Logo image
Hybrid machine learning approaches for performance modelling of mechanical systems
Dissertation   Open access

Hybrid machine learning approaches for performance modelling of mechanical systems

Kantu Thomas Kabengele
Doctor of Philosophy (PHD), University of Johannesburg
2025
Handle:
https://hdl.handle.net/10210/520827

Abstract

Mechanical and energy systems such as fused deposition modelling (FDM) 3D printing, combined-cycle plants and solar-hybrids exhibit nonlinear, interaction-rich behaviour that challenges both first-principles models and ad-hoc machine learning (ML). Results for artificial neural network (ANN) or adaptive neuro-fuzzy inference system (ANFIS) and metaheuristic hybrids with genetic algorithm (GA), particle swarm optimisation (PSO), antlion optimisation (ALO), and grey wolf optimisation (GWO) are promising yet fragmented, with inconsistent data handling, limited uncertainty reporting and little guidance on compute–accuracy trade-offs. This thesis proposes a general, reproducible framework for data-driven prediction of nonlinear mechanical systems. It enforces leakage-safe data practice, compares ANN/ANFIS and hybrids under a common rubric, integrates uncertainty and physics-consistency diagnostics, and reports cost–accuracy frontiers for deployment decisions. The framework is comprised of four pillars: 1) Protocol: audited datasets, group/time-aware splits, candidate set (ANN, ANFIS, + GA/PSO/ALO/GWO), fixed seeds and hyperparameter logs; 2) Evaluation: unified regression value or coefficient of correlation (R or R²), mean square error (MSE) or root mean square error (RMSE) with 95% confidence intervals (CIs) residuals and calibration; 3) Interpretability and Physics: global/local sensitivity, ANFIS rule/surface inspection, monotonicity checks (e.g., net power vs ambient temperature and condenser vacuum pressure) and 4) Cost– Accuracy: wall-time, iterations, model size/latency and Pareto charts to justify (or reject) hybridisation. The playbook is exercised on four systems, as follows: FDM 3D printing, combined cycle power plant (CCPP), Parabolic-dish heater and Integrated solar combined cycle power plant (ISCCPP). Each follows the same template: leakage-safe split; strong baselines (ANN with Levenberg-Marquardt [LM]/Bayesian regularisation [BR]/Scaled conjugate gradient [SCG], classical ANFIS); then hybrids with transparent settings; standardised reports (train/validation/test/overall), residuals, sensitivity, and compute profiles...
pdf
Kabengele_Corrected Final version4.42 MBDownloadView
Open Access

Metrics

1 Record Views

Details

Logo image