Predictive maintenance is vital for improving manufacturing efficiency by predicting equipment failures. This study presents a new framework for fore—casting machine failures using a dataset of machine records, combining machine learning with explainable AI. Ten machine learning classifiers, including TabNet and NGBoost, were tested, with Random Forest achieving the highest accuracy at 99.83%. To improve prediction reliability, three techniques were used: K-means clustering to group machines by operational patterns, anomaly detection using Isolation Forest as a feature, and cost-sensitive learning to focus on critical failure types. Five explainable AI methods, SHAP, Permutation Importance, Anchors, Partial Dependence Plots, and LIME, were applied to interpret the Random Forest model. LIME was the most effective based on a composite evaluation score. This approach enhances prediction accuracy and provides clear insights into failure causes, supporting informed maintenance decisions. The findings contribute to reliable and interpretable predictive maintenance systems in manufacturing.