Active Learning and Synthetic Data Generation Methodology: Applications for Industrial Use Cases

Abstract: Data-driven technologies are becoming an essential component of modern industrial processes, supporting tasks ranging from quality control to predictive maintenance and process optimization. Despite their growing relevance, many industrial environments still suffer from limited labeled data, costly experimentation, and highly imbalanced datasets. These constraints make the development of robust machine learning models particularly challenging, especially in tabular settings where data collection often requires physical measurements, expert intervention, or expensive simulations.

Active Learning and synthetic data generation have emerged as promising strategies to alleviate these limitations. Active Learning aims to identify the most informative samples to label, reducing annotation costs and improving model efficiency, while synthetic data techniques can expand the available dataset, mitigate imbalance, and enhance generalization in scenarios where real data are scarce. Together, these methods offer a powerful framework for building more reliable and data-efficient industrial AI.

This thesis investigates how these techniques can be applied and combined to address real-world industrial challenges. The work is supported by three peer-reviewed publications. The first study analyzes machine learning models for predicting femtosecond-laser groove depths in a tribological application, illustrating the difficulties associated with limited industrial data and motivating the need for more data-efficient learning approaches. The second study explores the benefits of integrating variance-based Active Learning with several synthetic data generation methods—SMOTE, GANs, and random oversampling—to improve classification performance in imbalanced datasets. Finally, the third study proposes an Active Learning–driven methodology for enhancing Design of Experiments through targeted data acquisition and region-specific synthetic data generation using GANs, VAEs, and LLMs. This last study demonstrates how combining uncertainty estimation with synthetic sampling can reduce model uncertainty while minimizing the number of required physical experiments.