DATA SEGMENTATION ALGORITHM FOR VEHICLE FUEL LEVEL MEASUREMENTS

Abstract

This paper presents a data segmentation algorithm designed for analyzing fuel level measurements in vehicles. The proposed approach aims to identify and classify noise caused by sensor inaccuracies, fuel surface oscillations, vibrations, and other external factors that complicate signal interpretation. To achieve this, the algorithm employs a combination of smoothing, Z-score analysis, the sum of absolute changes, and linear approximation techniques, which enable the detection of homogeneous segments within the data. These segments correspond to different operating modes of a vehicle, such as refueling, motion, or idling.

The preprocessing stage includes the removal of non-relevant points and the application of a Gaussian smoothing filter, optimized with a 200-point window. This step reduces short-term fluctuations while preserving significant variations. Subsequently, the segmentation process relies on calculating Z-scores and the sum of absolute changes within overlapping windows to identify potential transition points between modes. Threshold-based selection ensures robustness against insignificant fluctuations, preventing over-fragmentation of the dataset. Detected segments are further classified through linear approximation, which determines the slope and trend type (increasing, decreasing, or stable). Similar neighboring segments are merged to reduce redundancy while maintaining accuracy.

Experimental validation on datasets containing up to one million points demonstrates the reliability of the algorithm. The segmentation results were compared with visual (manual) assessments, showing a high degree of consistency. Notably, the method accurately detects refueling events and effectively distinguishes between noisy motion data and stable idle periods. Statistical analysis confirms that the algorithm provides minimal deviation from empirical boundaries, particularly for stable and decreasing trends, while intentionally extending growing segments to avoid misclassification.

The proposed algorithm proves adaptive, efficient, and resilient to common types of noise. It enables precise division of large datasets into meaningful segments, facilitating subsequent noise analysis and the development of advanced denoising techniques. This, in turn, contributes to improving the accuracy of fuel consumption monitoring and optimization of vehicle fuel systems.