Convolutional neural network for identification of plastic solid waste using near infrared spectroscopy

Abstract Plastic waste is a relevant challenge for waste management sector and further technological means has to be urgently researched. Near infrared spectroscopy (NIR), as a non-destructive, cost-effective and mature technology, is currently operational in some waste-sorting facilities. However, NIR remains challenging to discriminate different black plastics because black targets have low reflectance in the its spectral region. Therefore, this study aimed to address the problem in black plastic classification. Seven kinds of plastics: High Impact Polystyrene (HIPS), Acrylonitrile Butadiene Styrene (ABS), High Density Polyethylene (HDPE), Polyethylene terephthalate (PET), Polyamide 66 (PA66), Polycarbonate (PC) and Polypropylene (PP), were prepared for analysis, in which black plastic accounted for half of the total samples (84/159). Some methods such as partial least squares discrimination analysis (PLS-DA), soft independent modelling of class analogy (SIMCA), linear discriminant analysis (LDA) and convolutional neural network (CNN) were used to classify the plastics. Coupling with proper preprocessing methods, PLS-DA and LDA could provide an accuracy of 57.00%, and SIMCA of 69.98%. Meanwhile, the NIR spectra fed into the CNN model as one-dimensional data, the accuracy of CNN model could reach 98.00%. Compared to the SIMCA, PLS-DA and LDA, the accuracy of CNN provided an average improvement of 28-41%. Results demonstrated that NIR spectra combined with CNN gave an excellent accuracy which was potential to solve the bottleneck problem of black plastic discrimination.