Plastic recycled material purity monitoring method based on artificial intelligence

Abstract

The invention relates to a recycled plastic material purity monitoring method based on artificial intelligence, and the method comprises the steps: collecting surface visual image data and spectral signal data of a recycled plastic material, environment parameter data and corresponding recycled plastic material purity data, and carrying out the data processing; constructing a double-branch deep neural network model, and respectively extracting a surface visual image feature vector and a spectral signal feature vector from the surface visual image data and the spectral signal data and fusing the surface visual image feature vector and the spectral signal feature vector; in combination with the plastic recycled material fusion feature vector, a plastic feature data set is adopted to train and optimize a double-branch deep neural network model; and collecting real-time surface visual image data, real-time spectral signal data and real-time environmental parameter data of the to-be-monitored plastic recycled material, and inputting the data into the plastic recycled material purity monitoring model to obtain a real-time purity value of the plastic recycled material. The accuracy and efficiency of monitoring the purity of the recycled plastic material can be improved, and the problems that manual detection is low in efficiency and large in error are solved.

Description

Technical Field

[0001] The present application relates to the field of material detection technology, and specifically to a method for monitoring the purity of recycled plastics based on artificial intelligence. Background Art

[0002] Purity monitoring of recycled plastics refers to the process of testing the proportion of target plastic components, the types and contents of impurities contained in recycled plastics. Its purpose is to evaluate the quality of recycled materials and determine whether they meet the standards for reuse. It is of great significance to ensuring the quality of plastic products and the sustainable development of the plastic recycling industry.

[0003] Among the current technologies for monitoring the purity of recycled plastics, traditional near-infrared spectroscopy has difficulty identifying black plastic due to its strong absorption of near-infrared light. Although medium-wave infrared spectroscopy can solve this problem, it is difficult to widely use due to the high cost of equipment and reliance on specialized hardware. In addition, spectral technology lacks sensitivity for detecting low-proportion mixed materials, making it difficult to accurately distinguish plastics with similar chemical compositions. The computer vision technology used is significantly restricted by environmental factors. Contamination, wear and tear on the plastic surface, and changes in lighting will seriously affect the accuracy of visual features, making it difficult for the recognition model to operate stably in complex environments. Moreover, traditional visual methods rely solely on appearance features and cannot effectively distinguish plastics with similar chemical compositions, making misjudgment prone. Summary of the Invention

[0004] To address the above-mentioned problems in the prior art, the present application aims to provide an artificial intelligence-based method for monitoring the purity of recycled plastics. This artificial intelligence-based method can improve the accuracy and efficiency of monitoring the purity of recycled plastics, solving the problems of low efficiency and large errors in manual detection.

[0005] The present application discloses an artificial intelligence-based method for monitoring the purity of recycled plastics, comprising the following steps:

[0006] S1. Obtain multiple batches of historically monitored plastic recycled materials, collect surface visual image data and spectral signal data of the plastic recycled materials, as well as environmental parameter data and corresponding plastic recycled material purity data, and perform data processing to obtain a plastic feature dataset;

[0007] S2. Construct a dual-branch deep neural network model, extract a surface visual image feature vector and a spectral signal feature vector from the surface visual image data and the spectral signal data, respectively, and fuse them to obtain a fused feature vector of the plastic recycled material;

[0008] S3. Combine the fusion feature vector of the recycled plastics and use the plastic feature dataset to train and optimize the dual-branch deep neural network model to obtain a recycled plastics purity monitoring model;

[0009] S4. Collect real-time surface visual image data and real-time spectral signal data, as well as real-time environmental parameter data of the plastic recycled material to be monitored, and input them into the plastic recycled material purity monitoring model to obtain the real-time purity value of the plastic recycled material and generate the plastic recycled material purity monitoring result.

[0010] Preferably, the step S1 specifically includes:

[0011] The surface visual image data includes a surface image of the recycled plastic material and plastic color information, wherein the plastic color information is hue, saturation, and brightness information corresponding to an HSV color space obtained based on the RGB color value of each pixel in the surface image of the recycled plastic material;

[0012] The spectral signal data includes near infrared spectral data of plastic recycled materials in the wavelength range of 780-2526nm and 200-3500cm -1 Raman spectral data in the wavenumber range;

[0013] The environmental parameter data includes temperature data, humidity data and light intensity data of the environment in which the plastic recycled material is located;

[0014] The collected surface visual image data, the spectral signal data and the environmental parameter data are transmitted to an edge computing gateway.

[0015] Preferably, the step S1 further includes:

[0016] Obtaining the surface image of the plastic recycling material from the edge computing gateway, performing median filtering denoising and histogram equalization processing, converting the surface image of the plastic recycling material into a grayscale image, and calculating a gray-level co-occurrence matrix;

[0017] Extracting a texture pattern of the surface image of the recycled plastic material using a local binary pattern to obtain texture features of the surface of the recycled plastic material;

[0018] The contour of the plastic recycling material is obtained through edge detection algorithm, and the perimeter, area, circularity and rectangularity of the contour are calculated, and the shape features are extracted using Fourier descriptors;

[0019] Detecting and analyzing the surface image of the recycled plastic material to obtain the length and width of scratches, the length and direction of cracks, the area and diameter of holes, and the number and volume of bubbles on the surface of the recycled plastic material;

[0020] The near-infrared spectral data were filtered, smoothed, and subjected to principal component analysis and dimensionality reduction processing to obtain the absorbance within the wavelength range of 780-2526 nm and the absorbance peaks at wavelengths of 1100 nm, 1450 nm, and 1900 nm;

[0021] Acquire Raman spectra 200-3500cm -1 The scattered light intensity in the wavenumber range and polyethylene 2840cm -1 , Polypropylene 800-850cm -1 , polyvinyl chloride 600-700cm -1 The intensity and position of characteristic peaks;

[0022] Combined with the temperature data, the humidity data and the light intensity data, the processed surface visual image data and the spectral signal data are calibrated, and the calibrated surface visual image data, spectral signal data, and environmental parameter data are associated, labeled, and integrated with the plastic recycled material purity data to form a plastic feature data set.

[0023] Preferably, the step S2 specifically includes:

[0024] Constructing a dual-branch deep neural network model including a surface visual image data processing branch and a spectral signal data processing branch, wherein the surface visual image data processing branch is composed of multiple convolutional layers, pooling layers, and fully connected layers, and the spectral signal data processing branch is composed of multiple one-dimensional convolutional layers, global average pooling layers, and fully connected layers;

[0025] Inputting the surface visual image data in the plastic feature dataset into the surface visual image data processing branch, and sequentially passing through a convolutional layer, a pooling layer, and a fully connected layer to obtain a surface visual image feature vector;

[0026] Inputting the spectral signal data in the plastic feature dataset into the spectral signal data processing branch, and sequentially passing through a one-dimensional convolutional layer, a global average pooling layer, and a fully connected layer to obtain a spectral signal feature vector;

[0027] The surface visual image feature vector and the spectral signal feature vector are fused by splicing to form a plastic recycling material fusion feature vector.

