Micro-plastic type and aging state synchronous identification method and system based on dual-wavelength laser Raman spectrum

Abstract

The invention discloses a micro-plastic type and aging state synchronous identification method and system based on a dual-wavelength laser Raman spectrum, and belongs to the technical field of environment detection and analysis. The method comprises the following steps: collecting Raman spectrum data of the same micro-plastic sample under 532 nm laser and 785 nm laser; preprocessing and stacking along channel dimensions to construct a two-channel one-dimensional fusion tensor; and inputting the tensor into a pre-trained multi-task deep learning model. The model adopts an architecture of a shared feature extraction backbone network and double parallel classification heads, and can synchronously output types and aging state recognition results of the micro-plastics. According to the method, complementary information of the dual-wavelength spectrum is fully utilized, high-precision, automatic and synchronous identification of the types and aging states of the micro-plastics is realized through feature fusion and multi-task learning, the detection efficiency and accuracy are remarkably improved, and a reliable technical means is provided for precise evaluation of environmental risks of the micro-plastics.

Description

Technical Field

[0001] This application relates to the field of environmental monitoring and analysis technology, and in particular to a method and system for simultaneous identification of microplastic types and aging states based on dual-wavelength laser Raman spectroscopy. Background Technology

[0002] Microplastics, as a typical emerging pollutant, are widely distributed in marine, freshwater, soil, and even atmospheric environments, and their potential ecological risks and health threats are increasingly attracting global attention. It is worth noting that the environmental risk of microplastics is not solely determined by their abundance, but is closely related to their intrinsic characteristics such as type, particle size, surface properties, and aging state. For example, different polymer types exhibit significantly different toxic effects, and the aging process can alter the surface charge, adsorption capacity, and bioavailability of microplastics, thereby affecting their environmental migration, biological ingestion, and co-polluting effects. Therefore, accurately identifying the type and aging degree of microplastics is a crucial prerequisite for scientifically assessing their environmental fate and risk level.

[0003] Raman spectroscopy, as a molecular vibrational spectroscopy technique, has become one of the mainstream methods for microplastic identification due to its unique advantages of non-destructive testing, rapid response, and ability to provide "fingerprint information" of substances. However, facing the complex microplastic contamination scenarios in real-world environmental media, traditional Raman technology is facing two fundamental challenges:

[0004] First, there is the challenge of accurate identification across multiple categories. Environmental samples contain a wide variety of microplastics, and the Raman characteristic peaks of common polymers such as polyethylene (PE), polypropylene (PP), and polystyrene (PS) are adjacent or even overlapping, making spectral analysis difficult. Traditional machine learning algorithms rely on limited spectral features selected manually, making it difficult to fully extract deep identification information. Especially when the number of categories to be identified increases, the model complexity rises sharply, and the classification accuracy often drops precipitously, failing to meet the needs of high-precision screening of environmental samples. Second, there are limitations in judging aging states. Microplastics undergo aging processes such as photo-oxidation, chemical oxidation, and mechanical wear under the influence of multiple environmental factors, leading to significant changes in surface functional groups, crystallinity, and molecular chain structure. However, traditional Raman detection often uses a single wavelength excitation source (such as 785 nm), and the spectral differences of microplastics before and after aging are only manifested as weak peak shifts or intensity changes, resulting in a low signal-to-noise ratio and easy confusion with background interference. More importantly, traditional machine learning algorithms lack the ability to sensitively capture and deeply analyze these aging-related subtle features. Not only can they fail to effectively distinguish the aging stage, but they may also cause misclassification due to spectral distortion caused by aging, thus resulting in systematic bias in risk assessment.

[0005] The aforementioned technical bottlenecks directly restrict the accurate assessment of the environmental risks of microplastics: misidentification of species can lead to inaccurate assessment of toxicity effects, while the lack of identification of aging states makes it impossible to effectively quantify the environmental persistence, migration capacity, and compound pollution effects of microplastics.

[0006] Therefore, how to achieve high-precision, automated, and simultaneous identification of microplastic types and aging states, and significantly improve detection efficiency and accuracy, is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0007] This application provides a method and system for simultaneous identification of microplastic types and aging states based on dual-wavelength laser Raman spectroscopy. It relates to a precise identification technology for microplastic contaminants, which can solve and overcome the problems of isolated identification of microplastic types and assessment of aging states, insufficient accuracy, and weak anti-interference ability in the prior art.

[0008] The first aspect of this application provides a method for simultaneous identification of microplastic species and aging state based on dual-wavelength laser Raman spectroscopy, including: The same microplastic sample was excited by lasers with wavelengths of 532 nm and 785 nm respectively, and the raw Raman spectra of the microplastic sample were collected. The raw Raman spectral data of dual wavelengths are preprocessed, and the two preprocessed spectral data are stacked along the channel dimension to construct a dual-channel one-dimensional fusion tensor. The 532 nm wavelength is used to obtain spectral information sensitive to the chemical aging of potassium persulfate, and the 785 nm wavelength is used to obtain spectral information of the unaged state. A dual-channel one-dimensional fusion tensor is input into a pre-trained multi-task deep learning model. This model includes a shared feature extraction backbone network, a first fully connected classification head, and a second fully connected classification head. The shared feature extraction backbone network extracts features from the dual-channel one-dimensional fusion tensor and outputs a shared feature vector. The first fully connected classification head outputs the microplastic sample type identification result based on the shared feature vector. The second fully connected classification head outputs the microplastic sample aging state identification result based on the shared feature vector. The multi-task deep learning model is jointly trained end-to-end using a weighted multi-task loss function, allowing the shared feature extraction backbone network to be simultaneously modulated by the gradients of both the type identification task and the aging state identification task during backpropagation, thus learning a decoupled shared feature representation.

