EVA high-concentration masterbatch detection method based on multispectral imaging

Abstract

The present invention relates to a method for detecting high-concentration EVA masterbatch based on multispectral imaging, comprising the following steps: S1: setting an adjustable multicolor LED light source on a production line, and using a multispectral imager to obtain a multispectral image of a masterbatch to obtain a spectral data set; S2: extracting a spectrum from the spectral data set based on a masterbatch component recognition model to identify spectral features of different components or impurities; S3: constructing a masterbatch color difference detection model and a defect recognition model, and performing color difference detection and defect recognition based on spectral features; S4: based on real-time image color difference detection and defect recognition results, a control system of a production line is controlled based on intelligent feedback to achieve automatic adjustment; S5: visualizing based on color difference detection and defect recognition results, so that an operator can obtain real-time masterbatch detection results, and alarm processing is performed on unqualified products. This scheme realizes efficient and accurate monitoring of high-concentration EVA masterbatch, ensuring intelligent production and high product quality.

Description

Technical Field

[0001] The invention relates to the field of intelligent detection, and in particular to a method for detecting EVA high-concentration masterbatch based on multi-spectral imaging. Background Art

[0002] In modern manufacturing, the quality control of EVA (ethylene-vinyl acetate copolymer) high-concentration masterbatch is crucial. Masterbatch is a concentrated pigment used to color plastic products and is widely used in packaging, automotive, electronics and other industries. Traditional masterbatch detection methods mainly rely on manual visual inspection and simple optical instruments. These methods are not only time-consuming, but also easily affected by human factors, resulting in inconsistent test results.

[0003] With the advancement of science and technology, multispectral imaging technology has gradually become an innovative solution for masterbatch detection. Multispectral imaging can obtain image information of samples at multiple wavelengths, providing richer spectral data than traditional RGB imaging. This technology can capture subtle color differences and impurity features that are invisible to the naked eye, thereby improving the accuracy and reliability of detection. Summary of the invention

[0004] In order to solve the above problems, the purpose of the present invention is to provide a method for detecting EVA high-concentration masterbatch based on multispectral imaging, so as to achieve efficient and accurate monitoring of EVA high-concentration masterbatch and ensure intelligent production and high product quality.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for detecting EVA high-concentration masterbatch based on multispectral imaging comprises the following steps:

[0007] S1: Set an adjustable multi-color LED light source on the production line, and use a multi-spectral imager to obtain a multi-spectral image of the masterbatch to obtain a spectral data set, specifically:

[0008] Installing adjustable multi-color LED light sources on the production line, according to preset wavelength combinations, including red, green, blue and near-infrared, to ensure adequate reflection in different bands;

[0009] Use a diffuser to ensure that the light evenly covers the entire sample area and reduce the effects of shadows and uneven reflections;

[0010] Use a multispectral imager equipped with filters of different wavelengths to collect reflectance spectrum data of masterbatch and set imaging parameters, including exposure time and gain;

[0011] The collected multispectral image is corrected, and the spectral vector of each pixel is extracted from the corrected multispectral image to form a complete spectral data set, as follows:

[0012] The correction includes geometric correction and spectral correction, wherein the spectral correction uses a standard white plate image W(x,y) to perform white balance correction to eliminate the influence of ambient light:

[0013] ;

[0014] Among them, I(x,y) is the original image, I corrected (x,y) is the original image after spectral correction;

[0015] Geometric correction uses the geometric transformation matrix T to correct the image:

[0016] ;

[0017] in, is the image after geometric correction;

[0018] Extract the spectral vector of each pixel from the rectified image to form a complete spectral dataset;

[0019] S2: Extract spectra from the spectral data set based on the masterbatch component recognition model to identify the spectral characteristics of different components or impurities;

[0020] S3: Constructing a masterbatch color difference detection model and a defect recognition model, and performing color difference detection and defect recognition based on spectral features;

[0021] S4: Based on the real-time image color difference detection and defect recognition results, the control system of the production line is controlled based on the intelligent agent feedback to achieve automatic adjustment;

[0022] S5: Visualization based on color difference detection and defect recognition results enables operators to obtain real-time masterbatch detection results and issue alarms for unqualified products.

