Interpretable AI-Based Railway Communication Security Analysis Using Contextual Text Embeddings and Hybrid Rule-Based Models

Abstract

Modern railway communication systems require secure and reliable data exchange to ensure operational safety and efficiency. Critical processes such as signal transmission, train coordination, and automated control systems continuously generate large volumes of textual and event-driven communication data. Real-time analysis of this data is essential to detect potential threats, anomalies, or security breaches that could disrupt railway operations. However, conventional approaches based on static rule-based mechanisms or manual monitoring are insufficient to handle the complexity and contextual variability of such data streams. Traditional machine learning models, including Decision Tree with Cost Complexity Pruning (DTCCP) and Deep Neural Decision Tree (DNDT), offer some level of interpretability but often fail to capture deep semantic relationships in communication sequences. This limitation results in lower detection accuracy and higher false alarm rates. To overcome these challenges, an interpretable AI-based framework is proposed that integrates contextual text embeddings with hybrid rule-based learning. The system employs Sentence-BERT (SBERT) to generate meaningful semantic representations of railway communication data, capturing contextual dependencies and linguistic nuances. These embeddings are then processed using a RuleFit classifier, which combines linear modeling with decision rule extraction to ensure both predictive accuracy and interpretability. This approach enables the system to learn complex decision patterns while providing transparent explanations for its predictions. Experimental evaluation using accuracy, precision, recall, and F1-score demonstrates that the proposed model outperforms existing methods. It achieves reliable classification of communication data into secure and insecure categories, reduces false positives, and enhances decision-making for railway operators.