Microplastic detection system and detection method

Abstract

The application discloses a micro-plastic detection system and a detection method, and belongs to the technical field of micro-plastic detection; the application comprises a confocal photoelectric coupling module, microscopic imaging and coaxial alignment of Raman spectrum are realized by constructing a shared optical axis; a double-flow processing architecture combining a visual representation engine and a spectrum representation engine is used to obtain the physical form and chemical composition of the candidate target in the sample; wherein the visual representation engine uses an improved YOLOv8 network to quickly locate the candidate target and combines an edge detection algorithm to extract the size and shape parameters of the candidate target; the spectrum representation engine analyzes the Raman spectrum sequence of the candidate target based on a Transformer encoder, and accurately identifies the polymer type; then the system fuses the physical form data obtained by the visual representation engine and the chemical composition data obtained by the spectrum representation engine through spatial coordinate mapping, solves the problem of physical form and chemical composition analysis fragmentation in the prior art, and realizes efficient and accurate integrated detection of micro-plastics.

Description

Technical Field

[0001] This invention belongs to the field of microplastic detection technology, specifically relating to a microplastic detection system and detection method. Background Technology

[0002] With the widespread use of plastic products, microplastics (plastic particles with a diameter <5mm) have become a global environmental pollutant. Groundwater, as an important source of drinking water and industrial and agricultural water, is directly affected by microplastic pollution, impacting ecological security and human health. Furthermore, compared to ocean or surface water, groundwater environments have unique characteristics: complex sample matrices, relatively low microplastic abundance but smaller particle size, and long-term influence from groundwater hydrogeological conditions, often resulting in complex biofilms or aging and degradation on microplastic surfaces. Therefore, detecting microplastics in groundwater is more challenging. Traditional detection methods typically involve step-by-step processing, such as first performing microscopic imaging, then manually searching for targets for spectral analysis, or using two separate sets of equipment. This approach is not only inefficient, but also separates the morphological characterization and compositional analysis of microplastics, making it difficult to establish a precise "morphology-composition" correspondence at the single-particle level. This hinders in-depth analysis of the migration patterns and toxicological characteristics of different types of microplastics. Summary of the Invention

[0003] One of the core objectives of this invention is to provide a microplastic detection system that can integrate the physical morphology (size, shape, color) and chemical composition (polymer type) of the target being detected.

[0004] The second core objective of this invention is to provide a microplastic detection method that integrates the physical morphology (size, shape, color) and chemical composition (polymer type) of the target being detected.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A microplastic detection system, comprising:

[0006] The confocal optocoupler module includes a visual imaging optical path and a Raman spectroscopy acquisition optical path, and the visual imaging optical path and the Raman spectroscopy acquisition optical path are constructed in the same physical space with a shared optical axis that can be confocal aligned.

[0007] A precision motion control unit is used to carry the sample to be tested and to perform micron-level displacement in a plane perpendicular to the shared optical axis according to a preset scanning path or target coordinates.

[0008] The visual representation engine is used to receive the microscopic image stream acquired by the visual imaging optical path, locate the candidate target region in the image in real time, extract the bounding box coordinates of the candidate target, and extract the physical morphological feature vector of the candidate target by combining the edge detection algorithm.

[0009] The spectral characterization engine is used to receive the spectral data sequence acquired by the Raman spectroscopy acquisition optical path for the candidate target, analyze the molecular fingerprint features, and output the classification probability distribution of the polymer type.

[0010] The multimodal data fusion and decision module associates and matches the physical morphology feature vector with the classification probability distribution through a spatial coordinate mapping matrix to generate a composite feature fingerprint containing the physical morphology features and chemical composition of a single particle.

[0011] The intelligent report generation terminal outputs a multi-dimensional statistical report on contamination based on the composite feature fingerprint.

[0012] A further embodiment is that the confocal optocoupler module includes: an illumination source, an excitation source, a dichroic beam splitter, a microscope objective, an area array image sensor, and a spectrometer;

[0013] The illumination source is used to illuminate the object to be tested, so that the sample emits reflected light. The reflected light passes through the microscope objective and is transmitted to the area array image sensor through the dichroic beam splitter to form a visual imaging optical path.

[0014] The excitation light source is transmitted to the microscope objective via the dichroic beam splitter and focused onto the sample to be tested, exciting the sample to emit Raman scattered light; the Raman scattered light is transmitted to the spectrometer via the microscope objective and the dichroic beam splitter, forming the Raman spectral acquisition optical path.

