Conference Agenda

Urban railways are critical logistics systems that move large passenger flows through time-sensitive and capacity-constrained networks. When disruptions occur, their effects can propagate rapidly across stations, trains, operators, and passenger journeys, requiring coordinated decisions among operations, maintenance, and customer-service teams. For advanced logistics management, disruption records should therefore support not only retrospective reporting but also evidence-based severity assessment, operational learning, and future integration with simulation or digital-twin environments. However, in Thailand, rail disruption narratives are often recorded as mixed Thai–English free text with inconsistent spacing, abbreviations, and operator-specific expressions, limiting the usefulness of structured incident codes alone for large-scale analytics.

This study proposes a machine-learning-based natural language processing workflow for classifying disruption severity using daily records from the Department of Rail Transport dataset during 2021–2025 (N = 770). Severity labels were deterministically derived from delay minutes and grouped into three operational classes: A for delays below 5 minutes, B for delays from 5 to below 15 minutes, and S for delays of 15 minutes or more. These labels are interpreted as a proxy for passenger and operational impact. The DETAIL field was used as the main narrative input. Text preprocessing included removal of repeated spaces and Thai word segmentation using PyThaiNLP with a custom dictionary developed from station names, station codes, and disruption-cause terms to improve recognition of railway-specific expressions in mixed Thai–English text.

The experiment compared four text feature representations: Bag-of-Words, TF-IDF, Trigram features, and WangchanBERTa sentence embeddings. Each representation was evaluated with Logistic Regression and XGBoost. Two validation settings were used. A stratified split divided the data into 70% training, 20% validation, and 10% testing while preserving severity-class proportions. A chronological split used the same proportion but sorted records by date to evaluate temporal generalization for future deployment.

The dataset was imbalanced, with Class S accounting for 429 records (55.71%), Class B for 268 records (34.81%), and Class A for 73 records (9.48%). The average delay was 30.02 minutes, with a median of 15 minutes and a maximum of 4,320 minutes. Under the stratified split, BoW + XGBoost achieved the best balanced performance, with accuracy = 0.870 and F1-score = 0.859. BoW + Logistic Regression also performed strongly, with accuracy = 0.844 and F1-score = 0.838. Under the chronological split, TF-IDF + XGBoost achieved the highest accuracy of 0.857, while WangchanBERTa embeddings + XGBoost achieved the highest F1-score of 0.607. However, the chronological test set did not contain Class A examples; therefore, chronological results must be interpreted carefully and should be considered together with the stratified-split results.

These findings show that mixed Thai–English disruption narratives contain meaningful severity signals and that lightweight, interpretable NLP models can provide a practical baseline for disruption monitoring, severity validation, escalation support, and future rail logistics decision-support to an operator or digital-twin applications.