Abstract
The global proliferation of plastic waste presents significant environmental challenges, with effective sorting of complex waste streams being a critical bottleneck for recycling. Conventional sorting methods struggle with dark-colored plastics, a major limitation for near-infrared (NIR) systems, and require costly pre-cleaning of contaminated items. This study develops a robust methodology using Attenuated Total Reflectance-Fourier Transform Infrared (ATR-FTIR) spectroscopy combined with optimized machine learning to overcome these key limitations. Two models were established. Model 1 focused on the high-accuracy identification of 10 common plastic types, demonstrating 97.1% accuracy on an independent test set that included challenging dark and black samples. Model 2 addresses the pivotal challenge of identifying oil-contaminated plastics without any physical pre-cleaning. It innovatively employs Independent Component Analysis (ICA) for spectral unmixing, successfully separating the plastic's signal from the oil contaminant's. The extracted plastic spectra were then processed through an optimized workflow, achieving a remarkable accuracy of 92.5%. These results demonstrate that ATR-FTIR, empowered by advanced chemometric strategies like ICA and optimized machine learning, provides a powerful, non-destructive solution for sorting diverse and complex plastic waste. This work pioneers a viable pathway for the direct, algorithm-driven characterization of contaminated plastics, offering a promising approach to enhance the automation and efficiency of plastic recycling systems.