[0015] A further embodiment is that the excitation source is selected from one of a 785nm, 532nm, or 1064nm frequency-stabilized laser; and the cutoff wavelength of the dichroic beam splitter is matched with the center wavelength of the excitation source.

[0016] A further aspect is that the visual representation engine includes a deep convolutional neural network for locating candidate target regions and extracting candidate target bounding boxes. This deep convolutional neural network employs an improved variant of the YOLO architecture, including:

[0017] The backbone feature extraction network uses the CSPDarknet structure for multi-scale feature extraction.

[0018] The feature fusion network employs a path aggregation network structure to enhance feature transfer for small targets.

[0019] The detection head is decoupled, separating the target confidence prediction from the bounding box regression task; the bounding box regression is trained and optimized using the CIoU (Complete Intersection over Union) loss function, which is defined as:

[0020]

[0021] In the formula, IoU is the intersection-union ratio, and ρ is the distance between the predicted box center b and the ground truth box center b. gt The Euclidean distance is given by c, where c is the diagonal length of the minimum bounding rectangle, α is the weighting coefficient, and v is the aspect ratio consistency parameter.

[0022] A further embodiment of the scheme is that the visual representation engine also includes a morphological parameter calculation unit, used to perform the following operations within the candidate target region located by the deep convolutional neural network: performing contrast-limited adaptive histogram equalization on the region of interest (ROI); applying Gaussian filtering to remove high-frequency noise; extracting edge contours using a dual-threshold Canny operator; and calculating the area A, perimeter P, roundness C, and aspect ratio of the candidate target based on the contour, wherein the roundness calculation formula is C = 4πA / P. 2 .

[0023] A further embodiment involves a spectral characterization engine comprising a Transformer encoder network based on a self-attention mechanism. This Transformer encoder network includes a sequentially connected input embedding layer, a position encoding layer, a multi-layer encoder stack, and a classification output layer. The input embedding layer maps one-dimensional Raman spectral data into a high-dimensional feature vector. The position encoding layer superimposes position signals reflecting spectral wavenumber order information onto the high-dimensional feature vector. Each layer in the multi-layer encoder stack contains a multi-head self-attention module and a feedforward neural network. The multi-head self-attention module calculates long-range dependencies within the spectral sequence. The classification output layer outputs the probability distribution and confidence score of the polymer. The formula for calculating the attention matrix of the multi-head self-attention module is:

[0024]

[0025] Where Q, K, and V are the query matrix, key matrix, and value matrix, respectively, and d k This is the scaling factor.

[0026] A further aspect is that the spectral characterization engine also includes a spectral preprocessing module, which performs Savitzky-Golay smoothing filtering on the raw Raman spectrum before inputting it into the Transformer encoder network, and performs baseline correction using an adaptive iterative reweighted penalized least squares method.

[0027] A further solution is that the multimodal data fusion and decision module is used to maintain the affine transformation matrix between the image pixel coordinate system and the physical coordinate system of the precision motion control unit, and uses the nearest neighbor matching algorithm to map the center coordinates of the image target acquired at time t1 to the physical position command that triggers spectral acquisition at time t2.

[0028] A method for detecting microplastics, using the aforementioned dual-flow microplastic detection system, includes the following steps:

[0029] S1: Place the pre-treated sample on the precision motion control unit and start the panoramic scanning mode;

[0030] S2: The precision motion control unit moves along a preset path, and the visual imaging optical path in the confocal optocoupler module synchronously acquires the microscopic image stream of the sample and transmits the microscopic image stream to the visual characterization engine in real time.

[0031] S3: The visual representation engine uses an improved variant of the YOLO architecture to detect candidate targets in the image, outputs the bounding box coordinates of each target, and extracts the physical morphological parameters of each target.

[0032] S4: Generate a list of points to be measured based on the bounding box coordinates of each target obtained in S3; drive the precision motion control unit to move each target to the center of the laser spot in sequence according to the list of points to be measured;

[0033] S5: Excite the laser. The spectrometer collects the Raman spectrum of each target through the spectral acquisition optical path and transmits the Raman spectral data to the spectral characterization engine.

[0034] S6: The spectral characterization engine preprocesses the received raw spectral data by Savitzky-Golay filtering and AirPLS baseline correction, and then inputs it into the Transformer encoder network.

