A Systematic Framework for Data Quality Assessment in Industrial Process Monitoring

High-quality data is the cornerstone of reliable machine learning in industrial process monitoring, yet industrial datasets frequently suffer from complex degradations such as sensor failures, variable operating conditions, and label noise. Existing quality assessment methods typically rely on aggregate statistical metrics, failing to capture sample-level nuances or the interplay between data characteristics and model learning dynamics. To address these shortcomings, we propose a comprehensive, hierarchical framework that systematically quantifies data quality at both dataset and sample levels. Our approach integrates three fundamental quality dimensions: completeness, consistency, and diversity. The framework further combines information-theoretic measures with gradient-based importance estimation to deliver unified and interpretable quality assessments. Experimental evaluations on five synthetic datasets simulating industrial scenarios demonstrate that our quality-guided data cleaning strategies significantly enhance downstream model performance. The framework further provides robust quality tracking in online monitoring scenarios, validating the necessity of multi-scale quality assessment in data-centric industrial applications.