EVA and POE composite material irradiation quality evaluation method based on near infrared hyperspectrum

Abstract

This invention relates to the field of materials measurement and analysis technology, and particularly to a method for assessing the irradiation quality of EVA / POE composite materials based on near-infrared hyperspectral imaging. Addressing the dual dominance of microscopic crosslinking / degradation competition mechanisms and macroscopic spatial heterogeneity in the quality of irradiated modified materials, a multi-index comprehensive evaluation system integrating microscopic chemical characteristics and macroscopic spatial homogeneity is constructed. Its innovation lies in: utilizing spectral decoupling technology to separate competing reaction signals and quantitatively analyze microscopic chemical evolution; transforming spatial heterogeneity into statistical indicators, filling the technical gap in quantifying macroscopic dose distribution; and achieving deep integration of "spectral-spatial" dual-dimensional characteristics through a nonlinear model. This invention, through the synergistic complementarity of the above-mentioned multi-dimensional information, achieves a multi-dimensional, non-destructive quantitative assessment of the quality of EVA / POE composite materials after irradiation, effectively solving the problem that single-dimensional indicators cannot accurately reflect the true irradiation quality of materials.

Description

Technical Field

[0001] This invention relates to the field of materials science, which studies or analyzes materials by measuring their chemical or physical properties, and particularly to a method for evaluating the irradiation quality of EVA and POE composite materials based on near-infrared hyperspectral imaging. Background Technology

[0002] Ethylene-vinyl acetate (EVA) and polyolefin elastomer (POE) blends are widely used in high-performance footwear materials and other fields, and electron beam irradiation is a key modification method to improve their mechanical properties. However, the irradiation modification process is complex, microscopically controlled by the competitive balance between "crosslinking enhancement" and "degradation damage," and macroscopically limited by electron beam energy decay and material thickness fluctuations, easily leading to spatial non-uniformity of absorbed dose. This dual uncertainty of microscopic chemical structure evolution and macroscopic spatial distribution directly affects the consistency of product performance.

[0003] Currently, irradiation quality assessment mainly relies on methods such as gel content analysis, mechanical tensile testing, or Fourier transform infrared spectroscopy (FTIR). These methods are mostly destructive tests, cumbersome and time-consuming, and cannot meet the needs of industrial online testing. More importantly, these methods are limited to single-point sampling and cannot obtain spatial distribution information of the material across the entire field, making it difficult to assess the overall uniformity of irradiation treatment.

[0004] While near-infrared hyperspectral imaging (NIR-HSI) offers advantages such as being non-destructive, rapid, and combining image and spectrum, its application in evaluating irradiation modification still has limitations: First, conventional modeling strategies struggle to accurately decouple the competing mechanisms of crosslinking and degradation, resulting in weak physical interpretability; second, existing applications largely remain at the qualitative image display level, lacking quantitative indicators of spatial heterogeneity and failing to accurately reflect the true quality fluctuations of large-size industrial samples. Therefore, establishing an evaluation system capable of simultaneously characterizing microscopic chemical evolution and macroscopic spatial homogeneity is particularly important.

[0005] In summary, establishing an evaluation system capable of simultaneously characterizing microscopic chemical evolution and macroscopic spatial distribution has become an urgent need in the industry. Addressing the challenge that existing single models cannot accurately describe the complex chemical competition mechanisms and spatial inhomogeneities during irradiation, this invention proposes a multi-dimensional fusion solution: by constructing a comprehensive model integrating crosslinking / degradation spectral characteristics and spatial uniformity parameters, accurate evaluation of the irradiation quality of EVA and POE composite materials is achieved. Summary of the Invention

[0006] To address the problem that existing evaluation methods cannot simultaneously consider both microscopic chemical competition mechanisms and macroscopic spatial heterogeneity, leading to inaccurate evaluation results, this invention provides a method for evaluating the irradiation quality of EVA and POE composite materials based on near-infrared hyperspectral imaging.

[0007] To solve the above-mentioned technical problems, the solution adopted by the present invention is: a method for evaluating the irradiation quality of EVA and POE composite materials based on near-infrared hyperspectral imaging, characterized by comprising the following steps:

[0008] Step 1: Acquire hyperspectral images:

[0009] In a darkroom environment, a line scanning device was used to acquire hyperspectral images. The line scanning device included a camera, a halogen lamp light source, and a conveyor belt. The EVA and POE composite material sample was placed on the conveyor belt and photographed under illumination. The spectral images were recorded in the range of 900–1700 nm. The generated hyperspectral data cube had a spatial dimension of 640 × 1800 pixels and contained 224 spectral bands.

[0010] Using dark current Correction and White Reference The dual correction mechanism, combining calibration and verification, will adjust the original intensity data. Convert to relative reflectance As shown in the formula below:

[0011] ;

[0012] After reflectance correction, the region of interest is defined and spectral features are extracted. The extracted spectral features include the intensity of the crosslinking characteristic band. Degradation characteristic band intensity and auxiliary band intensity ;

[0013] Step 2: Spectral preprocessing and feature extraction:

[0014] The hyperspectral data were first preprocessed using multivariate scattering correction, followed by spectral smoothing and signal enhancement using Savitzky–Golay first-derivative filtering. The 0 kGy spectrum was subtracted from the average spectrum of each irradiation dose group to generate the average difference spectrum. Then, the peaks, troughs, and zero-point crossings in the Savitzky–Golay first-derivative spectrum were identified as characteristic spectra. These characteristic spectra are specific spectral markers of irradiation-induced chemical changes and can distinguish overlapping functional group vibrations.