[0035] S7: The Transformer encoder network decodes the Raman spectrum and outputs the classification probability of the polymer type;

[0036] S8: The multimodal data fusion and decision module uniquely matches the physical morphology parameters of the same target with the polymer type, generating a test report that includes physical morphology and chemical composition.

[0037] S9: The intelligent report generation terminal outputs a multi-dimensional test report based on S8, including microplastic abundance, particle size distribution, shape proportion, and polymer type statistics.

[0038] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method steps for detecting microplastics.

[0039] The beneficial effects of this invention are as follows:

[0040] 1. This invention includes a confocal optocoupler module, which achieves coaxial alignment of microscopic imaging and Raman spectroscopy by constructing a shared optical axis. It employs a dual-stream processing architecture combining a visual characterization engine and a spectral characterization engine. The visual characterization engine utilizes an improved YOLOv8 network to quickly locate individual targets in the image of the sample and extracts the size and shape parameters of each target using an edge detection algorithm. The spectral characterization engine, based on a Transformer encoder, parses the Raman spectral sequences of each target and accurately identifies the polymer type using a self-attention mechanism. Finally, the system fuses the physical morphology and chemical composition data of each target through spatial coordinate mapping, solving the problem of fragmented morphology and composition analysis in existing microplastic detection technologies. This achieves efficient and accurate integrated detection of microplastics and provides a crucial chain of evidence for microplastic traceability.

[0041] 2. Compared with traditional full-scan Raman imaging, the "visual guidance-fixed-point spectroscopy" mode of the microplastic detection system in this invention only collects spectra at the locations where particles are present. That is, during the spectral acquisition process, each microplastic target is moved to the laser spot area for detection by a precision motion control unit. Only the microplastic target area is detected, and there is no need to detect the entire sample area, which greatly shortens the detection time. The detection time for a single sample can be reduced from tens of hours to 1-2 hours.

[0042] 3. The visual representation engine in this invention uses a deep convolutional neural network to locate the candidate target region of microplastics in real time and extract its bounding box coordinates; at the same time, it uses the CIoU loss function to optimize the bounding box regression, which not only considers the overlapping area, but also introduces the center point distance and aspect ratio penalty terms, which significantly improves the ability to recognize stacked particles in the image, making the network more accurate in regressing the detection box of slender fibrous microplastics.

[0043] 4. The spectral characterization engine in this invention adopts the Transformer architecture, which enables the model to focus on the fingerprint region of Raman characteristic peaks, significantly improving the ability to distinguish similar polymers and exhibiting excellent recognition accuracy; especially when processing Raman spectra against complex matrix backgrounds, its accuracy is significantly better than traditional machine learning methods. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the hardware composition of Embodiment 1 of the present invention;

[0045] Figure 2 This is a framework diagram of the microplastic detection system in this invention;

[0046] Figure 3 This is a flowchart of the microplastic detection method in this invention.

[0047] The components include: 1. Light source; 2. Displacement stage; 3. Lens; 4. Industrial camera; 5. Raman probe; 6. Laser; 7. Spectrometer; 8. Host computer; 9. Objective lens; 10. Dichroic mirror; 11. Visual representation engine; 110. Deep convolutional neural network; 111. Morphological parameter calculation unit; 12. Spectral representation engine; 120. Spectral preprocessing module; 121. Transformer encoder network; 13. Multimodal data fusion and decision-making module; 14. Intelligent report generation terminal; 15. Confocal optocoupler module; 16. Precision motion control unit. Detailed Implementation

[0048] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0049] In the description of this invention, it should be understood that the terms "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0050] Example 1

[0051] This embodiment provides a microplastic detection system, such as Figure 2As shown, the system mainly includes a confocal optocoupler module 15, a precision motion control unit 16, a visual characterization engine 11, a spectral characterization engine 12, a multimodal data fusion and decision-making module 13, and an intelligent report generation terminal 14. The confocal optocoupler module 15 includes a visual imaging optical path and a Raman spectroscopy acquisition optical path, and these two paths share a common optical axis within the same physical space, achieving coaxial alignment between microscopic imaging and Raman spectroscopy. The confocal optocoupler module 15 is used to acquire images of the sample under test and excite Raman scattering light from target particles in the sample. The precision motion control unit 16 carries the sample under test and, according to a preset scanning path or target coordinates, moves the sample under test in a plane perpendicular to the shared optical axis, performing a micrometer-level displacement to comprehensively scan the sample. The confocal optocoupler module 15 includes: an illumination source, an excitation source, a dichroic spectrometer, a microscope objective, an area array image sensor, and a spectrometer. The microscope objective and illumination source can be coaxial optical paths, and the precision motion control unit 16 can be located between the illumination source and the microscope objective. The illumination source illuminates the sample, causing it to emit reflected light. This reflected light passes through the microscope objective and is propagated through a dichroic beam splitter to the area array image sensor, forming a visual imaging optical path. The excitation source propagates through the dichroic beam splitter to the microscope objective and is focused onto the sample, exciting it to emit Raman scattered light. This Raman scattered light propagates through the microscope objective and the dichroic beam splitter to the spectrometer, forming a Raman spectroscopy acquisition optical path. The illumination source uses a multi-band illumination source, providing broadband illumination covering the visible light band. The laser source uses a monochromatic excitation source, emitting narrowband laser light of a specific wavelength.

