Learning from a Noisy PLC-to-Database Monitoring Pipeline: A Controlled Study of Label-Noise Robustness Using a Physically Derived Noise-Rate Predictor

Industrial monitoring pipelines that log programmable logic controller (PLC) states to a database for machine learning inherit the noise introduced by the sampling process itself, yet noisy-label learning research is almost always evaluated against synthetically corrupted labels rather than real sampling noise. This study evaluates a physically derived noise-rate predictor, Pedge, by training decision tree and logistic regression classifiers on real, non-synthetic noisy labels from a Festo Modular Production System–Process Automation (MPS PA) industrial automation testbed—a four-station laboratory production line used for Industry 4.0 research and teaching—and assessing them against independently logged true PLC states across four stations. A controlled temporal-subsampling experiment, which varies the effective observation interval while holding the classification task fixed, shows that F1 score (the harmonic mean of precision and recall, used here in place of raw accuracy because of class imbalance in the target signals) against ground truth declines substantially for one process (0.951 to a sweep minimum of 0.679, 0.723 at the highest sampling condition tested), replicated across two target signals and both classifiers, while two other processes remain robust (F1 at or above 0.94), consistent with strong correlated features (correlations at or above 0.98), and one target signal shows no learnable baseline (F1 at or below 0.652) regardless of noise. These results indicate that a physically derived sampling-risk index can help anticipate machine learning sensitivity in industrial automation, but only when the target task is both learnable and lacks a redundant shortcut feature.