A plastic species rapid identification and sorting system and method based on raman spectrum

Abstract

The present application belongs to the technical field of plastic detection and sorting, and particularly relates to a plastic type rapid discrimination and sorting system and method based on Raman spectrum. The present application comprises a Raman spectrum acquisition module, a spectrum pretreatment module, a feature extraction module, a model discrimination module, and a sorting execution module. The adaptive iteratively reweighted penalized least squares (airPLS) is used for baseline correction, which improves the signal-to-noise ratio and purity of the spectrum data. In the feature extraction stage, principal component analysis (PCA) dimension reduction is introduced as an alternative strategy, which reduces the operation load of the model and meets the sorting demand with extremely high real-time requirement. A multi-model library containing support vector machine (SVM), random forest (RF), XGBoost, one-dimensional convolutional neural network (1D-CNN) and multi-modal CNN is constructed, and the performance of the models is compared and optimized, which improves the universality and landing success rate of the sorting system and method in different industrial scenarios.

Description

Technical Field

[0001] This invention belongs to the field of plastic detection and sorting technology, specifically relating to a rapid plastic type identification and sorting system and method based on Raman spectroscopy. Background Technology

[0002] Plastic products are widely used in packaging, agriculture, medical, electronics, and many other fields, but the environmental problems caused by their large-scale waste are becoming increasingly serious. Different types of plastics have significant differences in chemical composition, physical properties, and degradation characteristics. For example, common biodegradable plastics mainly include polylactic acid (PLA), polybutylene adipate terephthalate (PBAT), polybutylene succinate (PBS), polyhydroxyalkanoates (PHA), and polycaprolactone (PCL), while traditional non-biodegradable plastics include polyethylene (PE), polypropylene (PP), polystyrene (PS), polyethylene terephthalate (PET), polyvinyl chloride (PVC), and acrylonitrile-butadiene-styrene copolymer (ABS). In actual waste logistics, various types of plastics are often mixed together, with biodegradable and non-biodegradable plastics coexisting being particularly prominent. Without effective classification and sorting, not only will recycling efficiency and the quality of recycled products be reduced, but non-biodegradable plastics may also be mixed into the processing of biodegradable plastics, causing secondary pollution; at the same time, the mixing of different types of non-biodegradable plastics will also affect high-value recycling. Therefore, establishing a rapid and accurate method for identifying and sorting plastic types is of great significance for improving the recycling rate of plastics, reducing environmental pollution, and promoting the development of a circular economy.

[0003] Raman spectroscopy is widely recognized as an effective tool for plastic classification due to its unique spectral fingerprint. However, its application in plastic classification has also shown some limitations: First, in actual industrial scenarios, due to factors such as complex material additives, sample surface contamination, and aging, Raman spectra often exhibit baseline drift, high-frequency noise interference, and inconsistent spectral intensity scales, resulting in poor raw data quality. Second, the information content of a single spectral feature is insufficient, making it difficult to comprehensively characterize material properties. Third, traditional manual analysis or simple peak matching methods for spectra are highly subjective, inefficient, and have limited accuracy, making it difficult to meet industrial needs. To overcome these limitations, Chinese patent CN117893812A proposes a method and recycling device for rapidly identifying biodegradable plastics. It employs a spectral preprocessing workflow (smoothing, baseline removal, and normalization), which to some extent solves the problems of baseline drift and noise interference, improving the quality of the original data. By extracting peak parameters and combining them with GADF image encoding, it achieves the fusion of numerical features and two-dimensional image features, enabling multi-dimensional and in-depth feature capture of the vibrational and rotational information of biodegradable plastic molecules, solving the problem of insufficient information from a single feature. It utilizes a multimodal deep learning model to replace manual spectral analysis, leveraging its powerful nonlinear fitting capabilities to achieve high-precision, objective, and rapid automatic classification. However, while the above-mentioned existing technical solutions address the aforementioned fundamental limitations, they still have the following three shortcomings when facing more complex industrial demands:

[0004] (1) Insufficient baseline correction accuracy: Existing technologies use the traditional "sliding window averaging method" for spectral preprocessing. When dealing with complex fluorescence backgrounds, the correction is not accurate enough and background interference is easily left behind, affecting the purity of feature extraction.