[0009] Optionally, the raw dual-wavelength Raman spectral data are preprocessed, including: Denoising and baseline correction were performed on the raw dual-wavelength Raman spectral data. The baseline-corrected spectral data are normalized to a uniform intensity range.

[0010] Optionally, the shared feature extraction backbone network is a neural network containing multiple layers of one-dimensional convolutions, specifically including: Five sequentially connected one-dimensional convolutional modules, each performing one-dimensional convolution, batch normalization, ReLU activation, and pooling operations in sequence; the first four one-dimensional convolutional modules use max pooling, and the fifth one-dimensional convolutional module uses adaptive average pooling. Three fully connected layers are connected in sequence. They receive the features output by adaptive average pooling and perform nonlinear transformations to generate a shared feature vector.

[0011] Optionally, the method also includes: Construct a training dataset in which each sample includes a two-channel one-dimensional fusion tensor, a microplastic type label, and a microplastic aging state label. The training dataset is divided into a training set and a test set; Construct a multi-task deep learning model and initialize the model parameters; Define a multi-task loss function, which is a weighted sum of the first classification loss function and the second classification loss function. The first classification loss function corresponds to the category recognition task, and the second classification loss function corresponds to the aging state recognition task. The multi-task deep learning model is jointly trained end-to-end using the training set, and the multi-task loss function is minimized through the backpropagation algorithm until the multi-task deep learning model converges. The accuracy of the trained multi-task deep learning model in species recognition and aging state recognition was evaluated using a test set.

[0012] Optionally, constructing the training dataset includes: Microplastics include polyethylene, polypropylene, polystyrene, polyvinyl chloride, polyethylene terephthalate, polycarbonate, and polylactic acid; The aging states of microplastics include unaged state, UV-aged state, and potassium persulfate-aged state. For each combination of type and aging state, at least one hundred valid Raman spectral data are collected to generate the corresponding dual-channel one-dimensional fusion tensor.

[0013] Optionally, the first fully connected classification head is used to map the shared feature vector to seven categories, corresponding to polyethylene, polypropylene, polystyrene, polyvinyl chloride, polyethylene terephthalate, polycarbonate, and polylactic acid, respectively; the second fully connected classification head is used to map the shared feature vector to three categories, corresponding to the unaged state, UV-aged state, and potassium persulfate-aged state, respectively.

[0014] Optionally, the method further includes: Obtain test samples containing microplastics of unknown types and aging states; A method for simultaneous identification of microplastic types and aging states based on dual-wavelength laser Raman spectroscopy was implemented to process and identify microplastic test samples, thereby obtaining the identification results of microplastic types and aging states of the microplastic test samples. Based on the identification results of multiple microplastic test samples and real labels, the accuracy of type identification, aging state identification, and dual-label synchronous identification of the multi-task deep learning model were calculated.

[0015] The second aspect of this application provides a system for simultaneously identifying the type and aging state of microplastics based on dual-wavelength laser Raman spectroscopy, including: The data acquisition module is used to excite the same microplastic sample with lasers of 532 nm and 785 nm wavelengths respectively, and acquire the raw Raman spectral data of the microplastic sample with dual wavelengths. The data processing and fusion module is used to preprocess the raw Raman spectral data of dual wavelengths and stack the two preprocessed spectral data along the channel dimension to construct a dual-channel one-dimensional fusion tensor. The 532 nm wavelength is used to obtain spectral information sensitive to the chemical aging of potassium persulfate, and the 785 nm wavelength is used to obtain spectral information of the unaged state. A multi-task recognition module is used to input a dual-channel one-dimensional fusion tensor into a pre-trained multi-task deep learning model. This model includes a shared feature extraction backbone network, a first fully connected classification head, and a second fully connected classification head. The shared feature extraction backbone network extracts features from the dual-channel one-dimensional fusion tensor and outputs a shared feature vector. The first fully connected classification head outputs the microplastic sample type recognition result based on the shared feature vector. The second fully connected classification head outputs the microplastic sample aging state recognition result based on the shared feature vector. The multi-task deep learning model is jointly trained end-to-end using a weighted multi-task loss function, allowing the shared feature extraction backbone network to be simultaneously modulated by the gradients of both the type recognition task and the aging state recognition task during backpropagation, thus learning a decoupled shared feature representation.

[0016] A third aspect of this application provides a device for simultaneously identifying the type and aging state of microplastics based on dual-wavelength laser Raman spectroscopy, comprising: One or more processors; A memory on which one or more programs are stored; When the one or more programs are executed by the one or more processors, the one or more processors implement the method for simultaneous identification of microplastic types and aging states based on dual-wavelength laser Raman spectroscopy as described above.

[0017] The fourth aspect of this application provides a computer storage medium for storing a program, which, when executed, is used to implement the method for synchronous identification of microplastic types and aging states based on dual-wavelength laser Raman spectroscopy as described in any of the preceding claims.

[0018] Beneficial effects: High recognition accuracy: By fusing complementary spectral information from 532 nm laser (sensitive to aging state) and 785 nm laser (strong resistance to fluorescence interference), and utilizing a multi-task deep learning model to automatically learn deep identification features, the bottleneck of limited information from a single light source is effectively overcome. This achieves 100% accuracy in identifying microplastic types and up to 99.2% accuracy in identifying aging states, providing an efficient and reliable technical means for the precise assessment of the environmental behavior and ecological risks of microplastics.

