Phemelo Thapedi Selomane is a Mechanical Engineer and MEng researcher at the University of Pretoria. He completed a BEng in Mechanical Engineering from 2021 to 2024 and a BEng (Hons) in Mechanical Engineering in 2025, and is now completing an MEng. His research focuses on mechanised mining systems, particularly conveyor modelling and monitoring for condition-based maintenance.
Research focus
Mining conveyors are critical to continuous material transport. Mechanical degradation can affect production continuity, efficiency and maintenance needs. Their behaviour depends on interacting factors such as belt tension, friction, rotational speed, alignment and transient dynamics. Conventional condition-monitoring methods can identify deviations from normal operation, but may not reveal their physical causes. Phemelo’s research investigates a digital-twin framework that links a physics-based model with measured responses and data-driven methods to interpret changes in conveyor behaviour.
Methods and current progress
The work combines physics-based modelling, simulation, experimental measurement and data-driven inference. A longitudinal conveyor driveline model has been developed in MATLAB/Simulink using Simscape Driveline, and its behaviour is evaluated against analytical relationships where appropriate. Planned experimental validation will measure pulley rotational speed, belt speed, belt tension and motor electrical energy or power. Controlled, non-permanent perturbations on a conveyor test rig are intended to produce varied operating conditions without relying on payload variation as the main excitation source.
The data-driven work investigates variational autoencoders for healthy-state anomaly detection, supervised neural networks for fault classification, and inverse neural networks for estimating physical parameters from measured responses. A key question is whether the approach can remain useful when the dynamic model or its parameter values differ from those represented during training.
The longitudinal model has been validated, while development of a three-dimensional multibody representation continues. Preliminary results from synthetic degradation scenarios show that diagnostic performance depends strongly on model architecture: a feed-forward classifier achieved approximately 67% accuracy and a one-dimensional convolutional neural network approximately 99%; support-vector machine and logistic-regression approaches reached approximately 100% on the evaluated synthetic test data. A healthy-only variational autoencoder achieved approximately 91% healthy-state detection, with lower sensitivity to wear and early bearing-fault cases. These results are simulation-based and have not yet been validated with measured conveyor data; they should not be interpreted as field performance.
Applications and next steps
The work may support condition monitoring, maintenance decisions and interpretation of conveyor behaviour. A validated digital twin could provide a physics-based reference, while machine-learning components could assist with anomaly detection, fault classification and estimating changes in physical parameters. Current limitations include the realism and range of simulated degradation scenarios, measurement noise and the need for experimental validation. Next steps include completing model validation, refining the three-dimensional representation, evaluating the methods with measured data and testing generalisation across unseen conditions and model structures.