Studying the causes of humidity fluctuations in food products within industrial environments
2772-5022
Abstract
The production of processed foods and animal nutrition requires precise control over product characteristics to ensure consistent quality. Unexpected fluctuations in key parameters, particularly humidity, can lead to significant financial losses and reduced product quality. This study investigates the root causes of humidity variations in the final product and evaluates the predictability of this metric using variables with potential causal influence as predictors. We apply advanced causal discovery methods, including the Peter–Clark Momentary Conditional Independence (PCMCI) algorithm, and validate the findings with Granger causality and the Temporal Causal Discovery Framework (TCDF). To conduct these analyses, we develop a novel computer vision model to estimate the mass of organic material entering industrial furnaces from RGB imagery, enabling low-cost real-time monitoring directly on the main production-line conveyor. Experiments on production-line data show that causal filtering improves humidity-forecasting accuracy in three of four independent production runs (mean absolute-error improvement of 12.4 ×, Diebold–Mariano p < 0.05), while our vision-based mass estimation achieves a mean absolute coverage error of 0.81 percentage points (RMSE = 0.90 pp) validated against manually annotated ground truth. The integration of causal inference with machine learning and low-cost sensing offers a practical approach for enhancing process control and product consistency in industrial environments.