[0028] Preferably, the step S3 specifically includes:

[0029] Dividing the plastic feature dataset into a training set and a test set in a ratio of A1:A2, and initializing the parameters of each layer of the dual-branch deep neural network model;

[0030] Setting a training round Q, combining the fusion feature vector of the recycled plastic material, and using the training set to train the dual-branch deep neural network model, performing Q iterative training;

[0031] Inputting the test set into the trained two-branch deep neural network model for testing, combining the plastic recycled material fusion feature vector for forward propagation, and obtaining the plastic recycled material purity prediction value of all samples in the test set;

[0032] The dual-branch deep neural network model is evaluated based on the predicted purity value of the plastic recycled material.

[0033] Preferably, the model evaluation includes:

[0034] The plastic recycling material purity prediction value includes a first purity prediction value P based only on surface visual image features. v , the second purity prediction value P based only on spectral signal characteristics s , and the third purity prediction value P of the fusion feature f , calculate the mean inter-modal consistency score C J And the weighted evaluation index M C :

[0035]

[0036] Among them, C i represents the inter-modal consistency score of the i-th sample in the test set; P v (i) represents the first purity prediction value of the i-th sample in the test set; P s (i) represents the second purity prediction value of the i-th sample in the test set; P f (i) represents the third purity prediction value of the i-th sample in the test set; y (i) represents the actual purity value of the i-th sample in the test set; N represents the total number of samples in the test set; i=1, 2...N;

[0037] According to the inter-modality consistency score C i And the sample prediction variance of the test set, dynamically scale the confidence interval, and calculate the confidence interval:

[0038]

[0039] Where Z represents the dynamic adjustment coefficient of the dynamic scaling confidence interval; Z0 represents the standard confidence level quantile; γ represents the adaptive coefficient; represents the prediction variance of the i-th sample in the test set obtained by Monte Carlo calculation; max(σ 2 ) represents the maximum value of the variance of all samples in the test set; CI i represents the confidence interval of the i-th sample in the test set;

[0040] If M C ≤0.5, C J ≥0.7 and CI i If the proportion of samples that do not cover the true value is ≤ 20%, the two-branch deep neural network model is determined to meet the requirements and is used as a plastic recycled material purity monitoring model;

[0041] If M C ≤0.5, C J ≥0.7 and CI i The proportion of samples that do not cover the true value is ≤ 20%. If any one of them is not met, it is determined that the dual-branch deep neural network model does not meet the requirements, and the model is optimized until M C ≤0.5, C J ≥0.7 and CI i The proportion of samples that do not cover the true value is ≤ 20%.

[0042] Preferably, the model optimization includes:

[0043] According to the obtained weighted evaluation index M C , the mean inter-modality consistency score C J and confidence interval CI i , analyzing and optimizing the dual-branch deep neural network model;

[0044] When M C > 0.5, it is determined that the feature extraction is insufficient, and the feature extraction optimization strategy is executed until M is satisfied. C ≤0.5;

[0045] When C J <0.7, it is determined that the feature fusion mechanism has defects, and the feature fusion mechanism optimization strategy is executed until C is satisfied. J ≥0.7;

[0046] When CI i If the proportion of samples that do not cover the true value is greater than 20%, it is determined that the confidence interval cannot effectively contain the true value, and the confidence interval optimization strategy is executed until the CI is satisfied. i The proportion of samples that do not cover the true value is ≤ 20%.

[0047] Preferably, the step S4 specifically includes:

[0048] Collect the real-time surface visual image data and real-time spectral signal data of the plastic recycling material to be monitored, as well as the real-time environmental parameter data, perform data processing, and input the data into the plastic recycling material purity monitoring model after the data processing is completed to obtain the preliminary purity value P of the plastic recycling material pre ;

[0049] According to the pre-established calibration sample library, the similarity matching algorithm is used to retrieve multiple historical samples that are most similar to the characteristics of the current plastic recycled material, and the calibration purity mean P is calculated. J , and the preliminary purity value P pre Calibrate to get the real-time purity value P of plastic recycling material S :

[0050] P S =P pre +α*(P J -P pre )

[0051] Among them, α represents the adaptive weight, which is dynamically adjusted according to the sample matching degree and ranges from 0 to 1;

[0052] According to the real-time purity value P S Perform purity analysis on recycled plastics and generate purity monitoring results for recycled plastics.

[0053] Preferably, the purity analysis includes:

[0054] If P S ≥95%, the purity grade of the plastic recycling material is determined to be level one, and the plastic recycling material will continue to go through the original processing flow;

[0055] If 90%≤P S If the purity level of the recycled plastic material is less than 95%, the recycled plastic material is judged to be of Grade II purity, and an alarm message is issued, and the recycled plastic material is subject to a second re-inspection.

[0056] If 80%≤P S If the purity level of the recycled plastic material is less than 90%, the recycled plastic material is judged to be of grade three purity, and an alarm message is issued to sort the recycled plastic material to the downgrading processing area;

[0057] If P S If the purity level of the recycled plastic material is less than 80%, the recycled plastic material is judged to be of grade 4 purity, and an alarm message is issued, and the recycled plastic material is subjected to enhanced impurity removal treatment.

[0058] Preferably, the plastic recycled material purity monitoring results include:

[0059] The real-time surface visual image data, the real-time spectral signal data, the real-time environmental parameter data and the real-time purity value of the same batch of plastic recycled materials are integrated and processed, and the plastic recycled material purity monitoring results including the real-time purity value, purity grade and purity fluctuation trend are generated through visualization technology.

[0060] The advantages of the artificial intelligence-based plastic recycling purity monitoring method described in this application are:

[0061] The present application discloses a method for monitoring the purity of recycled plastics based on artificial intelligence. By synchronously collecting surface visual image data and spectral signal data, combined with environmental parameter data, a dual-branch deep neural network is used to achieve multi-dimensional feature fusion, thereby avoiding the one-sidedness of a single modality and significantly improving the accuracy and anti-interference ability of purity monitoring. By introducing environmental parameter data, the interference of environmental factors on the monitoring results can be corrected. The dual-branch deep neural network automatically learns high-dimensional abstract features to replace traditional artificial feature engineering, while being able to capture subtle differences that are difficult for the human eye to perceive. The fusion feature vector is directly mapped to the purity value to avoid error accumulation in multi-stage processing and significantly shorten the monitoring cycle. Real-time monitoring can provide immediate feedback, timely adjust the recycling process, improve efficiency, and automated monitoring reduces manual intervention and reduces costs. The method for monitoring the purity of recycled plastics based on artificial intelligence can improve the accuracy and efficiency of monitoring the purity of recycled plastics, and solve the problems of low efficiency and large errors in manual detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is a flow chart of an artificial intelligence-based plastic recycled material purity monitoring method described in this application. DETAILED DESCRIPTION

[0063] like Figure 1 As shown, the artificial intelligence-based plastic recycling material purity monitoring method described in this application includes the following steps:

[0064] S1. Obtain multiple batches of historically monitored purity recycled plastics, collect surface visual image data and spectral signal data of the recycled plastics, as well as environmental parameter data and corresponding recycled plastics purity data, and process the data to obtain a plastic feature dataset; the recycled plastics purity data includes the purity value of each recycled plastic, which is the purity value that has been monitored historically;

[0065] S2. Construct a dual-branch deep neural network model to extract surface visual image feature vectors and spectral signal feature vectors from the surface visual image data and spectral signal data respectively, and fuse them to obtain a fused feature vector of the plastic recycled material;

[0066] S3. Combine the fusion feature vector of recycled plastics and use the plastic feature dataset to train and optimize the dual-branch deep neural network model to obtain a recycled plastics purity monitoring model.