[0015] Step 3: Development of a comprehensive quality scoring model, including the following steps:

[0016] Step 3.1, Spectral Feature Verification:

[0017] After identifying the characteristic spectra, a spectral feature verification model was established to confirm the quantitative correlation between the characteristic spectra and absorbed dose and tensile strength. Specifically, firstly, an absorbed dose model was constructed based on partial least squares regression, and secondly, a tensile strength correlation model was constructed based on multiple linear regression, as follows:

[0018] First, partial least squares regression is used to establish a quantitative mapping relationship between characteristic spectral intensity and actual absorbed dose. The high-dimensional spectral independent variable matrix X and the dose dependent variable matrix Y are simultaneously projected into a low-dimensional latent variable space. This extracts the main variation information of the data while maximizing the covariance between the independent and dependent variables, thereby overcoming multicollinearity and filtering out spectral noise. The partial least squares regression algorithm performs bilinear decomposition on the spectral matrix X and the dose matrix Y, as shown in the following formula:

[0019] ;

[0020] ;

[0021] Where T and U are the latent variable matrices of X and Y, respectively; P and Q are the loading matrices of X and Y, respectively; and E and F are the residual matrices of X and Y, respectively.

[0022] Based on the above decomposition, a linear regression model of X on Y is established, as shown in the following formula:

[0023] ;

[0024] In the formula, βPLSR is the regression coefficient matrix; F* is the residual matrix of the linear regression model, representing the deviation between the actual measured value and the model prediction value; the model parameters are optimized based on leave-one-out cross-validation, with the goal of minimizing the root mean square error of the prediction. Finally, the optimal number of latent variables is determined to be 11, and the absorbed dose prediction model is constructed.

[0025] Secondly, multiple linear regression was used to quantify the independent contribution of specific spectral markers to the change in macroscopic tensile strength of the material, establishing a linear combination relationship between multiple independent variables and one dependent variable, thereby verifying whether the chemical changes in the selected bands are the direct cause of the change in mechanical properties; the verification model is shown in the following formula:

[0026] ;

[0027] In the formula, ΔY represents the change in tensile strength; n represents the number of characteristic spectral markers; It is the intensity change of the i-th characteristic spectral marker; It is the partial regression coefficient corresponding to the i-th feature; It is the intercept term; ε represents the model residuals;

[0028] Step 3.2, Spatial Uniformity Quantification:

[0029] Utilizing the "image-spectrum integration" characteristic of near-infrared hyperspectral imaging technology, the traditional single-point dose assessment is upgraded to a two-dimensional spatial distribution assessment. The specific operation is as follows:

[0030] The absorbed dose model trained in step 3.1 is applied to the hyperspectral data of each pixel within the region of interest of each sample. For each coordinate point in the image, its hyperspectral data is input into the absorbed dose model to calculate the predicted dose value corresponding to each coordinate point, thereby converting the original hyperspectral image into a dose quantization distribution map with spatial resolution. Based on the generated dose quantization distribution map, the predicted dose values ​​of all pixels are extracted, and their statistical indices are calculated to quantify spatial uniformity. The coefficient of variation (CV) is introduced as a spatial uniformity index. The larger the CV value, the greater the difference in absorbed dose at various points on the sample surface and the worse the spatial uniformity; conversely, a smaller CV value indicates more uniform irradiation. The calculation formula is shown in the following formula:

[0031] ;

[0032] Wherein, SD refers to the dispersion of the predicted dose of all pixels within the region of interest of the sample relative to the average value, and is calculated using the following formula: , where n is the total number of pixels; It refers to the arithmetic mean of the predicted doses of all pixels in the region of interest of the sample, representing the overall dose level received by the sample;

[0033] Step 3.3: Construct a comprehensive quality scoring model:

[0034] The random forest regression algorithm is used, taking the chemical characteristic indicators extracted from the spectrum and the CV value calculated in step 3.2 as input, and outputting the change in the macroscopic tensile strength of the material. A nonlinear prediction model was constructed; the chemical characteristic index is specifically the intensity of the cross-linking characteristic band. Degradation characteristic band intensity and auxiliary band intensity During model training, the average impurity reduction mechanism built into the random forest algorithm is used for feature importance analysis. This is achieved by calculating the cross-linked feature band intensity across all decision trees in the random forest. Degradation characteristic band intensity and auxiliary band intensity The contribution of this feature to the tensile strength prediction result is quantified by the sum of the reduction in mean square error brought about by node splitting.

[0035] Step 3.4: Systematically evaluate the above spectral feature verification model and comprehensive quality scoring model to ensure the accuracy and reliability of the evaluation method.

[0036] For the absorbed dose model, tensile strength model, and comprehensive quality scoring model, the coefficient of determination was calculated using leave-one-out cross-validation. Root mean square error and mean absolute error Assessment purpose Used to characterize the model's ability to explain data variations and to confirm whether the extracted spectral features can truly reflect dose changes and mechanical property changes; and The prediction error used to quantify the model is shown in the following formula:

[0037] ;

[0038] ;

[0039] ;

[0040] In the formula, This represents the actual measured value of the i-th sample; This represents the model prediction value for the i-th sample; This represents the arithmetic mean of the true values ​​of all samples.

[0041] A further improvement is made to the specific operation of defining the region of interest and extracting spectral features in step one, which is as follows:

[0042] By utilizing the difference in reflectance between the sample and the background conveyor belt in a specific band, an intensity threshold is set to generate a binary mask, which automatically segments the pixel region of the EVA / POE composite material sample, removes background noise interference, and extracts spectral features from each pixel point in the mask region as an independent spectral sample.

[0043] The specific implementation of the threshold-based mask method for separating the background conveyor belt and the EVA / POE composite material sample is as follows:

[0044] Set up a hyperspectral image Its size is Where M and N are the number of rows and columns of the image, respectively, and L is the number of spectral bands; each pixel Indicates position In the The spectral values ​​of each band are analyzed; the spectral features of each pixel are observed to identify the regions where the spectral values ​​of the background and target areas differ significantly, and then a threshold is set for filtering; specific spectral features include:

[0045] Mean spectrum: ;

[0046] Standard deviation spectrum: ;

[0047] Spectral values ​​of a specific band: directly select a specific band. spectral values The threshold setting section is divided into global thresholds. or multiple local thresholds The global threshold is ,in and These are the mean and standard deviation of the background, respectively. It is a constant; the local threshold is ,in and These are the mean and standard deviation of the local background.

[0048] After separating the background conveyor belt from the EVA and POE composite sample, the updated pixels have a mask value of 1, while the pixels with a mask value of 0 remain unchanged.