[0052] This embodiment also provides the hardware components of a microplastic detection system, such as... Figure 1As shown, the confocal optocoupler module 15 comprises a light source 1, a displacement stage 2, a lens 3, an industrial camera 4, a Raman probe 5, a laser 6, a spectrometer 7, a host computer 8, an objective lens 9, a dichroic mirror 10, and the host computer 8. The displacement stage 2 (equivalent to a precision motion control unit 16) carries the filter membrane sample or glass slide sample to be tested. The light source 1 (equivalent to the aforementioned illumination source) is located below the displacement stage 2 and provides broadband illumination; the light source 1 can be a halogen source. The objective lens 9 (equivalent to the aforementioned microscope objective) is located above the displacement stage 2. The objective lens 9 is a high numerical aperture microscope objective, with switchable magnification, and is used to focus the illumination light and excitation light onto the same sample focal plane on the displacement stage 2 during the detection process. The dichroic mirror 10 is tilted at 45 degrees above the objective lens 9, and the lens 3 and the industrial camera 4 (equivalent to an area array image sensor) are located sequentially on one side of the dichroic mirror 10. Raman probe 5 is located directly above dichroic mirror 10, while laser 6 and spectrometer 7 are located on either side of dichroic mirror 10. The displacement stage 2 is electrically connected to the host computer 8 and is controlled by the host computer 8 to perform micrometer-level displacement in the horizontal plane. Spectrometer 7, laser 6, and industrial camera 4 are all electrically connected to the host computer 8. The host computer 8 controls the activation of laser 6; spectrometer 7 transmits the acquired spectral data to the host computer 8; and the industrial camera 4 also transmits the real-time acquired microscopic image stream to the host computer 8. This embodiment consists of two optical paths: a visual imaging optical path and a Raman spectral acquisition optical path. In the visual imaging optical path, light source 1 illuminates the sample to be tested on displacement stage 2. The sample generates reflected light, which is focused by objective lens 9 and propagates to dichroic mirror 10. The reflected light is then directionally reflected by high-performance dichroic mirror 10 to lens 3, and finally focused by lens 3 onto industrial camera 4, forming a high-resolution microscopic image stream. Industrial camera 4 acquires the microscopic image stream in real time and transmits it to host computer 8. In the Raman spectroscopy acquisition optical path, laser 6 emits a narrowband laser of a specific wavelength (e.g., 785nm). The laser beam propagates from Raman probe 5 to dichroic mirror 10, and after passing through dichroic mirror 10 and objective lens 9, it illuminates the sample on the displacement stage 2, causing the sample to emit Raman scattered light. The Raman scattered light passes sequentially through objective lens 9 and dichroic mirror 10, is collected by Raman probe 5, and propagates to spectrometer 7. Spectrometer 7 converts the light signal into spectral data and transmits it to host computer 8. Raman probe 5 is existing technology. Raman probe 5 is equipped with a beam splitter, which can reflect the laser emitted by laser 6 back to dichroic mirror 10, and can also receive the Raman scattered light reflected back from the sample and transmit it to spectrometer 7. The internal principle of Raman probe 5 will not be elaborated here. In addition, the dichroic mirror 10 in this embodiment is set on a shared optical axis, which couples the visual imaging optical path and the spectral acquisition optical path on the same optical axis. Furthermore, the spectral separation characteristics of the dichroic mirror 10 strictly follow the wavelength selective reflection equation, ensuring that the focal planes of the visual imaging optical path and the Raman spectral acquisition optical path are precisely coincident.This design ensures "what you see is what you measure," meaning that the object at the center of the industrial camera's field of view is within the laser spot's range, eliminating parallax errors.