[0019] High efficiency of synchronous identification: It adopts a "one-core dual-head" multi-task learning model to complete the synchronous identification of two key attributes, namely type and aging status, in one go, avoiding the problems of error accumulation and low efficiency caused by step-by-step analysis, and significantly improving the detection throughput.

[0020] Highly targeted and automated: For the first time, the Raman spectral characteristics of two typical environmental aging processes, UV aging and potassium persulfate chemical aging, are systematically incorporated into a machine learning model, enabling effective discrimination of aging type and degree; the entire process, from spectral processing to result output, is highly automated, providing a core solution for high-throughput and intelligent screening of environmental samples.

[0021] The model exhibits good robustness: through end-to-end joint training and weighted multi-task loss optimization, the features learned by the shared backbone network are more generalizable, improving the reliability of sample detection in real-world complex environments.

[0022] Advanced feature fusion method: It abandons the traditional method of manually designing fusion rules or back-end decision fusion, and adopts "early channel stacking" fusion, which enables the model to automatically learn the deep correlation and complementary information between different wavelength spectra from the input layer, thereby improving the discriminative power of feature representation from the source. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1A schematic flowchart illustrating a method for simultaneously identifying microplastic types and aging states based on dual-wavelength laser Raman spectroscopy, provided for embodiments of this application; Figure 2 Raman spectra of seven original microplastics provided for embodiments of this application; Figures 3(a) and 3(b) show the dimensionality reduction results of seven microplastic PCA methods provided in the embodiments of this application; Figure 4 The confusion matrix provided in this application provides a schematic diagram illustrating the classification effect of microplastic types under single laser and dual laser modes; Figure 5 The confusion matrix provided in this application provides a schematic diagram illustrating the classification effect of microplastic aging under single laser and dual laser modes; Figure 6 A schematic diagram of a microplastic type and aging state synchronous identification system based on dual-wavelength laser Raman spectroscopy provided in this application embodiment; Figure 7 This is a schematic diagram of the structure of a device provided in an embodiment of this application. Detailed Implementation

[0025] This application provides a method and system for simultaneous identification of microplastic species and aging status based on dual-wavelength laser Raman spectroscopy, which can solve the problems of isolated identification of microplastic species and assessment of aging status, insufficient accuracy, and weak anti-interference ability in the prior art.

[0026] See Figure 1 This figure is a flowchart illustrating a method for simultaneously identifying microplastic types and aging states based on dual-wavelength laser Raman spectroscopy, provided in an embodiment of this application. The method for simultaneously identifying microplastic types and aging states based on dual-wavelength laser Raman spectroscopy provided in this application can be implemented, for example, through the following steps S101-103.

[0027] S101: The same microplastic sample was excited using lasers with wavelengths of 532 nm and 785 nm, respectively, and the raw Raman spectra of the microplastic sample were collected.

[0028] In this embodiment, based on a LabRAM Odyssey Raman spectrometer, dual-wavelength excitation light of 532 nm and 785 nm was used to systematically collect Raman spectra of seven common microplastics (polyethylene PE, polypropylene PP, polystyrene PS, polyvinyl chloride PVC, polyethylene terephthalate PET, polycarbonate PC, and polylactic acid PLA) and their three unaged and aged treatments (unaged, UV-aged, and potassium persulfate-aged). Hundreds of valid spectral data were collected for each treatment state of the microplastics. The wavenumber range for the 532 nm wavelength acquisition was set to 400-3200 cm⁻¹. -1 The wavenumber range for 785 nm wavelength acquisition is set to 400-2900 cm⁻¹. -1 .

[0029] S102: Preprocess the raw Raman spectral data of the two wavelengths, and stack the two preprocessed spectral data along the channel dimension to construct a two-channel one-dimensional fusion tensor.

[0030] In this embodiment, the raw dual-wavelength Raman spectral data are denoised and baseline corrected; the baseline-corrected spectral data are then normalized to a uniform intensity range. Specifically, the 532 nm wavelength is used to acquire spectral information sensitive to potassium persulfate chemical aging, and the 785 nm wavelength is used to acquire spectral information of the unaged state. Specifically, all obtained spectra were preprocessed with denoising, baseline correction, and normalization to improve spectral quality. Subsequently, the two preprocessed spectral data obtained from the same sample under 785 nm and 532 nm lasers were combined, and their association with dual labels of "microplastic species" (7 categories) and "microplastic aging state" (3 categories) was established to build a high-quality database for subsequent model training.

[0031] This invention proposes an "early channel stacking" dual-modal fusion method. In the data processing stage, the 532 nm and 785 nm Raman spectra of the same microplastic sample are preprocessed separately and stacked along the channel dimension to form a dual-channel one-dimensional tensor, creating a structured fusion input. This tensor is then input into a one-dimensional convolutional neural network, where the first-layer convolutional kernel simultaneously performs weighted calculations on the dual-channel data during the sliding process, achieving end-to-end adaptive feature fusion starting from the input layer. This method abandons the traditional approach of manually designing fusion rules or making backend decisions for fusion, enabling the model to automatically learn the deep correlations and complementary information between different wavelength spectra, thereby improving the discriminative power of feature representation from the source.

[0032] S103: Input the dual-channel one-dimensional fusion tensor into a pre-trained multi-task deep learning model and output the category recognition result and aging state recognition result.