[0005] (2) Computational redundancy and poor real-time performance: Existing technologies rely on full peak parameters and GADF images in feature strategies, which may lead to data dimensional redundancy and excessive model computation, resulting in limited inference speed and failure to meet the requirements of scenarios with extremely high real-time requirements.

[0006] (3) Poor model adaptability: Existing technologies rely on a single deep learning model architecture for spectral analysis, which lacks flexibility and cannot meet the differentiated needs of different industrial scenarios for computing power, speed and accuracy (e.g., low-end equipment requires lightweight models, while high-end equipment requires high-precision models). Summary of the Invention

[0007] To address the shortcomings of the existing technologies, the technical problem to be solved by this invention is to provide a rapid plastic type identification and sorting system and method based on Raman spectroscopy. This system improves the purity of spectral data by enhancing the accuracy of baseline correction in spectral preprocessing; it introduces feature extraction and dimensionality reduction techniques to balance information content and computational efficiency; and it establishes a library of multiple algorithm models, configuring the optimal model for different scenarios. This achieves a comprehensive improvement in data quality, computational efficiency, and adaptability to engineering applications.

[0008] The technical solution adopted in this invention is as follows:

[0009] This invention provides a rapid plastic type identification and sorting system based on Raman spectroscopy, including a Raman spectroscopy acquisition module, a spectral preprocessing module, a feature extraction module, a model discrimination module, and a sorting execution module;

[0010] The Raman spectroscopy acquisition module is used to acquire the Raman spectral signal of the plastic sample;

[0011] The spectral preprocessing module has a built-in preprocessing algorithm for performing cosmic ray removal, baseline correction, spectral smoothing, and normalization on the acquired Raman spectra.

[0012] The feature extraction module has built-in peak feature extraction, spectral image encoding and PCA dimensionality reduction algorithms, which are used to extract peak feature parameters from the preprocessed spectrum and / or generate spectral image encoding and / or construct feature vectors through PCA dimensionality reduction.

[0013] The model discrimination module has a built-in trained optimal machine learning classification model, which is used to identify the type of plastic from the input feature vector of the test sample.

[0014] The sorting execution module is used to drive the actuator to complete the automatic sorting of plastics based on the model's judgment results.

[0015] This invention also provides a rapid plastic type identification and sorting method based on Raman spectroscopy, implemented based on the above-mentioned identification and sorting system, including the following steps:

[0016] Step S1: Establish a standard spectral database;

[0017] Raman spectra of known types of plastic standard samples were collected, and the plastic type label corresponding to each spectrum was recorded to establish a standard Raman spectrum database.

[0018] Preferably, in step S1, the types of plastics include six conventional plastics: PP, PE, PS, PET, PVC, and ABS, and six biodegradable plastics: PLA, PCL, PHA, PBS, PBAT, and PGA.

[0019] Preferably, in step S1, when acquiring Raman spectra, the excitation wavelength is selected from 532nm, 633nm, 785nm, and 1064nm, the wavenumber range for spectral acquisition is 500–3000 cm⁻¹, and at least three different positions are selected for measurement for each sample to enhance data representativeness.

[0020] Preferably, in step S1, when acquiring the Raman spectrum, the excitation wavelength is selected as 785nm or 1064nm to effectively suppress fluorescence background interference.

[0021] Step S2: Preprocess the Raman spectra in the database established in step S1;

[0022] Preferably, in step S2, each Raman spectrum is processed sequentially as follows:

[0023] S201, spectral smoothing;

[0024] Savitzky-Golay convolutional smoothing filter is used, with the window size selected from 5 to 15 data points according to the spectral resolution, and the polynomial order selected as 2 or 3, which effectively reduces high-frequency random noise while preserving characteristic peak information.

[0025] S202, Baseline Correction;

[0026] The adaptive iterative reweighted penalized least squares method airPLS is adopted to automatically fit the spectral baseline by iteratively adjusting the weights, and to subtract the fluorescence background and baseline drift from the original spectrum;

[0027] S203, normalization processing;

[0028] By employing area normalization or vector normalization methods, the spectral intensity is unified to the [0,1] range, eliminating the intensity scale differences caused by laser power fluctuations, acquisition time differences, and different sample surface conditions.