[0067] S4. Collect real-time surface visual image data and real-time spectral signal data, as well as real-time environmental parameter data of the plastic recycled material to be monitored, and input them into the plastic recycled material purity monitoring model to obtain the real-time purity value of the plastic recycled material and generate the plastic recycled material purity monitoring results.

[0068] Furthermore, in this embodiment, step S1 specifically includes:

[0069] The surface visual image data includes a surface image of the plastic recycling material and plastic color information. The plastic color information is the hue, saturation, and brightness information corresponding to the HSV color space obtained based on the RGB color value of each pixel in the surface image of the plastic recycling material.

[0070] The spectral signal data includes the near infrared spectrum data of plastic recycled materials in the wavelength range of 780-2526nm and 200-3500cm -1 Raman spectral data in the wavenumber range;

[0071] Environmental parameter data include temperature data, humidity data and light intensity data of the environment in which the plastic recycled materials are located;

[0072] Transmit the collected surface visual image data, spectral signal data and environmental parameter data to the edge computing gateway;

[0073] Specifically, for the collection of surface visual image data, an industrial camera, such as a CCD camera, can be used to collect images of the surface of the plastic recycled material. An image processing algorithm is then used to convert the RGB color value of each pixel in the surface image of the plastic recycled material into the HSV color space to obtain the corresponding hue, saturation, and brightness information.

[0074] For the collection of spectral signal data, a near-infrared spectrometer can be used to collect near-infrared spectral data of plastic recycled materials in the wavelength range of 780-2526nm, and a Raman spectrometer can be used to collect near-infrared spectral data of plastic recycled materials in the wavelength range of 200-3500cm -1 Raman spectral data in the wavenumber range;

[0075] The temperature data is collected by the temperature sensor to collect the ambient temperature value, the humidity data is collected by the humidity sensor to collect the relative humidity percentage of the environment, and the temperature data and humidity data can be collected by the temperature and humidity sensor. The light intensity data is collected by the light intensity sensor to collect the ambient lux light intensity value:

[0076] The collected surface visual image data, spectral signal data and environmental parameter data are transmitted to the edge computing gateway via wired or wireless transmission.

[0077] Furthermore, in this embodiment, step S1 further includes:

[0078] Obtain the surface image of the plastic recycling material from the edge computing gateway, perform median filtering denoising and histogram equalization processing, convert the surface image of the plastic recycling material into a grayscale image, and calculate the gray-level co-occurrence matrix;

[0079] The local binary pattern is used to extract the texture pattern of the surface image of the plastic recycling material, and the texture characteristics of the surface of the plastic recycling material are obtained;

[0080] The contour of the plastic recycling material is obtained through edge detection algorithm, and the perimeter, area, circularity and rectangularity of the contour are calculated, and the shape features are extracted using Fourier descriptors;

[0081] Detect and analyze the surface images of recycled plastics to obtain the length and width of scratches, length and direction of cracks, area and diameter of holes, and number and volume of bubbles on the surface of recycled plastics;

[0082] The near-infrared spectral data were filtered, smoothed, and subjected to principal component analysis to reduce dimensionality, obtaining the absorbance within the wavelength range of 780-2526 nm and the absorbance peaks at wavelengths of 1100 nm, 1450 nm, and 1900 nm.

[0083] Acquire Raman spectra 200-3500cm -1 The scattered light intensity in the wavenumber range and polyethylene 2840cm -1 , Polypropylene 800-850cm -1 , polyvinyl chloride 600-700cm -1 The intensity and position of characteristic peaks;

[0084] Combined with temperature data, humidity data, and light intensity data, the processed surface visual image data and spectral signal data are calibrated. The calibrated surface visual image data, spectral signal data, environmental parameter data, and plastic recycled material purity data are associated, labeled, and integrated to form a plastic feature data set.

[0085] Specifically, after obtaining the surface image of the recycled plastic material from the edge computing gateway, median filtering is first performed to remove discrete noise points such as salt and pepper noise in the image, making the image smoother. Histogram equalization is then performed to adjust the image's grayscale histogram, redistribute the image's grayscale values, and enhance the image's overall contrast.

[0086] Convert the processed color image into a grayscale image to facilitate subsequent analysis and calculate the grayscale co-occurrence matrix, which can describe the joint distribution of grayscale values ​​of two pixels with a specific spatial relationship in the image. By analyzing the eigenvalues ​​of the grayscale co-occurrence matrix, such as the energy, entropy, contrast, and correlation of the image, the texture characteristics of the image can be obtained.

[0087] The local binary pattern (LBP) is used to extract the surface image texture pattern of the plastic recycled material. The basic principle of the local binary pattern is to take a certain pixel point in the image as the center, set a circular neighborhood (for example, the radius is 1, including 8 neighborhood points), and use the gray value of the central pixel as the threshold. It is compared with the gray value of the neighboring pixels. If the gray value of the neighboring pixel is greater than or equal to the gray value of the central pixel, it is recorded as 1, otherwise it is recorded as 0. These comparison results are combined into a binary number in a certain order (for example, clockwise), and then converted into a decimal number as the LBP value of the central pixel.

[0088] By calculating the LBP values ​​of all pixels in the image, the LBP texture pattern of the image is obtained, thereby further obtaining the texture characteristics of the surface of the plastic recycled material;

[0089] Use edge detection algorithms (such as the Canny edge detection algorithm) to obtain the contour of the plastic recycling material, calculate the contour's perimeter, area, circularity, and rectangularity parameters, and use Fourier descriptors to extract shape features. The Fourier descriptor converts the coordinate information of the object's contour into the frequency domain through Fourier transform, and describes the object's shape using a series of Fourier coefficients. These coefficients can reflect the object's shape characteristics. Even if the object is translated, rotated, and scaled, the consistency of the shape description can be maintained by properly processing the Fourier coefficients.

[0090] Surface images of recycled plastics are inspected and analyzed, and image recognition and measurement algorithms are used to determine surface details such as scratch length and width, crack length and direction, hole area and diameter, and bubble quantity and volume. For example, for scratch detection, edge detection and morphological operations can be used to determine the scratch edge, and pixel counting methods can be used to measure its length and width. For holes, a region growing algorithm can be used to mark the hole area, and then its area and diameter can be calculated.

