Hyperspectral Imaging Based Detection of PVC During Sellafield Repackaging Procedures

Traditionally, special nuclear material (SNM) at Sellafield has been stored in multilayered packages, consisting of metallic cans and an overlayer of plasticized polyvinyl chloride (PVC) as an intermediate layer when transitioning between areas of different radiological classification. However, it has been found that the plasticized PVC can break down in the presence of both radiation and heat, releasing hydrochloric acid, which can corrode these metallic containers. Therefore, internal repackaging procedures at Sellafield have focused recently on the removal of these PVC films from containers, where as much degraded and often adhered PVC as possible is manually removed based on visual inspection. This manual operation is time-consuming, and it is possible that residual fragments of PVC could remain, leading to corrosion-related issues in future. In this work, hyperspectral imaging (HSI) was evaluated as a new tool for detecting PVC on metallic surfaces. The Samples of stainless steel type 1.4404—also known as 316L, the same as is used to construct SNM cans—and PVC were imaged in our experiments, and support vector machine (SVM) classification models were used to generate detection maps. In these maps, pixels were classified into either PVC or 316L based on their spectral responses in the range 954–1700 nm of the electromagnetic spectrum. Results suggest that the HSI could be used for an effective automated detection and quantification of PVC during repackaging procedures, detection and quantification that could be extended to other similar applications.

a result of gas generation during storage [1]. Both these mechanisms potentially challenge packages' structural integrity. Sellafield Ltd. is currently constructing a new facility on-site, known as the Sellafield product and residue retreatment plant (SRP), to condition SNM packages for safe long-term storage.
Sellafield Ltd. considers that this potential degradation is a significant challenge and proposes to retrieve all the PVC containing packages from the store to undertake interim remediation. The operators take the plastic-coated can out of the overpack, remove as much PVC as possible, and then reseal it in an external can. Fig. 1 shows the images of an SNM can before and during the PVC removal process. The task is currently undertaken manually, and a human visual inspector decides the point at which no PVC remains on the can and the cleaning process can be stopped. This translates into time-consuming, subjective, manual operations, where residual fragments of PVC could remain on the can surface with subsequent corrosion risks for the future. An automated inspection tool able to effectively detect PVC on stainless steel 316L surface to a determined, acceptable level will provide a more robust, quantitative means of determining whether a can is clean, and may reduce the potential for further HClinduced corrosion in future years, thus leading to increased storage confidence and reduced risk of future rework.
Hyperspectral imaging (HSI) is a technology able to capture an image in hundreds of different contiguous wavelengths across the electromagnetic spectrum, providing fine spectral detail of the samples under study. The HSI data usually cover part of the near-infrared (NIR) spectrum and can be used to characterize several physical and chemical properties of the samples under inspection to detect features that would, otherwise, be invisible to the human eye.
Each pixel in an HSI image contains the spectral response or signature of the material captured by that pixel. In this context, a spectral signature, potentially unique for each material in nature, is a vector array made of hundreds of values representing the reflectance intensity for each wavelength or spectral band. With such comprehensive data, HSI technology has already been applied in a wide range of applications, including remote sensing [2], [3], food quality analysis [4], raw material classification [5], medical [6], [7], counterfeit detection [8], and others [9], [10], [11]. However, this technology is still expandable to many other fields. Particular applications of interest are the one proposed in [12], where HSI was used for plastic waste characterization, and the one in [13], where the same technology was used for the detection of fine metal particles in shredded electronic waste. In similar terms, HSI images could be used at Sellafield to detect PVC and residual PVC fragments on metallic surfaces, becoming a nonintrusive automated inspection tool if combined with appropriate signal and image processing algorithms and techniques.
In this work, the potential use of HSI to detect the presence of PVC on stainless steel 316L was evaluated. Experiments used real PVC and 316L samples, where a machine learning classifier known as support vector machine (SVM) [14], [15] was trained to classify the image pixels-based on the spectral information they contain-into PVC or 316L, leading to detection maps, which could be used for decision support in the PVC removal process. Different scenarios were tested,  which included PVC samples over a 316L background and vice versa. The results demonstrated the effectiveness of HSI in discriminating PVC from 316L, detecting and quantifying its presence.
The rest of this manuscript is organized as follows. Section II describes the PVC and stainless steel 316L samples used in the experiments, along with the HSI system used for data acquisition and the SVM classifier implemented to obtain the detection maps. Then, Section III presents the different experiments undertaken with related results. Finally, conclusions are drawn in Section IV.

