Defect Detection in Manufacturing: A Comparative Study of Machine Learning and Fuzzy Rule-Based Systems

Defects occurring in manufacturing processes can lead to customer dissatisfaction, reduced product quality, and increased operational costs. Accurately predicting product defects is critical for enhancing productivity and maintaining quality standards. In this study, classical machine learning (ML) models and a fuzzy inference system (FIS) are comparatively analyzed for the classification of manufacturing defects. The publicly available “Predicting Manufacturing Defects” dataset from Kaggle is utilized. The methodology consists of three main stages: (i) prediction of the DefectStatus using Logistic Regression, Decision Tree, Random Forest, Gradient Boosting, Support Vector Machine, and K-Nearest Neighbors models; (ii) determination of feature importance scores through Random Forest importance ranking and SHAP analysis; and (iii) retraining ML models using only the top-ranked features to evaluate performance changes. Based on these important features, a rule-based FIS model is developed using triangular membership functions and rules automatically extracted from a decision tree. Among the ML models, Random Forest achieved the highest accuracy in the first stage (Accuracy: 95.06%). According to both Random Forest and SHAP analyses, MaintenanceHours (0.23, 0.09), DefectRate (0.22, 0.07), QualityScore (0.15, 0.05), and ProductionVolume (0.10, 0.04) were identified as the most critical features. When trained using only these four features, the Random Forest model again yielded the highest classification performance with an accuracy of $\mathbf{9 5. 5 2 \%}$. The FIS model, built with eight fuzzy rules, achieved $\mathbf{9 4 \%}$ accuracy, demonstrating performance comparable to classical models. These findings indicate that a simplified and interpretable model based on a few key features can be as effective as more complex machine learning approaches in manufacturing defect classification.