Development of an intelligent monitoring system for the technical condition of industrial equipment based on machine learning methods and streaming data analysis

Abstract

The operation of industrial equipment is accompanied by a continuous stream of heterogeneous telemetry measurements whose information capacity far exceeds the capabilities of threshold control based on fixed setpoints. The discrepancy between the volume of recorded data and the depth of its interpretation produces a gap: a failure is often recognized only after a unit has entered a pre-failure state, whereas the leading signs of degradation are present in the signal considerably earlier. The aim of the work is to construct an intelligent monitoring system that reconciles stream processing of measurements with trainable models of state recognition and remaining useful life prediction. An author's computational model has been implemented: a model telemetry dataset of a rotating unit across five measurement channels was formed, on which state classification algorithms (logistic regression, random forest, gradient boosting, a GRU recurrent network) and a regression forecast of the residual resource were compared, and the characteristics of real-time stream processing and sensitivity to distribution drift were assessed. Gradient boosting provided an F1 of 0.914 at a false alarm rate of 3.8%, while the recurrent architecture anticipated the moment of failure on average by 41.6 h with a median detection latency of 0.67 s. The response of quality to drift proved nonlinear: a moderate feature shift reduced F1 by 6.1 percentage points, whereas recalibration over a sliding window restored the metric to 0.881. The leading character of vibration and current features relative to temperature ones means that early recognition of degradation is achieved by matching the forecast horizon to the inertia of the measurement channels rather than by increasing the number of sensors.