[0033] In this embodiment, a dual-channel one-dimensional fusion tensor is input into a pre-trained multi-task deep learning model. The multi-task deep learning model includes a shared feature extraction backbone network, a first fully connected classification head, and a second fully connected classification head. The shared feature extraction backbone network extracts features from the dual-channel one-dimensional fusion tensor and outputs a shared feature vector. The first fully connected classification head outputs the microplastic sample type identification result based on the shared feature vector. The second fully connected classification head outputs the microplastic sample aging state identification result based on the shared feature vector.

[0034] The shared feature extraction backbone network is a neural network containing multiple layers of one-dimensional convolutions, specifically five sequentially connected one-dimensional convolutional modules. Each one-dimensional convolutional module sequentially performs one-dimensional convolution, batch normalization, ReLU activation, and pooling operations. The first four one-dimensional convolutional modules use max pooling, while the fifth one-dimensional convolutional module uses adaptive average pooling. Three sequentially connected fully connected layers receive the features output by the adaptive average pooling operation and perform nonlinear transformations to generate a shared feature vector.

[0035] The first fully connected classification head maps the shared feature vector to seven categories, corresponding to polyethylene, polypropylene, polystyrene, polyvinyl chloride, polyethylene terephthalate, polycarbonate, and polylactic acid, respectively; the second fully connected classification head maps the shared feature vector to three categories, corresponding to the unaged state, UV-aged state, and potassium persulfate-aged state, respectively.

[0036] Specifically, the network structure adopted in this invention is as follows: the input dual-channel spectral data first passes through a shared backbone network containing five layers of one-dimensional convolutions, where batch normalization, ReLU activation, and pooling operations are performed after each convolution layer. The first four layers use max pooling, and the fifth layer uses adaptive average pooling to unify the feature dimension. Subsequently, high-level shared features are further extracted through three fully connected layers. Finally, the shared features are simultaneously fed into two parallel fully connected classification heads. One is used to predict aging states (such as UV, KO, virgin), and the other is used to identify microplastic types (such as PET, PE, PP, etc.). The entire network is jointly trained end-to-end using a weighted multi-task loss function, achieving unified optimization of feature fusion and dual-task classification.

[0037] This invention employs a "one-core, dual-head" hard parameter-sharing multi-task learning architecture. This architecture consists of a shared feature extraction backbone network and two parallel task-specific classification heads. The shared backbone comprises multiple layers of one-dimensional convolutional and fully connected layers, responsible for extracting general high-level features from the fused spectrum that contribute to both aging and species classification. Based on these general features, two lightweight and parameter-independent classification heads (fully connected networks) respectively perform classification outputs for aging states and microplastic species. During training, joint optimization is performed using a weighted multi-task loss function, allowing the shared backbone to be simultaneously modulated by the gradients of both tasks during backpropagation, thereby learning more generalizable feature representations. This design not only significantly improves computational efficiency through parameter sharing but also enhances the overall robustness and classification accuracy of the model by leveraging the correlation between tasks.

[0038] In one implementation of this application, a training dataset is constructed, wherein each sample in the training dataset includes a dual-channel one-dimensional fusion tensor, a microplastic category label, and a microplastic aging state label; the training dataset is divided into a training set and a test set; a multi-task deep learning model is constructed and its parameters are initialized; a multi-task loss function is defined, which is a weighted sum of a first classification loss function and a second classification loss function, wherein the first classification loss function corresponds to the category recognition task and the second classification loss function corresponds to the aging state recognition task; the multi-task deep learning model is jointly trained end-to-end using the training set, and the multi-task loss function is minimized using the backpropagation algorithm until the multi-task deep learning model converges; the accuracy of the trained multi-task deep learning model in category recognition and aging state recognition is evaluated using the test set.

[0039] The training dataset is constructed as follows: the types of microplastics include polyethylene, polypropylene, polystyrene, polyvinyl chloride, polyethylene terephthalate, polycarbonate, and polylactic acid; the aging states of microplastics include unaged state, UV-aged state, and potassium persulfate-aged state; for each type and each combination of aging states, no less than one hundred valid Raman spectral data are collected to generate the corresponding dual-channel one-dimensional fusion tensor.

[0040] Specifically, the database is divided into training and testing sets in an 8:2 ratio. A multi-channel, multi-task deep learning model is trained until the loss converges. The model can then distinguish the type and aging state of microplastics. Preferred models include, but are not limited to, artificial neural network models, convolutional neural network models, and Transformers.

[0041] In one implementation of this application, microplastic test samples containing unknown types and aging states are obtained; steps S101-S103 are executed to process and identify the microplastic test samples to obtain the type identification results and aging state identification results of the microplastic test samples; based on the identification results of multiple microplastic test samples and the real labels, the type identification accuracy, aging state identification accuracy, and dual-label synchronous identification accuracy of the multi-task deep learning model are calculated.

[0042] Specifically, the performance of the model is verified using a test set, and its accuracy, precision, recall, and other metrics in category identification and aging state classification are evaluated to comprehensively assess the model's recognition reliability.

[0043] For unknown microplastic samples, their dual-laser Raman spectra are collected according to steps S101 and S102, combined, and input into a trained machine learning model to simultaneously obtain the predicted results of the polymer type and aging state of the sample.