[0053] The monochromatic excitation source in this invention can be any one of a 785nm, 532nm, or 1064nm frequency-stabilized laser. It is worth noting that the cutoff wavelength of the dichroic beam splitter in this invention needs to match the center wavelength of the monochromatic excitation source. In this embodiment, a 785nm laser is chosen as the excitation source, as this wavelength is outside the visible light fluorescence region, effectively suppressing fluorescence interference while retaining a sufficient Raman scattering cross-section.

[0054] In this embodiment, the host computer 8 integrates a visual representation engine 11, a spectral representation engine 12, a multimodal data fusion and decision-making module 13, and an intelligent report generation terminal 14. The visual representation engine 11 receives the microscopic image stream acquired by the visual imaging optical path, locates candidate target regions in the image in real time, extracts the bounding box coordinates of the candidate targets, and extracts the physical morphological feature vectors of the candidate targets using an edge detection algorithm. The spectral representation engine 12 receives the spectral data sequence acquired by the Raman spectroscopy acquisition optical path for the candidate targets, analyzes the molecular fingerprint features, and outputs the classification probability distribution of the polymer type. The multimodal data fusion and decision-making module 13 correlates and matches the physical morphological feature vectors with the classification probability distribution through a spatial coordinate mapping matrix, generating a composite feature fingerprint containing the physical morphological features and chemical composition of a single particle. The intelligent report generation terminal 14 outputs a multidimensional statistical report on contamination based on the composite feature fingerprint.

[0055] In this embodiment, a microscopic image stream of the sample under test is acquired via a visual acquisition optical path. This image stream is transmitted by an industrial camera to a visual characterization engine 11 in a host computer 8. The visual characterization engine 11 receives the image stream, identifies candidate target regions (i.e., particles in the image), extracts the bounding box coordinates of each target, and calculates the physical dimensions of each target. Then, multiple bounding box coordinates are used to generate a list of points to be detected. The host computer 8 drives a displacement stage 2 to move each target sequentially to the center of the laser excitation spot. The host computer 8 controls a laser 6 to emit a laser beam, which is then irradiated onto the sample under test via a spectral acquisition optical path, exciting Raman scattering light. This Raman scattering light is then transmitted to a spectrometer 7 via the spectral acquisition optical path. The spectrometer 7 converts the light signal and transmits it to a spectral characterization engine 12. The spectral characterization engine 12 decodes the Raman spectrum and outputs the polymer type. Finally, a multimodal data fusion and decision module correlates and combines the data from the visual characterization engine 11 and the spectral characterization engine 12 to generate a detection report.

[0056] The visual representation engine 11 includes a deep convolutional neural network 110, which receives the microscopic image stream acquired by the visual imaging optical path, locates candidate target regions in the image in real time, and extracts their bounding box coordinates. The visual representation engine 11 also includes a morphological parameter calculation unit 111, which combines edge detection algorithms to extract the physical morphological feature vectors of candidate targets. The deep convolutional neural network 110 adopts an improved variant of the YOLO architecture, specifically an improved variant of YOLOv8. Its backbone feature extraction network uses a CSPDarknet structure for multi-scale feature extraction; the feature fusion network (Neck) uses a path aggregation network (PANet) structure to enhance feature transfer for small targets; the decoupled detection head separates target confidence prediction from bounding box regression tasks; and the bounding box regression is trained and optimized using the CIoU (Complete Intersection over Union) loss function, where the loss function is defined as: (In the formula, IoU is the intersection-union ratio, and ρ is the ratio of the predicted box center b to the ground truth box center b) gt The Euclidean distance is given by c, where c is the diagonal length of the minimum bounding rectangle, α is the weighting coefficient, and v is the aspect ratio consistency parameter. This embodiment uses the CIoU loss function to optimize the bounding box regression. This CIoU loss function not only considers the overlapping area but also introduces center point distance and aspect ratio penalty terms, significantly improving the ability to identify stacked particles and making the network more accurate in regressing the detection boxes of slender fibrous microplastics.