[0029] Step S3, spectral feature extraction;

[0030] Preferably, in step S3, extracting features from the Raman spectrum after preprocessing in step S2 includes a combination of one or more of the following methods:

[0031] (1) Peak feature parameter extraction;

[0032] Peak detection is performed on the preprocessed normalized spectrum, a threshold is set, and effective feature peaks are screened. Multi-peak Gaussian fitting is performed on the screened feature peaks to extract the position, peak intensity, peak area, and full width at half maximum (FWHM) parameters of each peak, and a multi-dimensional numerical feature vector is constructed.

[0033] (2) Spectral image encoding;

[0034] A one-dimensional Raman spectral sequence is mapped to a polar coordinate system through Gram angular difference field (GADF) or Gram angular field (GAF) transformation to generate a two-dimensional spectral image. This encoding method can transform the one-dimensional spectral problem into a two-dimensional image recognition problem while preserving the temporal and spatial correlation of the spectral sequence, making it easier to process using mature convolutional neural network architectures.

[0035] (3) Principal component analysis (PCA) for dimensionality reduction;

[0036] Principal component analysis is performed on the preprocessed full-spectrum data to extract the principal component components whose cumulative variance contribution rate exceeds a predetermined threshold. These components are then used as dimensionality-reduced feature vectors, effectively reducing data dimensionality and model computation.

[0037] Preferably, in step S3 (3), the predetermined threshold for the cumulative variance contribution rate is set to 95%.

[0038] Step S4: Training and optimization of the machine learning classification model;

[0039] Preferably, step S4 includes:

[0040] S401, After preprocessing all spectra in the standard spectral database in step S2 and extracting features in step S3, randomly divide them into training set and test set according to a preset ratio;

[0041] S402, using the feature vectors of the training set as input and the category labels as output, train a machine learning classification model, wherein the machine learning classification model is one or more of the following models in combination:

[0042] (1) Support Vector Machine Model: Experiments were conducted using linear kernel, polynomial kernel, Gaussian radial basis function (RBF) kernel and sigmoid kernel function respectively. The penalty parameter C and kernel parameter γ were optimized by grid search and cross-validation.

[0043] (2) Random Forest Model: Optimized by adjusting the number of decision trees and the maximum depth parameter;

[0044] (3) XGBoost model: Under the gradient boosting framework, the learning rate and the maximum depth of the tree are optimized by tuning the parameters;

[0045] (4) One-dimensional convolutional neural network model: Construct a network structure containing one-dimensional convolutional layers, pooling layers, fully connected layers and Dropout layers;

[0046] (5) Multimodal deep learning model: Construct a dual-branch convolutional neural network, which takes numerical feature vectors and GADF / GAF spectral images as inputs respectively, and outputs classification results after joint learning through feature fusion layer;

[0047] S403. Evaluate the performance of the model described in S402 using the test set. Evaluation metrics include classification accuracy, precision, recall, F1 score, and confusion matrix. Select the optimal classification model through comprehensive comparison and save it.

[0048] Step S5: Online identification and sorting of the samples to be tested.

[0049] Preferably, step S5 includes:

[0050] S501, collects Raman spectra of plastic samples to be tested on the sorting production line;

[0051] S502, after the spectrum to be measured is preprocessed in step S2 and feature extracted in step S3, it is input into the classification model that is preferred and saved in step S4, and the plastic type to which the sample belongs is output.

[0052] S503: Based on the plastic type identification result output by S502, a control signal is generated to drive the execution mechanism of the sorting execution module to complete automatic sorting; the execution mechanism sorts different types of plastics into the corresponding collection channels by air blowing or robotic arm.

[0053] The beneficial effects obtained by adopting the above technical solution are as follows:

[0054] (1) The present invention uses the adaptive iterative reweighted penalized least squares (airPLS) method for baseline correction. This method automatically fits and subtracts the fluorescence background curve at the bottom of the spectrum by iteratively adjusting the weights; it solves the problem of inaccurate correction of complex fluorescence background, can more accurately remove background interference, improve the signal-to-noise ratio and purity of spectral data, and provide higher quality input for subsequent feature extraction.

[0055] (2) In the feature extraction stage, this invention introduces principal component analysis (PCA) dimensionality reduction as an alternative strategy. By using linear transformation, the original high-dimensional peak parameters and spectral data are mapped to a low-dimensional space to extract the main components. This solves the problems of data dimensionality redundancy and excessive computation. While retaining the core feature information, the data dimensionality is greatly reduced, and the computational load of the model is reduced, thereby meeting the sorting scenarios in industrial sites with extremely high real-time requirements.

