning
Journal article   Open access   Peer reviewed

Subway door fault prediction employing stacking ensemble learning

Hongkang Song, Shaohu Tang, Jinghui Xia, Liang Zhang, Hailin Kang and Pengyu Li
Scientific reports, v 16(1), 12876
23 Mar 2026
DOI: https://doi.org/10.1038/s41598-026-43371-5
PMID: 41872356
url
https://doi.org/10.1038/s41598-026-43371-5View
Published, Version of Record (VoR) Open Open Access (License Unspecified)

Abstract

Subway door faults Data-driven models Fault prediction Ensemble learning Intelligent transportation systems Predictive maintenance
This study investigates the prediction of metro door failures, which are low-frequency events characterized by severe class imbalance, limited fault samples, and feature redundancy. Traditional machine learning and deep learning approaches face limitations in such scenarios, including strong reliance on feature engineering and limited generalization ability, which hinder the practical application of predictive maintenance. To address these challenges, a stacking-ensemble-based fault prediction method is proposed. The approach first employs a physically constrained data augmentation strategy to expand the sample set while strictly adhering to kinematic consistency. Subsequently, Spearman’s rank correlation coefficient and the variance inflation factor are combined for feature screening, and key variables are selected based on the eXtreme Gradient Boosting (XGBoost) gain. An improved random forest and XGBoost serve as first-level classifiers to output fault probabilities, which are then fused and probabilistically calibrated using logistic regression. Finally, a dynamic threshold optimization strategy based on F1-score maximization is introduced to balance precision and recall. On the test set, the proposed method demonstrates superior overall performance, achieving a receiver operating characteristic area under the curve of 0.977 and a precision–recall area under the curve of 0.913. Under the optimized threshold, the method achieves an accuracy, precision, recall, and F1 score of 0.937, 0.815, 0.810, and 0.812, respectively, outperforming traditional machine learning and deep learning models. SHapley Additive exPlanations analysis confirms that the model decision logic is consistent with physical failure mechanisms. These results validate the effectiveness and practicality of the proposed method for low-frequency, imbalanced scenarios. This study provides a high-precision fault prediction tool for subway door systems and offers a reference technical pathway for intelligent operation and predictive maintenance of key rail transit equipment under sample imbalance conditions, with practical engineering significance for improving train operation reliability and efficiency.

Metrics

1 Record Views

Details

Logo image