Design and implementation of a TINML-based predictive maintenance system for ball mill condition monitoring

Abstract

Ball mills are critical, high-value assets in mineral processing plants, and their unplanned failure causes costly downtime and safety hazards, particularly for the small-scale mining operations common in Zimbabwe that lack access to cloud-based monitoring infrastructure. This project presents the design and implementation of a low-cost, TinyML-based predictive maintenance system for ball mill condition monitoring that performs fault classification and severity assessment directly on an embedded microcontroller. A Raspberry Pi Pico W acquires vibration, temperature, current, and voltage data, from which thirteen time- and frequency-domain features are extracted. A Random Forest classifier, trained offline and exported to run on the device, distinguishes normal operation from imbalance, looseness, and overheating faults, while rule-based thresholds detect electrical and thermal anomalies. The model achieved 91.98 percent cross-validated accuracy and a 91.65 percent macro F1-score while using only 16.7 percent of the available flash memory and inferring in real time. Telemetry is published to the Adafruit IO cloud platform for remote monitoring, and local light-emitting diodes and a liquid crystal display provide on-site alerts. The results confirm that accurate, edge-based predictive maintenance is feasible on resource-constrained, low-cost hardware.