Using ATR-FTIR spectra and convolutional neural networks for characterizing mixed plastic waste

Highlights

• Present a convolutional neural net (CNN) framework for plastic classification.

• Approach uses ATR-FTIR spectra to classify plastics.

• Combination of CNN and ATR-FTIR provides avenue to enable sorting of mixed plastic waste.

Abstract

We present a convolutional neural network (CNN) framework for classifying different types of plastic materials that are commonly found in mixed plastic waste (MPW) streams. The CNN framework uses experimental ATR-FTIR (attenuated total reflection-Fourier transform infrared spectroscopy) spectra to classify ten different plastic types. An important aspect of this type of spectral data is that it can be collected in real-time; as such, this approach provides an avenue for enabling the high-throughput characterization of MPW. The proposed CNN architecture (which we call PlasticNet) uses a Gramian angular representation of the spectra. We show that this 2-dimensional (2D) matrix representation highlights correlations between different frequencies (wavenumber) and leads to significant improvements in classification accuracy, compared to the direct use of spectra (a 1D vector representation). We also demonstrate that PlasticNet can reach an overall classification accuracy of over 87% and can classify certain plastics with 100% accuracy. Our framework also uses saliency maps to analyze spectral features that are most informative.

Introduction

Plastics are essential materials that are used in a wide range of applications such as food packaging, construction, transportation, health care, and electronics. Since 1856 (when the first plastic celluloid was invented), the plastics industry has grown rapidly not only in terms of volume, but also in terms of the variety of materials produced. This rapid expansion has resulted in a massive environmental footprint; to give some perspective, in 2015, nearly 381 million tons of mixed plastic waste (MPW) were produced, this is more than the total weight of humans on earth (316 million tons). Notably, only 20% of all plastics produced were recycled (Ritchie and Roser, 2018); this recycling rate is notably low compared to that of other materials (e.g., aluminum has a recycling rate of nearly 100%). Most MPW end up in landfills and incinerators; landfills are not sustainable, especially when land availability is constrained (Abdel-Shafy and Mansour, 2018). MPW incineration reduces the need for landfills, but this process can release hazardous substances into the atmosphere (Hopewell et al., 2009).

MPW recycling is essential for mitigating the environmental impact of plastics, but this practice faces many obstacles (Schlesinger, 2013). Most of the recycled plastic is reprocessed into downgraded products (of lower value); for instance, plastics used for food packaging are often converted into cheaper building materials such as plastic lumber (Awoyera and Adesina, 2020). In other words, recycled plastic products are less valuable and thus there are limited incentives to produce them. Another key factor that hinders plastic recycling is our limited ability to effectively characterize and sort MPW streams, which can be quite complex (Milios et al., 2018). Traditionally, plastic components in MPW can only be partially identified based on techniques such as coding, density differences, and froth-flotation (Gundupalli et al., 2017). These technologies are easy to implement but are low-throughput and have several other limitations (Zhu et al., 2019); for example, density separation in water can effectively separate polypropylene (PP) and polyethylene (PE) from polyvinyl chloride (PVC), polyethylene-terephthalate (PET), and polystyrene (PS); however, PVC cannot be removed from PET in this manner because their density ranges overlap (Hopewell et al., 2009). Automated sorting with high-throughput, high-accuracy, and low-labor is necessary for effective MPW management.

