Printer Friendly

Identification of plastic waste using spectroscopy and neural networks.

INTRODUCTION

The economic viability of recycling plastic materials depends on developing an inexpensive and fast method for sorting dirty, crushed plastic bottles and containers. Most sources of recyclable material provide a random mixture of various plastic types, but recycling processes generally require a single polymer to be used. Therefore, the first step of the recycling process is to sort the input waste stream into its components. At present, this sorting process is labor-intensive and represents a significant portion of the costs associated with the recycling process. The value of recycled polymers is limited by the cost of the virgin plastics, which in turn is governed by the prevailing oil price (1). Automating the sorting process will improve profit margins for the existing small-scale recycling operations and is essential for any large-scale effort. A variety of identification methods have been proposed, developed, and commercialized (2). However, none of these available technologies is very fast, and most involve expensive or complicated instrumentation.

We have demonstrated the feasibility of a new approach for automatic plastic sorting. In this paper, we show that rapid and reliable sorting can be achieved using a combination of fixed-filter near-infrared (NIR) spectroscopy and neural network data analysis. In the next section, we examine the NIR absorbance spectra for high density polyethylene (HDPE), poly(ethylene terephthalate)(PET), and poly(vinyl-chloride)(PVC) and identify significant spectral features for polymer identification. The third section will introduce the concept of a neural network. We will show that the output of a three-filter NIR spectrometer can be processed by a neural network to positively identify HDPE, PET, and PVC. In the last section of this paper, we discuss a proposed compact, rugged instrument based on the new sorting method.

NIR ABSORBANCE SPECTRA

The near-infrared (NIR) region of the electromagnetic spectrum extends from about 0.75 [[micro]meter] to 2.5 [[micro]meter] in wavelength. A variety of chemical bonds (e.g., O-H, C-H) absorb light in this region, and it is well known that an analysis of NIR absorbance can be used to obtain information about the chemical structure of a sample (3). By measuring the absorption of light at a few well-chosen wavelengths, it is possible to obtain a unique signature for each type of plastic. The signature for an unknown sample can then be compared to those of known plastics in order to determine the polymer type.

Samples of crushed, dirty plastic were obtained from a recycling plant in Philadelphia. Plastic bottles and containers were removed directly from the sorting chutes and subsequently cut into 3 inch by 3 inch squares (both single- and double-wall thickness). We collected samples of the following polymer types: clear PET, green PET, PVC, nature HDPE, and colored HDPE. A total of 59 samples were prepared. The samples were not cleaned and in many cases included labels or creases. NIR absorbance spectra of these samples from 1100 nm to 2500 nm were measured using a commercial spectrometer (NIR Systems model 6500). We show examples of these spectra in Figs. 1 through 3.

The spectra collected for various samples of a given polymer type proved to be remarkably consistent in spite of the fact that the samples were dirty. This observation suggests that the levels of contamination typically found in plastic waste streams will not obscure the characteristic spectral features. The only variation noted between spectra of a given polymer was in the total absorbance values. Absolute absorbance is directly proportional to the thickness of the sample, whereas relative absorbance (the ratio between absorbance values at two distinct wavelengths) is independent of sample thickness. The thicknesses of the plastic samples to be identified are not known a priori; for this reason, we must analyze the spectra on the basis of relative measurements.

In the region between 1000 nm and 2000 nm, the location and relative height of several absorbance peaks were found to be unique for each polymer type. As shown on Fig. 1, the dominant absorbance peak in PET is at 1660 nm. In PVC [ILLUSTRATION FOR FIGURE 2 OMITTED], this peak has shifted to 1716 nm. This feature has been observed (4) in a number of polymers and is attributed to the first overtone of C-H stretching. PVC also exhibits two small, broad peaks at 1196 nm and 1422 nm. Figure 3 shows a typical spectrum of HDPE, in which there is a well-defined peak at 1214 nm, two poorly-resolved peaks near 1400 nm, and a prominent peak at 1732 nm. From these observations, we chose to focus on absorbance data at 1214, 1660, and 1716 nm. It is noteworthy that these wavelengths lie in relatively flat portions of the absorption spectrum of water (5). NIR absorption by water is of concern because of the presence of moisture both in the plant environment and adsorbed on the plastic itself.

In this investigation, we have assumed that the absorbance data would ultimately be provided by a simple, fixed-filter spectrometer. Such devices are more rugged and less expensive than the instrument that produced the continuous spectra shown in Figs. 1 through 3. We describe below a suitable instrument that uses interference filters to select the wavelengths of interest, viz., 1214, 1660, and 1716 nm. We have emulated its performance by convolving the NIR spectra with a Gaussian function representing the filter spectral transmission. This convolution produces a single value to represent the output of each filter/detector combination. In order to correct for variations in sample thickness, we normalized the data by dividing the measurements at 1660 nm and 1716 nm by the reading at 1214 nm (in effect giving relative absorbances). These two ratios were then used as the input to the neural network, which is described in the following section.