II. MATERIALS AND METHODS A. Samples
Three different types of plasticized PVC in the form of flexible films were provided by Sellafield Ltd. for this study. These included not only PVC films in apparent good condition (or at least nonthermally degraded) but also PVC films thermally degraded at the temperatures of 85 • C and 100 • C. These PVC films (see Fig. 2) were cut down into smaller pieces of different sizes for the experiments.
Stainless steel samples were also provided by Sellafield Ltd. These were sheets of austenitic stainless steel type 1.4404 (known as 316L) in a mesh form. This mesh structure was preferred for the experiments simply because it is easier to handle (cutting down into smaller items required no machinery). A 2-× 2-cm sample of 316L mesh is shown in Fig. 2(d).

B. HSI System
The system used to capture hyperspectral images was the RedEye 1.7 from INNO-SPEC [16] [see Fig. 3(a)], a camera with InGaAs detector able to acquire spectral data in the NIR range. After discarding some noisy bands inherent to the camera operation, a final range of 954-1700 nm was available, covered by a total of 236 spectral bands at a spectral resolution of around 3.2 nm. A system covering this spectral range was selected for PVC detection as, according to the literature, plastics exhibit clear features across the NIR range [12], [17], [18], [19].
The standoff distance between the camera and the samples was set at 50 cm, leading to a spatial resolution of ≈1 mm 2 per pixel. The spatial size of the captured images varied depending on the experiment, between 60 × 60 and 110 × 190 pixels after cropping to the area of interest, i.e., the size of the scenes undergoing inspection varied between 6 × 6 and 11 × 19 cm.
The system required two sets of halogen lamps with reflective boards [see Fig. 3(a)] as light source for homogeneous illumination of the samples, where reflectance intensity per pixel was obtained following a standard calibration procedure with dark and white reference images [20]. This calibration used a white tile made of Spectralon for white reference, normalizing reflectance in the range [0-1] (0-black and 1-white). Images were captured by line scanning (pushbroom technique [20], [21]), where the linear translational stage shown in Fig. 3 (Zolix TSA200-BF [22]) was used to introduce the relative movement between camera and samples. All images were captured indoors under the same conditions. It is worth mentioning here that deployment in the real-world would require a rotational stage applied to the can, rather than a linear stage, to line scan the surface of the cans. However, as the hyperspectral system works based on line scanning, the same approach and calibration procedure would apply.

C. Data Classification
Detection maps indicating the presence or absence of PVC were generated by SVM [14]. This machine learning technique exploits a margin-based criterion and has been widely used for hyperspectral data classification, avoiding the curse of dimensionality or Hughes phenomenon [23]. SVM is a supervised classifier, requiring labeled samples during the training process to build an SVM model, and can be implemented by several libraries that are publicly available [24], [25], [26]. In this work, the popular LibSVM library [24] was selected for SVM implementation.
In the experiments, SVM used a Gaussian kernel, tuning the related "penalty" and "gamma" parameters by means of a grid search procedure, with an internal validation using twofold cross validation. An SVM model was obtained by supervised training on 316L and PVC extracted features, leading to a two-class model able to classify the pixels in the images into either PVC or 316L. Visualization, processing, and classification of data were implemented in MATLAB (version R2017b), using a computer with Windows 10 (64 bit), i7 2.8GHz, four core, and 16-GB RAM. With these specifications, the generation of detection maps required between 5 and 10 s.