[0049] A further improvement is made: the specific steps in step 3.2 of converting the original hyperspectral image into a dose quantization distribution map with spatial resolution are as follows:

[0050] Step 3.2.1: Extract each spatial pixel from the corrected hyperspectral data one by one. Complete spectral curve The image resolution is 640×1800, and 1,152,000 independent spectral curves are processed.

[0051] Step 3.2.2: For each extracted spectral curve Perform the preprocessing and feature extraction operations in step two above, and calculate the spectral feature values ​​of the pixels for the extracted and identified feature bands;

[0052] Step 3.2.3: Use the partial least squares regression model: The single-pixel spectral features obtained in step 3.2.2 are then... Input the model to calculate the predicted radiation dose value for that pixel. Repeat this operation for each pixel in the hyperspectral image to convert the spectral intensity matrix into a dose value matrix.

[0053] Step 3.2.4: Map the obtained dose value matrix to the RGB color space to generate a pseudo-color image, and obtain a dose quantization distribution map with spatial resolution.

[0054] By adopting the aforementioned technical solution, the beneficial effects of the present invention are:

[0055] This invention, through the innovative integration of near-infrared hyperspectral imaging technology and multi-algorithm modeling, successfully solves many pain points in the traditional irradiation quality assessment of EVA and POE composite materials, achieving a leapfrog upgrade from "local destructive detection" to "full-space non-destructive and accurate assessment." Its beneficial effects are mainly reflected in the following aspects:

[0056] I. Breaking through the limitations of traditional technologies to achieve non-destructive rapid testing

[0057] Compared to traditional destructive assessment methods such as gel content analysis, tensile testing, and FTIR, this invention employs a non-destructive testing mode using near-infrared hyperspectral imaging. This eliminates the need for cumbersome pretreatment or destructive sampling of EVA and POE composite material samples, fundamentally avoiding sample loss during the testing process. It is particularly suitable for quality sampling and full inspection of finished and semi-finished products. Furthermore, through a line scanning device and automated data acquisition process, combined with efficient spectral preprocessing algorithms, the full spectrum acquisition and analysis of a single sample can be completed in a very short time, significantly shortening the testing cycle. This solves the core problems of long processing times and low efficiency associated with traditional methods, providing efficiency support for industrial-scale batch testing.

[0058] II. Filling the gap in spatial uniformity assessment to achieve full-dimensional quality characterization

[0059] Addressing the critical limitation of traditional techniques that can only perform local sampling and cannot assess the spatial uniformity of irradiation dose, this invention innovatively combines the "spectral dimension" and "spatial dimension" of hyperspectral imaging. Unlike existing hyperspectral techniques that typically only use spatial information as a qualitative, visualized image (mapping), this invention further explores the in-depth value of the spatial dimension by parameterizing spatial uniformity characteristics. A spatially resolved dose distribution map is generated through a pixel-by-pixel dose quantification model, and the coefficient of variation (CV) is extracted as a key quantitative parameter input to the model. This breakthrough enables the evaluation system to simultaneously cover the two core quality influencing factors: "chemical equilibrium of crosslinking and degradation" and "spatial dose uniformity," filling the gap in existing technologies that cannot achieve full-dimensional quality assessment. Experimental results have confirmed that this method can clearly distinguish the spatial uniformity differences of different formulation materials (e.g., the EVA7470 series exhibits severe heterogeneity with a CV > 33% in the 30-60 kGy dose range; the EVA7350 series shows good uniformity with a CV < 25%).

[0060] III. Construct a multi-dimensional comprehensive evaluation model to improve the accuracy and reliability of quality prediction.

[0061] This invention establishes a multi-dimensional evaluation framework covering both chemical and spatial characteristics through a progressive design of "spectral feature verification - spatial homogeneity quantification - comprehensive model construction." Unlike conventional hyperspectral applications that primarily predict static, independent component contents (such as moisture and sugar content), this invention establishes a specific spectral fingerprint mapping to address the complex "dynamic competition mechanism" during irradiation modification—the interplay between molecular crosslinking (enhancing performance) and degradation (damaging performance). First, the absorbed dose model constructed using partial least squares regression achieved a validation set determination coefficient R² of 0.973 and a root mean square error (RMSE) of only 4.111 kGy, accurately retrieving the actual irradiation dose of the material. Second, the multiple linear regression model demonstrated a strong correlation between spectral characteristics and tensile strength (validation set R² = 0.939), ensuring a precise correspondence between chemical characteristics and macroscopic properties. Finally, a comprehensive quality scoring model based on the random forest algorithm integrated dynamically changing chemical characteristic indicators with quantified spatial homogeneity parameters. Through leave-one-out cross-validation, the model achieved predictive correlations of 0.987 and 0.931 for the tensile strength changes of EVA7470 / POE and EVA7350 / POE materials, respectively, enabling accurate ranking of the final mechanical properties of the materials. Simultaneously, feature importance analysis using the RF algorithm clarified that spatial homogeneity is a key factor dominating quality (contributing 40.4%-43.8%), providing clear guidance for optimizing irradiation processes and enhancing the scientific rigor and guidance of quality assessment.

[0062] IV. Promote the upgrading of irradiation modification quality assessment technology to help optimize processes and control costs.

[0063] The implementation of this invention not only provides a novel technical solution for the irradiation quality assessment of EVA and POE composite materials, but also promotes the iterative upgrading of assessment technologies in the field of irradiation modification of polymer materials. By accurately quantifying the synergistic effects of chemical changes and spatial homogeneity, it can guide enterprises to optimize electron beam irradiation process parameters (such as dose gradient and irradiation intensity distribution), reducing material performance non-compliance caused by improper processes, and lowering scrap rates and production costs. Simultaneously, the non-destructive testing mode avoids sample waste caused by traditional destructive testing, further reducing quality control costs, and has significant economic value and industry-wide application significance. Attached Figure Description

[0064] Figure 1 This is a schematic diagram of a near-infrared hyperspectral imaging system used for spectral acquisition.

[0065] Figure 2 This is a visualization histogram of the raw hyperspectral data values ​​of the 79th band of the pixel.

[0066] Figure 3 , 4The Savitzky-Golay first derivative spectrum of the irradiated EVA / POE composite material shows the characteristic band diagrams of the crosslinking and degradation mechanisms.

