Machine Learning-Based Classification of Plastic Combustion Emissions Using Multi-Sensor Data 10.32526/ennrj/25/20250290

Abstract

Open burning of plastic waste releases particulate and gaseous pollutants whose profiles vary with polymer composition and combustion behavior. This study evaluated a multi-sensor machine-learning approach for classifying emissions from six common plastics: high-density polyethylene (HDPE), low-density polyethylene (LDPE), polyethylene terephthalate (PET), polypropylene (PP), polystyrene (PS), and polyvinyl chloride (PVC). The dataset comprised 2,160 observations (360 per class). Classification used six emission variables: PM₁, PM_2.5, PM₁₀, CO₂, CO, and total volatile organic compounds (TVOC), while relative humidity and temperature were analyzed separately as smoke-collection-chamber conditions. Kruskal-Wallis tests showed significant differences among plastic classes for all particulate and gaseous variables. Gradient Boosting (GB), Random Forest (RF), and k-nearest neighbor (k-NN) were optimized using the same randomized hyperparameter-search framework with stratified five-fold cross-validation. GB achieved the highest performance, with a cross-validated macro-F1 of 95.95%, test accuracy of 96.76%, and test macro-F1 of 96.75%. Cross-validated permutation importance identified CO2 as the most influential predictor, followed by PM₁₀, CO, PM_2.5, PM₁, and TVOC. Exhaustive evaluation of 57 feature combinations identified PM₁, PM_2.5, and CO₂ as the most parsimonious subset; the corresponding three-feature GB model retained 96.06% test accuracy and 96.05% macro-F1. The learning curve showed a 3.26-percentage-point train-validation gap at the largest training size. These results demonstrate strong discrimination of controlled plastic-combustion emission patterns, although mixed-waste, variable-combustion, and long-term sensor-drift validation is required before field application.