Early Warning Predictive Models for Ammonia Refrigeration: A Machine Learning Approach in Brewing Processes

Resumen

Ammonia (NH3) refrigeration systems are critical infrastructures in the brewing industry, yet they pose significant safety risks due to the toxic nature of the refrigerant. Traditional monitoring systems, reliant on fixed-threshold alarms (typically set at 25 ppm), often fail to detect precursor anomalies in dynamic operational environments, leading to delayed responses and increased safety risk. This study proposes an early warning predictive framework leveraging Internet of Things (IoT) sensor fusion and Machine Learning (ML) algorithms. By integrating chemical concentration data with operational variables—such as discharge pressure, temperature, vibration, and compressor state—two models were developed and evaluated: Random Forest (RF) and Long Short-Term Memory (LSTM) networks. Results demonstrate that the LSTM model, optimized for temporal sequence analysis, achieved a Recall of 0.74 (vs 0.97 for RF) and an AUC of 0.999 (LSTM)vs 0.93 (RF), providing an average early warning lead time of 42 minutesbefore the safety threshold (25 ppm) is breached. Furthermore, explainability analysis using SHAP (SHapley Additive exPlanations) confirmed that pressure drops and thermal variability are key predictors of leakage events. This research bridges the gap between theoretical ML applications and real-world industrial safety.