TY - JOUR

T1 - Inline classification of polymer films using Machine learning methods

AU - Koinig, Gerald

AU - Kuhn, Nikolai Emanuel

AU - Fink, Thomas

AU - Grath, Elias

AU - Tischberger-Aldrian, Alexia

PY - 2023/12/9

Y1 - 2023/12/9

N2 - Improving the sortability of plastic packaging film waste (PPFW) is crucial for increasing the recycling rate in Austria as they account for 150,000 t of the annually produced 300,000 t of plastic packaging waste. Currently PPFW is thermally recovered, as it is impossible to separate the mechanically recyclable monomaterial films from the non mechanically-recyclable multimaterial films. In this study, machine learning models capable of classifying inline into monolayer and multilayer films of PPFW according to their spectral fingerprint taken in transflection were created. Feature selection methods, like PCA and MRMR F-Tests, identified the most relevant spectral ranges for classification, that show the least redundancy and highest relevance. This effective subset of features decreases the required complexity of the model while reducing prediction time without compromising accuracy. The resulting models achieved a prediction accuracy of 85 % on unseen specimens with minimal prediction latency, effectively showing the inline applicability of these models in sorting aggregates.

AB - Improving the sortability of plastic packaging film waste (PPFW) is crucial for increasing the recycling rate in Austria as they account for 150,000 t of the annually produced 300,000 t of plastic packaging waste. Currently PPFW is thermally recovered, as it is impossible to separate the mechanically recyclable monomaterial films from the non mechanically-recyclable multimaterial films. In this study, machine learning models capable of classifying inline into monolayer and multilayer films of PPFW according to their spectral fingerprint taken in transflection were created. Feature selection methods, like PCA and MRMR F-Tests, identified the most relevant spectral ranges for classification, that show the least redundancy and highest relevance. This effective subset of features decreases the required complexity of the model while reducing prediction time without compromising accuracy. The resulting models achieved a prediction accuracy of 85 % on unseen specimens with minimal prediction latency, effectively showing the inline applicability of these models in sorting aggregates.

U2 - 10.1016/j.wasman.2023.11.028

DO - 10.1016/j.wasman.2023.11.028

M3 - Article

VL - 174.2024

SP - 290

EP - 299

JO - Waste management

JF - Waste management

SN - 0956-053X

IS - 15 February

ER -