Machine Learning-Based Detection of True Refuelling Events in Motorcycle Fuel System

Abstract

This thesis presents a machine learning-based approach for accurately detecting true refuelling events in motorcycle fuel systems, addressing the limitations of traditional threshold-based methods. Sensor inaccuracies and dynamic riding conditions often lead to false refuelling detections, compromising the reliability of fuel range estimations. To overcome this, a supervised classification model was developed using real-world sensor data, including fuel volume, wheel speed, lean angle, pitch angle, and fuel consumption. The data underwent extensive preprocessing, feature selection, and manual labeling to ensure model robustness. A Coarse Decision Tree classifier was selected for deployment based on its high accuracy and computational efficiency. The model was integrated into the Dashboard Control Unit (DCU) using Simulink and deployed on embedded hardware. Multiple integration strategies were tested, with the final implementation combining machine learning output and control logic within a state machine to ensure real-time performance and reliability. System-level testing demonstrated significant improvements in filtering out false positives and correctly identifying true refuelling events. Additional features, such as delta fuel rate, were introduced to enhance generalizability across different motorcycle models. The results confirm the effectiveness of machine learning in improving fuel monitoring systems, offering a scalable and intelligent solution for anomaly detection in automotive applications.

Date of Award	2025
Original language	English
Supervisor	Harald Kirchsteiger (Supervisor)