Maintaining product quality and reducing production defects are critical challenges in modern manufacturing systems. Variations in process parameters, equipment behavior, and operational performance often create complex relationships that the traditional quality control approach cannot capture effectively. This study employs interpretable machine learning (ML) frameworks to predict manufacturing defects and support the data-driven quality improvements. Two predictive scenarios are examined, including Adaptive Synthetic Sampling (ADASYN)-based oversampling for class imbalance handling, and ADASYN combined with nested grid search cross-validation for robust hyperparameter tuning. Eight ML models are evaluated, including AdaBoost, Support Vector Machine (SVM), Random Forest (RF), XGBoost, Light Gradient Boosting Machine (LGBM), Multilayer Perceptron (MLP), Stacking Ensemble, and Artificial Neural Network (ANN). Results show that ensemble models, particularly LGBM, RF, stacked ensemble, and XGBoost, provide the strongest predictive performance and reliability. SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) further identify the features that are most responsible for defect formation, supporting early intervention and continuous quality improvement in smart manufacturing.