A Screening Framework for Data Quality Estimation in Manufacturing

In the context of Industry 4.0, Machine Learning (ML) and Deep Learning (DL) offer transformative potential for manufacturing, including defect reduction and process optimization. However, challenges such as noise, class imbalance, and limited high-quality labeled data often undermine the effectiveness of ML/DL models in industrial settings. It is often not clear whether these challenges will prevent an effective model from being formed. These challenges highlight the need for robust data quality assessment methodologies to guide the data preprocessing and model training processes. This paper introduces the Tao Index, a novel metric designed to comprehensively assess the quality of industrial data by integrating both local metrics, such as distance, density, and intrinsic dimension, and global measures, including Maximum Mean Discrepancy (MMD) and Kolmogorov-Smirnov (K-S) distance. The Tao Index can serve as a predictor of model performance, and by identifying low separability early, it helps avoid costly investments in low-performance models and guides data cleanup, collection, and trend analysis efforts. Extensive experiments on benchmark and real-world datasets demonstrate that the Tao Index offers a realistic assessment of data quality, with deviations from empirical expectations less than $10 \%$ in all separable benchmark and real datasets, outperforming existing metrics in multiple scenarios. Moreover, the Tao Index shows strong generalization power across different industrial domains with consistent performance on different industrial datasets.