Separation, identification and detection method for micro-plastics in sediments
By combining the optimized density separation method and convolutional neural network, the problems of low extraction rate, high cost and low efficiency in the microplastic separation and recognition process are solved, and high accuracy and high efficiency microplastic detection is achieved.
Patent Information
- Application Number
- CN202510292799.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-12
- Publication Date
- 2025-06-24
AI Technical Summary
The prior art has problems such as low extraction rate, high cost, environmental pollution and excessive time-consuming in the separation, identification and detection of microplastics. The traditional machine learning methods have different effects and the evaluation results are not accurate enough.
The method based on optimized density separation and convolutional neural network is used to separate, identify and detect microplastics in the sediment. Specific steps include sediment acquisition and processing, surface-enhanced Raman spectral acquisition of microplastics, standard spectral pre-processing, creation of Raman spectral databases, and establishment and training convolutional neural network models.
High accuracy identification and quantitative analysis of microplastics is achieved, the time and cost of the detection process is reduced, the detection efficiency and accuracy are improved, and a high accuracy detection of a variety of microplastics can be detected in different water environments.
Smart Images
Figure CN120195145A_ABST
Abstract
Description
Technical Field
[0001] The present invention discloses a method for separating, identifying and detecting microplastics in sediments, belonging to the technical field of machine learning. Background Art
[0002] Microplastics generally refer to plastic particles with a diameter less than 5 millimeters. As pollutants affecting the environment and human health, they have received increasing attention in recent years. As a new type of pollutant, microplastics have been found in different sediment environments and have an impact on human health through the food chain. Developing a rapid, simple and accurate detection method is of great significance for the treatment and traceability of microplastics.
[0003] Currently, the commonly used methods for separating microplastics include density separation method, physical screening method, electrostatic separation method, magnetic separation method, oil extraction method, etc. These methods generally have some deficiencies, such as low extraction rate, high cost, environmental pollution, long time consumption, etc. The density separation method is to extract microplastics in sediments by using the density difference between sediment samples and microplastics. Sodium chloride (NaCl) solution, as the most commonly used microplastic extractant, has the advantages of low cost and no toxicity.
[0004] Currently, the commonly used methods for digesting microplastics include oxidation method, acid digestion, alkali digestion, etc. Hydrogen peroxide (H2O2) is the most commonly used oxidant for removing organic matter, with the advantages of low cost and little pollution after decomposition. However, in the actual operation process, the stirring time, standing time, digestion times and time will all affect the separation effect, and further optimization and discussion are needed for the extraction of microplastics using the density separation method and H2O2 digestion.
[0005] Currently, the common methods for detecting microplastics include visual identification method, mass spectrometry, vibrational spectroscopy, etc. Among them, Raman spectroscopy in vibrational spectroscopy has the advantages of simple, efficient and repeatable, and is a relatively reliable and accurate detection method. Raman spectroscopy is a scattering spectroscopy, and the information comes from the frequency difference between the incident light and the scattered light. It was first discovered by Chandrasekhara Venkata Raman in 1928. Through spectral analysis, the vibration and rotation information of molecules can be obtained to determine the chemical structure of materials. Raman spectroscopy has been used to detect from 800 cm -1 to 2900 cm -1Characteristic peaks of microplastics within a certain range (e.g., C-O stretching, C=O stretching, and CH2 bending) are used to determine the types of microplastics contained in the sediment. The spectral analysis method relies on the spectral peak characteristics of the extracted substances. In the traditional identification process, it is usually necessary to compare the spectral characteristics of substances in the database with those of the measured substances one by one. Due to the large variety of plastic materials and complex chemical compositions, and the large amount of spectral data collected, the one-to-one comparison method will be very time-consuming and produce large recognition errors. Introducing machine learning technology is the key to solving the challenges of microplastic detection. Compared with traditional methods, machine learning can not only improve the accuracy and efficiency of detection, but also detect and analyze microplastic particles more accurately, overcoming the limitations of traditional methods. Summary of the Invention
[0006] Currently, commonly used machine learning methods include random forest, logistic regression, support vector machine, K-nearest neighbor algorithm, decision tree, etc. Due to other reasons such as data algorithms, the effects of different machine learning methods in the identification of microplastics vary.