[0091] The near-infrared spectral data can be filtered and smoothed using the Savitzky-Golay filtering method and the principal component analysis (PCA) dimensionality reduction process. By performing eigendecomposition on the covariance matrix of the spectral data, the most important direction of change in the data is found, and the high-dimensional spectral data is projected into a low-dimensional space. While retaining most of the data information, the data dimension is reduced and the amount of calculation is reduced.

[0092] Obtain absorbance within the wavelength range of 780-2526nm, as well as absorbance peaks at 1100nm, 1450nm, and 1900nm. The absorbance peaks at these specific wavelengths are important indicators for identifying the type and purity of plastics.

[0093] Acquire Raman spectra 200-3500cm -1 The scattered light intensity in the wavenumber range and polyethylene 2840cm-1 , Polypropylene 800-850cm -1 , polyvinyl chloride 600-700cm -1 The intensity and position of characteristic peaks. Different plastics have unique characteristic peaks in Raman spectra. By measuring the intensity and position of these characteristic peaks, the composition of the plastic and the presence of impurities can be determined, thus providing a basis for purity analysis.

[0094] Combined with temperature data, humidity data, and light intensity data, the processed surface visual image data and spectral signal data are calibrated. Because environmental factors (such as temperature changes may affect the physical form of plastics and thus affect spectral characteristics, and humidity and light intensity may affect image acquisition quality) will interfere with the data, the visual and spectral data need to be adjusted according to environmental parameters;

[0095] The calibrated surface visual image data, spectral signal data, environmental parameter data and plastic recycled material purity data are associated, labeled and integrated, giving clear correspondences to various data types of each sample, and finally forming a plastic feature dataset;

[0096] An example of step S1 is as follows:

[0097] Obtain 20 historical batches of PE plastic recycled materials, each batch containing 100 samples, for a total of 2,000 samples;

[0098] Using an industrial CCD camera with a resolution of 6 megapixels, we captured surface images of 2,000 samples one by one at a distance of 25 cm from the recycled plastic material and in an environment with a light intensity of 1,300 Lux.

[0099] A near-infrared spectrometer with a wavelength range of 780–2526 nm and 6 scans was used, as well as a wavenumber range of 200–3500 cm -1 , a Raman spectrometer with a laser power of 90mW to obtain the spectral signal data of each sample;

[0100] A temperature and humidity sensor with a measurement accuracy of ±0.4°C and ±1.8% RH was used to record an ambient temperature of 27°C and a humidity of 40% RH. A light intensity sensor with a measurement accuracy of ±3% FS was used to record a light intensity of 1300 Lux.

[0101] The purity data of these 20 batches of plastic recycled materials have been determined by conventional methods (such as chromatography), for example, the purity of batch 1 was 91%, and the purity of batch 2 was 89%.

[0102] After the collected data is transmitted to the edge computing gateway, the surface visual image is processed: a 3×3 window median filter is used to remove noise, and histogram equalization is used to enhance contrast; the image is converted to grayscale and the gray-level co-occurrence matrix is ​​calculated;

[0103] The LBP algorithm with a neighborhood radius of 1 and a number of neighborhood points of 8 was used to extract texture patterns. The Canny edge detection (with high and low thresholds of 0.35 and 0.12) was used to obtain contours. Shape parameters such as perimeter and area were calculated, and scratches with a length of 3 mm were detected in some samples.

[0104] In terms of spectral data processing, the near-infrared spectrum was smoothed using Savitzky-Golay filtering with a window size of 5, and the dimension was reduced by principal component analysis to obtain the absorbance peak at a specific wavelength;

[0105] Extracted Raman spectrum 200-3500cm -1 The scattered light intensity within the wavenumber range and the PE characteristic peak (2840cm -1 )’s strength and location;

[0106] The visual and spectral data were calibrated in combination with the environmental parameters of light intensity, temperature and humidity. The calibrated image features, spectral features, and environmental parameters were associated and labeled with the purity values ​​of each sample in the corresponding batch, and integrated to form a plastic feature dataset containing 2,000 records.

[0107] Furthermore, in this embodiment, step S2 specifically includes:

[0108] Construct a dual-branch deep neural network model, including a surface visual image data processing branch and a spectral signal data processing branch. The surface visual image data processing branch consists of multiple convolutional layers, pooling layers, and fully connected layers. The spectral signal data processing branch consists of multiple one-dimensional convolutional layers, global average pooling layers, and fully connected layers.

[0109] The surface visual image data in the plastic feature dataset is input into the surface visual image data processing branch, and passes through the convolution layer, pooling layer, and fully connected layer in sequence to obtain the surface visual image feature vector; the surface visual image data passes through the convolution layer in sequence for feature extraction, the pooling layer reduces the data dimension, and the fully connected layer integrates the features to obtain the surface visual image feature vector;

[0110] The spectral signal data in the plastic feature dataset is input into the spectral signal data processing branch and sequentially passes through a one-dimensional convolutional layer, a global average pooling layer, and a fully connected layer to obtain a spectral signal feature vector. The spectral signal data is sequentially passed through a one-dimensional convolutional layer to extract local features of the spectral data, a global average pooling layer to aggregate features, and a fully connected layer to further adjust and combine features to obtain a spectral signal feature vector.

[0111] The surface visual image feature vector and the spectral signal feature vector are fused by splicing to form a fusion feature vector of the plastic recycling material;

[0112] An example of step S2 is as follows:

[0113] Construct a dual-branch deep neural network model, for the surface visual image data processing branch:

[0114] Convolutional layer settings: Determine the number of convolutional layers, for example, set it to 3 convolutional layers, each with a different number of convolution kernels. The first convolutional layer has 32 3×3 convolution kernels, a stride of 1, a padding of 1, and uses the ReLU activation function, which extracts basic image features such as edges, textures, and other primary information through convolution operations.

[0115] The second convolutional layer has 64 3×3 convolution kernels with a stride of 1 and a padding of 1. It also uses the ReLU activation function to further extract more complex features.

[0116] The third convolutional layer has 128 3×3 convolution kernels with a stride of 1, a padding of 1, and a ReLU activation function to make feature extraction more in-depth;

[0117] Pooling layer settings: Set up two pooling layers, using the maximum pooling method. The first pooling layer window size is 2×2, and the stride is 2. The function is to downsample the convolution layer output to reduce the amount of data while retaining the main features. The second pooling layer window size is 2×2, and the stride is 2 to further reduce the data dimension.

[0118] Fully connected layer settings: Set up two fully connected layers. The first fully connected layer has 256 neurons and uses the ReLU activation function to integrate and transform the features output by the pooling layer. The second fully connected layer has 128 neurons to prepare for the subsequent output feature vector.

[0119] For the spectral signal data processing branch:

[0120] One-dimensional convolution layer settings: Determine the number of one-dimensional convolution layers. For example, set it to 3 one-dimensional convolution layers. The first one-dimensional convolution layer is equipped with 32 convolution kernels of length 3, with a stride of 1, and uses the ReLU activation function to perform preliminary feature extraction on the spectral signal data.

