This study proposes a hybrid predictive maintenance framework for Computer Numerical Control machining systems by integrating risk-based prioritization with deep learning-based time-series forecasting. The framework combines Fine-Kinney risk assessment and multi-criteria decision-making using the Technique for Order Preference by Similarity to Ideal Solution to identify and rank critical failure modes. Subsequently, Recurrent Neural Network, Long Short-Term Memory, and Gated Recurrent Unit models are developed to predict failure intervals based on historical maintenance data. The methodology is validated using real-world failure data collected from 17 CNC machines over a two-year period. The results indicate that Long Short-Term Memory and Gated Recurrent Unit models outperform classical Recurrent Neural Networks in terms of prediction accuracy, achieving significantly lower error rates and higher correlation with actual failure patterns. In addition, the integration of risk prioritization ensures that predictive modeling efforts are focused on the most critical failure modes, enhancing both analytical efficiency and practical relevance. To assess real-world applicability, a numerical cost–benefit analysis is conducted. The findings show that even a conservative reduction of 20% in unplanned downtime can yield a return on investment exceeding 200%, demonstrating the strong economic feasibility of the proposed approach. Overall, the study provides a comprehensive and scalable predictive maintenance solution that combines technical performance with economic impact, offering a practical decision-support tool for small and medium-sized manufacturing enterprises.

Keywords

Predictive Maintenance, Deep Learning, Cost Reduction, CNC Machines, Risk Identification.