From Raw Data to OEE Forecasting: A Predictive Approach for Beverage Filling Lines

Abstract

In beverage manufacturing, the operational efficiency of filling lines is a critical determinant of productivity and profitability. This study proposes a time-series forecasting framework for predicting Overall Equipment Effectiveness (OEE) by modeling two key sub-components, Availability and Performance, while treating Quality as a constant, consistent with the facility’s strict quality standards. Two forecasting horizons are evaluated: a 1-day horizon for short-term fault detection and a 7-day horizon for production planning. The framework leverages a two-year historical production dataset alongside a seven-month machine sensor dataset, both processed through a structured data cleansing pipeline. To capture latent temporal behavior from high-dimensional sensor streams, the TS2Vec self-supervised representation learning model was employed. Eight machine learning algorithms were benchmarked: Support Vector Regression, XGBoost, Random Forest, Extra Trees Regressor, Histogram-based Gradient Boosting, Ridge Regression, ElasticNet, and an Ensemble Voting model. Two prediction strategies were compared: Direct OEE, which predicts the aggregate OEE score as a single target, and Composite OEE, which predicts Availability and Performance independently before combining them. Experimental results demonstrate that the Composite OEE strategy establishes predictive advantages primarily over extended horizons. The Extra Trees Regressor achieved a mean absolute error (MAE) of 9.39 ( $R^{2}$ : 0.48) for the 7-day horizon under the Composite layout, outperforming the Direct baseline. Conversely, for the 1-day horizon, the Direct Extra Trees Regressor achieved a lower error profile (MAE: 9.35) than its Composite counterpart. Crucially, non-parametric validation confirmed that there is no statistically significant difference between the error ranks of the top Direct and Composite frameworks. Furthermore, while self-supervised sensor representations introduce a critical micro-context of equipment health that stabilizes predictions within data-constrained environments, the marginal accuracy lift does not statistically justify the execution overhead of deep representation learning over history-only baselines. Consequently, a Composite modeling approach running on extended horizons is recommended for manufacturing diagnostics and root-cause monitoring.