Research on Manufacturing Defect Prediction Based on XGBoost Model and Multi-source Data Fusion, along with Model Interpretability Analysis

In modern manufacturing, predicting and preventing defects is crucial. Traditional methods fall short due to inefficiency and high costs. This paper employs the XGBoost algorithm (eXtreme Gradient Boosting, an advanced implementation of gradient boosting) to build a defect prediction model, optimized via hyperparameter tuning (e.g., tree count, depth, learning rate). Bayesian Optimization (BO) aids in finding the best model. Cross-validation assesses generalization. Models using Random Forest, Logistic Regression, and SVM are also tested. BO-XGBoost achieves high accuracy (0.9589), precision (0.9584), and recall (0.9940). SHAP (SHapley Additive exPlanations) helps interpret results. Waterfall, bee swarm, and interaction plots explain feature contributions and interactions. This practical method supports industrial defect prediction.