[0121] The second one-dimensional convolutional layer has 64 convolution kernels of length 3, a step size of 1, and a ReLU activation function to further extract the characteristics of the spectral signal;

[0122] The third one-dimensional convolutional layer has 128 convolution kernels of length 3, a stride of 1, and a ReLU activation function to further explore the spectral features;

[0123] Global average pooling layer setting: The output of the one-dimensional convolution layer is processed by the global average pooling layer, and the average value of each feature map is used as the output, which reduces the data dimension and highlights the overall features;

[0124] Fully connected layer settings: Set up two fully connected layers. The first fully connected layer has 256 neurons and uses the ReLU activation function to integrate and transform the pooled spectral features. The second fully connected layer has 128 neurons and prepares for outputting the spectral signal feature vector.

[0125] Processing surface visual image data:

[0126] Preprocess the surface visual image data in the plastic feature dataset, adjust the image size to a uniform specification, such as 224×224 pixels, and then input it into the surface visual image data processing branch;

[0127] The network passes through the first convolutional layer, where the convolution kernel is convolved with the image to extract preliminary features, and then passes through the ReLU activation function to increase nonlinearity. It then enters the first pooling layer for downsampling, and repeats through subsequent convolutional and pooling layers to gradually extract and filter features. Finally, it passes through the fully connected layer, where the features are integrated and transformed to obtain a surface visual image feature vector of length 128.

[0128] Processing spectral signal data:

[0129] Normalize the spectral signal data in the plastic feature data set to make its value range uniform, and then input it into the spectral signal data processing branch;

[0130] After passing through the first one-dimensional convolution layer, the convolution kernel is convolved with the spectral signal to extract preliminary features, and then the ReLU activation function is used to increase nonlinearity. After the global average pooling layer reduces the dimension, the fully connected layer is used for feature integration transformation to obtain a spectral signal feature vector with a length of 128.

[0131] Fuse feature vectors:

[0132] The obtained surface visual image feature vector and spectral signal feature vector are fused by splicing them in dimension. For example, two 128-dimensional vectors are spliced ​​into a 256-dimensional fused feature vector of recycled plastics.

[0133] For example, the plastic feature dataset contains 2000 samples, and surface visual image data is extracted from each sample. After the extracted surface visual image data is resized to 224×224 pixels, it is input into the surface visual image data processing branch;

[0134] Features are extracted and filtered through the convolution layer and pooling layer in sequence, and then a 128-dimensional surface visual image feature vector is obtained through the fully connected layer. At the same time, the spectral signal data of each sample is normalized and input into the spectral signal data processing branch. After passing through the one-dimensional convolution layer, the global average pooling layer and the fully connected layer, a 128-dimensional spectral signal feature vector is obtained. Finally, these two 128-dimensional vectors are spliced ​​to form a 256-dimensional plastic recycled material fusion feature vector.

[0135] Furthermore, in this embodiment, step S3 specifically includes:

[0136] The plastic feature dataset is divided into training and test sets in the ratio of A1:A2, and the parameters of each layer of the two-branch deep neural network model are initialized;

[0137] Set the training round Q, combine the fusion feature vector of the plastic recycled materials, use the training set to train the two-branch deep neural network model, and perform Q iterative training;

[0138] The test set is input into the trained two-branch deep neural network model for testing. The forward propagation is combined with the fusion feature vector of the plastic recycled material to obtain the predicted purity value of the plastic recycled material for all samples in the test set.

[0139] The dual-branch deep neural network model was evaluated based on the predicted purity value of plastic recycled materials;

[0140] Specifically, when dividing the training set and test set according to the ratio of A1:A2, stratified sampling can be considered. Stratified sampling can ensure that the distribution of the training set and test set in each category is similar, avoiding model training bias caused by uneven data distribution.

[0141] For initialization of parameters of each layer of the dual-branch deep neural network model, a random initialization method (such as Xavier initialization) can be selected.

[0142] Furthermore, in this embodiment, model evaluation includes:

[0143] The purity prediction value of the plastic recycling material includes a first purity prediction value P based only on the surface visual image features. v , the second purity prediction value P based only on spectral signal characteristics s , and the third purity prediction value P of the fusion feature f , calculate the mean inter-modal consistency score C J And the weighted evaluation index M C :

[0144]

[0145] Among them, C irepresents the inter-modal consistency score of the i-th sample in the test set; P v (i) represents the first purity prediction value of the i-th sample in the test set; P s (i) represents the second purity prediction value of the i-th sample in the test set; P f (i) represents the third purity prediction value of the i-th sample in the test set; y (i) represents the actual purity value of the i-th sample in the test set; N represents the total number of samples in the test set; i = 1, 2...N;

[0146] According to the inter-modality consistency score C i And the sample prediction variance of the test set, dynamically scale the confidence interval, and calculate the confidence interval:

[0147]

[0148] Where Z represents the dynamic adjustment coefficient of the dynamic scaling confidence interval; Z0 represents the standard confidence level quantile; γ represents the adaptive coefficient; represents the prediction variance of the i-th sample in the test set obtained by Monte Carlo calculation; max(σ 2 ) represents the maximum variance of all samples in the test set; CI i represents the confidence interval of the i-th sample in the test set;

[0149] If M C ≤0.5, C J ≥0.7 and CI i If the proportion of samples that do not cover the true value is ≤ 20%, the two-branch deep neural network model is judged to meet the requirements and is used as a model for monitoring the purity of recycled plastics;

[0150] If M C ≤0.5, C J ≥0.7 and CI i The proportion of samples that do not cover the true value is ≤ 20%. If any one of them is not met, the dual-branch deep neural network model is judged to be unqualified and the model is optimized until M C ≤0.5, C J ≥0.7 and CI i The proportion of samples that do not cover the true value is ≤ 20%.

[0151] Furthermore, in this embodiment, model optimization includes:

[0152] According to the obtained weighted evaluation index M C , mean inter-modality consistency score C J and confidence interval CI i, analyze and optimize the dual-branch deep neural network model;

[0153] When M C > 0.5, it is determined that the feature extraction is insufficient, and the feature extraction optimization strategy is executed until M is satisfied. C ≤0.5; Feature extraction optimization strategies include increasing the number of layers or neurons in the dual-branch deep neural network to enhance the model's expressiveness;

[0154] When C J <0.7, it is determined that the feature fusion mechanism has defects, and the feature fusion mechanism optimization strategy is executed until C is satisfied. J ≥0.7; The feature fusion mechanism optimization strategy includes adding a cross-modal contrast learning module in the feature extraction stage to enhance modality consistency by minimizing the distance between the two modal features in the embedding space;

[0155] When CI i If the proportion of samples that do not cover the true value is greater than 20%, it is determined that the confidence interval cannot effectively contain the true value, and the confidence interval optimization strategy is executed until the CI is satisfied. i The proportion of samples that do not cover the true value is ≤ 20%. The confidence interval optimization strategy includes fine-tuning the parameters of the calculation formula of the dynamic adjustment coefficient z based on the deviation between the actual prediction result and the true value, and optimizing the confidence interval generation mechanism.

