Self-training-based approach with improved XGBoost for aluminum alloy casting quality prediction

Abstract

The revolutionary advances in integrated components in current automotive industry have led to a sharply rising demand for aluminum alloy castings. Targeted quality inspection is thus proposed for components manufacturers to achieve high responsiveness and low operational cost. This suggests casting machine manufacturers to integrate advanced quality prediction functions into the next generation of intelligent casting machines. However, acquiring ample quality inspection data is essential for implementing such functions, which is often challenging, if not infeasible, due to practical issues such as data proprietorship or privacy. Self-training is a good candidate for dealing with scarce labeled data, and XGBoost is commonly used as the base classifier. However, misclassification of unlabeled data happens using XGBoost, which could lead to incorrect pseudo-label assignments, eventually resulting in higher misclassification rate. To address this challenge, a self-training and improved XGBoost-based aluminum alloy casting quality prediction approach is proposed. This approach integrates the classification loss of unlabeled data in the objective function as a new regularization term and considers first and second partial derivatives of the classification loss function for unlabeled data in the leaf node's weight score. The proposed approach penalizes those classification models that misclassify unlabeled data, thereby improves quality prediction performance. To evaluate the effectiveness of our approach, a casting machine manufacturer was collaborated to conduct a case study. The results on three-type casting quality prediction demonstrate that our approach could achieve an accuracy, precision, recall and F1 score of 93.2 %, 90 %, 64.2 %, and 0.75, respectively, outperforming all compared approaches. The approach supports casting machine manufacturers to pre-train a casting quality prediction models with scarce labeled data, enabling swift deployment and customization for targeted quality inspection.