Implementación de un módulo inteligente para la detección de robo de combustible en vehículos de carga integrado a la plataforma web FuelControl

Abstract

Fuel theft in cargo vehicles represents a critical challenge for the Colombian logistics sector, generating significant economic losses and affecting operational efficiency. Traditional monitoring systems, based on fixed thresholds, are often ineffective at detecting subtle methods such as "siphoning" and generate a high volume of false positives due to natural sensor fluctuations. This thesis presents the development and implementation of an intelligent anomaly detection module, integrated into the FuelControl web platform, for the precise identification of fuel theft in cargo fleets. Following the CRISP-DM methodology, 733,022 telematic records from four vehicles over six months were analyzed. A robust preprocessing pipeline was implemented, including trajectory segmentation, calculation of vehicle-specific adaptive thresholds using Median Absolute Deviation (MAD), and feature engineering that integrated contextual variables (spatial, temporal, and behavioral). Machine learning models were trained and evaluated, notably a calibrated Random Forest that achieved a precision of 65.8% with a false positive rate of 5%, outperforming the rule-based system by 43.7%. The strategy of employing specialized models by temporal theft pattern (short, extended, and post-journey) enabled improvements of up to 275% in recall. In a pilot test, the system demonstrated an 83% reduction in alert volume and a 164% increase in the detection of actual thefts, confirming its operational utility for efficient and secure fleet management.