Enhancing Railway Safety in Indonesia: A Data-Driven Approach to Track Irregularity Detection using In-Service Train Accelerometers

Abstract

Addressing the critical need for efficient railway track irregularity detection in Indonesia, this article presents a novel data-driven approach for continuous track condition monitoring. By leveraging on-board accelerometer measurements from in-service trains, rigorously validated against traditional Track Recording Vehicle (TRV) data, this work offers a significant advancement over conventional periodic inspections. The methodology uniquely utilizes vibration data from both sides of the train body, enabling precise identification and classification of various track irregularities. Among several evaluated machine learning algorithms, a hyperparameter-tuned Random Forest model demonstrated superior performance, achieving an accuracy of 96.62% and a macro F1-Score of 47.77%. While achieving an overall classification accuracy of 96.62%, the macro F1-Score of 47.77% highlights the challenges posed by the inherent class imbalance in track defect data, where the model performs well at identifying normal track conditions but struggles to detect rare yet critical anomaly classes. Crucially, its high recall for critical irregularities, such as Twist over 3m (40.75%) and Track Gauge (46.10%), is paramount for safety-critical railway applications, effectively minimizing dangerous false negatives and ensuring comprehensive detection of potential hazards. This research highlights the significant potential of integrating on-board accelerometer data with advanced machine learning to enable proactive, cost-effective railway asset management, thereby enhancing operational safety and efficiency.