Software implementation of an intelligent resource management system for a robotics enterprise based on predictive machine learning models

Abstract

Resource management at a robotics enterprise is complicated by the fact that the condition of the manipulator fleet, the demand for spare parts and the workload of maintenance crews change nonlinearly and remain poorly predictable when calendar regulations and reactive responses to failures are relied upon. The discrepancy between the actual operating time of drive units and the normative inter-service intervals leads either to premature replacement of serviceable components or to unscheduled stoppages of technological stations. The aim of the work is to construct and verify in software an intelligent control loop in which decisions on maintenance, stock replenishment and personnel allocation are made on the basis of forecasts of remaining useful life and failure probability. Using model data for a fleet of 184 robotic stations, four forecasting algorithms were compared; gradient boosting yielded a mean absolute error of remaining-life estimation of 8,7 days with a coefficient of determination of 0.901. Switching control to the predictive loop raised the technical utilization coefficient from 0.782 to 0.861 and reduced unscheduled downtime by 50,6%, with the gain fading once predictive maintenance covers roughly three quarters of the fleet. The control result is achieved by reconfiguring the thresholds of telemetry already being collected rather than by increasing the dimensionality of the forecasting model.