A. Spectral Response of PVC and Stainless Steel 316L
The first experiment conducted was aimed at evaluating and comparing the spectral response of PVC and 316L mesh in the NIR range to ensure it would be possible to characterize and differentiate these two materials. Fig. 4 shows plots that allow a comparison between the spectral responses of 316L and PVC [the hyperspectral system captured an image of a small piece of 316L on top of a PVC layer, see Fig. 4(a) and (b)].
As can be seen in Fig. 4, the 316L response is relatively flat (no significant peaks) across the spectral range of the sensor, while PVC exhibits some curves around 1200, 1400, and 1650 nm, which can be used to differentiate it from 316L. These findings correlate with the literature, where a similar In these images, PVC looks darker than 316L due to the feature extraction.

B. Processing Data for Amplifying Discriminative Features
Based on the spectral responses shown in Fig. 4, there are some clear spectral features that could be used to distinguish PVC from stainless steel 316L. Data processing for feature extraction was implemented to exploit and amplify these features effectively, improving subsequent data analysis and automated classification by SVM.
The feature extraction proposed in this work was based on two different techniques: 1) singular spectrum analysis (SSA) [28] and 2) first derivative computation. SSA is a technique traditionally used for time-series analysis, which can decompose a 1-D signal into the following: 1) main trend; 2) periodic components; and 3) noise [29], [30], [31]. Therefore, SSA was selected in order to remove the noise and highfrequency content from the original spectral responses [28], extracting only their main trends.
SSA was configured based on [28] (window size of ten elements, selecting the first component in the eigenvalue decomposition). After that, the first derivative of the SSA main trends was computed to obtain the final features (this was implemented using the MATLAB "gradient" function). The first derivative is expected to intensify those regions in which PVC and 316L become more distinct. This procedure was applied to all the pixels before undergoing classification. Fig. 5(a) shows the resulting spectral responses after feature extraction for both PVC and 316L. While the 316L features are relatively flat across the NIR spectrum, PVC features present some peaks, which are clearly identifiable. Fig. 5(b) shows a conventional RGB image containing PVC samples on top of a 316L layer (top) alongside respective spectral images from our HSI dataset at 1380 nm (middle) and 1651 nm (bottom) after feature extraction. As can be seen, the processed spectral images show significantly higher contrast between the 316L background and the PVC samples when compared with the RGB data. In fact, while seeing some of these PVC samples in the RGB image can be difficult (see highlighted area), they are easily visible in the spectral images.

C. Detection Maps
The resulting spectral responses after feature extraction were used to train and validate an SVM model able to classify any pixel in a hyperspectral image into either 316L or PVC. A single image (see Fig. 4) was used for training, where a 10 × 10-pixel area for 316L and another 10 × 10-pixel area for PVC were selected (after feature extraction) to train and generate the SVM model. A total of 200 pixels is a relatively small amount of training data; however, this modest training was selected to demonstrate the capabilities of the model. Significantly more training data can be gathered in future should there be a desire to adopt this technology for practical use.
Two validation experiments were carried out after training the model. In these experiments, new hyperspectral images, unseen during the training, were evaluated by the model.
In Experiment #1, a total of three 316L samples were placed on top of a layer of PVC (several individual PVC sheets were used to form this layer), making the PVC the background of the image acquired. In Experiment #2, 316L was used as the background of the image, where small pieces of PVC (a total of eight) were placed on top of the 316L sample. The detection maps obtained in both experiments are shown in Figs. 6 and 7. As shown in the figures, the SVM model was able to detect the three 316L stainless steel samples in Experiment #1 (Fig. 6) and the eight PVC samples in Experiment #2 (Fig. 7), with some classification statistics available in Table I.

