Fault detection in semiconductor manufacturing is complex due to class imbalance, missing values, outliers, and high data dimensionality. This study employs advanced machine learning techniques on the SECOM dataset to enhance fault detection accuracy and interpretability. Using oversampling methods like SMOTE and ADASYN, along with rigorous data cleaning and feature selection, we reduced features from 590 to 33. Principal Component Analysis (PCA) retained 91% of the dataset’s variance. Several classifiers, including Random Forest, SVM, KNN, Logistic Regression, Gradient Boosting, and XGBoost, were employed. A Voting Classifier combining top-performing models achieved superior results, with an accuracy of 99.68%. Emphasizing interpretability, we applied Permutation Feature Importance (PFI) and Partial Dependence Plots (PDP) to understand the model’s behavior and the significance of different sensors. Our approach offers a robust, accurate, and interpretable solution for fault detection in semiconductor manufacturing.