NEURAL NETWORKS

In 1943, McCulloch and Pitts (6) showed that Boolean functions could be computed using simple "neural" processing elements loosely modeled after biological neurons. This remarkable result marked the birth of a new scientific discipline, which has come to be known as "neural networks" (7). In recent years this discipline has developed into a mature field with important contributions from physics, computer science, and neurobiology. New developments in the theory of neural computation have led to impressive real-world applications in process control, process modeling, and forecasting.

Neural networks are composed of many simple processors (called "neurons" after their biological counterparts), which pass information to one another through weighted connections. Neural network computers are not programmed; they learn about their environment by repeated exposure to examples of desired behavior. Typically, each neuron applies a nonlinear transformation to a weighted sum of its inputs and adapts to its environment according to a learning rule, which determines how the interconnecting weights are to be modified.

The back propagation network (BPN) is the best studied and most often applied method of neural computation (8). While the BPN is not necessarily the most powerful neural computer, it has been shown to be suitable in a wide array of applications. In this study, we sorted HDPE, PET, and PVC using the three-layer back propagation network shown in Fig. 4.

We wanted to examine the performance of a hypothetical three-filter spectrometer with filter wavelengths of 1214, 1660, and 1716 nm. Before developing the neural network models, the NIR data were preprocessed in two ways. We estimated the transmission of the three filters by convolving the raw sample data with 15 nm bandwidth Gaussian functions and assuming 65% transmission through the filters. This convolution produces three numbers for each sample. As noted above, we can compensate for differing polymer sample thicknesses by taking ratios of the filter responses. Thus, the data preprocessing produced two response ratios for each sample.

The neural network has two inputs. The first input, [I.sub.1], is the ratio of the response at 1660 nm to that at 1214 nm. [I.sub.2] is the ratio of the response at 1716 nm to that at 1214 nm. The three outputs of the network can be thought of as indicator lights. The system is trained so that the first "light" turns on if the polymer sample is HDPE, the second "light" turns on for PET, and the third "light" turns on for PVC.

As shown in Fig. 4, the inputs are presented to two input layer neurons which pass [I.sub.1] and [I.sub.2] through weighted connections to two hidden layer neurons. Each hidden neuron has two input connections - a total of four connections from the input layer to the hidden layer. These connections have associated weights which can be arranged as a 2 x 2 real matrix, W1.

The [i.sup.th] hidden unit neuron computes a weighted sum of its inputs and subtracts a threshold [Theta] [1.sub.i],

[h.sub.i] = ([summation of] W[1.sub.t, j] [I.sub.j] where j = 1 to 2) - [Theta] [1.sub.i] (1)

Hidden unit i completes its operation by passing [h.sub.i] through a sigmoid nonlinear function producing an output value,

f([h.sub.i]) = 1/1 + exp([-h.sub.i]) (2)

In turn, these values are passed to three output layer neurons through six weighted connections represented as a 3 x 2 real matrix, W2. The outputs [O.sub.1], [O.sub.2], and [O.sub.3] are given by

[O.sub.i] = f{([summation of] W[2.sub.i, j] f([h.sub.j]) where j = 1 to 2) - [Theta][2.sub.i]} i = 1, 3 (3)

[O.sub.1], [O.sub.2], and [O.sub.3] can be regarded as the likelihood of a sample being PET, HDPE, or PVC, respectively. The network was "trained" by adjusting the weight matrices W1 and W2 so that a given input [I.sub.1] and [I.sub.2] produced a correct classification. Thirty-seven sets of sample data were selected for training purposes. This data consisted of measured values of [I.sub.1] and [I.sub.2] and associated desired outputs [O.sub.1], [O.sub.2], and [O.sub.3] (e.g., if the sample were HDPE, the desired outputs would be [O.sub.1] = 1.0, [O.sub.2] = [O.sub.3] = 0.0). The training data was repetitively presented to the network and the actual network outputs were compared with the desired output values. The difference between the actual and desired outputs was used to change the weight matrices W1 and W2 according to a proprietary ("stiff") training method (9).

After training was completed, the network was tested using the 22 samples not seen in the training phase. They were all classified perfectly, indicating that the network had learned to distinguish HDPE, PET, and PVC. A good gauge of a model's performance is the statistical [R.sup.2] parameter, which is the ratio of the variance of the model to that of the testing data. If [R.sup.2] is 1, the model is perfect; and if [R.sup.2] is less than 0.5, the model is very poor. For this problem, the trained neural network has [R.sup.2] = 0.9996. Therefore, we can be confident that the trained network will positively identify PET, HDPE, and PVC.

