Machine learning based nominal root stress calculation model for gears with a progressive curved path of contact

Abstract

The study aims to investigate the possibility of employing machine learning models in the design of non-involute gears. Such a model would be useful for design calculations of non-standard gears, where there are no available guidelines. The aim is to create a decision-support model accompanying the Finite Element Method (FEM) simulations, from which the data for training was collected. Multiple models for numerical prediction were tested, i.e. linear regression, Support Vector Machine, K-nearest neighbour, neural network, AdaBoost, and random forest. The models were firstly validated with N-fold cross-validation. Further validation was done with new FEM simulations. The results from the simulations and the models were in good agreement. The best-performing ones were random forest and AdaBoost. Based on the validation results, a machine learning constructed model for calculating nominal root stress in gears with a progressive curved path of contact is proposed. The model can be used as an alternative to FEM simulations for determining the nominal root stress in real-time, and is able to calculate the stress for gears with different number of teeth, widths, modules, paths of contact, materials, and loads. Therefore, many combinations of gear geometries can be analysed and the most suitable can be chosen.