The study objective is to develop applied approaches to predictive analytics of automobile fleet maintenance based on the selection of models and algorithms, as well as the description of the learning contour and model settings to achieve sustainable forecasting accuracy. The task to which the paper is devoted is to form applied approaches for predictive analytics of automobile fleet maintenance with a focus on the selection of models and algorithms, training on the prepared data, and parameter tuning to achieve sustainable forecasting accuracy. Research methods. Forecasting, mathematical modeling, statistical analysis, system analysis, reliability theory, probability theory, management theory. The novelty of the work includes theoretical and applied approaches to predictive analytics of automobile fleet maintenance based on the following sequence "problem statement → correct time sampling → model selection → training and configuration → maintenance implementation". Study results. It is shown that the specifics of the fleet (heterogeneity of operating modes, imbalanced events, censoring of observations, planned replacement vs failure competition) require the correct formulation of target events, time validation, calibration of probabilities and the transfer of forecasts into solutions taking into account the cost of errors. Conclusions: The technique "data → model → solution" is proposed, which includes interpreted time-to-event analysis models and high-precision models based on tabular features (gradient boosting), as well as practical schemes for implementing and monitoring quality in operation.
Zayats et al. (Wed,) studied this question.