[0023] Furthermore, the masterbatch component recognition model is constructed based on the support vector machine, as follows:

[0024] Acquire multispectral images of masterbatch, extract the spectral vector of each pixel from each multispectral image, and construct a spectral data set;

[0025] Use filters to remove noise from spectral data and normalize the spectral data;

[0026] Combining principal component analysis and linear discriminant analysis to extract features, and marking the extracted spectral features to obtain the final spectral feature set;

[0027] A masterbatch component recognition model was constructed based on support vector machine (SVM), and the model was trained using labeled spectral datasets to ensure that the model could accurately identify different components. The model performance was evaluated through cross-validation, and the model parameters were adjusted to improve the recognition accuracy to obtain the final masterbatch component recognition model.

[0028] Furthermore, feature extraction is performed by combining principal component analysis and linear discriminant analysis as follows:

[0029] Based on principal component analysis, the dimension of spectral data is reduced, the covariance matrix is ​​calculated, eigenvalue decomposition is performed, and the principal components are extracted:

[0030] ;

[0031] Where Q is the covariance matrix; The mean vector of the spectral data; is the i-th spectral vector; N is the number of spectral data vectors; the superscript T indicates transposition;

[0032] Select the first k eigenvectors to achieve dimensionality reduction, that is:

[0033] ;

[0034] in, is the eigenvector matrix of PCA; Spectral data after dimension reduction; is the spectral data after normalization;

[0035] Linear discriminant analysis (LDA) is used to extract features that can maximize inter-class differences and minimize intra-class differences, and calculate intra-class and inter-class scatter matrices:

[0036] ;

[0037] Among them, S W is the intra-class scatter matrix; S B is the between-class scatter matrix; represents the mean of the spectral vector of category c; represents the ith spectral vector belonging to category c; N c is the number of vectors of spectral data after dimension reduction; is the mean of the spectral vector; C is the number of categories;

[0038] Solution The characteristic values ​​of:

[0039] ;

[0040] Among them, V is the projection matrix of linear discriminant analysis, S LDAis the spectral feature data after linear discriminant analysis.

[0041] Furthermore, a masterbatch component recognition model is constructed based on the support vector machine SVM, as follows:

[0042] Divide the labeled spectral data into training sets and test sets, and standardize the spectral features of the training sets and test sets;

[0043] Use Bootstrap random sampling with replacement to generate multiple different subsets from the training set;

[0044] Train an independent SVM model for each subset, use cross-validation to evaluate the model performance for the SVM parameters selected during training, and adjust the SVM parameters of each model using the Bayesian optimization method;

[0045] The prediction results of each trained SVM model on the test sample are voted, and the final category of the sample is determined by majority voting.

[0046] Furthermore, the SVM model is as follows:

[0047] Using RBF kernel function:

[0048]

[0049] in, They are and samples; γ is the parameter of the RBF kernel function;

[0050] SVM objective function:

[0051] ;

[0052] in, is the penalty factor; For the The slack variable corresponding to the samples; is the weight vector; b is the bias term; N’ is the total number of samples;

[0053] By introducing Lagrange multipliers, the optimization problem of SVM is converted into a dual form and calculated using kernel functions, and the objective function is converted to:

[0054] ;

[0055] in, , is the Lagrange multiplier; , They are and The true labels of samples; represents the set of all Lagrange multipliers in the entire optimization process;

[0056] Finally, the classification decision function of SVM is:

[0057] ;

[0058] in, is a new input sample.

[0059] Furthermore, the masterbatch color difference detection model is as follows: the color-related spectral bands are extracted from the spectral features of the identified components, and the spectral features are converted into the CIE 1931 XYZ color space by combining the standard illuminant and the standard observer function:

[0060] ;

[0061] Where, S(λ) is the sample spectrum; is the CIE standard observer chromaticity function; X, Y, Z are the representations of colors in the XYZ color space; E(λ) is the standard illuminant; λ represents the wavelength;

[0062] Convert XYZ values ​​to CIE-LAB color values:

[0063] ;

[0064] in, Indicates brightness, and Indicates color contrast; is the standard white reference value;

[0065] where f(t) is defined as:

[0066] ;

[0067] Among them, t represents the input and δ is a constant;

[0068] use Formula to calculate color difference:

[0069]

[0070] in, and is the LAB value of the reference sample and the sample to be tested;

[0071] Set the standard color difference threshold. If the color difference exceeds this threshold, it will be judged as color difference abnormality.

