Uncertainty-Aware Feature Importance in Deep-Drawing Using Entropy-PFI on Production Data

Data-driven models are increasingly employed in manufacturing to support process monitoring and quality assurance. Beyond predictive accuracy, their reliability depends on understanding how input signals contribute to predictive confidence. This paper introduces Entropy-based Permutation Feature Importance (Entropy-PFI) as a method for uncetainty-ware interpretability in deep-drawing production. A dataset of over 46,000 production cycles was analyzed using a Gaussian Proces Regression model with 19 input features. The proposed method was systematically evaluated through baseline analysis, ablation benchmarking, noise injection, and correlation injection, and was compared with established interpretability techniques such as SHAP and PFI. Results demonstrate that Entropy-PFI provides a more faithful repentation of feature contributions to predictive uncrtainty than the alternatives. Specifically, Entropy-PFI identified critical process signals, detected the effects of noise and calibration drift d distinguished between harmful redundancy and beneficial robustness among correlated features. From an engineering perspective, these din how that Entropy-PFI can guide sensor monitoring, calibration, and placement, thereby supporting more robust and uncertainty-aware process control. The study concludes that Entropy-PFI provides actionable insights for interpretable and reliable AI applications in complex manufacturing systems.