[0057] The visual representation engine 11 also includes a morphological parameter calculation unit 111, specifically used to perform the following operations within the candidate target region located by the deep convolutional neural network 110: performing contrast-limited adaptive histogram equalization (CLAHE) on the region of interest (ROI); applying Gaussian filtering to remove high-frequency noise; extracting sub-pixel-level edge contours using a dual-threshold Canny operator; and calculating the area, perimeter, roundness, and aspect ratio of the target particle based on the contour. When calculating the actual size of the target particle based on the contour point set, the pixel area is converted to physical size using an optical resolution model; the roundness parameter is derived through the geometric relationship between perimeter and area (the roundness calculation formula is C = 4πA / P). 2 The aspect ratio is obtained through eigenvalue decomposition of the minimum bounding rectangle. The final output includes a morphological report containing size classification statistics and the proportion of fiber fragment particle shapes, thus solving the problem that traditional rectangular frames cannot describe the specific shape of particles.

[0058] The spectral characterization engine 12 in this embodiment includes a spectral preprocessing module 120 and a Transformer encoder network 121 based on a self-attention mechanism. The spectrometer 7 first transmits the Raman spectrum to the spectral preprocessing module 120, which performs Savitzky-Golay smoothing filtering on the received raw Raman spectrum to eliminate high-frequency noise and uses Adaptive Iterative Reweighted Penalized Least Squares (AirPLS) for baseline correction to remove background fluorescence. The preprocessed Raman spectrum is then input to the Transformer encoder. The Transformer encoder network 121 establishes a multi-layer feature parsing channel, specifically including an input embedding layer, a positional encoding layer, a multi-layer encoder stack, and a classification output layer connected in sequence. The input embedding layer maps the one-dimensional Raman spectral data into a high-dimensional feature vector. The positional encoding layer superimposes positional signals reflecting the spectral wavenumber order information onto the feature vector. In this embodiment, by preserving wavenumber information through positional encoding and combining it with AirPLS baseline correction preprocessing, the model can effectively extract fingerprint features from a noisy background. In addition, each layer in the multi-layer encoder stack contains a multi-head self-attention module and a feed-forward network; the multi-head self-attention module is used to calculate the long-range dependencies within the spectral sequence, and its attention matrix is ​​calculated using the following formula: (where Q, K, and V are the query matrix, key matrix, and value matrix, respectively, and d) k (This is a scaling factor.) The multi-head self-attention mechanism in the multi-layer encoder stack can capture the correlation between feature peaks that are far apart in the spectrum (e.g., polyethylene at 1064 cm⁻¹). -1 With 1295cm -1 (Peak intensity ratio at the specified location). This global contextual understanding capability is particularly important for identifying aging and degrading microplastics. Furthermore, the multi-head self-attention mechanism focuses the model on the fingerprint region of Raman feature peaks, significantly improving the ability to distinguish similar polymers. The feedforward neural network enhances nonlinear expressive power through dimensionality expansion. Finally, the classification output module outputs the probability distribution and confidence score of the polymer. The classification output module flattens the spectral sequence features and processes them sequentially through two linear layers. Between the two linear layers, the ReLU activation function and Dropout operation are executed sequentially, ultimately outputting the polymer probability distribution, and taking the maximum probability as the classification result.

[0059] The multimodal data fusion and decision module 13 in this invention maintains the affine transformation matrix between the image pixel coordinate system (u, v) and the physical coordinate system (X, Y) of the precision motion control unit, and uses a nearest neighbor matching algorithm to map the center coordinates of the image target acquired at time t1 to the physical position command that triggers spectral acquisition at time t2, ensuring the homology between physical morphology data and chemical composition data. In this embodiment, by establishing the affine transformation relationship between the physical coordinate system of the precision motion control unit and the image pixel coordinate system, the system can automatically match the morphology dataset acquired by the visual representation engine 11 with the chemical identifier acquired by the spectral representation engine 12 through coordinate indexing to form a complete physical and chemical profile of a single particle. The resulting detection report is divided into two dimensions: physical morphology and chemical composition. The physical report presents a histogram of size distribution and the proportion of shape composition, while the chemical report is classified by polymer type and labeled with the characteristic peak matching degree. Finally, the intelligent report generation terminal 14 outputs a multidimensional detection report containing microplastic abundance, particle size distribution, shape proportion, and polymer type statistics based on the above detection report.

[0060] Example 2

[0061] like Figure 3 As shown, this embodiment provides a microplastic detection method, which is executed using the microplastic detection system described in the above embodiment. The specific detection steps are as follows:

[0062] S1: Place the pre-treated sample to be tested on the precision motion control unit 16 and start the panoramic scanning mode;

[0063] S2: The precision motion control unit 16 moves along a preset path (such as a serpentine scan), and the visual imaging optical path in the confocal optocoupler module 15 synchronously acquires the microscopic image stream of the sample and transmits the microscopic image stream to the visual characterization engine 11 in real time.