[0072] Furthermore, the defect recognition model is built based on XGBOOST, as follows:

[0073] Start with the mean of all samples as the base model, and in each iteration, calculate the residual of the current model, that is, the difference between the predicted value and the actual value;

[0074] Using the residuals as the target variable, train a new decision tree to fit these residuals;

[0075] Add the output of the new tree to the current model and update the predicted value;

[0076] And add regularization terms to the objective function of XGBoost to control model complexity and prevent overfitting;

[0077] In each iteration, XGBoost selects the best tree structure and leaf node weights by optimizing the following objective function:

[0078] ;

[0079] in, is the loss function; is the actual value; It is Round The predicted value of samples; represents the g-th decision tree; is the objective function, the superscript Indicates wheel; The number of samples; G is the number of decision trees.

[0080] Furthermore, S4 is specifically:

[0081] Use proportional-integral-derivative (PID) controllers to adjust operating parameters of production equipment based on color difference and defect information;

[0082] ;

[0083] in, Control output; is the error, , are the color difference error and the number of defects, respectively, which are defined as the difference between the target value and the actual detection value; , , are the parameters of the PID controller, corresponding to the proportional, integral and differential gains respectively;

[0084] The control strategy is optimized using the reinforcement learning algorithm Q-learning. The intelligent agent learns the optimal control actions by interacting with the environment, and rewards are given according to the degree of color difference and defect reduction to guide the intelligent agent's learning.

[0085] Define the state space S, including the current color difference, number of defects, and equipment parameters; the action space A is the control action, including adjusting the temperature, speed, and material ratio; reward is given according to the degree of color difference and defect reduction, and the reward function is defined as:

[0086] ;

[0087] Among them, Defects is the number of defects; is the color difference, and is the weight coefficient;

[0088] Q value update formula:

[0089] ;

[0090] in, is the value of executing action a in state s; is the discount factor; s and Respectively represent the current state and the subsequent state to which the action is transferred after execution; a and Respectively represent the current state s and the subsequent state The possible actions to be taken.

[0091] The present invention has the following beneficial effects:

[0092] 1. The present invention combines image recognition algorithm and machine learning to extract spectral features from multispectral images and identify different components or impurities, providing a solid foundation for building color difference detection models and defect recognition models. Through the intelligent feedback control system, the production line can be automatically adjusted to further improve production efficiency and product quality.

[0093] 2. The present invention constructs a masterbatch component recognition model with enhanced generalization ability and robustness through the combination of Bagging framework and SVM, improves the classification performance of complex samples by using the integration strategy of diversity enhancement, and continuously improves the recognition accuracy of the model through cross-validation and parameter optimization procedures;

[0094] 3. The present invention combines spectral analysis and CIE-LAB color space conversion to effectively monitor the color difference of masterbatches, and by combining the masterbatch component identification results, the XGBoost model can effectively identify subtle defects in the masterbatch and provide highly accurate defect detection results;

[0095] 4. The present invention combines PID controller and Q-learning algorithm to realize dynamic optimization and adjustment of production equipment operating parameters. PID controller provides real-time error correction, while Q-learning further optimizes the control effect by learning the optimal strategy. The system can effectively reduce color difference and defects and improve the automation and intelligence level of the production line. BRIEF DESCRIPTION OF THE DRAWINGS

[0096] Figure 1 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0097] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0098] refer to Figure 1 In this embodiment, a method for detecting EVA high-concentration masterbatch based on multispectral imaging is provided, comprising the following steps:

[0099] S1: Setting an adjustable multi-color LED light source on the production line and using a multi-spectral imager to obtain a multi-spectral image of the masterbatch to obtain a spectral data set;

[0100] S2: Extract spectra from the spectral data set based on the masterbatch component recognition model to identify the spectral characteristics of different components or impurities;

[0101] S3: Constructing a masterbatch color difference detection model and a defect recognition model, and performing color difference detection and defect recognition based on spectral features;

[0102] S4: Based on the real-time image color difference detection and defect recognition results, the control system of the production line is controlled based on the intelligent agent feedback to achieve automatic adjustment;

[0103] S5: Visualization based on color difference detection and defect recognition results enables operators to obtain real-time masterbatch detection results and issue alarms for unqualified products.

