🤖 machine learning

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

This paper proposes an Automated Data Processing framework that utilizes reinforcement learning and SHAP-based feature selection to iteratively optimize machine learning model-feature combinations for FDM warpage detection, achieving significant improvements in predictive accuracy and stability compared to baseline full-feature configurations.

Original authors: Saleh Valizadeh Sotubadi, Nazanin Mahjourian, Vinh Nguyen

Published 2026-07-22

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Saleh Valizadeh Sotubadi, Nazanin Mahjourian, Vinh Nguyen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot to bake the perfect cake. You have a giant pantry full of ingredients: flour, sugar, eggs, spices, and even some weird stuff like glitter or hot sauce. You also have a whole library of different baking recipes. If you just throw everything into a bowl and hope for the best, the robot might get confused, waste time, or bake a disaster. This is exactly the problem engineers face with "smart manufacturing." They have massive amounts of data from sensors—like temperature gauges, microphones listening to the machine, and cameras watching the product—but they don't always know which pieces of data are actually helpful and which are just noise.

To solve this, scientists use "Machine Learning," which is like giving the robot a brain that learns from experience. But there's a catch: if you feed the robot too much junk data, it gets overwhelmed. This is where "Explainable AI" (XAI) comes in. Think of XAI as a super-smart tutor that doesn't just tell the robot the answer, but explains why it chose that answer, pointing out exactly which ingredients (data points) mattered most. The big question researchers are asking is: Can we build a system that automatically figures out the best recipe and the best ingredients without a human having to guess?

This paper tells the story of a team that built a digital "coach" to help robots learn how to spot a specific defect called "warpage" in 3D printing. Warpage is when a plastic part bends or curls up while it's being printed, ruining the final product. The researchers created a system called an Automated Data Processing (ADP) framework. Imagine this framework as a tireless coach that runs thousands of practice drills. It tries out different combinations of "recipes" (machine learning models) and different "ingredient lists" (feature sets).

Here is how the coach works: It uses a strategy inspired by a game called the "Multi-Armed Bandit." Picture a row of slot machines. Some machines pay out more than others, but you don't know which ones until you pull the lever. The coach pulls the levers (trains models) over and over. If a machine pays out a good reward (high accuracy), the coach remembers that combination and tries it more often. If it loses, the coach moves on. But there's a twist: the coach is also equipped with a special XAI tool called SHAP. This tool acts like a magnifying glass, looking at the robot's brain after every practice session to see which ingredients were actually doing the heavy lifting. The coach then creates a new, smaller ingredient list using only the top-rated items and tries the drills again.

The team tested this on 217 different printing scenarios using two common materials, PLA and ABS. They started by letting the robot use every piece of data they had collected. Then, they let the SHAP tool pick the best subset of data. The results were clear: the robot performed significantly better when it used the curated, smaller list of ingredients. When the robot used all the data, it got a score (called AUC) of 0.9248. But when the coach used the SHAP tool to pick the best features, the score jumped to 0.9731. That's a big improvement in reliability.

The paper also compared this smart coach to other methods. They tried using a technique called PCA, which is like squishing all the ingredients into a generic paste to save space, and another method called Mutual Information (MI), which looks for statistical connections. The coach found that PCA didn't work well because it smoothed out the important, complex details of the printing process. MI was okay, but it didn't understand the specific "personality" of the robot's brain as well as the SHAP tool did. The SHAP method consistently found the best features, leading to more stable and accurate predictions.

In short, the researchers showed that you don't need to feed a robot a mountain of data to get a great result. By using a smart, automated coach that learns which data points are truly important, you can build a system that is faster, more accurate, and easier to understand. The paper suggests that this approach is a solid step toward making smart manufacturing more efficient, though it notes that the system still needs diverse data to work perfectly in every possible situation. It's not a magic wand that solves everything instantly, but it is a powerful new tool for teaching machines how to see clearly.

Technical Summary: Automated Data Engineering and Feature Selection for Warpage Detection in FDM

Problem Statement
The integration of Machine Learning (ML) into smart manufacturing, specifically for monitoring Fused Deposition Modeling (FDM) processes, faces significant hurdles. Developing high-fidelity ML frameworks for tasks like warpage detection requires extensive domain expertise in both manufacturing processes and ML algorithms. Traditional approaches often rely on manual feature extraction and selection, which are time-consuming, prone to redundancy, and may result in suboptimal model performance. Furthermore, while Automated Machine Learning (AutoML) has advanced model selection and hyperparameter tuning, it often lacks systematic, intelligent mechanisms for input feature selection. In manufacturing IoT applications, adding excessive sensors can increase costs and degrade performance, whereas focusing on key sensors requires careful, often manual, feature engineering. Existing solutions often suffer from fragmented pipelines that fail to fully integrate feature interpretability with adaptive decision-making.