[0056] (3) This invention constructs a multi-model library including support vector machine (SVM), random forest (RF), XGBoost, one-dimensional convolutional neural network (1D-CNN) and multimodal CNN, and systematically trains and compares these models; it solves the problem that a single model cannot adapt to different needs. Users can flexibly choose according to the computing power limit of the actual production line (such as edge computing or cloud computing) and accuracy requirements, avoiding the blindness of model selection, and improving the universality and success rate of the sorting system and method in different industrial scenarios. Attached Figure Description

[0057] Figure 1 This is a flowchart of the rapid identification and sorting method for plastic types based on Raman spectroscopy according to the present invention.

[0058] Figure 2 These are Raman fingerprint images of twelve types of plastics in embodiments of the present invention.

[0059] Figure 3 This is a schematic diagram of the feature extraction module of the rapid plastic type identification and sorting system based on Raman spectroscopy of the present invention. Detailed Implementation

[0060] The technical solution of the present invention will now be described more clearly and completely with reference to the accompanying drawings.

[0061] This invention provides a rapid plastic type identification and sorting system based on Raman spectroscopy, including a Raman spectroscopy acquisition module, a spectral preprocessing module, a feature extraction module, a model discrimination module, and a sorting execution module;

[0062] The Raman spectroscopy acquisition module is used to acquire the Raman spectral signal of the plastic sample;

[0063] The spectral preprocessing module has a built-in preprocessing algorithm for performing cosmic ray removal, baseline correction, spectral smoothing, and normalization on the acquired Raman spectra.

[0064] The feature extraction module has built-in peak feature extraction, spectral image encoding and PCA dimensionality reduction algorithms, which are used to extract peak feature parameters from the preprocessed spectrum and / or generate spectral image encoding and / or construct feature vectors through PCA dimensionality reduction.

[0065] The model discrimination module has a built-in trained optimal machine learning classification model, which is used to identify the type of plastic from the input feature vector of the test sample.

[0066] The sorting execution module is used to drive the sorting mechanism to complete the automatic sorting of plastics based on the model discrimination results.

[0067] This invention also provides a rapid identification and sorting method for plastics based on Raman spectroscopy, implemented based on the aforementioned identification and sorting system. This embodiment uses six typical conventional plastics (PP, PE, PS, PET, PVC, ABS) and six typical biodegradable plastics (PLA, PCL, PHA, PBS, PBAT, PGA) as examples to describe the specific implementation method of this invention in detail. Figure 1 As shown, it includes the following steps:

[0068] Step 1: Standard Sample Preparation and Spectral Acquisition

[0069] Standard samples of twelve plastics, including PP, PE, PS, PET, PVC, ABS, PLA, PCL, PHA, PBS, PBAT, and PGA, were prepared or obtained. At least ten independent samples were prepared for each material, and the sample surfaces were cleaned with alcohol. A portable Raman spectrometer equipped with a 785 nm excitation wavelength was used. Within the wavenumber range of 50–3000 cm⁻¹, the integration time was set to 1–3 seconds, and the laser power was controlled within 0–30 mW. Five different locations were selected for each sample for acquisition, resulting in at least 50 Raman spectra for each material. A standard Raman spectral database was established (Table 1).

[0070] The above twelve types of plastic Raman fingerprint spectra are as follows: Figure 2 As shown.

[0071] The Raman spectra of PE plastics are mainly at 1064, 1130, 1396, and 1442 cm⁻¹. -1 There are several main vibration peaks at this location. Among them, the one at 1064 cm⁻¹ is the most significant. -1 The vibrational peak is caused by the asymmetric stretching vibration between carbon atoms, at 1130 cm⁻¹. -1 The vibrational peak is caused by the symmetrical stretching vibration between carbon atoms; 1396 cm⁻¹ -1 The vibrational peak is caused by the oscillating vibration between carbon and hydrogen atoms, at 1463 cm⁻¹. -1 The vibration peaks at that point exhibit in-plane oscillations of carbon and hydrogen atoms.