[0104] In this embodiment, S1 is specifically:

[0105] Installing adjustable multi-color LED light sources on the production line, according to preset wavelength combinations, including red, green, blue and near-infrared, to ensure adequate reflection in different bands;

[0106] Use a diffuser to ensure that the light evenly covers the entire sample area and reduce the effects of shadows and uneven reflections;

[0107] Use a multispectral imager equipped with filters of different wavelengths to collect reflectance spectrum data of masterbatch, set imaging parameters including exposure time and gain; optimize image quality to ensure high resolution and low noise;

[0108] Automatic Exposure Control (AEC):

[0109] Use the sensor to monitor the image brightness in real time and calculate the average brightness value of the current image.

[0110] Feedback adjustment: Dynamically adjust the exposure time based on the difference between the target brightness and the current brightness to ensure that the image brightness is within the optimal range.

[0111] Automatic Gain Control (AGC):

[0112] Signal Strength Assessment: Evaluate the signal strength of an image, especially in low-light conditions.

[0113] Gain adjustment: dynamically adjust the gain based on the difference between signal strength and target strength;

[0114] The collected multispectral images are corrected to eliminate the chromatic aberration caused by the imaging angle and surface characteristics; and the spectral vector of each pixel is extracted from the corrected multispectral images to form a complete spectral data set.

[0115] In this embodiment, the collected multispectral image is corrected, and the spectral vector of each pixel is extracted from the corrected image to form a complete spectral data set, as follows:

[0116] The correction includes geometric correction and spectral correction, wherein the spectral correction uses a standard white plate image W(x,y) to perform white balance correction to eliminate the influence of ambient light:

[0117] ;

[0118] Among them, I(x,y) is the original image, I corrected (x,y) is the original image after spectral correction;

[0119] Geometric correction uses the geometric transformation matrix T to correct the image:

[0120] ;

[0121] in, is the image after geometric correction;

[0122] The spectral vector of each pixel is extracted from the rectified image to form a complete spectral dataset.

[0123] In this embodiment, the masterbatch component recognition model is constructed based on a support vector machine, as follows:

[0124] Acquire multispectral images of masterbatch, extract the spectral vector of each pixel from each multispectral image, and construct a spectral data set;

[0125] Use filters to remove noise from spectral data and normalize the spectral data;

[0126] Combining principal component analysis and linear discriminant analysis to extract features, and marking the extracted spectral features to obtain the final spectral feature set;

[0127] A masterbatch component recognition model was constructed based on support vector machine (SVM), and the model was trained using labeled spectral datasets to ensure that the model could accurately identify different components. The model performance was evaluated through cross-validation, and the model parameters were adjusted to improve the recognition accuracy to obtain the final masterbatch component recognition model.

[0128] In this embodiment, feature extraction is performed by combining principal component analysis and linear discriminant analysis as follows:

[0129] Based on principal component analysis, the dimension of spectral data is reduced, the covariance matrix is ​​calculated, eigenvalue decomposition is performed, and the principal components are extracted:

[0130] ;

[0131] Where Q is the covariance matrix; The mean vector of the spectral data; is the i-th spectral vector; N is the number of spectral data vectors; the superscript T indicates transposition;

[0132] Select the first k eigenvectors to achieve dimensionality reduction, that is:

[0133] ;

[0134] in, is the eigenvector matrix of PCA; Spectral data after dimension reduction; is the spectral data after normalization;

[0135] Linear discriminant analysis (LDA) is used to extract features that can maximize inter-class differences and minimize intra-class differences, and calculate intra-class and inter-class scatter matrices:

[0136] ;

[0137] Among them, S W is the intra-class scatter matrix; S B is the between-class scatter matrix; represents the mean of the spectral vector of category c; represents the ith spectral vector belonging to category c; N c is the number of vectors of spectral data after dimension reduction; is the mean of the spectral vector; C is the number of categories;

[0138] Solution The characteristic values ​​of:

[0139] ;

[0140] Among them, V is the projection matrix of linear discriminant analysis, S LDA is the spectral feature data after linear discriminant analysis.