Methodology
The authors propose an Automated Data Processing (ADP) framework that leverages Reinforcement Learning (RL) and Explainable AI (XAI) to automate the selection of optimal model-feature combinations. The framework is structured into two primary units:

Decision-Making Unit (Multi-Armed Bandit): This unit operates as an RL agent using a Multi-Armed Bandit (MAB) algorithm. It explores a search space where each "state" represents a specific configuration of ML models and feature subsets. The agent selects actions (model-feature pairs) using an $\epsilon$ -greedy strategy combined with the Upper Confidence Bound (UCB) method to balance exploration and exploitation.
Classifier Unit: This unit trains and evaluates binary classification models (Random Forest, XGBoost, and Multi-Layer Perceptron) to predict warpage (warped vs. unwarped).
- Feature Engineering: Inputs include time-domain features extracted from sensor data (acceleration, acoustic emission, strain gauge, thermal imaging), printing parameters (temperature, height, infill density), and material properties.
- XAI-Driven Feature Selection: A key innovation is the integration of SHapley Additive exPlanations (SHAP). After initial training on full feature sets, SHAP is applied to rank feature importance based on the specific model's learned behavior. The framework then iteratively retrains models using only the top- $k$ ranked features to identify the most informative subset.

The framework operates through an iterative loop:

Execute ADP with full feature sets to establish baseline Q-values.
Apply SHAP XAI to rank features for each state.
Select top- $k$ features based on XAI rankings.
Re-execute ADP using these reduced feature subsets to update Q-values and converge on an optimal policy.

The reward function is a weighted combination of accuracy and F1-score, transformed via a sigmoid function to stabilize Q-value updates.

Key Contributions

ADP-XAI Framework: The development of an automated framework that integrates RL-based policy updating with SHAP-based feature selection to optimize ML model performance in manufacturing.
Intuitive Feature Selection: Moving beyond manual or purely statistical feature selection (like PCA) by using model-agnostic XAI to derive feature subsets that are specifically aligned with the predictive behavior of the trained models.
Systematic Evaluation: A comprehensive evaluation across 217 datasets derived from FDM processes, comparing the proposed ADP-XAI approach against baseline full-feature configurations and other feature selection methods (PCA and Mutual Information).

Results
The study utilized 217 datasets collected from PLA and ABS prints under various conditions. The results demonstrate the efficacy of the proposed framework:

Performance Improvement: The ADP framework utilizing SHAP-selected features improved the test-set Area Under the Curve (AUC) from 0.9248 (baseline full-feature) to 0.9731.
Reward Stability: The mean reward value increased by more than 50% compared to the baseline full-feature configuration. The optimal state (State 4) achieved a final Q-value of 0.5469 with SHAP features, compared to 0.4317 with all features.
Feature Selection Comparison:
- SHAP vs. PCA: PCA failed to improve performance, likely due to its linear nature and inability to capture the nonlinear dependencies inherent in FDM process data.
- SHAP vs. Mutual Information (MI): While MI showed improvement with larger feature sets ( $k \ge 20$ ), SHAP consistently outperformed MI, particularly at lower feature counts ( $k < 20$ ). SHAP identified a more discriminative subset because it accounts for the specific model's prediction behavior and nonlinear interactions, whereas MI relies on statistical dependency independent of the model.
Convergence: The framework converged to an optimal policy within 1000 iterations, with State 4 (Random Forest with SHAP-selected features) emerging as the most robust configuration.

Significance and Claims
The paper claims that the proposed ADP framework offers a scalable, automated solution for model selection in data-intensive manufacturing scenarios. By integrating XAI capabilities, the framework not only improves predictive accuracy and stability but also enhances model interpretability. The authors assert that leveraging SHAP allows the system to identify a physically meaningful and discriminative subset of features, reducing dimensionality without sacrificing fidelity.

The study positions this work as a step toward "intelligent, self-improving data pipelines" that support smart manufacturing. However, the authors maintain a modest tone regarding limitations, acknowledging that the framework's accuracy is dependent on data quality and diversity. They note that the current tabular RL approach faces computational hurdles as the search space expands, suggesting that future iterations will need to incorporate deep learning-based RL methods and multi-modal data integration (e.g., combining time-series, 2D images, and 3D point clouds) to handle more complex manufacturing environments.

Technical Summary: Automated Data Engineering and Feature Selection for Warpage Detection in FDM

More like this