ANALYSIS AND PROSPECTS OF ADAPTATION OF EXPERIENCE IN APPLYING MACHINE LEARNING IN THE SYSTEM OF TECHNICAL MAINTENANCE AND REPAIR OF CITY BUSES

In the context of accelerated digitalization of automotive transportation, machine learning methods are forming the foundation for transitioning from outdated reactive and preventive maintenance strategies to modern predictive approaches. The aim of this work is to systematize information and analyze the experience of applying machine learning methods in the maintenance and repair systems of vehicles, with the subsequent identification of directions for adapting these methods to the specific challenges of urban bus fleets. The research methodology is based on the analysis of scientific and technical literature using formalized selection criteria, a comparative analysis of the efficiency metrics of ML architectures, and the synthesis of the obtained data in order to assess their applicability and develop a conceptual framework for adapting predictive models to the operating conditions of city buses. A systematic review of scientific publications and industrial solutions on the application of ML models for diagnosing and forecasting the technical condition of vehicles during the 2020–2025 period has been conducted. Particular attention is paid to a comparative assessment of the effectiveness of classical algorithms (Random Forest, XGBoost, LightGBM) and deep architectures (LSTM, CNN, hybrid models) when processing data obtained from onboard systems. It has been established that hybrid CNN-LSTM models demonstrate the highest diagnostic accuracy (up to 99.02%) in predicting electric motor failures, outperforming both classical methods and individual deep architectures. The economic benefits of implementing predictive systems include a reduction in operational costs by 10–40%, a decrease in vehicle downtime by up to 50%, and an improvement in overall fleet reliability. However, the analysis has revealed key limitations to the practical application of ML solutions: imbalanced training datasets due to the rarity of failures, high sensor inaccuracies (exceeding 40% for certain parameters), interpretability issues associated with deep learning “black boxes,” and integration challenges with existing ERP/CMMS systems. The necessity of developing Explainable AI (XAI) methods to increase specialists’ trust in predictions is substantiated. The scientific novelty of the work lies in the development of a conceptual framework for adapting ML models to the cyclic operation mode of buses, which includes incorporating geospatial data, road conditions, and requirements for forecast interpretability. Based on this, directions for future research are proposed to enhance the accuracy of residual life prediction for critical bus components.