[0141] In this embodiment, a masterbatch component recognition model is constructed based on a support vector machine (SVM), as follows:

[0142] Divide the labeled spectral data into training sets and test sets, and standardize the spectral features of the training sets and test sets;

[0143] Use Bootstrap random sampling with replacement to generate multiple different subsets from the training set;

[0144] Train an independent SVM model for each subset, use cross-validation to evaluate the model performance for the SVM parameters selected during training, and adjust the SVM parameters of each model using the Bayesian optimization method;

[0145] The prediction results of each trained SVM model on the test sample are voted, and the final category of the sample is determined by majority voting.

[0146] In this embodiment, the SVM model is as follows:

[0147] Using RBF kernel function:

[0148]

[0149] in, They are and samples; γ is the parameter of the RBF kernel function;

[0150] SVM objective function:

[0151] ;

[0152] in, is the penalty factor; For the The slack variable corresponding to the samples; is the weight vector; b is the bias term; N’ is the total number of samples;

[0153] By introducing Lagrange multipliers, the optimization problem of SVM is converted into a dual form and calculated using kernel functions, and the objective function is converted to:

[0154] ;

[0155] in, , is the Lagrange multiplier; , They are and The true labels of samples; represents the set of all Lagrange multipliers in the entire optimization process;

[0156] Finally, the classification decision function of SVM is:

[0157] ;

[0158] in, is a new input sample.

[0159] In this embodiment, the masterbatch color difference detection model is as follows: color-related spectral bands are extracted from the spectral features of the identified components, and the spectral features are converted into the CIE 1931 XYZ color space by combining the standard illuminant and the standard observer function:

[0160] ;

[0161] Where, S(λ) is the sample spectrum; is the CIE standard observer chromaticity function; X, Y, Z are the representations of colors in the XYZ color space; E(λ) is the standard illuminant; λ represents the wavelength;

[0162] Convert XYZ values ​​to CIE-LAB color values:

[0163] ;

[0164] in, Indicates brightness, and Indicates color contrast; is the standard white reference value;

[0165] where f(t) is defined as:

[0166] ;

[0167] Among them, t represents the input and δ is a constant;

[0168] use Formula to calculate color difference:

[0169]

[0170] in, and is the LAB value of the reference sample and the sample to be tested;

[0171] Set the standard color difference threshold. If the color difference exceeds this threshold, it will be judged as color difference abnormality.

[0172] In this embodiment, the defect recognition model is built based on XGBOOST, as follows:

[0173] Start with the mean of all samples as the base model, and in each iteration, calculate the residual of the current model, that is, the difference between the predicted value and the actual value;

[0174] Using the residuals as the target variable, train a new decision tree to fit these residuals;

[0175] Add the output of the new tree to the current model and update the predicted value;

[0176] And add regularization terms to the objective function of XGBoost to control model complexity and prevent overfitting;

[0177] In each iteration, XGBoost selects the best tree structure and leaf node weights by optimizing the following objective function:

[0178] ;

[0179] in, is the loss function; is the actual value; It is Round The predicted value of samples; represents the g-th decision tree; is the objective function, the superscript Indicates wheel; The number of samples; G is the number of decision trees.

[0180] In this embodiment, S4 is specifically:

[0181] Use proportional-integral-derivative (PID) controllers to adjust operating parameters of production equipment based on color difference and defect information;

[0182] ;

[0183] in, Control output; is the error, , are the color difference error and the number of defects, respectively, which are defined as the difference between the target value and the actual detection value; , , are the parameters of the PID controller, corresponding to the proportional, integral and differential gains respectively;

[0184] The control strategy is optimized using the reinforcement learning algorithm Q-learning. The intelligent agent learns the optimal control actions by interacting with the environment, and rewards are given according to the degree of color difference and defect reduction to guide the intelligent agent's learning.

[0185] Define the state space S, including the current color difference, number of defects, and equipment parameters; the action space A is the control action, including adjusting the temperature, speed, and material ratio; reward is given according to the degree of color difference and defect reduction, and the reward function is defined as:

[0186] ;

[0187] Among them, Defects is the number of defects; is the color difference, and is the weight coefficient;

[0188] Q value update formula:

[0189] ;

[0190] in, is the value of executing action a in state s; is the discount factor; s and Respectively represent the current state and the subsequent state to which the action is transferred after execution; a and Respectively represent the current state s and the subsequent state The following possible actions can be taken.

