Identification Of Dismantled Plastics From Waste Household Appliances And Electronics Using Near-infrared Spectroscopy

The environmental crisis caused by plastic debris has received a great amount of attention recently.Recycling plastic is not only a disposal method of plastic waste,but also a practice that can reduce the market demand for virgin plastics at the first place.Plastic sorting is a critical link in plastic recycling.At this moment,plastic sorting is performed manually in e-waste recycling industry,which is neither efficient nor accurate.Near-infrared(NIR)automatic sorting can replace manual sorting and increase the sorting efficiency.This study collected a variety of plastics from waste household appliances and electronics and analyzed their near-infrared(NIR)spectra.The types included polypropylene(PP),acrylonitrile-butadiene-styrene plastic(ABS),polypropylene(PS),and ABS/polycarbonate blends(ABS/PC).By comparing the spectra of waste plastics from different sources and virgin plastics,it was found that the plastics of each type have very similar peak positions and shapes regardless of the sources,except that the collected flame-retardant ABS and virgin ABS have obvious differences in the NIR region.Thus,flame-retardant ABS can be an independent class in the classification.After optimizing the pre-processing method and selected the appropriate spectral interval,four classification algorithms were tested,which were the spectral angle mapper(SAM),partial least-squares discriminate analysis(PLSDA),linear discriminate analysis(LDA),and support vector machine(SVM).Spectral data from samples of different sources and those collected under different conditions were used to evaluate the performance of classifiers.It is found that the SAM classification is susceptible to noise interference and thus not suitable for predicting the spectra with relatively low signal-to-noise ratio,and its accuracy is lower than others;PLSDA cannot accurately classify test samples from complex sources when trained on limited data,but it is a method of high accuracy in other circumstances;LDA has stable prediction accuracy,but analysis of the ROC curve shows that it might not performance well the binary classification of PS plastics;and,SVM has a very robust performance under most of the evaluation,expect for classifying low-resolution spectra.Furthermore,genetic algorithm and simulated annealing were applied to optimize the calculation efficiency of the above classification methods,and 20 wavelengths were selected to perform the classification.Results show that the prediction accuracy will not decrease much after computational optimization.After screening the wavelength variables for the SVM classifier through genetic algorithm,the prediction accuracy on waste plastics can reach 100%,and the prediction accuracy on low signalto-noise spectral dataset still can reach 96.6%.Finally,an automatic on-line sorting system for dismantled plastics and a scanning system for identifying polymer compositions in plastics flakes were proposed.And,a two-step detection procedure was designed which consisted of the rejection of the unknown spectrum using SAM and the classification of SVM optimized by genetic algorithm.