Abstract
Poultry litter composting is a complex, multi-phase biological process that remains challenging to monitor and control using traditional methods. This study presents a machine learning framework that combines Gaussian Mixture Models (GMMs) for unsupervised soft clustering with supervised regression models for real-time phase inference. Using high-resolution sensor data from two full-scale poultry litter composting experiments, GMMs successfully identified four latent composting phases: pre-composting, mesophilic, thermophilic, and maturation, capturing gradual transitions and overlapping states that conventional temperature thresholds often fail to detect. A supervised learning module was then developed to map real-time sensor snapshots to the GMM-derived phase probabilities. Models trained via H2O AutoML achieved high predictive performance across all clusters (R² > 0.99, RMSE < 0.032), with ensemble tree methods outperforming other algorithms. The proposed approach enables low-latency, label-free compost phase monitoring using only instantaneous temperature and humidity data, making it well-suited for embedded control systems and real-time decision support. This framework provides a generalizable and interpretable tool for intelligent composting, with applications in process optimization, anomaly detection, and phase-aware control strategies in agro-industrial waste management.