[0007] Convolutional neural network (CNN) can effectively learn important features in Raman signals from spectral data and establish a correlation with the target variable. In the present invention, CNN is used to identify microplastics in sediment.
[0008] Aiming at microplastics in the environment, for the technical problems of low single flotation efficiency of the traditional density separation method and the inability to effectively remove invisible organic matter to the naked eye; for the problems of inconsistent effects of traditional machine learning methods and inaccurate evaluation results, the present invention proposes a method for separating, identifying and detecting microplastics in sediment based on optimized density separation and convolutional neural network.
[0009] The technical solution of the present invention is as follows:
[0010] A method for separating, identifying and detecting microplastics in sediment, comprising the following steps:
[0011] (1) Sediment collection
[0012] (2) Sediment treatment and microplastic separation
[0013] The collected sediment is dried, then sieved, subjected to two density flotations and organic matter digestion, and then the upper clarified liquid is taken and filtered onto a glass fiber filter membrane by vacuum-assisted filtration;
[0014] (3) Acquisition of surface-enhanced Raman spectrum of microplastics
[0015] By adding gold nanoparticles, stronger Raman characteristic peaks of microplastics are obtained;
[0016] (4) Pretreatment of standard spectrum
[0017] Collect Raman spectral data and perform three steps of processing on it: baseline correction, smoothing filtering, and normalization;
[0018] (5) Creation of a sediment microplastic Raman spectral database
[0019] Construct and train a convolutional neural network model by establishing a standardized Raman spectral database for checking microplastics in test samples;
[0020] Since there are mainly three types of interferences in the Raman spectra of microplastics in sediments, namely fluorescence, high-frequency noise signals, and the overlap of characteristic peaks of different microplastics, the adopted strategy is to collect multiple surface-enhanced Raman scattering (SERS) spectra of specific regions of the sample, first preprocess these spectra, and finally take the average value of each Raman shift of these spectra to obtain the average mapping spectrum (AMS). The AMS can reduce the peak variation of SERS;
[0021] (6) Establish a convolutional neural network-based sediment microplastic analysis model;
[0022] (7) Convolutional neural network model training
[0023] Train the convolutional neural network using the obtained database;
[0024] (8) Comparison of the effects of the convolutional neural network model
[0025] Collect the Raman spectra of the microplastics on the glass fiber membrane in step (2), input them into the convolutional neural network model in step (6) to obtain the predicted labels, and perform operations on the predicted labels to obtain the predicted labels of the distribution of microplastics in the sediment.
[0026] Preferably, the specific steps of the above step (2) are as follows:
[0027] (2-1) After drying the sediment, sieve it through a wire mesh with a pore size of 5 mm;
[0028] (2-2) Perform two density flotations and organic matter digestion;
[0029] For the first density flotation and organic matter digestion, add 30 mL of 35% H2O2 and 30 mL of 0.11 g / mL NaCl solution to the sediment, stir for 1 hour, precipitate for 3 hours, and then take the upper clear liquid; break the electrostatic adsorption between microplastics and other particles (inorganic mineral particles, organic matter particles, metal oxide particles, etc.) by stirring;
[0030] Second-density flotation and organic matter digestion. Add 30 mL of 35% H2O2 and an excessive amount of NaCl solution to the supernatant, stir for 1 hour, settle for 3 hours, then take the upper clear liquid and filter it onto a glass fiber filter membrane using the vacuum-assisted filtration method.