[0067] Figure 5 This is a performance graph of the PLSR prediction model based on near-infrared spectral characteristics.

[0068] Figure 6 This is a performance graph of the tensile strength multiple linear regression (MLR) prediction model based on spectral characteristics.

[0069] Figure 7 This is a spatially resolved dose quantization plot (Gaussian smoothed) at a dose level of 30 kGy. The top two samples (EVA7470 series) exhibit severe heterogeneity, while the bottom two samples (EVA 7350 series) are relatively homogeneous. This visualization directly confirms the quantitative CV analysis in Table 4.

[0070] Figure 8 shows the correlation analysis of leave-one-out cross-validation between the comprehensive quality scoring model evaluation results and the measured tensile property rankings. (Left) EVA7350 / POE (R² = 0.931); (Right) EVA7470 / POE (R² = 0.931). 2 = 0.987).

[0071] Figure 9 shows the feature importance analysis from the random forest (RF) model, quantifying the contributions of spectral indices (crosslinking, degradation, and auxiliary) and spatial uniformity index. Detailed Implementation

[0072] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments.

[0073] refer to Figures 1 to 9 The embodiments of this invention disclose the specific process of an irradiation quality assessment method for EVA and POE composite materials based on near-infrared hyperspectral imaging.

[0074] Preparation before the procedure: materials and irradiation

[0075] The EVA / POE composite material was synthesized and supplied by Maotai (Fujian) New Material Technology Co., Ltd. Two formulations were used in this study: EVA7470 / POE (EVA with 26% VA content) and EVA7350 / POE (EVA with 18% VA content). All samples were 60×15×20mm in size and were prepared by melt blending at a mass ratio of 70% EVA and 30% POE.

[0076] Electron beam irradiation was performed using an electron accelerator with an energy of 10.1 MeV and a power of 15 kW. Eight dose gradients were set up: 0 kGy as a blank control group, and dose points ranging from 20 kGy to 80 kGy (in 10 kGy intervals). After irradiation, the final tensile strength was determined using an electron universal testing machine according to GB / T1040.3-2006 standard. In this study, the nominal irradiation dose (kGy) applied by the accelerator was used as a reference value (calibration standard) for quantifying the average absorbed dose of the samples.

[0077] Step 1: Acquire hyperspectral images:

[0078] The near-infrared hyperspectral imaging system operates in reflective mode and employs a pushbroom configuration. The device includes a SPECIMFX 17 camera, a halogen lamp light source, and a conveyor belt; a schematic diagram of the system is shown in Figure 1.

[0079] All hyperspectral images were acquired in a darkroom environment. Key imaging parameters were as follows: distance from the light source to the sample 300 mm, distance from the camera to the sample 500 mm, exposure time 2 ms, and conveyor belt speed 0.75 cm / s. Spectral images were recorded in the range of 900–1700 nm, and the resulting hyperspectral data cube had a spatial dimension of 640 × 1800 pixels, containing 224 spectral bands.

[0080] Raw intensity data Through dark current Reference to White (Teflon plate) Image converted to relative reflectance As shown in formula (1):

[0081] ;

[0082] After reflectivity correction, the regions of interest (ROIs) are defined.

[0083] The specific steps for defining the region of interest and extracting spectral features are as follows: Utilizing the difference in reflectance between the sample and the background conveyor belt at a specific wavelength of 1210 nm, a binary mask is generated by setting an intensity threshold. This automatically segments the pixel regions of the EVA and POE composite material samples, removes background noise interference, and treats each pixel within the mask region as an independent spectral sample for spectral feature extraction. The extracted spectral features include the intensity of the cross-linking characteristic band. Degradation characteristic band intensity and auxiliary band intensity ;

[0084] The specific implementation of the threshold-based mask method for separating the background conveyor belt from the EVA and POE composite material sample is as follows:

[0085] Set up a hyperspectral image Its size is Where M and N are the number of rows and columns of the image, respectively, and L is the number of spectral bands. Each pixel Indicates position In the The spectral values ​​of each band are analyzed; the spectral features of each pixel are observed to identify the regions where the spectral values ​​of the background and target areas differ significantly, and then a threshold is set for filtering; specific spectral features include:

[0086] Mean spectrum: ;

[0087] Standard deviation spectrum: ;

[0088] Spectral values ​​of a specific band: directly select a specific band. spectral values The threshold setting section is divided into global thresholds. or multiple local thresholds The global threshold is ,in and These are the mean and standard deviation of the background, respectively. It is a constant; the local threshold is ,in and These are the mean and standard deviation of the local background.

[0089] After separating the background conveyor belt from the EVA / POE composite sample, pixels with a mask value of 1 are updated, while pixels with a mask value of 0 remain unchanged.

[0090] After separating the background and target using a threshold-based masking method, the background image needs to be updated. Pixels with a mask value of 1 need to be updated, while pixels with a mask value of 0 remain unchanged. The specific implementation formula is shown below:

[0091] ;

[0092] This study used a white background to acquire hyperspectral data from a gallium arsenide epitaxial wafer and visualized the raw hyperspectral data values ​​of the 79th band of each pixel using a histogram. The results showed that the gallium arsenide epitaxial wafer and the background could be effectively separated by setting a global threshold of 450. Specific results are shown below. Figure 2 As shown.

[0093] Step 2: Spectral preprocessing and feature extraction:

[0094] Hyperspectral data preprocessing aimed to reduce baseline drift and correct for scattering effects. Data were first preprocessed using multivariate scattering correction (MSC), with Savitzky–Golay filtering employing an 11-point window and a second-order polynomial, determined through cross-validation to minimize noise interference, and performing spectral smoothing and signal enhancement. To highlight irradiation-induced changes, average difference spectra were generated by subtracting the 0 kGy spectrum from the average spectrum of each irradiation dose group. Then, peaks, troughs, and zero-point crossings in the SG derivative spectrum were identified as characteristic bands. These features are specific spectral markers of irradiation-induced chemical changes, capable of distinguishing overlapping functional group vibrations.