PROPOSED HARDWARE IMPLEMENTATION

The excellent results discussed in the previous section indicate that only three points of the NIR absorbance spectrum are needed to identify correctly PET, PVC, and HDPE. A continuous spectrum is not needed; therefore, the spectrometer that provides the absorbance data can be relatively simple. In this section, we suggest how to implement an instrument capable of identifying these three polymers.

Absorbance at a particular wavelength [Lambda] is defined (10) to be

A([Lambda]) = [-log.sub.10][I([Lambda])/[I.sub.0]([Lambda])] (4)

where [I.sub.0]([Lambda]) is the intensity of light incident on the sample and I([Lambda]) is the intensity of the light transmitted through it. The ratio I([Lambda])/[I.sub.0]([Lambda]) is defined as the transmission of the sample, so absorbance is determined by the common logarithm of the transmission of the sample. The wavelength can be selected by an appropriate interference filter, which has a narrow passband centered on the wavelength of interest. The present application requires three separate wavelength channels, so three filters are needed.

Figure 5 shows a simple three-filter spectrometer suitable for sorting crushed plastic bottles by type. A broadband source of near-infrared light illuminates the plastic sample via a fiber-optic bundle. We have determined in our laboratory that a 15-watt tungsten bulb generates sufficient broadband NIR illumination for this purpose. NIR sources are also available commercially. In any case, the color temperature of the source must be stabilized so that the spectral output remains constant. Matching lenses are used to collimate the light before it passes through the plastic sample. Collimating the light allows a separation of several inches between the fiber-optic bundle that illuminates the sample and the bundle that receives the transmitted light. This separation is necessary in sorting applications to provide room for passage of the crushed bottles. Note that some of the transmitted light will be scattered out of the receiving aperture by the bottles. In practical applications, efficient collection of the transmitted light can be problematic, but we have found that the simplicity of collimated optics justifies the inherent loss of the scattered light component. The use of fiber optics allows the instrument to be mounted away from the sorting chutes.

Light transmitted by the plastic bottle returns to the instrument via a fiber-optic bundle that splits into three separate bundles. Each of these smaller bundles is terminated by a filter/detector combination. The interference filters select the wavelengths of interest, so that the corresponding detector measures the light intensity at that wavelength only. We have experimented with InAs detectors (operating in DC mode at ambient temperature) in conjunction with 15 nm bandpass filters. In general, we have observed adequate signal strength, but there is a significant dependence of the dark current on temperature variations. The resulting shifts in signal level preclude reliable measurements in an industrial setting. For this reason, the detectors in Fig. 5 should be cooled. InAs detectors are available with two-stage thermoelectric coolers and would be suitable for use in this instrument. Alternatively, the light source could be modulated, and the detectors could be AC couped to the amplifiers. Synchronous detection would remove the problem of DC drift but would bring additional complexity to the device.

The outputs from the detectors are conditioned by logarithmic amplifiers in order to make the signal linear with respect to absorbance. The linearized signals are then digitized and recorded by the computer. Absorbance measurements require the intensity [I.sub.0]([Lambda]) of incident light to be recorded for all three wavelengths by measuring the detector signals without any sample in the light path. Subsequent readings with samples in place are then used to calculate the absorbance, as in Eq 4. The computer that controls the instrument is programmed to implement the neural network model that was described above. Once the absorbances have been calculated, the identification is performed by the neural network software. The result of the analysis determines the disposition of the bottle, which is routed to the appropriate chute by a mechanical device.

A simplified version of the device shown in Fig. 5 was built in our lab to demonstrate the concepts just discussed. That device was able to sort PET and PVC successfully: aside from some difficulties with the instrumentation (e.g., DC drift), the method has been shown to be feasible and inexpensive. Nonetheless, it should be stated that several vendors of fixed-filter NIR spectrometers manufacture devices similar to that in Fig. 5 (minus the neural network). It should also be noted that the neural network could be implemented either in software (running on the processor that controls the spectrometer) or in hardware (e.g., Inters 80170NX neural net IC). Therefore we do not advocate building our system from the component level; instead, Fig. 5 should be viewed as a conceptual guide for understanding the method and applying it to commercial instrumentation.

The instrument shown in Fig. 5 has no moving parts and is very compact. Compared to conventional scanning spectrometers, it should be inexpensive, compact, robust, and fast. With a fixed-filter design and a hardware implementation of the neural network, it should be possible to attain several thousand identifications per second.