[0072] The Raman spectra of PP plastic are mainly at 809, 842, 1154, 1331, and 1460 cm⁻¹. -1 There are several main vibration peaks at 809 and 842 cm⁻¹. -1 The vibrational peaks at 1154 cm⁻¹ are all caused by the rocking vibrations between carbon and hydrogen atoms. -1 Caused by symmetrical stretching vibrations between carbon atoms, 1322 cm -1 Caused by the stretching vibrations between carbon atoms, 1460 cm -1 It is caused by the bending vibrations between hydrocarbon atoms.

[0073] The Raman spectra of PHA plastics are mainly at 812 and 1453 cm⁻¹. -1 There are several main vibration peaks at 812 cm. -1 The characteristic peak represents the symmetric stretching vibration between carbon and oxygen atoms, while the 1453 cm⁻¹ peak... -1 The characteristic peak represents the bending vibration of the carbon-hydrogen bond in the methyl group.

[0074] The Raman spectra of PVC plastic are mainly at 1324 and 1430 cm⁻¹. -1 There are several main vibration peaks at 1324 cm. -1 The characteristic peak represents the stretching vibration between hydrocarbon atoms, at 1430 cm⁻¹. -1 Characteristic peaks represent bending vibrations between hydrocarbon atoms.

[0075] The Raman spectra of PET plastic are mainly at 1615 and 1731 cm⁻¹. -1 There are several main vibration peaks at 1615 cm. -1 The characteristic peak is caused by the stretching vibration of the carbon-carbon bonds in the benzene ring, while the 1731 cm⁻¹ peak is caused by the stretching vibration of the carbon-carbon bonds in the benzene ring. -1 The characteristic peak represents the stretching oscillation of the carbon-oxygen bond.

[0076] The Raman spectra of PHAB plastics are mainly at 1712, 1615, 1513, 1365, 1252, 1167, 1022, and 805 cm⁻¹. -1 The main vibration peak is located at 1712 cm. -1 The vibration peak is caused by the C=O stretching vibration, 1615 cm⁻¹ -1 The vibration peak is caused by the stretching vibration of the benzene ring skeleton, 1513 cm⁻¹ -1 The vibration peak is caused by the stretching vibration of the benzene ring skeleton, at 1365 cm⁻¹. -1 The vibration peak is caused by in-plane bending vibration of the OH group, at 1252 cm⁻¹. -1 The vibration peak is caused by the stretching vibration of CO, at 1167 cm⁻¹. -1 The vibration peak is caused by the stretching vibration of COC, at 1022 cm⁻¹. -1 The vibration peak is caused by in-plane deformation vibration of the benzene ring, 805 cm⁻¹ -1 The vibration peak is caused by the out-of-plane deformation vibration of the benzene ring.

[0077] The Raman spectrum of PS plastic is mainly at 1002 cm⁻¹. -1 The main vibration peak is located at 1002 cm. -1 The characteristic peaks represent the breathing vibrations of the benzene ring.

[0078] The Raman spectrum of PGA plastic is mainly at 919 cm⁻¹. -1 The main vibration peak is located at 919 cm.-1 The vibration peak at that point is caused by the symmetrical stretching vibration between carbon and oxygen atoms.

[0079] The Raman spectrum of ABS plastic is mainly at 1002 cm⁻¹. -1 The main vibration peak is located at 1002 cm. -1 The peak represents the breathing vibration of the benzene ring.

[0080] The Raman spectra of PVA plastics are mainly at 857, 1082, 1362, 1414, and 1447 cm⁻¹. -1 There are several main vibration peaks at this location. Among them, the one at 857 cm⁻¹ is... -1 The vibrational peak is caused by the coiling vibrations between hydrocarbon atoms, at 1082 cm⁻¹. -1 The characteristic peak represents the deformation vibration of the oxygen-hydrogen bond, at 1352 cm⁻¹. -1 The characteristic peak is generated by the vibration of the carbon-oxygen bond, at 1445 cm⁻¹. -1 The characteristic peaks are generated by the shearing mode of carbon-hydrogen bonds.

[0081] The Raman spectra of PCL plastics are mainly at 1725, 1143, 1075, and 1446 cm⁻¹. -1 Several major vibrational peaks are observed, including one at 1725 cm⁻¹. -1 The vibration peak is caused by the C=O stretching vibration, 1143 cm. -1 The vibration peak is caused by the stretching vibration of COC, at 1075 cm⁻¹. -1 The vibration peak is caused by the stretching vibration of the CC skeleton, at 1446 cm⁻¹. -1 The vibration peak is caused by the bending vibration of CH2.