[0044] like Figure 2 As shown, Figure 2 The Raman spectra of the seven original microplastics provided in this application embodiment are shown, with the left side acquired by a 785 nm laser and the right side acquired by a 532 nm laser. All seven microplastics exhibited distinct Raman characteristic peaks under dual laser excitation. For example, PE showed peaks at 1064, 1133, 1297, and 1443 cm⁻¹. -1 There are strong peaks at these points, corresponding to the skeletal stretching vibration and CH2 bending vibration modes. The characteristic peaks of PP are at 811, 844, 1154, and 1461 cm⁻¹. -1 At this location, it corresponds to its CH2 rocking vibration and C–C skeleton vibration. PS at 1002 cm -1 The vicinity shows a benzene ring respiration vibration peak, and at 1604 cm⁻¹ -1 The presence of a characteristic peak for the C=C stretching vibration of the benzene ring at 1727 cm⁻¹ is an important basis for identifying PS. PET shows a peak at 1727 cm⁻¹. -1 A strong C=O stretching vibration peak is observed nearby, and at 1616 cm⁻¹... -1 The presence of skeletal vibrations of the benzene ring at this point is the most typical characteristic peak for PET. PVC shows a peak at 638 cm⁻¹. -1 and 695 cm -1 The characteristic double peak of C-Cl stretching vibration is observed at 890 cm⁻¹. The characteristic peak of PC appears at 890 cm⁻¹. -1 Nearby, corresponding to the CH bending vibration of the benzene ring. PLA at 876 cm⁻¹ -1 With 1452 cm -1 There is a characteristic vibrational peak nearby, with its C=O stretching vibration located at approximately 1765 cm⁻¹. -1It is worth noting that the spectral ranges covered by different excitation wavelengths are complementary. The 785 nm laser primarily obtains wavelengths in the 500–2500 cm⁻¹ region. -1 The "fingerprint region" spectrum, containing subtle vibrational information of various chemical bonds, is suitable for the preliminary identification of most microplastics. The 532 nm laser extends the spectral range to 3200 cm⁻¹. -1 Completely covers the C–H stretching vibration region (2800–3000 cm). -1 This provides additional vibrational spectroscopic evidence for distinguishing different polymers (especially C–H structures with different degrees of saturation).

[0045] As shown in Figures 3(a) and 3(b), these figures present the PCA dimensionality reduction results for seven microplastics provided in this application embodiment, used for data pattern analysis. After PCA dimensionality reduction, Raman spectra obtained under 785 nm laser excitation show a significant separation trend between the chemical aging state of PLA plastic and the UV aging state of PVC and the corresponding other two states, indicating that this method has a certain ability to distinguish spectral changes caused by some aging mechanisms. However, the distinction between different aging states for other types of microplastics is low, and cluster overlap is significant, indicating that the three states of a single plastic cannot be effectively separated in the existing feature space. Similarly, under 532 nm laser excitation, the UV aging states of PC, PET, PS, and PVC can also be clearly distinguished from the other two states, but there is still significant overlap between the three states (unaged, chemically aged, and UV-aged) for other types of microplastics, making accurate qualitative identification impossible. The above results collectively indicate that reliably distinguishing multiple aging states of a single plastic remains challenging, and further integration of feature engineering, nonlinear dimensionality reduction, or supervised machine learning methods is needed to improve state recognition capabilities.

[0046] In the identification of microplastic species, models based on a single laser (785 nm or 532 nm) can achieve accurate identification of microplastics. However, in the more challenging task of identifying aging states, the single laser model shows significant differences and limitations: using the 785 nm laser-based model, the total number of tested samples is 133, and the aging accuracy is 92.5%; using the 532 nm laser-based model, the total number of tested samples is 337, and the accuracy is 97.3%. The dual-laser spectral feature fusion strategy proposed in this invention, by integrating the complementary information from 785 nm and 532 nm lasers, reduces the total number of tested samples to 135 and successfully improves the overall accuracy of aging state identification to 99.2%. This result not only verifies the synergistic enhancement effect of dual-laser information in characterizing the chemical changes of microplastic aging, but also demonstrates the unique advantages and significant progress of the method of this invention in achieving simultaneous and high-precision identification of species and aging states, as shown in Table 1.

[0047] Table 1. Comparison of machine learning classification performance in single-laser and dual-laser modes.

[0048] Starting with the confusion matrix, a more detailed analysis of the model's classification results and misclassified samples can be performed. (See [link to relevant documentation]). Figure 4 and Figure 5 , Figure 4 The confusion matrix provided in this application provides a schematic diagram illustrating the classification effect of microplastic types under single laser and dual laser modes; Figure 5 The confusion matrix provided in this application illustrates the classification effect of microplastic aging under single-laser and dual-laser modes. Details are as follows: (1) In the identification of microplastics, the single excitation wavelength has limitations in dealing with subtle spectral similarities between specific polymers. For example, the single 785 nm laser model has an extremely high overall accuracy (0.992) in the identification of 7 types of polymers, but there is still one specific misjudgment: the spectrum of a polylactic acid (PLA) sample was misjudged as polyvinyl chloride (PVC). This misjudgment may be due to the local overlap or similar relative intensity of some Raman characteristic peaks of PLA and PVC under the 785 nm laser, which causes the model to confuse in very few cases. Both the single 532 nm laser model and the dual laser fusion model can achieve accurate identification of species, and the confusion matrix is ​​perfectly diagonally distributed. This shows that for the identification of microplastic species, the 532 nm laser can provide sufficiently discriminative spectral features; while the dual laser fusion strategy adopted in this invention inherits and consolidates this perfect performance without introducing any additional noise or confusion due to information fusion.

