A Study on a Random Forest-Based Algorithm for Predicting Yield Rates in Process-Oriented Production Lines

Abstract

Industrial production line data poses three intertwined difficulties: the features are high-dimensional and anonymized, the process is strongly sequential, and the class distribution is extremely skewed (171:1). This paper presents a Process-based Cascaded Random Forest (P-RF) algorithm tailored to these conditions. Instead of flattening the feature space, P-RF maps the physical order of workstations onto a hierarchy of forest modules and lets a suspicion probability flow forward along the line, so that cumulative error effects are captured rather than ignored. To cope with the severe imbalance, SMOTE resampling, cost-sensitive learning, and grid search are combined into a single optimization scheme. Experiments show that the final model reaches an overall accuracy of 91.21% with a macro-average F1 score of 0.4839, while cutting false alarms to 29,361—a trade-off that favors low disruption on high-volume lines. A global feature-importance traceback further confirms that the transmission probabilities from preceding workstations dominate model decisions, revealing the “error snowball effect” at work. This not only supports the soundness of the P-RF architecture, but also gives production managers a concrete handle for quality control and precise traceability.