[0031] Preferably, the specific steps of the baseline correction, smoothing filtering, and normalization in the above step (3) are as follows:
[0032] (3-1) Baseline correction
[0033] Perform baseline correction using the adaptive iterative reweighted penalized least squares model, and update the weights of each point according to the difference between the baseline value and the original signal fitted by the previous loop, expressed as:
[0034]
[0035] In the formula, x is the original signal, z is the fitted baseline, and d t is the sum of the absolute values less than the variable in the t-th iteration, and x represents the eigenvalue of the i-th sample, i and represents the reference value of the i-th sample in the previous round (round t-1);
[0036] (3-2) Smoothing filtering
[0037] The invention uses a Savitzky-Golay filter to smooth the noise spectrum. The Savitzky-Golay filter performs a least squares fit on each window of the data through a polynomial to obtain the smoothing coefficient h corresponding to each point in the window i , expressed as:
[0038]
[0039] In the formula, w is the window width, and x k represents the smoothed value at position k, and
[0040] (3-3) Normalization
[0041] To perform a linear transformation on the original data, normalize the data to obtain scaled data in the range (0, 1). The conversion formula is as follows:
[0042]
[0043] Among them, x, x max , and x min are the original spectral data, the maximum value in the spectral data, and the minimum value in the spectral data, respectively.
[0044] Preferably, the specific steps of step (5) above are as follows:
[0045] Use a one-dimensional convolutional neural network model, including two convolutional layers and three fully connected layers, improve the used convolutional neural network model, remove the max pooling layer, increase the convolutional kernel and convolutional stride, so that there are fewer operations and the training time is reduced;
[0046] The output of the output layer is classified using the Sigmoid function, the activation function of other layers uses the rectified linear unit, and the binary cross-entropy loss function is used as the objective function. To prevent overfitting of the model, an L2 regularization term is added. Therefore, the objective equation of the convolutional neural network is expressed as:
[0047]
[0048] In the formula, n is the total number of sample trainings, y i is the actual label of the sample, is the predicted label of the sample, λ is the regularization coefficient, and ω is the weight to be regularized.
[0049] Preferably, the specific steps of step (6) above are as follows:
[0050] Divide the database into 1024 training sets and 256 validation sets in a ratio of 8:2; set the learning rate, batch size, maximum number of iterations, and L2 regularization weight to 0.002, 126, 1000, and 0.006 respectively; in order to obtain the best state of the model and prevent overfitting, early stopping is adopted, and the validation set accuracy of the validation set is used as the termination judgment condition.
[0051] The beneficial effects of the present invention are:
[0052] Using the optimized CNN algorithm proposed by the present invention, it has higher accuracy in microplastic identification compared with other machine learning algorithms. Machine learning mainly focuses on distinguishing differences between features. A wider effective detection range means more differences in data features. To verify the superiority of CNN in the quantitative analysis of PE, different machine learning methods such as random forest, logistic regression, support vector machine, K-nearest neighbor algorithm, decision tree, etc. were used for comparison. The results show that the determination coefficient of CNN is higher than that of other machine learning models, and the root mean square error is lower than that of other models, proving its advantage in the quantitative analysis of microplastics. The determination coefficient of CNN for identifying the concentration of PE solutions of different sizes in different water environments reaches 0.9972, and the root mean square error is as low as 0.33, showing a significant advantage compared with other machine learning models. The proposed method can achieve high-precision and wide-concentration-range detection of the concentration of microplastic solutions without manual judgment.
[0053] Traditional Raman quantitative analysis is restricted by feature selection. In traditional methods, features related to the target variable need to be manually selected and extracted. This process may require prior knowledge and experience. However, Raman spectral data usually contains many features, some of which may contribute little to quantitative analysis. Therefore, methods relying on manual feature selection may be limited by the representation of a limited number of features, resulting in reduced analysis accuracy. Conventional methods may also be affected by interference and noise. Raman spectral data may be interfered by signals from various sources, such as background noise and scattering noise. These interferences may obscure or confuse the true features of the sample, leading to increased errors in quantitative analysis.
[0054] CNN can effectively learn the important features in Raman signals from spectral data and establish correlations with the target variable. By leveraging the power of CNN, quantitative analysis using Raman spectra can achieve more accurate prediction and estimation of target variables (such as sample concentration or quality). The accuracy and stability are superior to other machine learning algorithms, and the complexity of the model can be adjusted according to the quantity and complexity of the data. It is completely different from the traditional manual processing of spectral data and can effectively save the time of SERS spectral analysis.
