(515b) Rapid and Real-Time Mixed-Plastic Waste Analysis Using Infrared Spectroscopy and Machine Learning | AIChE

(515b) Rapid and Real-Time Mixed-Plastic Waste Analysis Using Infrared Spectroscopy and Machine Learning

Authors 

Jiang, S. - Presenter, University of Wisconsin-Madison
Zavala, V. M., University of Wisconsin-Madison
The production of plastics continues to climb but their recycling rate remains quite low. In 2015, nearly 381 million tons of plastic were produced, but only 20% of the world's plastic waste was recycled [1]. Mixed plastic waste (MPW) is putting unprecedented pressure on landfills [2] and it is critical to convert MPW into clean and usable materials to enable a more circular system. A fundamental challenge that hinders this is the difficulty in analyzing the composition of MPW; such information can enable the development of more appropriate treatment strategies.

Traditionally, offline infrared (IR) spectroscopy has been used to identify the chemical components in MPW [3]. Offline IR spectra are easy to identify but are labor- and time-intensive to obtain; as such, this method has limited use in real-time plastic composition assessment. In contrast, online IR spectroscopy allows for rapid detection of plastic composition with a single scan time of 11 microseconds [4]. However, the spectra obtained by this method can be highly noisy (due to interference).

Recently, machine learning methods such as convolutional neural networks have been used to analyze spectral data (in the context of soil characterization) and they have been shown to provide high predictive accuracies and to be resistant to noise [5]. Conventionally, IR spectra are treated as one-dimensional vectors that are analyzed using one-dimensional (1D) CNNs [6]. The 1D CNN works by convolving the spectrum using filters to extract hidden patterns. Although IR spectra already carry a rich amount of information, the hidden correlations between the signal intensities at different wavenumbers are not explicitly encoded.

In this work, we propose to encode the signal correlation information using Gramian angular fields (GAFs) and analyze GAF matrices using 2D CNNs [7]. GAFs have been used to transform time-series vectors into matrices to capture temporal correlations and thus improve classification accuracy. To facilitate fast training of 2D CNN-based algorithms, we use a Piecewise Aggregate Approximation (PAA), which allows us to reduce the dimension of the field matrix [8]. We also used saliency analysis [9] to understand the most important signal intervals and correlations between signals. We show that this approach provides a robust, low-cost, and rapid method for analyzing the composition of MPW and can enable future, high-throughput recycling of plastics.

[1] Ritchie, Hannah, and Max Roser. "Plastic pollution." Our World in Data (2018).
[2] Abdel-Shafy, Hussein I., and Mona SM Mansour. "Solid waste issue: Sources, composition, disposal, recycling, and valorization." Egyptian journal of petroleum 27.4 (2018): 1275-1290.
[3] Zhu, Shichao, et al. "Plastic solid waste identification system based on near infrared spectroscopy in combination with support vector machine." Advanced Industrial and Engineering Polymer Research 2.2 (2019): 77-81.
[4] “S2050 2.0-5.0µm Fiber Spectrometer - NLIR.” [Online]. Available: https://www.nlir.com/products/s2050-2-0-5-0µm-fiber-spectrometer/. [Accessed: 15-Mar-2021].
[5] Ng, Wartini, et al. "Convolutional neural network for simultaneous prediction of several soil properties using visible/near-infrared, mid-infrared, and their combined spectra." Geoderma 352 (2019): 251-267.
[6] Chen, Xiaoyi, et al. "1D convolutional neural network for the discrimination of aristolochic acids and their analogues based on near-infrared spectroscopy." Analytical Methods 11.40 (2019): 5118-5125.
[7] Wang, Zhiguang, and Tim Oates. "Encoding time series as images for visual inspection and classification using tiled convolutional neural networks." Workshops at the twenty-ninth AAAI conference on artificial intelligence. Vol. 1. 2015.
[8] Keogh, Eamonn J., and Michael J. Pazzani. "Scaling up dynamic time warping for datamining applications." Proceedings of the sixth ACM SIGKDD international conference on Knowledge discovery and data mining. 2000.
[9] Sundararajan, Mukund, Ankur Taly, and Qiqi Yan. "Axiomatic attribution for deep networks." International Conference on Machine Learning. PMLR, 2017.