Recent innovations in recycling technology include increasingly reliable detection instruments and improved materials identification algorithms; these have improved the accuracy and productivity of automated sorting. Methods such as spectroscopy, hyperspectral imaging (HSI), ultrasonic techniques, X-ray diffraction (XRD), thermal imaging or infrared imaging, combined with machine learning (ML) algorithms, have been successful in accurately identifying plastics that are commonly found in MPW (da Silva and Wiebeck, 2020; Karlsson et al., 2016; Siddiqui et al., 2008; Signoret et al., 2020, 2019; Wu et al., 2015). Michel et al. (2020) analyzed four different spectroscopic methods with various machine learning (ML) algorithms, such as k-nearest neighbors (KNN), linear discriminant analysis (LDA), and support vector machines (SVM), to identify marine plastic debris and consumer plastic. Among the four spectroscopic methods, the attenuated total reflection-Fourier transform infrared spectroscopy (ATR-FTIR) technique performed best, with an accuracy of 89–98%. Da Silva et al. (2020) developed a method to identify nine different types of plastics, including polyamide (PA) and polycarbonate (PC), based on μFTIR hyperspectral imaging and ML. Roh et al. (2018) used laser-induced breakdown spectroscopy with an algorithm-based radial basis function neural network to identify black plastics, including PP, PS, and acrylonitrile-butadiene-styrene (ABS), and achieved an accuracy of over 95%. Allen et al. (1999) proposed an automated sorting system using near-infrared spectroscopy to identify waste from electronic and electrical equipment (WEEE). Gundupalli et al. (2017) used a thermal imaging camera integrated with a robotic manipulator to classify recyclable materials in MSW and achieved a 90% accuracy. While these results are highly encouraging, these methods are slow and low-throughput and are not tailored to real-time sorting (they rely on manual sample collection).

ATR-FTIR can analyze plastic components found in MPW in real-time; as such, one can envision the development of fast, online ML techniques that can analyze ATR-FTIR spectra to characterize MPW streams. Recently, ML methods such as convolutional neural networks (CNNs) have been used to analyze spectral data (Ng et al., 2019). A key advantage of CNNs over other ML methods is their ability to automatically extract and organize discriminative features directly from raw data (without the need to pre-compute hand-crafted features). The training of powerful CNN models can be facilitated by the availability of advanced computing hardware (e.g., GPUs) and of vast data streams found in online systems. The integration of online ATR-FTIR and CNNs thus provides a potential avenue to sort plastic waste with high accuracy and throughput in real-time.

In this work, we propose a computational framework to characterize plastic components of MPW by analyzing ATR-FTIR spectra using CNNs. Experimental data was obtained by preparing small sheets of plastics of different shapes and used ATR-FTIR to scan sheets for 10 different types; this data collection approach mimics how rigid waste plastics are found in online processing of MPW streams. The proposed framework uses CNNs to analyze the spectra and sort/classify plastic components. The spectra collected can be represented as 1D vectors and analyzed by using 1D CNNs (Chen et al., 2019). The 1D CNN extracts features of a spectrum by convolving it with different filters. A limitation of this approach, however, is that it might fail to capture correlations across frequencies (wavenumbers which may compromise the prediction accuracy). To deal with this issue, we present a new data representation that captures signal correlation information; specifically, we represent a spectrum as Gramian angular fields (GAFs). GAFs are matrices (2D data objects) that can be analyzed using 2D CNNs (Oates, 2015) and these data objects can better capture spectral correlations. A problem with this approach, however, is that the training of 2D CNNs is significantly more computationally expensive than that of 1D CNNs. To ameliorate this issue, we use a Piecewise Aggregate Approximation (PAA) approach to reduce the dimension of the input GAF matrices (Keogh and Pazzani, 2000). This framework also uses saliency analysis (Sundararajan et al., 2017) to understand the most important features of spectra that can help identify different plastic components. We demonstrate that this CNN framework (which we call PlasticNet) can reach an overall classification accuracy of over 87% and can classify certain plastics with 100% accuracy. The conjunction of ATR-FTIR and CNN creates a powerful, low-cost, and rapid method for analyzing the composition of plastic waste and enables future recycling and reproduction of high-quality plastics.

The paper is structured as follows. In Section 2 we describe the experimental data collection and preparation. In Section 3 we discuss computational framework, including Gramian angular fields, 1D and 2D convolutional neural networks. In Section 4 we describe the results and discussion. In Section 5 we present conclusions and suggests directions of future work.