CONCLUSION

We have proposed a near-infrared (NIR) spectroscopic method for quickly identifying polymer type in plastics recycling operations. The method is based on the observation that a relatively small number of selected wavelengths in the NIR spectrum are needed to correctly distinguish certain polymer types. A novel aspect of the data analysis is that a neural network is used to make the determination from absorption measurements at these wavelengths. This method has been used successfully to identify polymer type in samples of crushed, dirty bottles collected from a recycling plant. A proposed implementation of this scheme in hardware would use three stationary filters to provide the data needed to separate PET, PVC, and HDPE; it would be quite inexpensive and very fast.

A noteworthy feature of this method is that additional polymer types could be handled by retraining the network. In some cases it would not be necessary to install additional NIR channels; if however a large number of polymer types were to be identified, then additional channels would be needed. In any case, the neural network can be trained to accept new polymer types, whereas the typical decision-tree method of identification requires significant effort to add new types. It should also be noted that the time required for the neural network to identify the polymer is independent of the number of polymers it knows, in contradistinction to strategies that search through a library of known polymer types. The combination of a fixed-filter NIR spectrometer with a hardware-based neural network would provide a very fast yet flexible identification system suitable for plastic identification and sortation in recycling operations (11).

ACKNOWLEDGEMENTS

The NIR spectra for this work were provided by A. M. Brearley and her assistants; we wish to thank her not only for the data but also for several helpful discussions about NIR spectroscopy.

REFERENCES

1. In July 1991, virgin PET resin was selling for 67[cents]/lb, recycled PET was 38[cents]/lb and unprocessed PET bottles were 8[cents]/lb. Earlier in 1991 at the height of the Persian Gulf War, virgin PET resin was selling for 80[cents]/lb, recycled PET was 48[cents]/lb (Chem. Eng. News, July 8, 1991).

2. ASOMA Instruments Inc. (Watertown, Mass.) sells a radioisotope-based sensor (model 652-D) suitable for detecting PVC or other chlorine-bearing polymers. Automated Industrial Controls Inc. (Baltimore, Md.) is developing a FTIR-based system for sorting polymer materials.

3. C. S. Creaser and A. M. C. Davies, eds., Analytical Applications of Spectroscopy, The Royal Society of Chemistry, London (1988); D. L. Wetzel, Analytical Chem., 55, 1165A (1983).

4. E. W. Crandall and A. N. Jagtap, J. Appl. Polym. Sci., 21, 449 (1977).

5. K. F. Palmer and D. Williams, J. Opt. Soc. Am., 64, 1107 (1974).

6. W. S. McCulloch and W. Pitts, Bull. Math. Biophys., 5, 115 (1943).

7. An excellent introductory text is J. A. Hertz, A. S. Krogh, and R. G. Palmer, Introduction to the Theory of Neural Computation, Addison-Wesley, Reading, Mass. (1991).

8. D. E. Rumelhart, G. E. Hinton and R. J. Williams, Nature, 323, 533 (1986).

9. U.S. Patent No. 5,046,020 issued to D. L. Filkin, assigned to the DuPont Company (1991).

10. J. E. Stewart, Infrared Spectroscopy, Marcel Dekker, New York (1970).

11. Note added in proof: After this manuscript was written, we learned that Sandia National Laboratory has recently been working on a similar technique; see M. K. Alam, S. L. Stanton, and G. A. Hebner, Spectroscopy, 9, 31 (1994).
COPYRIGHT 1995 Society of Plastics Engineers, Inc.
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 1995 Gale, Cengage Learning. All rights reserved.

Article Details
Printer friendly Cite/link Email Feedback
Author:Scott, D.M.; Waterland, R.L.
Publication:Polymer Engineering and Science
Date:Jun 1, 1995
Words:3305
Previous Article:Photocrosslinking of polyethylene for production of thin wall insulated wire.
Next Article:Thermal studies of blends of poly(phenylene sulfide)(PPS) with poly(phenylene sulfide sulfone)(PPSS) and with poly(phenylene sulfide ether)(PPSE).
Topics:


Related Articles
Automatic flat dies gain artificial intelligence.
What plastic is it? New ways for recyclers to solve the mystery.
Paper or plastic? New test has the answer.
Laser-Induced Plasma Spectroscopy for Plastic Identification.
Plastics Recycling: Technology Chasing a Reluctant Market.
Derivation of optimal processing parameters of polypropylene foam thermoforming by an artificial neural network.
Artificial neural networks based fingerprint authentication with clusters algorithm.
An intelligent system for classification of the communication formats using PSO.
Wavelet neural network-based fault diagnosis in air-handling units.

Terms of use | Privacy policy | Copyright © 2019 Farlex, Inc. | Feedback | For webmasters