[0156] An example of step S3 is as follows:

[0157] Get 2,000 samples contained in the plastic feature dataset, all of which are PE (polyethylene) plastic recycled materials;

[0158] The dataset is divided into a training set and a test set according to the ratio of A1:A2=8:2. Then the training set contains 2000×0.8=1600 samples, and the test set contains 2000×0.2=400 samples.

[0159] During the segmentation process, a stratified sampling method was used to stratify the plastic recycled materials according to factors such as the purity range and production batch, ensuring that the sample distribution of the training set and the test set in each dimension is similar;

[0160] The Xavier initialization method is used to initialize the parameters of each layer of the dual-branch deep neural network model. The surface visual image data processing branch is constructed with 4 convolutional layers, 2 pooling layers, and 2 fully connected layers. The first convolutional layer is set with 32 3×3 convolution kernels, a stride of 1, and a padding of 1.

[0161] The spectral signal data processing branch is constructed as three one-dimensional convolutional layers, one global average pooling layer, and two fully connected layers. The first one-dimensional convolutional layer is set with 32 convolution kernels of length 3 and a stride of 1.

[0162] The training round number Q was set to 60, the Adam optimization algorithm was selected, the initial learning rate was set to 0.001, and the dual-branch deep neural network model was trained using the training set in combination with the fusion feature vector of the plastic recycled materials. The loss function was calculated in each round of training. If the loss function decreased by less than 0.0003 for four consecutive rounds, the training was terminated early.

[0163] The data of the 400 test samples was preprocessed. The surface visual image data was normalized and resized to 224×224 pixels, and the spectral signal data was standardized. The processed data was input into the trained two-branch deep neural network model and forward propagated in combination with the fusion feature vector of the plastic recycled material. For example, sample P-001 was processed separately by the surface visual image data processing branch and the spectral signal data processing branch. After fusion, it was output through the fully connected layer and the purity prediction value was 87%. This operation was repeated for all 400 samples to obtain the prediction value.

[0164] Calculate the inter-modality consistency score C i , taking sample P-002 as an example, the first purity prediction value P based only on the surface visual image features v (002) =85%, the second purity prediction value P based only on the spectral signal characteristics s (002) =86%, according to the formula C 002 =0.99, calculate C for 400 samples i And calculate the average to get the mean inter-modal consistency score C J =0.8;

[0165] Calculate the weighted evaluation index M C , the third purity prediction value P of the known sample P-002 based on the fusion feature f (002) =88%, actual purity value y (002) =89%, substitute into the formula (where N = 400), after calculating M C =0.48;

[0166] Calculate confidence interval CI i , calculate the sample prediction variance through Monte Carlo simulation For example, sample P-002 The standard confidence level quantile Z0 = 1.96, the adaptive coefficient γ = 0.5, the maximum value of the variance of all samples in the test set max(σ 2)=0.006, substitute it into the formula to calculate the dynamic adjustment coefficient Z, and get the confidence interval CI2 of sample P-002=[86%,90%], and the statistical result shows that the proportion of samples that CI2 does not cover the true value is 16%;

[0167] Because M C =0.48≤0.5, C J =0.8≥0.7, the proportion of samples that CI2 does not cover the true value is 16%≤20%, and the two-branch deep neural network model is judged to meet the requirements and is used as a plastic recycled material purity monitoring model.

[0168] Furthermore, in this embodiment, step S4 specifically includes:

[0169] Collect the real-time surface visual image data and real-time spectral signal data of the monitored plastic recycled materials, as well as the real-time environmental parameter data, and process the data. After the data processing is completed, input it into the plastic recycled material purity monitoring model to obtain the preliminary purity value P of the plastic recycled material. pre ; The data collection and data processing method of the plastic recycled materials to be monitored are consistent with the collection and processing method in step S1;

[0170] According to the pre-established calibration sample library, the similarity matching algorithm is used to retrieve multiple historical samples that are most similar to the characteristics of the current plastic recycled material, and the calibration purity mean P is calculated. J , and the initial purity value P pre Calibrate to get the real-time purity value P of plastic recycling material S :

[0171] P S =P pre +α*(P J -P pre )

[0172] Among them, α represents the adaptive weight, which is dynamically adjusted according to the sample matching degree and has a value range of 0 to 1. The adaptive weight α is selected according to the matching degree. When the matching degree is high, a large α is selected to make the calibration purity mean P J The initial purity value P pre The calibration effect is stronger. When the matching degree is low, a small α is selected to reduce the calibration amplitude and avoid over-calibration.

[0173] According to the real-time purity value P S Perform purity analysis on recycled plastics and generate purity monitoring results for recycled plastics.

[0174] Furthermore, in this embodiment, the purity analysis includes:

[0175] If P SIf the purity level is ≥95%, the plastic recycled material is judged to be of Grade 1 purity and will continue to undergo the original processing flow. Grade 1 purity indicates that the impurity content of the plastic recycled material is extremely low and does not affect its physical and chemical properties. Therefore, no additional treatment is required and the subsequent operations of the original processing flow can be carried out.

[0176] If 90%≤P S If the purity level is less than 95%, the plastic recycled material is judged to be of Grade II purity, and an alarm message is issued, and the plastic recycled material is subject to a second re-inspection. When the purity level is Grade II, it means that the plastic recycled material contains a small amount of impurities, which may have a certain impact on the quality of subsequent products. Therefore, a second re-inspection is required to more accurately detect the type and content of impurities, and further adjust the processing technology of the plastic recycled material based on the re-inspection results.

[0177] If 80%≤P S If the purity level is less than 90%, the plastic recycling material is judged to be of Grade III purity, and an alarm message is issued to sort the plastic recycling material to the downgrading treatment area. When the purity level is Grade III, it means that the impurity content of the plastic recycling material is relatively high and its quality can no longer meet the production requirements of conventional high-quality products. Therefore, it is necessary to sort it to the downgrading treatment area in order to use the plastic recycling material in the production of products with lower purity requirements and realize the rational utilization of resources.

[0178] If P S If the purity level is less than 80%, the plastic recycled material is judged to be of grade four purity, and an alarm message is issued, and the plastic recycled material is subjected to enhanced impurity removal treatment; when the purity level is grade four, it means that the plastic recycled material contains a large amount of impurities and cannot be directly used for the production of plastic products. Therefore, enhanced impurity removal treatment can be carried out to remove impurities through enhanced impurity removal technology (such as pyrolysis technology) to improve the purity of the plastic recycled material.

[0179] Furthermore, in this embodiment, the purity monitoring results of the recycled plastics include:

[0180] The real-time surface visual image data, real-time spectral signal data, real-time environmental parameter data and real-time purity value of the same batch of plastic recycled materials are integrated and processed, and the purity monitoring results of the plastic recycled materials including real-time purity value, purity grade and purity fluctuation trend are generated through visualization technology;

[0181] An example of step S4 is as follows:

[0182] The PE plastic recycling materials to be monitored are monitored in real time. An industrial CCD camera with a resolution of 6 million pixels is used to collect real-time surface visual image data of the plastic recycling materials at a distance of 25 cm and a light intensity of 1300 Lux.

