A comprehensive methodology for processing fuel–level telemetry dat of shunting locomotives: nonlinear noise analysis and refueling event detection

Abstract

Analysis of recent research and publications. The authors' previous work proposed an algorithm for time–series segmentation [1], differentiated adaptive smoothing with refueling detection [2], and classification of nonlinear noise by spectral and statistical type [3]. Each was validated on its own sample, and the noise classification is not accompanied by any assessment of noise–removal effectiveness, which motivates merging these results into a single methodology.

Purpose of the study. To combine three previously published algorithms into a single fuel–level telemetry processing pipeline and to validate their joint operation on a common sample, with independent validation via a fuel–balance model.

Presentation of the main material. Three previously published algorithms were combined into a single pipeline, with a formalized confidence score added for every detected refueling event, accounting for the linearity and monotonicity of the level rise, level stability around the event, and a movement penalty. Noise classification was extended to the full segment sample, and four noise–removal methods were assessed by signal–to–noise gain. Validation used six locomotives: 280 refueling events at a mean confidence of 0.83, and 27,792 segments classified.

Conclusions. Colored noise predominates over white and the Gaussian type over the rest; the drift type that dominated in [3] was not observed in this sample. The moving average and Kalman filter give a stable gain of 1.3–4.2 dB, and the hybrid filter stays within 1.3–2.0 dB, whereas the formally highest wavelet figure reflects a property of the metric, which rewards the smallest change to the input noise. The balance model shows moderate linear agreement with the actual hourly consumption (R² = 0.19–0.52) for five of six locomotives — a control rather than a predictive tool.