Intelligent Detection of Communication Anomalies in Railway Systems Using Deep Sequential Modeling Paradigms

Abstract

The growing dependence on digital communication systems in railway infrastructure has created a critical need for intelligent methods to ensure communication security and operational reliability. Railway communication data, including logs, alerts, and textual messages, is highly unstructured and contains valuable indicators of system status, making accurate classification a challenging task due to contextual ambiguity and linguistic variability. The problem addressed in this work is the development of an automated system capable of classifying communication data based on the target column which represents two classes: Secure and Not Secure. The traditional system relies entirely on manual analysis, where experts interpret communication data based on experience, which is time-consuming, inconsistent, error-prone, and not scalable for large datasets. These limitations highlight the need for an efficient automated solution. To address this, the proposed system integrates semantic embedding techniques with interpretable machine learning (ML) models. Text data is transformed into dense vector representations using Sentence Bidirectional Encoder Representations from Transformers (SentenceBERT or SBERT) based on the Multi-Purpose Encoder Network (MPNet) architecture, enabling effective capture of contextual and semantic information. These embeddings are used to train multiple classifiers, including Differentiable Neural Decision Trees (DNDT), Decision Tree with Cost Complexity Pruning (Decision Tree CCP), and RuleFit, where RuleFit is selected as the final model due to its ability to combine rule-based reasoning with linear modeling. The proposed system ensures accurate, scalable, and interpretable predictions, thereby enhancing railway communication security monitoring and supporting faster, reliable decision-making in critical environments.