A Data-Driven Framework Integrating Image Analysis, Feature Engineering, and Explainable AI for Overheating Management in Additive Manufacturing: From Quantification to Mitigation
Guardat en:
| Publicat a: | ProQuest Dissertations and Theses (2025) |
|---|---|
| Autor principal: | |
| Publicat: |
ProQuest Dissertations & Theses
|
| Matèries: | |
| Accés en línia: | Citation/Abstract Full Text - PDF |
| Etiquetes: |
Sense etiquetes, Sigues el primer a etiquetar aquest registre!
|
| Resum: | Additive Manufacturing (AM) or 3D printing, particularly Laser Powder Bed Fusion (L-PBF), is a transformative technology offering design flexibility and precision. However, challenges like overheating threaten the reliability of the printed components and that of the printer as well. Overheating is a significant challenge as it is the cause of several types of defects in AM. This proposal presents a comprehensive framework addressing this challenge by combining image analysis, predictive modeling, and optimization strategies. In the literature addressing overheating or thermal anomalies in AM, there is no unified approach to identify and quantify this challenge. Since the first step in solving any problem is to identify and preferably measure it, our framework introduces the Consolidated Overheating Ratio (COR), a novel metric that leverages pyrometry data and image analysis techniques to quantify overheating in printed regions with exceptional precision, reliability, and, most importantly, practicality and ease of application. This approach standardizes overheating assessment, enabling accurate identification and robust quantification of this risk.Beyond measuring overheating, accurate prediction plays a critical role in effective overheating management. Reliable forecasting of overheating enables early anticipation of potential issues and provides opportunities for proactive mitigation. Existing methods used to predict overheating or thermal defects typically rely on computationally intensive physics-based or simulation-based methods that often use overly simplified assumptions. We develop a hybrid CNN-LSTM model informed by geo-sequential feature engineering to predict overheating based on the geometrical and sequential dynamics of AM data. By implicitly integrating domain-specific knowledge, the model captures the complex interplay of heat accumulation and dissipation implicitly without explicitly modeling these processes. This approach enhances predictive accuracy while enabling rapid predictions that can be utilized during printing. To establish a comprehensive control system for mitigating overheating, the process must extend beyond quantification and prediction to include actionable strategies. Once the likelihood of overheating is predicted, the next step is to implement an action plan that adjusts printing parameters influencing the thermal behavior of printed layers. Optimizing these parameters is a well-researched area, widely explored as a means to address various types of printing defects, including overheating. Our framework leverages Explainable AI (XAI), specifically Counterfactual Explanations (CE). CE generates counterfactuals, which are hypothetical scenarios identifying the necessary changes to feature values to achieve a desired outcome. We have enhanced this approach to produce counterfactuals that are not only realistic but also require minimal and logically consistent adjustments to feature values, while respecting the causal relationships between these features. This enhanced method, called SHANCE (SHAP and Neighbor-Enhanced Counterfactual Explanations), combines SHAP-based feature importance with nearest neighbor constraints to ensure that the counterfactuals are actionable and feasible. Furthermore, a physics-based validation step integrates volume energy density (VED) thresholds and logical printing patterns, ensuring that the counterfactuals adhere to the causal dependencies between features, thereby maintaining manufacturability and effectively mitigating overheating. For further evaluation of SHANCE performance, a comparative study of SHANCE is conducted against established CE methods using the MNIST dataset. This study quantitatively assesses key performance metrics of CE methods, focusing primarily on sparsity and plausibility in the generated counterfactual. Further, SHANCE is compared to established parameter optimization methods to evaluate the value of adopting an XAI-based approach for printing parameter optimization. This performance evaluation examines the behavior of SHANCE, Genetic Algorithm (GA), and Neural Network (NN)-based optimization across a diverse set of real-world factors. Leveraging a structured Design of Experiments framework, both One-Factor-at-a-Time and Full Factorial analyses are conducted to investigate trade-offs between explainability, computational cost, and optimization quality. These additions provide deeper insight into SHANCE’s robustness and offer practical guidance for selecting optimization methods under varying data conditions, model uncertainties, and implementation constraints. This integrated approach bridges data-driven insights and physical validation, offering a robust solution for controlling overheating in AM. By enhancing reliability, interpretability, and efficiency, it paves the way for advancing AM processes and broadening their industrial adoption. |
|---|---|
| ISBN: | 9798297990036 |
| Font: | ProQuest Dissertations & Theses Global |