D. Spectral Response of Thermally Degraded PVC
One of the reasons why PVC coatings are being currently removed from the metallic cans is that thermal degradation of PVC under heat could be considered a corrosion agent. Therefore, during removal procedures, operators are expected to find PVC at different stages of thermal degradation. In this section, the spectral response of PVC in good condition or at least nonthermally degraded (as used in Sections III-A-C) is compared with the response of PVC thermally degraded at different temperatures (85 • C and 100 • C). Fig. 8(c) shows the spectral response of nonthermally degraded and thermally degraded PVC, from a random pixel within each highlighted area. As can be seen, the spectral response in the NIR range (954-1700 nm) is almost identical for the three cases. Only small variations in the spectral intensity were found, but these are likely to be due to the different spatial location of the samples in the image, which slightly affects the amount of reflectance light received by the camera.

E. Detection Maps Including Thermally Degraded PVC
Final experiments included thermally degraded PVC to assess the performance of the SVM model in the presence of these degraded samples. The same model used in Experiments #1 and #2 was also used here, i.e., the model was not further trained with any thermally degraded PVC samples. The reason for this is that, according to Section III-D, the responses for thermally degraded and nonthermally degraded PVC (as used in training) are nearly identical.
In Experiment #3, a total of four 316L samples were placed on a background of PVC (layer made of several PVC sheets), and in Experiment #4, 316L was used as the background of the image again, placing small pieces of PVC (a total of eleven) on top of the 316L. Therefore, the number of samples was increased with relation to previous experiments, while the size of them was reduced to challenge the detection. PVC used here included the nonthermally degraded and thermally degraded types (at 85 • C and 100 • C).
Detection maps in Figs. 9 and 10 demonstrate the robustness of the proposed method, where all samples were detected (see Table II), even the smallest one occupying only a few pixels in    6 and 7, and 9 and 10), it is clear that all PVC samples were detected and highlighted, which is the main purpose of this application. However, the exact quantification of the detection accuracy in terms of number of pixels was not possible, as the available ground truth was subjective and based on human observation. Thus, a visual comparison of input and output is provided as opposed to more quantitative analysis, e.g., accuracy, precision, recall, and so on, which will be the focus of future work. One of the challenges associated with this application is the spatial resolution of the hyperspectral images and the related pixel mixing effect in the boundaries of the samples. Lower resolution implies that the exact shape of the samples may be not captured accurately. However, as long as the size of the PVC samples is above the limit of detection, they will be highlighted. The resolution in this application was ≈1 mm 2 per pixel, which covers samples that small. Smaller limit of detection could be achieved by reducing the standoff distance between camera and samples.

IV. CONCLUSION
The SNM packages containing plasticized PVC have been used at Sellafield Ltd. for decades. However, when this plasticized PVC is exposed to radiation and/or heat, it degrades releasing corrosive chloride products (e.g., HCl), and Sellafield Ltd. is undertaking procedures for the removal of this PVC coating. At present, these removal operations are conducted manually with the risk of missing residual PVC fragments through human visual inspection. In this work, HSI combined with intelligent signal processing has been evaluated as potential tool for automated and effective PVC detection using the spectral features of both stainless steel 316L and PVC materials for their detection and differentiation.
An HSI system capturing data in the NIR range (954-1700 nm) was used to investigate potential features able to discriminate PVC from 316L, where key differences were found in the regions around 1200, 1400, and 1650 nm. These key features were processed and amplified by a feature extraction method based on the first derivative of the SSA main trend. Then, a two-class SVM model, trained on the processed features, was proven effective in distinguishing between PVC and 316L samples under diverse conditions. These conditions included different numbers of samples, different sample sizes, random location of samples, and different PVC states (nonthermally degraded PVC and thermally degraded PVC at 85 • C and 100 • C).
These results show that HSI could be used for effective PVC detection, quantification, and decision support in the PVC removal process. Next steps include the following: 1) a quantitative evaluation of the accuracy of detection (pixel-wise); 2) moving from laboratory-based experiments to on-site evaluation before implementation for industrial validation; and 3) optimizing the data acquisition and analysis processes for real-time operation.