Abstract
The article presents the results of the pilot implementation of the predictive equipment maintenance system at the SaryarkaAvtoProm LLP automotive industry enterprise. The system is based on an ensemble of machine learning algorithms trained on multichannel sensor data (vibration, temperature, pressure, flow rate of process fluids) and failure logs. The prototype of the online monitoring system was integrated with two priority assembly lines of the machine assembly shop (engine and transmission units). For two months, the system analyzed sensor data streams in real time and generated warnings when the failure probability threshold of 0.7 was exceeded. During this period, 14 alarm notifications were generated, nine of which corresponded to the actual pre-failure conditions of the equipment, and in five cases pronounced signs of wear were detected, eliminated as part of scheduled repairs. It is shown that the introduction of a predictive maintenance system has reduced total maintenance costs from 1,000,000 to 720,000 conventional units (savings of about 28%) and reduced the duration of unplanned downtime from 340 to 200 hours (a decrease of ≈41%). The economic impact factors are analyzed: advance planning of repairs, reduction of the number of emergency situations, optimization of stocks and reduction of the volume of emergency purchases. The pilot implemented sensor-data acquisition and preprocessing, ensemble-based risk estimation, an alert interface, and feature-contribution explanations. Digital twins and federated learning were not part of the experimental system and are considered only as directions for future scaling.