DETECTION OF REFUELING EVENTS AND PROCESSING OF FUEL LEVEL DATA OF SHUNTING LOCOMOTIVES

Abstract

The paper proposes an algorithm for automatic processing of fuel level telemetry data of shunting diesel locomotives aimed at reliable detection of refueling events and accurate estimation of fuel consumption. Raw fuel level signals obtained from onboard sensors are characterized by a high level of vibration noise, measurement errors, and abnormal spikes caused by operating conditions, which significantly complicates automated analysis. To address this problem, a comprehensive data processing pipeline is developed that combines adaptive anomaly removal, event detection, signal segmentation, and differentiated smoothing.

The algorithm first segments the time series into continuous sessions by detecting large temporal gaps in telemetry in order to prevent distortion of fuel consumption trends. Anomaly filtering is then performed using an adaptive threshold based on local statistical characteristics of the signal. Outliers are identified by comparing deviations from a weakly smoothed trend with a locally computed standard deviation and are replaced using time-aware linear interpolation, which preserves the overall signal shape.

Refueling events are detected based on analysis of the smoothed fuel level derivative. Candidate events are filtered using constraints on minimum volume increase, duration, and locomotive operating state, including engine power and motion indicators. Detected refueling intervals are merged when temporally adjacent, allowing robust identification of complete refueling processes.

A key contribution of the proposed approach is differentiated adaptive smoothing. Strong Gaussian smoothing with a variable window width is applied during normal operation and idling to suppress noise and form a stable fuel consumption trend. In contrast, minimal smoothing is applied within refueling intervals, preserving sharp fuel level rise fronts and ensuring high accuracy of refueled volume estimation. Smoothing parameters are adapted to locomotive speed and current fuel level.

The algorithm was validated on real telemetry data from five shunting locomotives comprising 500,000 measurement points collected over several months of operation. Experimental results demonstrate 100% accuracy in detecting 43 refueling events, with no false positives or missed events. The average signal-to-noise ratio improvement achieved by the proposed processing reached 32.6 dB, confirming the effectiveness of the differentiated smoothing strategy. The developed method can be used for automated fuel accounting, detection of non-target fuel usage, and improving the reliability of operational reporting.