Automated Data Engineering and Feature Selection for the Case Study of Warpage Detection in Fused Deposition Modeling

Research Question and Approach

This research addresses the challenge of developing high-fidelity machine learning (ML) models for in-situ monitoring in smart manufacturing, specifically for detecting warpage in Fused Deposition Modeling (FDM). Developing such models typically requires extensive domain expertise for manual feature engineering and model selection. To automate this process, the authors developed an Automated Data Processing (ADP) framework that integrates reinforcement learning (RL) with Explainable AI (XAI).

The framework operates as a two-level system: a decision-making unit and a classifier unit. The decision-making unit uses a Multi-Armed Bandit (MAB) algorithm to select the best combination of ML models and feature subsets. It evaluates these combinations based on predictive accuracy and F1-scores, using the results as rewards to update its policy. To manage feature dimensionality, the framework employs SHAP XAI to identify and select the most informative features from the raw sensor data, which includes vibration, acoustic emission, thermal imaging, and strain gauge measurements.

The proposed ADP framework successfully converges toward optimal model-feature configurations through iterative episodes. By leveraging SHAP-based feature selection, the framework effectively reduces redundancy and focuses on the most critical data signals. The experimental results demonstrate a clear performance gain: the framework achieved a test-set AUC of 0.9731, compared to 0.9248 for the baseline configuration using all available features. Furthermore, the mean reward value increased by over 50%, indicating that the reinforcement learning policy effectively learned to prioritize high-performing, efficient model-feature pairs.

Manual feature engineering and model selection are significant bottlenecks in deploying ML for industrial applications. This study demonstrates that integrating XAI into an automated pipeline not only improves predictive accuracy but also enhances the interpretability of the resulting models. By automating the selection of relevant features and models, this framework lowers the barrier for implementing reliable, data-driven quality control in manufacturing environments where domain expertise may be limited.