This research proposes a lightweight misbehavior detection system for communication-based train control using ensemble learning models. The study evaluates Bagging-based methods, including Random Forest and k-Nearest Neighbors with Bagging, alongside Boosting-based approaches such as AdaBoost, XGBoost, and LightGBM. The models were tested on the CBTCSet dataset, addressing data imbalance and assessing performance based on accuracy, precision, recall, F1-score, testing time, and fit stability to meet real-time CBTC requirements.
The results indicate that Random Forest with the Weighted Imbalance method provides the best balance between detection performance and computational efficiency, achieving 92% accuracy, 92% precision, 92% recall, and a 92% F1-score. The total testing time for 15% of the dataset, consisting of 173,843 data entries, was 12.21 seconds, resulting in an average processing time of 70.23 µs per entry. While other models demonstrated specific advantages, some suffered from overfitting, underfitting, or excessive processing time, limiting their feasibility for real-time deployment.
These findings confirm that Bagging-based models, particularly Random Forest, offer the most effective trade-off between detection accuracy and computational feasibility, making them the most viable choice for real-time CBTC operations to enhance safety and system resilience