[0049] (2) In the identification of microplastic aging types, the single excitation wavelength is not sensitive enough to the changes in substances caused by aging. For example, in the single 785 nm laser mode, the main misclassification of this model occurs between KO aging and the original state, reflecting its recognition limitations. Four KO aging samples were misclassified as the original state, indicating that the 785 nm laser is not sensitive to certain molecular structure changes caused by potassium persulfate oxidation, resulting in the omission of some oxidation features. Five UV aging samples were misclassified as KO aging, indicating that the model is confused when distinguishing between two different types of aging (photo-oxidation and chemical oxidation). The original sample has a high recognition rate, with only one misclassified as UV aging. In the single 532 nm laser model, the misclassification pattern of this model is significantly different from that of the 785 nm laser model, mainly reflected in the confusion between UV aging and the original state. Six UV aging samples were misclassified as the original state, while three original samples were misclassified as UV aging. This indicates that while the 532 nm laser has extremely strong discriminative power (100% accuracy) for KO aging, it may overlap with some spectral features of the original state when capturing UV aging or specific degrees of molecular changes. No misclassifications were found for KO aging samples, verifying the specific response of the 532 nm laser to this aging type. The dual-laser fusion model, after fusing information from both channels, significantly reduced the model's misclassification rate and fundamentally changed the error pattern. The most prominent misclassification of KO aging as the original state (a major problem with the 785 nm model) was completely eliminated. Bidirectional misclassifications between UV aging and the original state were greatly reduced, with only one UV aging misclassification remaining. Crucially, no misclassifications occurred between UV aging and KO aging, successfully resolving the core confusion problem of the single 785 nm model in this area.

[0050] This invention achieves simultaneous and automated identification of microplastic types and aging states by integrating multi-wavelength excitation, spectral preprocessing optimization, multi-modal feature fusion, and a deep neural network model. Specifically, this invention establishes a multi-wavelength Raman spectral database, utilizes a deep convolutional neural network to automatically learn the deep spectral features of different polymer types and aging stages, and combines an attention mechanism to enhance sensitivity to weak aging marker peaks. Ultimately, this forms an end-to-end, highly robust intelligent identification system, providing strong technical support for the accurate assessment and efficient management of microplastic environmental risks.

[0051] Based on the methods provided in the above embodiments, this application also provides a system for synchronously identifying microplastic types and aging states based on dual-wavelength laser Raman spectroscopy. The following describes the system for synchronously identifying microplastic types and aging states based on dual-wavelength laser Raman spectroscopy in conjunction with the accompanying drawings.

[0052] See Figure 6The figure is a schematic diagram of a microplastic type and aging state synchronous identification system based on dual-wavelength laser Raman spectroscopy provided in an embodiment of this application.

[0053] The microplastic type and aging state synchronous identification system 600 based on dual-wavelength laser Raman spectroscopy provided in this application includes: a data acquisition module 601, a data processing and fusion module 602, and a multi-task identification module 603.

[0054] The data acquisition module 601 is used to excite the same microplastic sample with lasers of 532 nm wavelength and 785 nm wavelength respectively, and acquire the raw Raman spectral data of the microplastic sample with dual wavelengths. The data processing and fusion module 602 is used to preprocess the raw dual-wavelength Raman spectral data and stack the two preprocessed spectral data along the channel dimension to construct a dual-channel one-dimensional fusion tensor. The 532 nm wavelength is used to acquire spectral information sensitive to potassium persulfate chemical aging, and the 785 nm wavelength is used to acquire spectral information of the unaged state. The multi-task recognition module 603 is used to input a dual-channel one-dimensional fusion tensor into a pre-trained multi-task deep learning model. The multi-task deep learning model includes a shared feature extraction backbone network, a first fully connected classification head, and a second fully connected classification head. The shared feature extraction backbone network extracts features from the dual-channel one-dimensional fusion tensor and outputs a shared feature vector. The first fully connected classification head outputs the microplastic sample type recognition result based on the shared feature vector. The second fully connected classification head outputs the microplastic sample aging state recognition result based on the shared feature vector. The multi-task deep learning model is jointly trained end-to-end using a weighted multi-task loss function, allowing the shared feature extraction backbone network to be simultaneously subjected to gradient modulation from both the type recognition task and the aging state recognition task during backpropagation, thus learning a decoupled shared feature representation.

[0055] In one possible implementation, the data processing and fusion module 602 is specifically used for: Denoising and baseline correction were performed on the raw dual-wavelength Raman spectral data. The baseline-corrected spectral data are normalized to a uniform intensity range.

[0056] In one possible implementation, the shared feature extraction backbone network is a neural network containing multiple layers of one-dimensional convolutions, specifically including: Five sequentially connected one-dimensional convolutional modules, each performing one-dimensional convolution, batch normalization, ReLU activation, and pooling operations in sequence; the first four one-dimensional convolutional modules use max pooling, and the fifth one-dimensional convolutional module uses adaptive average pooling. Three fully connected layers are connected in sequence. They receive the features output by adaptive average pooling and perform nonlinear transformations to generate a shared feature vector.

[0057] In one possible implementation, a training module is also included, which has the following functions: Construct a training dataset in which each sample includes a two-channel one-dimensional fusion tensor, a microplastic type label, and a microplastic aging state label. The training dataset is divided into a training set and a test set; Construct a multi-task deep learning model and initialize the model parameters; Define a multi-task loss function, which is a weighted sum of the first classification loss function and the second classification loss function. The first classification loss function corresponds to the category recognition task, and the second classification loss function corresponds to the aging state recognition task. The multi-task deep learning model is jointly trained end-to-end using the training set, and the multi-task loss function is minimized through the backpropagation algorithm until the multi-task deep learning model converges. The accuracy of the trained multi-task deep learning model in species recognition and aging state recognition was evaluated using a test set.