[0095] Step 3: Development of a comprehensive quality scoring model, including the following steps:

[0096] Step 3.1, Spectral Feature Verification:

[0097] After identifying the characteristic spectral bands associated with irradiation effects, mathematical models need to be established to verify the quantitative correlation between these microscopic spectral features and physical causes (absorbed dose) and macroscopic properties (tensile strength). In this step, partial least squares regression (PLSR) and multiple linear regression (MLR) are used to construct the verification models.

[0098] (1) Construction of absorbed dose model

[0099] First, partial least squares regression is used to establish a quantitative mapping relationship between characteristic spectral intensity and actual absorbed dose. Due to the high degree of collinearity (i.e., strong correlation between bands) among the extracted spectral characteristic bands, traditional linear regression methods are prone to failure. The partial least squares regression algorithm simultaneously projects the high-dimensional spectral independent variable matrix X and the dose dependent variable matrix Y into a low-dimensional latent variable (LVs) space. This extracts the main variation information of the data while maximizing the covariance between the independent and dependent variables, thereby effectively overcoming multicollinearity and filtering out spectral noise. The partial least squares regression algorithm performs bilinear decomposition on the spectral matrix X and the dose matrix Y, as shown in formulas (2) and (3):

[0100] ;

[0101] ;

[0102] Where T and U are the score matrices (i.e., latent variable matrices) of X and Y, respectively; P and Q are the loading matrices of X and Y, respectively; and E and F are the residual matrices of X and Y, respectively, representing the remaining variation information (such as noise) in the original data that was not explained by the latent variables.

[0103] Based on the above decomposition, a linear regression model of X on Y is established, as shown in formula (4):

[0104] ;

[0105] In the formula, βPLSR is the regression coefficient matrix; F* is the residual matrix of the linear regression model, representing the deviation between the actual measured value and the model prediction value. This embodiment optimizes the model parameters based on leave-one-out cross-validation (LOOCV) with the goal of minimizing the root mean square error of prediction (RMSE). The optimal number of latent variables was determined to be 11, thereby constructing a high-precision absorbed dose prediction model.

[0106] Traditional methods of proportionally dividing the training and test sets significantly reduce the number of training samples, leading to underfitting of the model. Meanwhile, ordinary k-fold cross-validation is susceptible to fluctuations due to random grouping. To obtain the most unbiased and stable estimate of the model's generalization performance under limited sample conditions, this study uniformly adopts a leave-one-out cross-validation strategy. In this method, only one sample is reserved as the validation set in each iteration, while the remaining n-1 samples are used as the training set. This process is repeated n times until all samples are predicted. The deterministic nature of leave-one-out cross-validation eliminates the uncertainty introduced by random grouping, and since each modeling iteration utilizes almost the entire dataset, it can approximate the model's true performance under full sample conditions to the greatest extent possible. It is the standard validation paradigm in the field of small-sample hyperspectral analysis.

[0107] (2) Construction of tensile strength correlation model

[0108] Secondly, multiple linear regression (MLR) was used to quantify the independent contribution of specific spectral markers to changes in the macroscopic tensile strength of the material. Multiple linear regression is used to directly establish a linear combination relationship between multiple independent variables and a dependent variable when the number of spectral features is small and their physical meaning is clear, thereby verifying whether the chemical changes in the selected band are the direct cause of the changes in mechanical properties. The verification model is shown in formula (5):

[0109] ;

[0110] In the formula, ΔY represents the change in tensile strength (i.e., the difference between the strength of the irradiated sample and the strength of the unirradiated sample); n represents the number of characteristic spectral markers. It is the i-th characteristic spectral marker (including the intensity of the crosslinking characteristic band). Degradation characteristic band intensity The change in intensity of (etc.); It is the partial regression coefficient corresponding to the i-th feature, reflecting the weight of the influence of the chemical group change on tensile strength; ε represents the intercept term; ε represents the model residuals.

[0111] Step 3.2, Spatial Uniformity Quantification:

[0112] This step utilizes the "image-spectrum integration" characteristic of near-infrared hyperspectral imaging technology to upgrade the traditional single-point dose assessment to a two-dimensional spatial distribution assessment, which is one of the core innovations of this invention. The specific operation is as follows: the absorbed dose model trained in step 3.1 is applied to the hyperspectral data of each pixel in the region of interest of each sample. That is, for each coordinate point in the image, its hyperspectral data is input into the model to calculate the predicted dose value corresponding to each coordinate point. Through this process, the original hyperspectral reflectance image is converted into a dose quantization distribution map with spatial resolution, which intuitively presents the distribution of irradiation dose on the material surface (such as local hot spots or cold spots). Based on the generated dose quantization distribution map, the predicted dose values ​​of all pixels are extracted, and their statistical indicators are calculated to quantify the spatial uniformity. The coefficient of variation (CV) is introduced as the spatial uniformity index UNIF, and the calculation formula is shown in formula (6): ;

[0113] In the formula, (Average): refers to the arithmetic mean of the predicted dose of all pixels within the ROI region of the sample, representing the overall dose level received by the sample; SD (Standard Deviation): refers to the dispersion of the predicted dose of all pixels within the ROI region of the sample relative to the average, calculated using the following formula: , where n is the total number of pixels.

[0114] A higher coefficient of variation (CV) indicates a greater difference in absorbed dose across the sample surface and poorer spatial homogeneity; conversely, a lower CV indicates more uniform irradiation. This index represents the first time that "spatial homogeneity" has been incorporated as a quantitative parameter into a quality assessment system.

[0115] 3.2.1 The specific process of converting the original hyperspectral image into a dose quantization distribution map with spatial resolution:

[0116] Extract each spatial pixel from the corrected hyperspectral data cube one by one. Complete spectral curve This means that if the image resolution is... We will handle Each extracted spectrum is a separate spectral curve. Perform the third spectral preprocessing and feature extraction operation mentioned above. Based on the extracted feature bands (1185 nm related to crosslinking and 1700 nm related to degradation), calculate the spectral feature value (peak height, peak area, or absorbance ratio of a specific band) of the pixel.

