Development of machine learning algorithms for predictive maintenance of equipment in the automotive industry

Abstract

The aim of this study was to develop and evaluate a machine-learning pipeline for predictive maintenance of industrial equipment in the automotive industry. The empirical basis comprised twelve months of multisensor observations collected from fifteen production lines, including vibration, temperature, pressure, and process-fluid-flow measurements, together with maintenance and failure logs. The proposed framework combined aligned tabular and sequential representations of sensor data and integrated Random Forest, XGBoost, and LSTM predictions through soft-voting ensemble fusion under a time-based validation protocol. The results showed that the ensemble outperformed the best standalone model, increasing Accuracy from 0.93 to 0.95 and F1-score from 0.88 to 0.91, while achieving ROC AUC = 0.96 on the held-out test interval. Pilot deployment in online monitoring mode confirmed the operational relevance of the predictive signal and showed that the system can support threshold-based early warning in real production conditions. The economic assessment further indicated that predictive maintenance based on the proposed pipeline reduced overall maintenance costs by approximately 28% and unplanned downtime by about 41%. These findings suggest that predictive maintenance in automotive production should be interpreted not only as a classification task, but as an integrated decision-support problem requiring  synchronisedheterogeneous data, time-aware evaluation, and deployment-oriented alert logic.