Understanding the Impact of Temporal Aggregation on Uncertainty in Quality Indicator Prediction for Industrial Processes

Abstract

Industrial predictive datasets often rely on coarse synchronized targets obtained by aggregating serial measurements over predefined segments. Although such aggregation is usually imposed by storage constraints and synchronization requirements, the resulting labels are commonly treated as deterministic in downstream modeling. This can be misleading when the underlying process is serially dependent, because aggregation then injects uncertainty at the label level before any model is trained and directly affects the performance that can realistically be achieved. This work examines that effect using segment-level coiling temperature prediction in hot strip steel manufacturing as a real-world example, where meter-level coiling temperature measurements are synchronized to tracked material segments and averaged to form prediction targets. A serial, dependence-aware, and deployable formulation is introduced to quantify the uncertainty associated with these aggregated targets and propagate it to the downstream predictive task. Results show that serial dependence persists in the meter-level coiling temperature series even after downsampling, and that slower sampling increases the uncertainty associated with the aggregated labels. The estimated aggregation uncertainty is further shown to be of the same order as the error achieved by an extensively optimized downstream predictor, indicating that a non-negligible portion of the apparent prediction error is attributable to noise introduced during target construction rather than to deficiencies of the predictive model alone. The findings highlight that aggregation design should not be treated as an innocent preprocessing choice. Instead, it should be considered as a factor that directly shapes attainable predictive performance, with sampling frequency emerging as an actionable lever for improving the performance ceiling of industrial datasets constructed from serial measurements.