[0117] Using a pre-trained and validated partial least squares regression (PLSR) model: , to extract single-pixel spectral features Input the model to calculate the predicted radiation dose value for that pixel. This process is repeated for each pixel in the hyperspectral image, thus converting the "spectral intensity matrix" into a "dose value matrix." The resulting dose value matrix is ​​then mapped to a color space to generate a pseudo-color image. Figure 7 In the graph, different colors directly represent different dose levels; red represents high-dose areas, and blue represents low-dose areas. This graph clearly shows the dose gradient on the sample surface increasing from the center to the edge, thus visually quantifying the spatial uniformity of irradiation.

[0118] Specifically, this conversion process achieves pixel-level model inversion through a programming algorithm. The system traverses every spatial coordinate point of the hyperspectral image, extracts its corresponding full-band spectral data, preprocesses it, and inputs it into a pre-trained dose prediction model. The model outputs the predicted dose value for that point, ultimately reconstructing a two-dimensional dose distribution matrix with the same spatial resolution as the original image. Through pseudo-color mapping, this matrix is ​​transformed into an intuitive dose distribution heatmap, thus achieving a visualization leap from microscopic spectral signals to macroscopic spatial distribution.

[0119] Step 3.3: Construction of the Integrated Quality Score (IQS) Model

[0120] To address the challenge of irradiation quality being controlled by both nonlinear factors such as the degree of chemical crosslinking / degradation and spatial uniformity, this step constructs an integrated comprehensive quality scoring model.

[0121] A nonlinear prediction model is constructed using the Random Forest (RF) regression algorithm.

[0122] Input variables (X): include chemical characteristic indicators extracted from the spectrum and spatial uniformity index. (i.e., CV value). The specific chemical characteristic index is the intensity of the cross-linking characteristic bands identified above. Degradation characteristic band intensity and auxiliary band intensity Output variable (Y): Change in the macroscopic tensile strength of the material. .

[0123] During the training of the comprehensive quality scoring model, the Mean Decrease Impurity mechanism built into the Random Forest algorithm is used for feature importance analysis. This mechanism calculates the importance of a given feature (in all decision trees of the Random Forest) by... , , The contribution of a feature to tensile strength prediction is quantified by the sum of the reduction in mean squared error brought about by node splitting. Based on this, the constructed comprehensive quality scoring model can not only predict tensile strength, but also clearly indicate whether chemical factors or spatial homogeneity factors dominate the current quality fluctuations by ranking the importance of features (for example, the analysis results show that the contribution of spatial homogeneity UNIF exceeds 40%, indicating that it is a key quality factor).

[0124] Step 3.4, Model Evaluation Indicators

[0125] The above-mentioned spectral feature verification model and comprehensive quality scoring model were systematically evaluated to ensure the accuracy and reliability of the evaluation methods:

[0126] For the absorbed dose model, tensile strength model, and comprehensive quality scoring model, the coefficient of determination was calculated using leave-one-out cross-validation. Root mean square error and mean absolute error Assessment purpose Used to characterize the model's ability to explain data variations and to confirm whether the extracted spectral features can truly reflect dose changes and mechanical property changes; and The prediction error of the quantification model is used as shown in formulas (7), (8), and (9):

[0127] ;

[0128] ;

[0129] ;

[0130] In the formula, This represents the actual measured value of the i-th sample; This represents the model prediction value for the i-th sample; This represents the arithmetic mean of the true values ​​of all samples.

[0131] 4.1 Spectral Analysis and Characterization

[0132] Irradiation-induced structural changes in EVA / POE composites produced characteristic responses in the near-infrared spectrum. After spectral preprocessing, including SG first-derivative filtering (Figures 3 and 4), we successfully extracted irradiation-sensitive bands by identifying peaks, troughs, and zero-point crossings. These features serve as robust spectral markers for resolving overlapping functional group vibrations and supporting subsequent quantitative assessments. The assignment, vibrational distribution, and chemical correlations of these sensitive bands have been validated based on existing literature and are detailed in Tables 1 and 2. Based on their corresponding chemical events, the core spectral markers were categorized into: crosslinking bands (e.g., 1185 nm and 1608 nm), degradation bands (e.g., 1610 nm and 1631 nm), and auxiliary bands.

[0133] Table 1. Molecular mechanism correlation of EVA7470 / POE series in irradiation-sensitive bands

[0134]

[0135] Table 2. Molecular mechanism correlation of EVA7350 / POE series in irradiation-sensitive bands

[0136]

[0137] 4.2 Spectral Characteristic Verification

[0138] Before constructing the final comprehensive quality scoring model, the validity of the spectral features identified in point 4.1 above must be verified. This study performs dual verification by confirming that these features are correlated with both "physical inducement" (absorbed dose) and "macroscopic properties" (tensile strength).

[0139] To this end, we established two independent validation models. First, a partial least squares regression model was used to correlate spectral features with absorbed dose, and this model performed excellently (Figure 5), with a validation set R0. 2 = 0.973, RMSE only 4.111 kGy. Secondly, a multiple linear regression model was used to correlate spectral features with tensile strength, which also achieved a high validation R-value. 2 =0.939 ( Figure 6 The high R-values ​​of these two models 2 The values ​​confirm that the selected spectral features are effective indicators, reflecting both the material's physical irradiation history and its final mechanical properties.

[0140] Subsequently, we examined the evolution of macroscopic mechanical properties with dosage (3). The EVA7470 / POE material (with higher VA content) exhibited a nonlinear transition, with tensile strength peaking at 7.08 MPa at 50 kGy, followed by accelerated degradation, causing the strength to drop to 5.10 MPa at 80 kGy. In contrast, EVA7350 / POE (with lower VA content) showed a stable performance improvement across the entire dosage range, reaching a maximum of 6.89 MPa at 80 kGy. This complex nonlinear behavior highlights the necessity of constructing a multifactor quality model and also verifies the correctness of the methods used in subsequent steps.

[0141] Table 3. Changes in tensile strength and properties of EVA / POE composites under different irradiation doses.

[0142]

[0143] 4.3 Quantification of Spatial Uniformity

[0144] After validating the effectiveness of the PLSR dose model, this study applied it to each pixel of a hyperspectral image to assess spatial homogeneity, a key factor mentioned in the introduction. This method generates spatially resolved dose quantization maps and smooths them using a Gaussian filter to visualize macroscopic dose distribution trends.

