An Integrated Autoencoder–SOM Framework for GPS-Based Driving Behaviour Analysis in Long-Distance Public Transport

Aggressive driving is a major contributor to road accidents, yet its detection remains challenging due to reliance on costly and intrusive sensing technologies, particularly in long-distance public transport operations with varying road conditions. This study proposes a scalable, GPS-based framework for detecting aggressive driving using non-intrusive bus trajectory data. A dataset of 5.9 million GPS records was cleaned and segmented into one-minute intervals, deriving eight descriptive features. Class imbalance caused by predominantly neutral driving was mitigated through resampling. The framework employs an unsupervised Autoencoder–SOM (AESOM) model to encode correlated features into a latent representation and cluster distinct behavioural patterns. Four categories emerged: Smooth/Neutral, Passive, Defensive, and Aggressive, validated through statistical analysis and visualisations. Segment-level classifications were aggregated into a Driving Behaviour Index (DBI), producing a trip-level aggressiveness measure based on weighted cluster assignments. Visualisations and statistical summaries confirm that DBI scores align consistently with acceleration and braking variability patterns across journeys. The proposed framework enables scalable, non-intrusive driver behaviour monitoring for fleet management and road safety applications.