Enhanced product defect forecasting using partitioned attributes and ensemble machine learning

This study addresses a critical challenge in industrial big data analytics for smart manufacturing: conventional machine learning methods often fail to account for data discontinuities caused by scrapped defective intermediates in multi-stage production processes, inadvertently treating non-conforming products as qualified during model training. We propose a novel process-aware data analytics framework specifically designed for process industries, featuring: (1) intelligent attribute partitioning based on information flow discontinuity points, and (2) an ensemble modelling approach combining Random Forest and C5.0 Decision Tree algorithms to generate interpretable prediction rules with quantified feature importance rankings. Validated using real-world production data from a Chinese rail steel manufacturer, our methodology demonstrates superior performance by explicitly incorporating process-specific data correlations. The proposed solution effectively mitigates information distortion caused by scrapped intermediates while maintaining operational interpretability – a crucial requirement for industrial implementation. The research results increased the accuracy rate of the test set of the random forest experiment from 88.39 % to 92.69 %, and the accuracy rate of the test set of the decision tree experiment from 71.89 % to 79.15 %. Additionally, the experimental results verify that, compared with the traditional methods, our framework has better applicability in capturing product quality in the manufacturing industry when process attributes are considered.