[0183] A near-infrared spectrometer with a wavelength range of 780–2526 nm and 6 scans was used, as well as a wavenumber range of 200–3500 cm -1 , Raman spectrometer with a laser power of 90mW to obtain real-time spectral signal data;

[0184] Using a temperature and humidity sensor with a measurement accuracy of ±0.4°C and ±1.8% RH, the system records the real-time ambient temperature of 27°C and humidity of 40% RH. A set of data is collected every 2 minutes, each containing relevant data for several samples of recycled plastics that passed through the production line during that time period.

[0185] For the collected real-time surface visual image data, a 3×3 window median filter is first used to remove noise, and the contrast is enhanced by histogram equalization;

[0186] After converting the image into grayscale, the gray-level co-occurrence matrix is ​​calculated, and the texture pattern is extracted using the LBP algorithm with a neighborhood radius of 1 and a number of neighborhood points of 8;

[0187] Canny edge detection (high and low thresholds of 0.35 and 0.12, respectively) was used to obtain the contours and calculate the perimeter and area;

[0188] For real-time spectral signal data, the near-infrared spectrum is smoothed using Savitzky-Golay filtering with a window size of 5, and the dimension is reduced by principal component analysis to obtain the absorbance peak at a specific wavelength;

[0189] Extracted Raman spectrum 200-3500cm -1 The scattered light intensity within the wavenumber range and the PE characteristic peak (2840cm -1 )’s strength and location;

[0190] After calibrating the visual and spectral data in combination with real-time environmental parameters (including light intensity 1300 Lux), the processed data is input into the plastic recycling material purity monitoring model trained and meeting the requirements in step S3 to obtain the preliminary purity value P of the plastic recycling material. pre For example, after the data collected in a certain period is processed and input into the model, the preliminary purity value P is obtained. pre =86%;

[0191] A calibration sample library containing 500 historical samples was pre-established. These samples cover PE plastic recycled materials of different purity ranges and different production batches. The five historical samples with the most similar characteristics to the current plastic recycled materials were retrieved through the cosine similarity matching algorithm. For example, the purity values ​​of these five historical samples were 86%, 89%, 90%, 87%, and 88%, respectively. The calibration purity mean was calculated.

[0192] The adaptive weight α is dynamically adjusted according to the sample matching degree. For example, if the matching degree is high this time, α is set to 0.6. According to the formula P S =88%+0.6*(88%-86%)=89.2%;

[0193] Because 80% ≤ P S =89.2%<90%, the purity level of the plastic recycling material is determined to be level three, and an alarm message is issued to sort the plastic recycling material to the downgrading processing area;

[0194] Integrate and process the real-time surface visual image data, real-time spectral signal data, real-time environmental parameter data and real-time purity value of each plastic recycling material in the monitored batch;

[0195] There are currently 150 pieces of plastic recycled materials in the batch to be monitored. After monitoring all 150 pieces, the number of pieces at each purity level is counted, including 50 pieces of first-level purity, 30 pieces of second-level purity, 50 pieces of third-level purity, and 20 pieces of fourth-level purity. The monitoring results of these 150 pieces of plastic recycled materials are visualized through data visualization technology. A line graph is used to show the purity fluctuation trend of the 150 monitored materials, and a bar graph is used to show the quantity distribution of each purity level. At the same time, the real-time purity value and purity level of each piece of plastic recycled material are marked, as well as the overall monitoring progress and abnormal status of the batch.

[0196] In the description of this application, it should be understood that the directions or positional relationships indicated by directional words such as "front, back, up, down, left, right", "horizontal, vertical, horizontal" and "top, bottom" are usually based on the directions or positional relationships shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description. Unless otherwise specified, these directional words do not indicate or imply that the device or element referred to must have a specific direction or be constructed and operated in a specific direction. Therefore, they cannot be understood as limiting the scope of protection of this application.

[0197] Those skilled in the art can make various other corresponding changes and deformations based on the technical solutions and concepts described above, and all of these changes and deformations should fall within the scope of protection of the claims of this application.

Claims

1. A method for monitoring the purity of recycled plastics based on artificial intelligence, characterized in that: The following steps are involved: S1. Obtain multiple batches of historically monitored plastic recycled materials, collect surface visual image data and spectral signal data of the plastic recycled materials, as well as environmental parameter data and corresponding plastic recycled material purity data, and perform data processing to obtain a plastic feature dataset; S2. Construct a dual-branch deep neural network model, extract a surface visual image feature vector and a spectral signal feature vector from the surface visual image data and the spectral signal data, respectively, and fuse them to obtain a fused feature vector of the plastic recycled material; S3. Combine the fusion feature vector of the recycled plastics and use the plastic feature dataset to train and optimize the dual-branch deep neural network model to obtain a recycled plastics purity monitoring model; S4. Collect real-time surface visual image data and real-time spectral signal data, as well as real-time environmental parameter data of the plastic recycled material to be monitored, and input them into the plastic recycled material purity monitoring model to obtain the real-time purity value of the plastic recycled material and generate the plastic recycled material purity monitoring result.

2. The artificial intelligence-based plastic recycling material purity monitoring method according to claim 1, characterized in that: The step S1 specifically includes: The surface visual image data includes a surface image of the recycled plastic material and plastic color information, wherein the plastic color information is hue, saturation, and brightness information corresponding to an HSV color space obtained based on the RGB color value of each pixel in the surface image of the recycled plastic material; The spectral signal data includes near infrared spectral data of plastic recycled materials in the wavelength range of 780-2526nm and 200-3500cm -1 Raman spectral data in the wavenumber range; The environmental parameter data includes temperature data, humidity data and light intensity data of the environment in which the plastic recycled material is located; The collected surface visual image data, the spectral signal data and the environmental parameter data are transmitted to an edge computing gateway.

3. The artificial intelligence-based plastic recycling material purity monitoring method according to claim 2, characterized in that: The step S1 further includes: Obtaining the surface image of the plastic recycling material from the edge computing gateway, performing median filtering denoising and histogram equalization processing, converting the surface image of the plastic recycling material into a grayscale image, and calculating a gray-level co-occurrence matrix; Extracting a texture pattern of the surface image of the recycled plastic material using a local binary pattern to obtain texture features of the surface of the recycled plastic material; The contour of the plastic recycling material is obtained through edge detection algorithm, and the perimeter, area, circularity and rectangularity of the contour are calculated, and the shape features are extracted using Fourier descriptors; Detecting and analyzing the surface image of the recycled plastic material to obtain the length and width of scratches, the length and direction of cracks, the area and diameter of holes, and the number and volume of bubbles on the surface of the recycled plastic material; The near-infrared spectral data were filtered, smoothed, and subjected to principal component analysis and dimensionality reduction processing to obtain the absorbance within the wavelength range of 780-2526 nm and the absorbance peaks at wavelengths of 1100 nm, 1450 nm, and 1900 nm; Acquire Raman spectra 200-3500cm -1 The scattered light intensity in the wavenumber range and polyethylene 2840cm -1 , Polypropylene 800-850cm -1 , polyvinyl chloride 600-700cm -1 The intensity and position of characteristic peaks; Combined with the temperature data, the humidity data and the light intensity data, the processed surface visual image data and the spectral signal data are calibrated, and the calibrated surface visual image data, spectral signal data, and environmental parameter data are associated, labeled, and integrated with the plastic recycled material purity data to form a plastic feature data set.

