A Study on AI-Based Monitoring Methods for Railway Signaling Impedance Bonds

This study investigates the feasibility of applying artificial intelligence (AI) techniques for monitoring railway impedance bonds, which play a critical role in isolating traction return currents from track circuit signals. Four sensing indicators—terminal voltage, traction return current, frequency-domain response, and thermal gradient between the interior and exterior of the bond housing—were used as representative parameters. To reproduce both normal and fault conditions, a Simulink-based simulation model was implemented, generating 400 datasets for AI model training. The generated dataset was utilized to evaluate four representative AI architectures: CNN, LSTM, MLP, and a hybrid CNN-LSTM model. Through 10-fold cross-validation, CNN and CNN-LSTM exhibited the highest accuracy and AUC values, both approaching 1.0, indicating superior classification performance. In contrast, the MLP model, designed as a lightweight baseline, showed limited discriminative power, while the LSTM effectively captured temporal dependencies but required significantly longer training time. These results demonstrate that AI-driven monitoring is applicable to impedance bond condition assessment, complementing conventional threshold-based diagnostic techniques. Moreover, the proposed approach can serve as a foundation for predictive maintenance systems and contribute to enhancing the overall safety and operational reliability of railway signaling infrastructure.