Deep learning-enabled chemometric analysis of spectral data for plastic waste sorting : innovation report

Proper sorting of mixed plastic waste into different plastic types into pure waste stream is important to ensure that recycling processes produces high quality recyclate that has good market value. However, current sorting technologies that are widely used include manual sorting, physical sorting or image recognition each suffer from their own drawbacks. Therefore, this research aims to utilise deep learning as a chemometric tool for plastic sorting using spectral data, developing robust chemometric models that can facilitate accurate sorting of mixed plastic waste. As the use of advanced deep learning models for chemometric analysis of spectral data has not been deeply explored in the literature, this innovation report first evaluates the effectiveness of several convolutional neural networks (CNN) for sorting of plastic waste into their types and degree of weathering from Fourier transform infrared (FTIR) and Raman data, achieving an accuracy of 0.941 and 0.964 for FTIR and Raman respectively. The accuracy could be further boosted to 0.949 and 0.967 for Raman and FTIR respectively through a novel tree-based neural network structure as PolymerSpectralDecisionNet (PSDN) developed in this work. Research in this field also suffer from limited datasets, which results in trained models that may not be generalisable. To address this issue, a novel deep learning framework was developed using a combination of cross-modal generative model and multi-modal deep learning. The multi-modal fusion of FTIR data with cross-modal generated synthetic Raman and laser-induced breakdown spectroscopy (LIBS) data saw an improvement in classification accuracy from 0.933 to 0.963. The data fusion methodologies were then further expanded to integrate FTIR data with hyperspectral imaging to perform semantic segmentation. The use of a 3D U-Net as a segmentation model achieved a mean intersection over union (IoU) of 0.950 but is computationally expensive. This research then develops a novel Spectral Aware U-Net which is significantly less computationally demanding, while still achieving a mean IoU of 0.891. Hence, the advanced deep learning methodologies developed in this innovation report provides an arsenal of tools for industrial adoption towards improving plastic waste sorting performance.