4. The artificial intelligence-based plastic recycling material purity monitoring method according to claim 3, characterized in that: The step S2 specifically includes: Constructing a dual-branch deep neural network model including a surface visual image data processing branch and a spectral signal data processing branch, wherein the surface visual image data processing branch is composed of multiple convolutional layers, pooling layers, and fully connected layers, and the spectral signal data processing branch is composed of multiple one-dimensional convolutional layers, global average pooling layers, and fully connected layers; Inputting the surface visual image data in the plastic feature dataset into the surface visual image data processing branch, and sequentially passing through a convolutional layer, a pooling layer, and a fully connected layer to obtain a surface visual image feature vector; Inputting the spectral signal data in the plastic feature dataset into the spectral signal data processing branch, and sequentially passing through a one-dimensional convolutional layer, a global average pooling layer, and a fully connected layer to obtain a spectral signal feature vector; The surface visual image feature vector and the spectral signal feature vector are fused by splicing to form a plastic recycling material fusion feature vector.

5. The artificial intelligence-based plastic recycling material purity monitoring method according to claim 4, characterized in that: The step S3 specifically includes: Dividing the plastic feature dataset into a training set and a test set in a ratio of A1:A2, and initializing the parameters of each layer of the dual-branch deep neural network model; Setting a training round Q, combining the fusion feature vector of the recycled plastic material, and using the training set to train the dual-branch deep neural network model, performing Q iterative training; Inputting the test set into the trained two-branch deep neural network model for testing, combining the plastic recycled material fusion feature vector for forward propagation, and obtaining the plastic recycled material purity prediction value of all samples in the test set; The dual-branch deep neural network model is evaluated based on the predicted purity value of the plastic recycled material.

6. The artificial intelligence-based plastic recycling material purity monitoring method according to claim 5, characterized in that: The model evaluation includes: The plastic recycling material purity prediction value includes a first purity prediction value P based only on surface visual image features. v , the second purity prediction value P based only on spectral signal characteristics s , and the third purity prediction value P of the fusion feature f , calculate the mean inter-modal consistency score C J And the weighted evaluation index M C : Among them, C i represents the inter-modal consistency score of the i-th sample in the test set; P v (i) represents the first purity prediction value of the i-th sample in the test set; P s (i) represents the second purity prediction value of the i-th sample in the test set; P f (i) represents the third purity prediction value of the i-th sample in the test set; y (i) represents the actual purity value of the i-th sample in the test set; N represents the total number of samples in the test set; i=1, 2...N; According to the inter-modality consistency score C i And the sample prediction variance of the test set, dynamically scale the confidence interval, and calculate the confidence interval: Where Z represents the dynamic adjustment coefficient of the dynamic scaling confidence interval; Z0 represents the standard confidence level quantile; γ represents the adaptive coefficient; represents the prediction variance of the i-th sample in the test set obtained by Monte Carlo calculation; max(σ 2 ) represents the maximum value of the variance of all samples in the test set; CI i represents the confidence interval of the i-th sample in the test set; If M C ≤0.5, C J ≥0.7 and CI i If the proportion of samples that do not cover the true value is ≤ 20%, the two-branch deep neural network model is determined to meet the requirements and is used as a plastic recycled material purity monitoring model; If M C ≤0.5, C J ≥0.7 and CI i The proportion of samples that do not cover the true value is ≤ 20%. If any one of them is not met, it is determined that the dual-branch deep neural network model does not meet the requirements, and the model is optimized until M C ≤0.5, C J ≥0.7 and CI i The proportion of samples that do not cover the true value is ≤ 20%.

7. The artificial intelligence-based plastic recycling material purity monitoring method according to claim 6, characterized in that: The model optimization includes: According to the obtained weighted evaluation index M C , the mean inter-modality consistency score C J and confidence interval CI i , analyzing and optimizing the dual-branch deep neural network model; When M C > 0.5, it is determined that the feature extraction is insufficient, and the feature extraction optimization strategy is executed until M is satisfied. C ≤0.5; When C J <0.7, it is determined that the feature fusion mechanism has defects, and the feature fusion mechanism optimization strategy is executed until C is satisfied. J ≥0.7; When CI i If the proportion of samples that do not cover the true value is greater than 20%, it is determined that the confidence interval cannot effectively contain the true value, and the confidence interval optimization strategy is executed until the CI is satisfied. i The proportion of samples that do not cover the true value is ≤ 20%.

8. The artificial intelligence-based plastic recycling purity monitoring method according to claim 7, characterized in that: The step S4 specifically includes: Collect the real-time surface visual image data and real-time spectral signal data of the plastic recycling material to be monitored, as well as the real-time environmental parameter data, perform data processing, and input the data into the plastic recycling material purity monitoring model after the data processing is completed to obtain the preliminary purity value P of the plastic recycling material pre ; According to the pre-established calibration sample library, the similarity matching algorithm is used to retrieve multiple historical samples that are most similar to the characteristics of the current plastic recycled material, and the calibration purity mean P is calculated. J , and the preliminary purity value P pre Calibrate to get the real-time purity value P of plastic recycling material S : P S =P pre +α*(P J -P pre ) Among them, α represents the adaptive weight, which is dynamically adjusted according to the sample matching degree and ranges from 0 to 1; According to the real-time purity value P S Perform purity analysis on recycled plastics and generate purity monitoring results for recycled plastics.

9. The artificial intelligence-based plastic recycling material purity monitoring method according to claim 8, characterized in that: The purity analysis includes: If P S ≥95%, the purity grade of the plastic recycling material is determined to be level one, and the plastic recycling material will continue to go through the original processing flow; If 90%≤P S If the purity level of the recycled plastic material is less than 95%, the recycled plastic material is judged to be of Grade II purity, and an alarm message is issued, and the recycled plastic material is subjected to a second re-inspection process; If 80%≤P S If the purity level of the recycled plastic material is less than 90%, the recycled plastic material is judged to be of grade three purity, and an alarm message is issued to sort the recycled plastic material to the downgrading processing area; If P S If the purity level of the recycled plastic material is less than 80%, the recycled plastic material is judged to be of grade 4 purity, and an alarm message is issued, and the recycled plastic material is subjected to enhanced impurity removal treatment.

10. The artificial intelligence-based plastic recycling material purity monitoring method according to claim 9, characterized in that: The purity monitoring results of the plastic recycled materials include: The real-time surface visual image data, the real-time spectral signal data, the real-time environmental parameter data and the real-time purity value of the same batch of plastic recycled materials are integrated and processed, and the plastic recycled material purity monitoring results including the real-time purity value, purity grade and purity fluctuation trend are generated through visualization technology.

Images