Data-Driven Modeling of Friction in Drawbead Test Through Advanced Machine Learning

Friction at the drawbead in metal forming operations directly affects the quality of drawpieces. However, identifying the complex effect of friction process parameters on the coefficient of friction (CoF) is difficult based on experimental results. The aim of this paper is to analyze the results of a drawbead simulator test using various machine learning (ML) methods to select the most appropriate algorithm and to analyze in detail the feature importance, permutation importance, and cumulative Shapley additive explanation values of predictors. The test material was DC04 low-carbon steel sheet. Experimental tests were conducted for varying friction process conditions. Of the three different ML algorithms (support vector machine, regression trees, and ensemble tress), the support vector machine (SVM) algorithm with a cubic kernel function provided the lowest root mean square error (0.0085) and the highest correlation coefficient R² (0.9657) for the test data. The predictors in descending order of permutation importance are friction conditions, drawbead height, sample width, Sa of countersamples, and sample orientation. A combined swarm-box chart presenting Shapley values for an SVM model with a cubic kernel function indicates that a low value of the drawbead height predictor has a strong, increasing effect on CoF. However, low values of the remaining explanatory parameters (sample width, mean roughness of countersamples, and sample orientation) have a decreasing effect on CoF.