Additive manufacturing, particularly fused deposition modeling (FDM), has evolved into highly automated processes that generate massive volumes of data. Each print produces variables such as temperature, speed, material flow, layer height, and dozens of parameters that, when properly analyzed, can predict the final part quality. However, data abundance does not automatically translate into useful knowledge. This is where automated data engineering and intelligent feature selection become the true drivers of optimization.
Feature selection is not a mere statistical filter; it is a strategic decision that determines which variables should be fed into machine learning models. A traditional approach involves training all models with the full set of variables, but this often leads to overfitting, higher computational cost, and suboptimal results. In contrast, techniques inspired by reinforcement learning —such as policy updates using Q-values— allow exploring model–feature combinations iteratively, evaluating in each episode the accuracy and F1-score to reward the most promising configurations. This process converges toward optimal combinations without requiring exhaustive prior domain knowledge.
Simultaneously, explainable artificial intelligence (XAI) based on Shapley values (SHAP) provides a metric for each feature’s importance, enabling the construction of reduced yet informative subsets. By integrating SHAP into the reinforcement cycle, the system learns not only which model works best but also why certain variables are relevant. Recent studies show that this synergy can raise AUC from 0.9248 to 0.9731, increasing the average reward by more than 50% compared to the initial full‑feature configuration. Performance stability, visualized through reward distributions, confirms the method is robust against FDM data variability.
In an industrial context, implementing such an infrastructure requires more than advanced algorithms. Companies need custom software that integrates data pipelines, model orchestration, and result visualization. An automated data processing (ADP) framework must be able to ingest real‑time data from printer sensors, clean it, run feature selection, and deploy the winning model into production. For this, the cloud becomes an indispensable ally. Cloud AWS/Azure services offer scalability, secure storage, and elastic computing power to handle training episodes without bottlenecks.
Cybersecurity also plays a critical role. Digital manufacturing environments are sensitive targets: an attack that alters printing parameters could compromise entire batches. Therefore, any data engineering solution must incorporate encryption protocols, multi‑factor authentication, and continuous auditing. Companies that embrace digital transformation trust technology partners that offer both cybersecurity and specialized software development. Additionally, integration with Business Intelligence tools such as BI/Power BI allows production managers to monitor quality trends, deviations, and model predictions on interactive dashboards.
Another key enabler is AI agents. These autonomous assistants can oversee the feature selection process, launch new training episodes when data changes are detected, and recommend real‑time adjustments. An agent trained with reinforcement can decide which model to run on each production batch, minimizing prediction error without human intervention. This automation reduces setup times and frees engineers for higher‑value strategic tasks.
At Q2BSTUDIO we understand that excellence in additive manufacturing is not achieved only with good machines, but with an intelligent software ecosystem. Our team develops solutions ranging from data capture to predictive model deployment, always with a modular, secure, and scalable approach. By combining our expertise in artificial intelligence, cloud computing, and process automation, we help companies reach levels of precision and efficiency that were previously unattainable.
In summary, automated data engineering and feature selection based on reinforcement and explainability represent the next step in the digital maturity of the FDM sector. Organizations that adopt these methodologies will not only improve their technical metrics (AUC, F1, average reward) but also build a solid foundation for continuous improvement. The key lies in integrating cutting‑edge tools —from AI agents to BI dashboards— within a custom software architecture that guarantees both flexibility and security. The future of additive manufacturing is already here, and it is written with intelligent data.
