Abstract
The utilization of wind resources for the generation of electricity has expanded as a result of the growing efforts to reduce carbon footprints through the global transition into renewable energy sources. However, the incessant failures of wind turbine (WT) gearbox still remain one of the major concerns in the reliable operation of wind power generation plants. The increasing financial implications of these gearbox failures have necessitated a proactive in-silico modelling strategy for investigations to improve its lifecycle. Also, investigations under non-stationary operating conditions experienced realistically still need to be explored in literature. At the same time, past studies mainly focused on individual faults of cracks and pits, with studies yet to be conducted on the occurrence of the two faults in the same tooth. Hence, this study employed an intelligent-finite element approach for failure investigations of this critical sub-system. Adequate knowledge of the current state of the research on the usage of the finite element method (FEM) for failure investigations and the artificial neural network (ANN) for intelligent fault diagnoses was first obtained to identify existing gaps, while motivating a nexus of both techniques. The most prevalent failures experienced by this critical sub-system were also documented and investigated using important failure parameters: time-varying meshing stiffness (TVMS), maximum stress, and maximum strain. With the efficiency of the techniques verified, cracks and pits were induced on the sun gear of the WT gearbox model, which is prone to failure, and were evaluated under dynamic conditions derived from seasonal wind resources in the Eastern Cape.
Failures were recognized with reductions in the TVMS findings, whereas peaks in the simulated sequences for maximum stresses and strains. The healthy and faulty models' results were compared, and their mean percentage deviations (MPD) were estimated. The MPD from the healthy TVMS response of the cracked, pitted, and simultaneous fault TVMS responses were 4.63%, 2.95%, and 6.10%, respectively; the MPD from the max stress responses were 26.15%, 59.08%, and 59.80%, respectively; and the MPD from the max strain responses were 22.90%, 54.15%, and 57.83%, respectively. This indicated that the simultaneous occurrence of both faults had an additive effect on TVMS while pitting failure dominated that of maximum stresses and strains. Finally, considering the healthy and faulty conditions, a long short-term memory (LSTM) network, which is a potent neural network model, was trained to predict the sequences of the studied failure parameters. The seasonal loads were found to have a minor impact on the LSTM network's forecast accuracy. The network predicted TVMS with an overall average accuracy of 89.39%, surpassing maximum stresses and
vi
strains, which were predicted with 75.84% and 75.97% accuracy, respectively. The LSTM network also performed well in forecasting the maximum stress and strain peaks, with accuracy greater than 85% and some attaining 100%. Finally, credible forecasts of the effects of wind resources on critical components can be factored into investment viability prior to wind site development. Failure investigations utilizing intelligent modelling approaches would be useful in the quest to minimize WT gearbox downtimes. Considering the high cost of gear failures, the results from this study give the first level of understanding a potential problem. The ability to understand distinct failure behaviors and predict faults before they reach a catastrophic stage would improve the reliability of WT gearboxes and wind power generation.
Keywords: ANN, Cracks, FEM, Maximum strain, Maximum stress, Pits, TVMS, Wind turbine gearbox.