[0058] In one possible implementation, a training module is included, which has the following functions: Microplastics include polyethylene, polypropylene, polystyrene, polyvinyl chloride, polyethylene terephthalate, polycarbonate, and polylactic acid; The aging states of microplastics include unaged state, UV-aged state, and potassium persulfate-aged state. For each combination of type and aging state, at least one hundred valid Raman spectral data are collected to generate the corresponding dual-channel one-dimensional fusion tensor.

[0059] In one possible implementation, a first fully connected classification head is used to map the shared feature vector to seven categories, corresponding to polyethylene, polypropylene, polystyrene, polyvinyl chloride, polyethylene terephthalate, polycarbonate, and polylactic acid, respectively; a second fully connected classification head is used to map the shared feature vector to three categories, corresponding to the unaged state, the UV-aged state, and the potassium persulfate-aged state, respectively.

[0060] In one possible implementation, a model validation module is also included, specifically for: Obtain test samples containing microplastics of unknown types and aging states; A method for simultaneous identification of microplastic types and aging states based on dual-wavelength laser Raman spectroscopy was implemented to process and identify microplastic test samples, thereby obtaining the identification results of microplastic types and aging states of the microplastic test samples. Based on the identification results of multiple microplastic test samples and real labels, the accuracy of type identification, aging state identification, and dual-label synchronous identification of the multi-task deep learning model were calculated.

[0061] Since the dual-wavelength laser Raman spectroscopy-based microplastic type and aging state synchronous identification system 600 is a system corresponding to the dual-wavelength laser Raman spectroscopy-based microplastic type and aging state synchronous identification method provided in the above method embodiments, the specific implementation of each unit of the dual-wavelength laser Raman spectroscopy-based microplastic type and aging state synchronous identification system 600 is based on the same concept as the above method embodiments. Therefore, for the specific implementation of each unit of the dual-wavelength laser Raman spectroscopy-based microplastic type and aging state synchronous identification system 600, please refer to the description of the dual-wavelength laser Raman spectroscopy-based microplastic type and aging state synchronous identification method in the above method embodiments, and will not be repeated here.

[0062] This application embodiment also provides a device for synchronously identifying the type and aging state of microplastics based on dual-wavelength laser Raman spectroscopy, the device including: a processor and a memory; The memory is used to store instructions; The processor is used to execute the instructions in the memory to perform the method for simultaneous identification of microplastic types and aging states based on dual-wavelength laser Raman spectroscopy mentioned in the above embodiments.

[0063] It should be noted that the hardware structure of the microplastic type and aging state synchronous identification device based on dual-wavelength laser Raman spectroscopy provided in this application embodiment can be as follows: Figure 7 The structure shown, Figure 7 This is a schematic diagram of the structure of a device provided in an embodiment of this application.

[0064] Please see Figure 7 As shown, device 700 includes: a processor 710, a communication interface 720, and a memory 730. The number of processors 710 in device 700 can be one or more. Figure 7 Taking a processor as an example, in this embodiment, the processor 710, communication interface 720, and memory 730 can be connected via a bus system or other means. Figure 7 Taking the connection between China and Israel via the 740 bus system as an example.

[0065] Processor 710 may be a central processing unit (CPU), a network processor (NP), or a combination of a CPU and an NP. Processor 710 may further include hardware chips. These hardware chips may be application-specific integrated circuits (ASICs), programmable logic devices (PLDs), or combinations thereof. The PLD may be a complex programmable logic device (CPLD), a field-programmable gate array (FPGA), generic array logic (GAL), or any combination thereof.

[0066] The memory 730 may include volatile memory, such as random-access memory (RAM); the memory 730 may also include non-volatile memory, such as flash memory, hard disk drive (HDD) or solid-state drive (SSD); the memory 730 may also include a combination of the above types of memory.

[0067] Optionally, the memory 730 stores an operating system and programs, executable modules, or data structures, or subsets thereof, or extended sets thereof. The programs may include various operation instructions for implementing various operations. The operating system may include various system programs for implementing various basic services and handling hardware-based tasks. The processor 710 can read the programs in the memory 730 to implement the method for simultaneous identification of microplastic types and aging states based on dual-wavelength laser Raman spectroscopy provided in this embodiment.

[0068] The bus system 740 can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The bus system 740 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 7The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0069] This application also provides a computer-readable storage medium including instructions that, when executed on a computer, cause the computer to perform the method for synchronous identification of microplastic types and aging states based on dual-wavelength laser Raman spectroscopy mentioned in the above embodiments.

[0070] This application also provides a computer program product containing instructions that, when run on a computer, cause the computer to execute the method for synchronously identifying the type and aging state of microplastics based on dual-wavelength laser Raman spectroscopy mentioned in the above embodiments.

[0071] Although the invention has been specifically shown and described in conjunction with preferred embodiments, those skilled in the art should understand that various changes in form and detail may be made to the invention without departing from the spirit and scope of the invention as defined in the appended claims, all of which shall be within the scope of protection of the invention.