[0145] Result graph ( Figure 7 This revealed significant differences between the two material systems. In the 30 kGy test group, it was clear that the two samples at the bottom (EVA 7350 series) exhibited a relatively uniform (mostly dark blue) low-dose surface with only slight local fluctuations; in stark contrast, the two samples at the top (EVA 7470 series) showed severe spatial heterogeneity (coexistence of red / yellow and blue regions).

[0146] This spatial heterogeneity was quantified using the coefficient of variation (CV) (Table 4). The results showed that EVA7350 maintained good homogeneity across all doses (CV < 25%). Conversely, EVA7470 exhibited severe non-homogeneity in the 30–60 kGy dose range (CV > 33%), confirming significant local dose anomalies. This CV value, representing spatial homogeneity, was retained as a key “UNIF” feature for constructing the final ensemble model.

[0147] Table 4. Irradiation dose uniformity analysis of EVA7470 and EVA7350 series samples based on pixel-by-pixel prediction

[0148]

[0149] 4.4 Integrated Quality Score (IQS) Model

[0150] The final comprehensive quality scoring model was constructed to integrate all identified quality influencing factors. Using the RF algorithm, the model uses the spectral feature indices (ICL, IDG, IAUX) in 4.1 and the CV uniformity index UNIF in 4.3 as inputs to predict the macroscopic tensile strength ΔY.

[0151] The performance of the comprehensive quality assessment model was calculated using the leave-one-out cross-validation method to determine the coefficient of determination. Root mean square error and mean absolute error Perform verification, such as Figure 8 As shown in the figure, the model achieved consistent rankings for both materials, with the left figure showing the highest R-value for EVA7350 / POE. 2 The value is 0.931, and the R value of EVA7470 / POE in the right figure is... 2 The value is 0.987. This confirms that the integrated quality scoring model, by incorporating chemical and spatial factors, can reliably translate complex nonlinear material variations into accurate quality rankings.

[0152] This study successfully established a non-destructive evaluation framework based on near-infrared hyperspectral imaging technology, constructing a comprehensive model capable of evaluating chemical changes and their spatial homogeneity. The effectiveness of this framework was systematically validated, as shown in Table 5. First, SG filtering was used to identify radiation-sensitive spectral bands. Then, by establishing a regression model, it was confirmed that these characteristics are related to absorbed dose (R0). 2 = 0.987 via PLSR) and tensile strength (R 2 = 0.931 via MLR) all showed strong correlation. Subsequently, this dose model was applied to pixel-by-pixel quantization to generate a dose distribution map, from which the coefficient of variation (CV) was derived as the spatial uniformity index UNIF.

[0153] Finally, this study used the Random Forest (RF) algorithm to integrate the aforementioned spectral characteristic indices with the CV uniformity index, constructing a final comprehensive irradiation quality scoring model. The evaluation results of this comprehensive quality scoring model were correlated with leave-one-out cross-validation calculations for the measured tensile performance ranking. The R-value of EVA7350 / POE was [data missing]. 2 Reaching 0.931, the R of EVA7470 / POE 2 The result reached 0.987, confirming the reliability of its sorting.

[0154] Table 5. Evaluation Results and Quality Level Classification of the Comprehensive Quality Scoring Model

[0155]

[0156] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions above are merely illustrative of the principles of the present invention. Various changes and modifications can be made to the present invention without departing from its spirit and scope. All such changes and modifications fall within the scope of the present invention as claimed, which is defined by the appended claims and their equivalents.

Claims

1. A method for assessing the irradiation quality of EVA and POE composite materials based on near-infrared hyperspectral imaging, characterized in that, Includes the following steps: Step 1: Acquire hyperspectral images: Hyperspectral images were acquired using a line scanning device in a darkroom environment; dark current was used. Correction and White Reference The dual correction mechanism, combining calibration and verification, will adjust the original intensity data. Convert to relative reflectance As shown in the formula below: ； After reflectance correction, the region of interest is defined and spectral features are extracted. The extracted spectral features include the intensity of the crosslinking characteristic band. Degradation characteristic band intensity and auxiliary band intensity ; Step 2: Spectral preprocessing and feature extraction: The hyperspectral data were first preprocessed using multivariate scattering correction, followed by spectral smoothing and signal enhancement using Savitzky–Golay first derivative filtering. The 0 kGy spectrum was subtracted from the average spectrum of each irradiation dose group to generate the average difference spectrum. Then, the peaks, troughs and zero-point crossings in the Savitzky–Golay first derivative spectrum were identified as characteristic spectra. These characteristic spectra are specific spectral markers of irradiation-induced chemical changes, capable of distinguishing overlapping functional group vibrations; Step 3: Development of a comprehensive quality scoring model, including the following steps: Step 3.1, Spectral Feature Verification: After identifying the characteristic spectra, a spectral feature verification model was established to confirm the quantitative correlation between the characteristic spectra and absorbed dose and tensile strength. Specifically, firstly, an absorbed dose model was constructed based on partial least squares regression, and secondly, a tensile strength correlation model was constructed based on multiple linear regression, as follows: First, partial least squares regression is used to establish a quantitative mapping relationship between characteristic spectral intensity and actual absorbed dose. The high-dimensional spectral independent variable matrix X and the dose dependent variable matrix Y are simultaneously projected into a low-dimensional latent variable space. This extracts the main variation information of the data while maximizing the covariance between the independent and dependent variables, thereby overcoming multicollinearity and filtering out spectral noise. The partial least squares regression algorithm performs bilinear decomposition on the spectral matrix X and the dose matrix Y, as shown in the following formula: ； ； Where T and U are the latent variable matrices of X and Y, respectively; P and Q are the loading matrices of X and Y, respectively; and E and F are the residual matrices of X and Y, respectively. Based on the above decomposition, a linear regression model of X on Y is established, as shown in the following formula: ； In the formula, βPLSR is the regression coefficient matrix, and F* is the residual matrix of the linear regression model, representing the deviation between the actual measured value and the model prediction value. Secondly, multiple linear regression was used to quantify the independent contribution of specific spectral markers to the change in macroscopic tensile strength of the material. The linear combination relationship between multiple independent variables and one dependent variable was established to verify the model, as shown in the following formula: ； In the formula, ΔY represents the change in tensile strength; n represents the number of characteristic spectral markers; It is the intensity change of the i-th characteristic spectral marker; It is the partial regression coefficient corresponding to the i-th feature; It is the intercept term; ε represents the model residuals; Step 3.2, Spatial Uniformity Quantification: The absorbed dose model trained in step 3.1 is applied to the hyperspectral data of each pixel within the region of interest of each sample. For each coordinate point in the image, its hyperspectral data is input into the absorbed dose model to calculate the predicted dose value corresponding to each coordinate point, thereby converting the original hyperspectral image into a dose quantization distribution map with spatial resolution. Based on the generated dose quantization distribution map, the predicted dose values ​​of all pixels are extracted, and their statistical indices are calculated to quantify spatial uniformity. The coefficient of variation (CV) is introduced as a spatial uniformity index. The larger the CV value, the greater the difference in absorbed dose at various points on the sample surface and the worse the spatial uniformity; conversely, a smaller CV value indicates more uniform irradiation. The calculation formula is shown in the following formula: ； Wherein, SD refers to the dispersion of the predicted dose of all pixels within the region of interest of the sample relative to the average value, and is calculated using the following formula: Where n is the total number of pixels; It refers to the arithmetic mean of the predicted doses of all pixels in the region of interest of the sample, representing the overall dose level received by the sample; Step 3.3: Construct a comprehensive quality scoring model: The random forest regression algorithm is used, taking the chemical characteristic indicators extracted from the spectrum and the CV value calculated in step 3.2 as input, and outputting the change in the macroscopic tensile strength of the material. A nonlinear prediction model was constructed. Step 3.4: Systematically evaluate the above spectral feature verification model and comprehensive quality scoring model to ensure the accuracy and reliability of the evaluation method. For the absorbed dose model, tensile strength model, and comprehensive quality scoring model, the coefficient of determination was calculated using leave-one-out cross-validation. Root mean square error and mean absolute error Assessment purpose Used to characterize the model's ability to explain data variations and to confirm whether the extracted spectral features can truly reflect dose changes and mechanical property changes; and The prediction error used to quantify the model is shown in the following formula: ； ； ； In the formula, This represents the actual measured value of the i-th sample; This represents the model prediction value for the i-th sample; This represents the arithmetic mean of the true values ​​of all samples.

