Surface Quality and Performance Prediction of Spinning Forming 20 Steel Cylindrical Part by Machine Learning Method

The ellipticity, straightness and residual stress of 20 steel cylindrical parts by spinning forming under different thinning rates （10%, 20%, 30%, 40%, 50%） and feed rates （0.55, 0.70, 0.85, 1.00, 1.15, 1.30, 1.45 mm·r⁻¹） were obtained by finite element simulation. The data set was constructed, and machine learning models were established by particle swarm optimization （PSO） algorithm optimizing support vector regression （SVR）, random forest （RF）, gradient boosting decision tree （GBDT）, and extreme gradient boosting tree （XGBoost） algorithms. The prediction effect were compared and analyzed, and the Shapley additive independence feature explanation （SHAP） analysis was conducted. The results show that with the increase of thinning rate, the ellipticity first increased, then decreased and then increased again, the straightness first increased and then decreased, and the residual compressive stress first decreased and then increased. With the increase of feed rate, the ellipticity increased, the straightness increased, and the residual stress did not change significantly. The prediction accuracy of the models for straightness, ellipticity and residual stress from high to low was GBDT model, XGBoost model, RF model and SVR model. The SHAP analysis results of the GBDT model were basically consistent with the actual results, proving the reliability of the GBDT model prediction. The GBDT model predication showed that the spinning process parameters that meet the requirements of straightness less than 0.5 mm, ellipticity less than 0.4 mm and residual stress less than 160 MPa were thinning rate of 40% and feed rate of 0.55−0.90 mm·r⁻¹.