[0082] Table 1. Molecular fingerprint database of twelve typical plastic components

[0083]

[0084] Step 2: Spectral Preprocessing

[0085] For each spectrum in the database, the following processing is performed sequentially: (1) Smoothing and noise reduction is performed using a Savitzky-Golay filter (11-point window, 3rd order polynomial); (2) Baseline correction is performed using the airPLS algorithm to eliminate fluorescence background; (3) Spectral intensity is unified to the range of [0,1] using the area normalization method.

[0086] Step 3: Feature Extraction

[0087] like Figure 3As shown, three types of features are extracted simultaneously on the preprocessed spectrum: (1) Peak detection is performed with a normalized intensity of 0.2 as the threshold. After screening the characteristic peaks, multi-peak Gaussian fitting is performed to extract the peak position, intensity, area and FWHM to form a feature vector; (2) The spectrum is transformed by GADF to generate a two-dimensional image of size 128×128; (3) The entire spectrum is reduced in dimension by PCA, and the first 20 principal components are retained.

[0088] Step 4: Model Training and Optimization

[0089] The dataset was randomly divided into training and test sets in a 7:3 ratio. Four algorithms were used for training: SVM (comparing linear and RBF kernels, optimizing C and γ parameters through grid search with 5-fold cross-validation), Random Forest (RF) (300 decision trees), XGBoost (learning rate 0.1, maximum depth 6), and Neural Network (NN) (containing 3 convolutional layers and 2 fully connected layers). Accuracy, precision, recall, and F1 score on the test set were used as evaluation metrics. The test results for each model are summarized in Table 2.

[0090] Table 2 Performance Comparison of Different Classification Models

[0091]

[0092] Through comprehensive comparison, the 1D-CNN model performed best across all metrics, and was therefore selected as the final discrimination model. This result is consistent with previous research reports that combining Raman spectroscopy with deep learning can achieve a qualitative classification accuracy of over 99% for plastic products.

[0093] Step 5: Online Identification and Sorting

[0094] The trained 1D-CNN model is deployed into the sorting system. On the sorting line, the plastic material to be tested is conveyed to the detection station via a conveyor belt. The Raman spectroscopy acquisition module acquires the spectral signal, which, after the aforementioned preprocessing and feature extraction processes, is input into the model in real time. The model outputs a label indicating the type of material. The control module drives the air-blowing sorting device based on the label information, blowing the plastic type into the corresponding collection box, thus achieving fully automated sorting.

Claims

1. A rapid plastic type identification and sorting system based on Raman spectroscopy, characterized in that, It includes a Raman spectroscopy acquisition module, a spectral preprocessing module, a feature extraction module, a model discrimination module, and a sorting execution module; The Raman spectroscopy acquisition module is used to acquire the Raman spectral signal of the plastic sample; The spectral preprocessing module has a built-in preprocessing algorithm for performing cosmic ray removal, baseline correction, spectral smoothing, and normalization on the acquired Raman spectra. The feature extraction module has built-in peak feature extraction, spectral image encoding and PCA dimensionality reduction algorithms, which are used to extract peak feature parameters from the preprocessed spectrum and / or generate spectral image encoding and / or construct feature vectors through PCA dimensionality reduction. The model discrimination module has a built-in trained optimal machine learning classification model, which is used to identify the type of plastic from the input feature vector of the test sample. The sorting execution module is used to drive the actuator to complete the automatic sorting of plastics based on the model's judgment results.

2. A rapid method for identifying and sorting plastic types based on Raman spectroscopy, implemented based on the identification and sorting system described in claim 1, characterized in that, Includes the following steps: Step S1: Establish a standard spectral database; Raman spectra of known types of plastic standard samples were collected, and the plastic type label corresponding to each spectrum was recorded to establish a standard Raman spectrum database. Step S2: Preprocess the Raman spectra in the database established in step S1; Step S3, spectral feature extraction; Step S4: Training and optimization of the machine learning classification model; Step S5: Online identification and sorting of the samples to be tested.

3. The discrimination and sorting method according to claim 2, characterized in that, In step S1, the types of plastics include six conventional plastics: PP, PE, PS, PET, PVC, and ABS, and six biodegradable plastics: PLA, PCL, PHA, PBS, PBAT, and PGA.