2. The method for evaluating the irradiation quality of EVA and POE composite materials based on near-infrared hyperspectral imaging according to claim 1, characterized in that, The specific steps for defining the region of interest and extracting spectral features in step one are as follows: By utilizing the difference in reflectance between the sample and the background conveyor belt in a specific band, an intensity threshold is set to generate a binary mask, which automatically segments the pixel region of the EVA / POE composite material sample, removes background noise interference, and extracts spectral features from each pixel point in the mask region as an independent spectral sample. The specific implementation of the threshold-based mask method for separating the background conveyor belt and the EVA / POE composite material sample is as follows: Set up a hyperspectral image Its size is Where M and N are the number of rows and columns of the image, respectively, and L is the number of spectral bands; each pixel Indicates position In the The spectral values ​​of each band are analyzed; the spectral features of each pixel are observed to identify the regions where the spectral values ​​of the background and target areas differ significantly, and then a threshold is set for filtering; specific spectral features include: Mean spectrum: ; Standard deviation spectrum: ; Spectral values ​​of a specific band: directly select a specific band. spectral values The threshold setting section is divided into global thresholds. or multiple local thresholds The global threshold is ,in and These are the mean and standard deviation of the background, respectively. It is a constant; the local threshold is ,in and These are the mean and standard deviation of the local background. After separating the background conveyor belt from the EVA and POE composite sample, the updated pixels with a mask value of 1 and the pixels with a mask value of 0 remain unchanged.

3. The method for evaluating the irradiation quality of EVA and POE composite materials based on near-infrared hyperspectral imaging according to claim 1, characterized in that, The specific steps in step 3.2 to convert the original hyperspectral image into a dose quantization distribution map with spatial resolution are as follows: Step 3.2.1: Extract each spatial pixel from the corrected hyperspectral data one by one. Complete spectral curve The image resolution is 640×1800, and 1,152,000 independent spectral curves are processed. Step 3.2.2: For each extracted spectral curve Perform the preprocessing and feature extraction operations in step two above, and calculate the spectral feature values ​​of the pixels for the extracted and identified feature bands; Step 3.2.3: Use the partial least squares regression model: The single-pixel spectral features obtained in step 3.2.2 are then... Input the model to calculate the predicted radiation dose value for that pixel. Repeat this operation for each pixel in the hyperspectral image to convert the spectral intensity matrix into a dose value matrix. Step 3.2.4: Map the obtained dose value matrix to the RGB color space to generate a pseudo-color image, and obtain a dose quantization distribution map with spatial resolution.

4. The method for evaluating the irradiation quality of EVA and POE composite materials based on near-infrared hyperspectral imaging according to claim 1, characterized in that, The line scanning device includes a camera, a halogen lamp light source, and a conveyor belt; the EVA and POE composite material sample is placed on the conveyor belt and photographed under light, and the spectral image is recorded in the range of 900–1700 nm. The generated hyperspectral data cube has a spatial dimension of 640 × 1800 pixels and contains 224 spectral bands.

5. The method for evaluating the irradiation quality of EVA and POE composite materials based on near-infrared hyperspectral imaging according to claim 1, characterized in that, In step 3.3, the chemical characteristic index specifically refers to the intensity of the cross-linking characteristic band. Degradation characteristic band intensity and auxiliary band intensity During model training, the average impurity reduction mechanism built into the random forest algorithm is used for feature importance analysis. This is achieved by calculating the cross-linked feature band intensity across all decision trees in the random forest. Degradation characteristic band intensity and auxiliary band intensity The contribution of this feature to the tensile strength prediction result is quantified by the sum of the reduction in mean square error brought about by node splitting.

Images