[0191] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0192] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0193] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0194] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0195] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

Claims

1. A method for detecting EVA high-concentration masterbatch based on multispectral imaging, characterized in that: The following steps are involved: S1: Set an adjustable multi-color LED light source on the production line, and use a multi-spectral imager to obtain a multi-spectral image of the masterbatch to obtain a spectral data set, specifically: Installing adjustable multi-color LED light sources on the production line, according to preset wavelength combinations, including red, green, blue and near-infrared, to ensure adequate reflection in different bands; Use a diffuser to ensure that the light evenly covers the entire sample area and reduce the effects of shadows and uneven reflections; Use a multispectral imager equipped with filters of different wavelengths to collect reflectance spectrum data of masterbatch and set imaging parameters, including exposure time and gain; The collected multispectral image is corrected, and the spectral vector of each pixel is extracted from the corrected multispectral image to form a complete spectral data set, as follows: The correction includes geometric correction and spectral correction, wherein the spectral correction uses a standard white plate image W(x,y) to perform white balance correction to eliminate the influence of ambient light: ； Among them, I(x,y) is the original image, I corrected (x,y) is the original image after spectral correction; Geometric correction uses the geometric transformation matrix T to correct the image: ； in, is the image after geometric correction; Extract the spectral vector of each pixel from the rectified image to form a complete spectral dataset; S2: Extract spectra from the spectral data set based on the masterbatch component recognition model to identify the spectral characteristics of different components or impurities; S3: Constructing a masterbatch color difference detection model and a defect recognition model, and performing color difference detection and defect recognition based on spectral features; S4: Based on the real-time image color difference detection and defect recognition results, the control system of the production line is controlled based on the intelligent agent feedback to achieve automatic adjustment; S5: Visualization based on color difference detection and defect recognition results enables operators to obtain real-time masterbatch detection results and issue alarms for unqualified products.

2. The method for detecting high-concentration EVA masterbatch based on multispectral imaging according to claim 1, characterized in that: The masterbatch component recognition model is constructed based on a support vector machine, as follows: Acquire multispectral images of masterbatch, extract the spectral vector of each pixel from each multispectral image, and construct a spectral data set; Use filters to remove noise from spectral data and normalize the spectral data; Combining principal component analysis and linear discriminant analysis to extract features, and marking the extracted spectral features to obtain the final spectral feature set; A masterbatch component recognition model was constructed based on support vector machine (SVM), and the model was trained using the final spectral feature set to ensure that the model could accurately identify different components. The model performance was evaluated through cross-validation, and the model parameters were adjusted to improve the recognition accuracy to obtain the final masterbatch component recognition model.

3. A method for detecting high-concentration EVA masterbatch based on multispectral imaging according to claim 2, characterized in that: The feature extraction by combining principal component analysis and linear discriminant analysis is specifically as follows: Based on principal component analysis, the dimension of spectral data is reduced, the covariance matrix is ​​calculated, eigenvalue decomposition is performed, and the principal components are extracted: ； Where Q is the covariance matrix; The mean vector of the spectral data; is the i-th spectral vector; N is the number of spectral data vectors; the superscript T indicates transposition; Select the first k eigenvectors to achieve dimensionality reduction, that is: ； in, is the eigenvector matrix of PCA; Spectral data after dimension reduction; is the spectral data after normalization; Linear discriminant analysis (LDA) is used to extract features that maximize inter-class differences and minimize intra-class differences, and calculate intra-class and inter-class scatter matrices: ； Among them, S W is the intra-class scatter matrix; S B is the between-class scatter matrix; represents the mean of the spectral vector of category c; represents the ith spectral vector belonging to category c; N c is the number of vectors of spectral data after dimension reduction; is the mean of the spectral vector; C is the number of categories; Solution The characteristic values ​​of: ； Among them, V is the projection matrix of linear discriminant analysis, S LDA is the spectral feature data after linear discriminant analysis.