4. The discrimination and sorting method according to claim 3, characterized in that, In step S1, when acquiring Raman spectra, the excitation wavelength is selected from 532nm, 633nm, 785nm, and 1064nm, and the wavenumber range of the spectrum acquisition is 500 to 3000cm⁻¹. At least three different positions are selected for measurement for each sample.

5. The discrimination and sorting method according to claim 4, characterized in that, In step S1, when acquiring the Raman spectrum, the excitation wavelength is selected as 785nm or 1064nm.

6. The discrimination and sorting method according to claim 5, characterized in that, In step S2, each Raman spectrum is processed sequentially as follows: S201, spectral smoothing; Savitzky-Golay convolutional smoothing filter is used, with a window size of 5 to 15 data points and a polynomial order of 2 or 3. S202, Baseline Correction; The adaptive iterative reweighted penalized least squares method airPLS is adopted to automatically fit the spectral baseline by iteratively adjusting the weights, and to subtract the fluorescence background and baseline drift from the original spectrum; S203, normalization processing; The spectral intensity is unified to the [0,1] interval by using area normalization or vector normalization methods.

7. The discrimination and sorting method according to claim 6, characterized in that, In step S3, features are extracted from the Raman spectrum after preprocessing in step S2, including a combination of one or more of the following methods: (1) Peak feature parameter extraction; Peak detection is performed on the preprocessed normalized spectrum, a threshold is set, and effective feature peaks are screened. Multi-peak Gaussian fitting is performed on the screened feature peaks to extract the position, peak intensity, peak area, and full width at half maximum (FWHM) parameters of each peak, and a multi-dimensional numerical feature vector is constructed. (2) Spectral image encoding; A one-dimensional Raman spectral sequence is mapped to a polar coordinate system through Gram angle difference field GADF or Gram angle field GAF ​​transformation to generate a two-dimensional spectral image. (3) Principal component analysis (PCA) for dimensionality reduction; Principal component analysis was performed on the preprocessed full-spectrum data to extract the principal component components whose cumulative variance contribution rate exceeded a predetermined threshold, which were then used as the dimensionality-reduced feature vectors.

8. The discrimination and sorting method according to claim 7, characterized in that, In step S3 (3), the predetermined threshold for the cumulative variance contribution rate is set to 95%.

9. The discrimination and sorting method according to claim 8, characterized in that, Step S4 includes: S401, After preprocessing all spectra in the standard spectral database in step S2 and extracting features in step S3, randomly divide them into training set and test set according to a preset ratio; S402, using the feature vectors of the training set as input and the category labels as output, train a machine learning classification model, wherein the machine learning classification model is one or more of the following models in combination: (1) Support Vector Machine Model: Experiments were conducted using linear kernel, polynomial kernel, Gaussian radial basis function (RBF) kernel and sigmoid kernel function respectively. The penalty parameter C and kernel parameter γ were optimized by grid search and cross-validation. (2) Random Forest Model: Optimized by adjusting the number of decision trees and the maximum depth parameter; (3) XGBoost model: Under the gradient boosting framework, the learning rate and the maximum depth of the tree are optimized by tuning the parameters; (4) One-dimensional convolutional neural network model: Construct a network structure containing one-dimensional convolutional layers, pooling layers, fully connected layers and Dropout layers; (5) Multimodal deep learning model: Construct a dual-branch convolutional neural network, which takes numerical feature vectors and GADF / GAF spectral images as inputs respectively, and outputs classification results after joint learning through feature fusion layer; S403. Evaluate the performance of the model described in S402 using the test set. Evaluation metrics include classification accuracy, precision, recall, F1 score, and confusion matrix. Select the optimal classification model and save it.

10. The discrimination and sorting method according to claim 9, characterized in that, Step S5 includes: S501, collects Raman spectra of plastic samples to be tested on the sorting production line; S502, after the spectrum to be measured is preprocessed in step S2 and feature extracted in step S3, it is input into the classification model that is preferred and saved in step S4, and the plastic type to which the sample belongs is output. S503: Based on the plastic type identification result output by S502, a control signal is generated to drive the execution mechanism of the sorting execution module to complete automatic sorting; the execution mechanism sorts different types of plastics into the corresponding collection channels by air blowing or robotic arm.

Images