[0064] S3: The visual representation engine 11 uses an improved YOLOv8 network to detect candidate targets in the image and outputs the bounding box coordinates of each target. For the target region, contrast-limited adaptive histogram equalization (CLAHE) is performed, and the edge contour is extracted using a dual-threshold Canny operator. Based on the contour set, the morphological parameters such as the area, roundness, and aspect ratio of the target particles are calculated.

[0065] S4: Generate a list of points to be measured based on the bounding box coordinates of each target obtained in S3, and drive the displacement stage 2 to move each target particle to the center of the laser spot in sequence.

[0066] S5: Start the laser 6 to emit a laser, and the spectrometer 7 collects the Raman spectrum of each target through the spectral acquisition optical path and transmits the Raman spectral data to the spectral characterization engine 12;

[0067] S6: The spectral characterization engine 12 inputs the received raw spectral data into the Transformer encoder network 121 after preprocessing by Savitzky-Golay filtering and AirPLS baseline correction.

[0068] S7: The Transformer encoder network 121 uses a multi-head self-attention mechanism to capture the long-range feature dependence of Raman spectra and decodes the Raman spectra to output the classification probability of polymer types (such as polyethylene, polypropylene, etc.).

[0069] S8: The multimodal data fusion and decision module uses the affine transformation matrix between the physical coordinates of the precision motion control unit 16 and the image pixel coordinates to uniquely match the physical morphological parameters (data obtained by the visual representation engine 11) and chemical composition results (data obtained by the spectral representation engine 12) of the same particle, and generates a detection report containing physical morphology and chemical composition.

[0070] S9: Intelligent Report Generation Terminal 14 outputs a multi-dimensional test report based on S8, including microplastic abundance, particle size distribution, shape proportion, and polymer type statistics.

[0071] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method steps for detecting microplastics.

[0072] Although the above embodiments have been shown and described, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Any changes, modifications, substitutions and variations made to the above embodiments by those skilled in the art are within the protection scope of the present invention.

Claims

1. A microplastic detection system, characterized in that, include: The confocal optocoupler module includes a visual imaging optical path and a Raman spectroscopy acquisition optical path, and the visual imaging optical path and the Raman spectroscopy acquisition optical path are constructed in the same physical space with a shared optical axis that can be confocal aligned. A precision motion control unit is used to carry the sample to be tested and to perform micron-level displacement in a plane perpendicular to the shared optical axis according to a preset scanning path or target coordinates. The visual representation engine is used to receive the microscopic image stream acquired by the visual imaging optical path, locate the candidate target region in the image in real time, extract the bounding box coordinates of the candidate target, and extract the physical morphological feature vector of the candidate target by combining the edge detection algorithm. The spectral characterization engine is used to receive the spectral data sequence acquired by the Raman spectroscopy acquisition optical path for the candidate target, analyze the molecular fingerprint features, and output the classification probability distribution of the polymer type. The multimodal data fusion and decision module associates and matches the physical morphology feature vector with the classification probability distribution through a spatial coordinate mapping matrix to generate a composite feature fingerprint containing the physical morphology features and chemical composition of a single particle. The intelligent report generation terminal outputs a multi-dimensional statistical report on contamination based on the composite feature fingerprint.

2. The microplastics detection system of claim 1, wherein, The confocal optocoupler module includes: an illumination source, an excitation source, a dichroic beam splitter, a microscope objective, an area array image sensor, and a spectrometer; The illumination source is used to illuminate the object to be tested, so that the sample emits reflected light. The reflected light passes through the microscope objective and is transmitted to the area array image sensor through the dichroic beam splitter to form a visual imaging optical path. The excitation light source is transmitted to the microscope objective via the dichroic beam splitter and focused onto the sample to be tested, exciting the sample to emit Raman scattered light; the Raman scattered light is transmitted to the spectrometer via the microscope objective and the dichroic beam splitter, forming the Raman spectral acquisition optical path.

3. The microplastics detection system of claim 2, wherein, The excitation source is selected from one of the following: a 785nm, 532nm, or 1064nm frequency-stabilized laser; the cutoff wavelength of the dichroic beam splitter is matched with the center wavelength of the excitation source.