Claims

1. A method for simultaneous identification of microplastic type and aging state based on dual-wavelength laser Raman spectroscopy, characterized in that, The method includes: The same microplastic sample was excited using lasers with wavelengths of 532 nm and 785 nm, and the raw Raman spectra of the microplastic sample were collected. The raw Raman spectral data of the two wavelengths are preprocessed, and the two preprocessed spectral data are stacked along the channel dimension to construct a two-channel one-dimensional fusion tensor. The 532 nm wavelength is used to obtain spectral information sensitive to the chemical aging of potassium persulfate, and the 785 nm wavelength is used to obtain spectral information of the unaged state. The dual-channel one-dimensional fusion tensor is input into a pre-trained multi-task deep learning model, wherein the multi-task deep learning model includes a shared feature extraction backbone network, a first fully connected classification head, and a second fully connected classification head; the shared feature extraction backbone network extracts features from the dual-channel one-dimensional fusion tensor and outputs a shared feature vector; the first fully connected classification head outputs the species identification result of the microplastic sample based on the shared feature vector; the second fully connected classification head outputs the aging state identification result of the microplastic sample based on the shared feature vector, wherein the multi-task deep learning model is jointly trained end-to-end using a weighted multi-task loss function, so that the shared feature extraction backbone network is simultaneously subjected to gradient modulation of the species identification task and the aging state identification task during backpropagation, and learns a decoupled shared feature representation.

2. The method according to claim 1, characterized in that, The preprocessing of the dual-wavelength raw Raman spectral data includes: The raw dual-wavelength Raman spectral data were denoised and baseline corrected. The baseline-corrected spectral data are normalized to a uniform intensity range.

3. The method according to claim 1, characterized in that, The shared feature extraction backbone network is a neural network containing multiple layers of one-dimensional convolutions, specifically including: Five sequentially connected one-dimensional convolutional modules, each performing one-dimensional convolution, batch normalization, ReLU activation, and pooling operations in sequence; the first four one-dimensional convolutional modules use max pooling, and the fifth one-dimensional convolutional module uses adaptive average pooling. The three fully connected layers are connected in sequence. They receive the features output by the adaptive average pooling operation and perform nonlinear transformation to generate the shared feature vector.

4. The method according to claim 1, characterized in that, The method further includes: Construct a training dataset, wherein each sample in the training dataset includes a dual-channel one-dimensional fusion tensor, a type label of the microplastic, and an aging state label of the microplastic; The training dataset is divided into a training set and a test set; Construct the multi-task deep learning model and initialize the model parameters; Define a multi-task loss function, which is a weighted sum of a first classification loss function and a second classification loss function. The first classification loss function corresponds to the category identification task, and the second classification loss function corresponds to the aging state identification task. The multi-task deep learning model is jointly trained end-to-end using the training set, and the multi-task loss function is minimized through the backpropagation algorithm until the multi-task deep learning model converges. The accuracy of the trained multi-task deep learning model in species recognition and aging state recognition was evaluated using the test set.

5. The method according to claim 4, characterized in that, The construction of the training dataset includes: The types of microplastics include polyethylene, polypropylene, polystyrene, polyvinyl chloride, polyethylene terephthalate, polycarbonate, and polylactic acid; The aging states of the microplastics include unaged state, UV-aged state, and potassium persulfate-aged state. For each combination of type and aging state, at least one hundred valid Raman spectral data are collected to generate the corresponding dual-channel one-dimensional fusion tensor.

6. The method according to claim 1, characterized in that, The first fully connected classification head is used to map the shared feature vector to seven categories, corresponding to polyethylene, polypropylene, polystyrene, polyvinyl chloride, polyethylene terephthalate, polycarbonate and polylactic acid respectively; the second fully connected classification head is used to map the shared feature vector to three categories, corresponding to the unaged state, UV-aged state and potassium persulfate-aged state respectively.

7. The method according to claim 1, characterized in that, The method further includes: Obtain test samples containing microplastics of unknown types and aging states; By performing the method described in claim 1, the microplastic test sample is processed and identified to obtain the type identification result and aging state identification result of the microplastic test sample; Based on the identification results of multiple microplastic test samples and the real labels, the accuracy of the multi-task deep learning model in terms of type identification, aging state identification, and dual-label synchronous identification is calculated.

8. A system for simultaneous identification of microplastic type and aging state based on dual-wavelength laser Raman spectroscopy, characterized in that, The system includes: The data acquisition module is used to excite the same microplastic sample with lasers of 532 nm and 785 nm wavelengths respectively, and acquire the raw Raman spectral data of the microplastic sample with dual wavelengths. The data processing and fusion module is used to preprocess the raw dual-wavelength Raman spectral data and stack the two preprocessed spectral data along the channel dimension to construct a dual-channel one-dimensional fusion tensor. The 532 nm wavelength is used to obtain spectral information sensitive to the chemical aging of potassium persulfate, and the 785 nm wavelength is used to obtain spectral information of the unaged state. A multi-task recognition module is used to input the dual-channel one-dimensional fusion tensor into a pre-trained multi-task deep learning model. The multi-task deep learning model includes a shared feature extraction backbone network, a first fully connected classification head, and a second fully connected classification head. The shared feature extraction backbone network extracts features from the dual-channel one-dimensional fusion tensor and outputs a shared feature vector. The first fully connected classification head outputs the species identification result of the microplastic sample based on the shared feature vector. The second fully connected classification head outputs the aging state identification result of the microplastic sample based on the shared feature vector. The multi-task deep learning model is jointly trained end-to-end using a weighted multi-task loss function, so that the shared feature extraction backbone network is simultaneously subjected to gradient modulation from both the species identification task and the aging state identification task during backpropagation, learning a decoupled shared feature representation.

9. An electronic device, characterized in that, The device includes: a processor and a memory; The memory is used to store instructions; The processor is configured to execute the instructions in the memory to perform the method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, Including instructions that, when run on a computer, cause the computer to perform the method described in any one of claims 1-7 above.

Images