4. The method for detecting high-concentration EVA masterbatch based on multispectral imaging according to claim 2, characterized in that: The masterbatch component recognition model is constructed based on the support vector machine SVM, as follows: The final spectral feature set is divided into a training set and a test set, and the spectral features of the training set and the test set are standardized; Use Bootstrap random sampling with replacement to generate multiple different subsets from the training set; Train an independent SVM model for each subset, use cross-validation to evaluate the model performance for the SVM parameters selected during training, and adjust the SVM parameters of each model using the Bayesian optimization method; The prediction results of each trained SVM model on the test sample are voted, and the final category of the sample is determined by majority voting.

5. The method for detecting high-concentration EVA masterbatch based on multispectral imaging according to claim 4, characterized in that: The SVM model is specifically as follows: Using RBF kernel function: in, They are and samples; γ is the parameter of the RBF kernel function; SVM objective function: ; in, is the penalty factor; For the The slack variable corresponding to the samples; is the weight vector; b is the bias term; N’ is the total number of samples; By introducing Lagrange multipliers, the optimization problem of SVM is converted into a dual form and calculated using kernel functions, and the objective function is converted to: ； in, , is the Lagrange multiplier; , They are and The true labels of samples; represents the set of all Lagrange multipliers in the entire optimization process; Finally, the classification decision function of SVM is: ； in, is a new input sample.

6. The method for detecting high-concentration EVA masterbatch based on multispectral imaging according to claim 1, characterized in that: The masterbatch color difference detection model is as follows: color-related spectral bands are extracted from the spectral features of the identified components, and the spectral features are converted into the CIE 1931 XYZ color space by combining the standard illuminant and the standard observer function: ； Where, S(λ) is the sample spectrum; is the CIE standard observer chromaticity function; X, Y, Z are the representations of colors in the XYZ color space; E(λ) is the standard illuminant; λ represents the wavelength; Convert XYZ values ​​to CIE-LAB color values: ； in, Indicates brightness, and Indicates color contrast; is the standard white reference value; where f(t) is defined as: ； Among them, t represents the input and δ is a constant; use Formula to calculate color difference: in, and is the LAB value of the reference sample and the sample to be tested; Set the standard color difference threshold. If the color difference exceeds this threshold, it will be judged as color difference abnormality.

7. The method for detecting high-concentration EVA masterbatch based on multispectral imaging according to claim 1, characterized in that: The defect recognition model is built based on XGBOOST, as follows: Start with the mean of all samples as the base model, and in each iteration, calculate the residual of the current model, that is, the difference between the predicted value and the actual value; Using the residuals as the target variable, train a new decision tree to fit these residuals; Add the output of the new tree to the current model and update the predicted value; And add regularization terms to the objective function of XGBoost to control model complexity and prevent overfitting; In each iteration, XGBoost selects the best tree structure and leaf node weights by optimizing the following objective function: ； in, is the loss function; is the actual value; It is Round The predicted value of samples; represents the g-th decision tree; is the objective function, the superscript Indicates wheel; The number of samples; G is the number of decision trees.

8. The method for detecting high-concentration EVA masterbatch based on multispectral imaging according to claim 1, characterized in that: The S4 is specifically: Use proportional-integral-derivative PID controller to adjust the operating parameters of production equipment based on color difference and defect information; ； in, Control output; is the error, , are the color difference error and the number of defects, respectively, which are defined as the difference between the target value and the actual detection value; , , are the parameters of the PID controller, corresponding to the proportional, integral and differential gains respectively; The control strategy is optimized using the reinforcement learning algorithm Q-learning. The intelligent agent learns the optimal control action by interacting with the environment. Rewards are given according to the degree of color difference and defect reduction to guide the intelligent agent's learning. Define the state space S, including the current color difference, number of defects, and equipment parameters; the action space A is the control action, including adjusting the temperature, speed, and material ratio; reward is given according to the degree of color difference and defect reduction, and the reward function is defined as: ； Where Defects is the number of defects; ΔE ab is the color difference, and is the weight coefficient; Q value update formula: ； in, is the value of executing action a in state s; is the discount factor; s and Respectively represent the current state and the subsequent state to which the action is transferred after execution; a and Respectively represent the current state s and the subsequent state The possible actions to be taken.

Images