4. The microplastics detection system of claim 1, wherein, The visual representation engine includes a deep convolutional neural network for locating candidate target regions and extracting candidate target bounding boxes. This deep convolutional neural network employs an improved variant of the YOLO architecture, including: The backbone feature extraction network uses the CSPDarknet structure for multi-scale feature extraction. The feature fusion network employs a path aggregation network structure to enhance feature transfer for small targets. The detection head is decoupled, separating the target confidence prediction from the bounding box regression task; the bounding box regression is trained and optimized using the CIoU loss function, which is defined as: In the formula, IoU is the intersection-union ratio, and ρ is the distance between the center b of the predicted bounding box and the center b of the ground truth bounding box. gt The Euclidean distance is given by c, where c is the diagonal length of the minimum bounding rectangle, α is the weighting coefficient, and v is the aspect ratio consistency parameter.

5. The microplastics detection system of claim 4, wherein, The visual representation engine also includes a morphological parameter calculation unit, which performs the following operations within the candidate target region located by the deep convolutional neural network: performs contrast-limited adaptive histogram equalization on the region of interest (ROI) of the image; and applies Gaussian filtering to remove high-frequency noise. Edge contours are extracted using a dual-threshold Canny operator; based on the contours, the area A, perimeter P, circularity C, and aspect ratio of the candidate targets are calculated, where the circularity calculation formula is C = 4πA / P. 2 .

6. The microplastics detection system of claim 1, wherein, The spectral characterization engine includes a Transformer encoder network based on a self-attention mechanism. This Transformer encoder network comprises an input embedding layer, a position encoding layer, a multi-layer encoder stack, and a classification output layer connected in sequence. The input embedding layer maps one-dimensional Raman spectral data into a high-dimensional feature vector. The position encoding layer superimposes position signals reflecting spectral wavenumber order information onto the high-dimensional feature vector. Each layer in the multi-layer encoder stack contains a multi-head self-attention module and a feedforward neural network. The multi-head self-attention module calculates long-range dependencies within the spectral sequence. The classification output layer outputs the probability distribution and confidence score of the polymer. The formula for calculating the attention matrix of the multi-head self-attention module is: wherein Q, K, V are query matrix, key matrix and value matrix respectively, d k is a scaling factor.

7. The microplastic detection system of claim 6, wherein, The spectral characterization engine also includes a spectral preprocessing module. Before inputting the raw Raman spectrum into the Transformer encoder network, it first performs Savitzky-Golay smoothing filtering on the spectrum and then performs baseline correction using an adaptive iterative reweighted penalized least squares method.

8. The microplastic detection system of claim 1, wherein, The multimodal data fusion and decision module is used to maintain the affine transformation matrix between the image pixel coordinate system and the physical coordinate system of the precision motion control unit, and uses the nearest neighbor matching algorithm to map the center coordinates of the image target acquired at time t1 to the physical position command that triggers spectral acquisition at time t2.

9. A microplastic detection method, using the microplastic detection system according to any one of claims 1-8, characterized in that, Includes the following steps: S1: Place the pre-treated sample on the precision motion control unit and start the panoramic scanning mode; S2: The precision motion control unit moves along a preset path, and the visual imaging optical path in the confocal optocoupler module synchronously acquires the microscopic image stream of the sample and transmits the microscopic image stream to the visual characterization engine in real time. S3: The visual representation engine uses an improved variant of the YOLO architecture to detect candidate targets in the image, outputs the bounding box coordinates of each target, and extracts the physical morphological parameters of each target. S4: Generate a list of points to be measured based on the bounding box coordinates of each target obtained in S3; drive the precision motion control unit to move each target to the center of the laser spot in sequence according to the list of points to be measured; S5: Excite the laser. The spectrometer collects the Raman spectrum of each target through the spectral acquisition optical path and transmits the Raman spectral data to the spectral characterization engine. S6: The spectral characterization engine preprocesses the received raw spectral data by Savitzky-Golay filtering and AirPLS baseline correction, and then inputs it into the Transformer encoder network. S7: The Transformer encoder network decodes the Raman spectrum and outputs the classification probability of the polymer type; S8: The multimodal data fusion and decision module uniquely matches the physical morphology parameters of the same target with the polymer type, generating a test report that includes physical morphology and chemical composition. S9: The intelligent report generation terminal outputs a multi-dimensional test report based on S8, including microplastic abundance, particle size distribution, shape proportion, and polymer type statistics.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in claim 9.

Images