Prediction method for spiral bevel gear tooth surface roughness based on machine learning

ObjectiveTo address the issues of parameter optimization and prediction accuracy of the surface roughness of spiral bevel gears, and to overcome the limitations of traditional methods in effectively handling complex nonlinear relationships and the influence of multiple variables, machine learning model is used to predict the surface roughness of the spiral bevel gears.MethodsFirstly, based on the spiral bevel gear grinding test dataset, decision tree (DT), support vector machine (SVR), and artificial neural network (ANN) methods are applied to construct roughness prediction models for the convex and concave surfaces of spiral bevel gears, and the prediction results of the three machine learning models are compared. Secondly, based on this, a multiple linear regression method is used to derive a tooth surface roughness prediction formula that considers the processing parameters of spiral bevel gears. Finally, the contribution of each input feature to the predicted tooth surface roughness is analyzed using the machine learning model explanation tool (SHAP), providing theoretical support for the application of machine learning in high-performance gear manufacturing.ResultsThe results show that the DT and SVR exhibit underfitting and overfitting, respectively, leading to poor prediction performance. The ANN accurately fits the data and predicts tooth surface roughness with high precision, but its computational efficiency is relatively low. The average relative errors in predicting the roughness of the convex and concave surfaces of the spiral bevel gear are 3.5％ and 6.09％, respectively. The influence of the input processing parameters on the predicted tooth surface roughness follows the order of grinding speed, grinding depth, and generating speed.