Estimating Weight of Pressed Carbide Inserts using Machine Learning

Abstract: In any mass-scale production pipeline, even the slightest efficiency improvement can have massive beneficial implications in terms of cost and time. Therefore, at Sandvik Coromant there is a drive to innovate and implement new technologies that can be integrated into their carbide insert production line. This work has investigated the possibility of using machine learning to reduce the time of one of the steps in this production line, namely the weighing of the pressed carbide insert. By utilizing sensor data obtained during pressing, an earlier study demonstrated that highly accurate estimations could be achieved for a certain press and insert, with RMSE of 0.0015 and R2 of 0.950. The aim of our report was to see if those results could be extrapolated to any press and any insert, as well as if the estimations could be further improved by using regularization and different training set sizes. Four different variations of algorithm and training/test split were used: OLS multiple regression with 50 data points for training (the same used in the previous study) and using 25% of the data for training, and ElasticNet with the same setup. The results showed that ElasticNet with 25% training data had the highest recorded average RMSE and R2 of 0.0022 and 0.760, respectively. It was thus concluded that the previous study’s results could not be fully replicated en-masse. ANOVA tests showed that using more training data had significant improvement on average model performance, while the inclusion of regularization was shown to be non-significant.