Analysis of machine learning methods efficiency for short-term forecasting of repair labor intensity based on limited telematics data

Abstract

Development and verification of an iterative methodology for short-term forecasting of repair labor intensity for adaptive maintenance planning at the early stages of digital transformation, when only limited initial operating-period data are available (up to 160 thousand km). A comparative analysis of machine learning methods was carried out on a dataset of 100 repair requests: gradient boosting (XGBoost, LightGBM), recurrent neural networks (LSTM, GRU), and fifth-degree polynomial regression. The key efficiency criterion was the forecast error for one scheduled service interval ahead (MAE_ext). The XGBoost algorithm demonstrated the best balance between interpolation accuracy (R²=0.89) and resistance to short-term extrapolation (MAE_ext=0.55), outperforming polynomial models. On this basis, a methodological framework was proposed that relies on the rolling forecasting horizon principle and cyclic retraining. The novelty of the study consists in adapting the online/incremental learning paradigm to the applied problem of fleet service management and in interpreting forecast coefficients as operational risk metrics. The practical significance lies in proposing a scientifically grounded and cautious approach to launching predictive maintenance, minimizing the risks of decisions based on incorrect extrapolation.