[0055] This method can also be applied to the water environment. Different types of microplastics (PET, PP, PS, PVC, PE, and PC) are dispersed in different water environments (pure water, tap water, rainwater, lake water, and seawater), and the accuracy is also high. Description of the Drawings
[0056] Figure 1 is the flow chart of sediment treatment and microplastic separation in step 2 of the present invention;
[0057] Figure 2 is the SEM image of the blank filter paper and the filter paper doped with gold nanoparticles of microplastic particles;
[0058] Figure 3 is the schematic diagram of data acquisition and database establishment;
[0059] Figure 4 is the effect diagram of baseline correction and smoothing filtering processing;
[0060] Figure 5 is the schematic diagram of noise and fluorescence;
[0061] Figure 6 is the schematic diagram of the one-dimensional CNN architecture;
[0062] Figure 7 is the comparison of the effects of different machine learning models;
[0063] Figure 8 is the schematic diagram of the overall process. Detailed Implementation Modes
[0064] Those skilled in the art will understand that the following embodiments are only used to illustrate the present invention and should not be construed as limiting the scope of the present invention. For those embodiments where specific technologies or conditions are not indicated, the technologies or conditions described in the literature in this field or according to the product specifications are followed. For reagents or instruments whose manufacturers are not indicated, they are all conventional products that can be obtained through market purchases.
[0065] Embodiment:
[0066] A method for separating, identifying and detecting microplastics in sediments, comprising the following steps:
[0067] 1. Sediment collection: All sediments used in the experiment were from Changdang Lake. A total of 10 representative sampling points were selected around the lake. At each sampling point, 1 dm 3 of lake sediments were collected. During the sampling process, a spade was used to collect the sediments, and glass bottles were used to store the sediments, avoiding the use of plastic products throughout the process to interfere with the experimental results.
[0068] 2. Sediment treatment and microplastic separation: As Figure 1 shown, after the sediments are dried, they are sieved through a wire mesh with a pore size of 5 mm to remove larger impurities such as stones, grass roots, and fiber impurities. In order to achieve the best separation effect, two density flotation and organic matter digestion are carried out. For the first density flotation and organic matter digestion, 30 mL of 35% H2O2 and 30 mL of 0.11 g / mL sodium chloride solution are added to the sediments, stirred for 1 hour, precipitated for 3 hours, and then the upper clear liquid is taken. To prevent small inorganic particles from floating together with the organic matter in the sediments, an unsaturated NaCl solution is used for the first time. The electrostatic adsorption between microplastics and other particles (inorganic mineral particles, organic matter particles, metal oxide particles, etc.) is broken by stirring. For the second density flotation and organic matter digestion, 30 mL of 35% H2O2 and an excessive amount of NaCl solution are added to the supernatant, stirred for 1 hour, and settled for 3 hours. Then the upper clear liquid is taken and filtered onto a glass fiber filter membrane by vacuum-assisted filtration.
[0069] 3. Obtaining the surface-enhanced Raman spectrum of microplastics
[0070] By adding gold nanoparticles, stronger Raman characteristic peaks of microplastics can be obtained;
[0071] In order to obtain stronger Raman characteristic peaks of microplastics, gold nanoparticles (AuNPs) need to be added to the sample to be tested. The present invention uses a silver ion-assisted seed-mediated method to prepare AuNPs. The following are the detailed steps for preparing the gold colloid solution.
[0072] (1) Dissolve chloroauric acid trihydrate (HAuCl4·3H2O) powder in ultrapure water to prepare a 2.5 mmol / L chloroauric acid solution.
[0073] (2) Prepare 10 g / L sodium citrate solution.
[0074] (3) Heat the HAuCl4·3H2O solution to boiling and immediately add 1.5 mL of sodium citrate solution.
[0075] (4) Continue heating until the color changes to transparent wine red with obvious Tyndall effect.
[0076] (5) Stop heating and stabilize for 30 min. At this point, AuNPs are synthesized.
[0077] In order to enhance the intensity of the SERS characteristic peak, 3 mL of gold colloid solution was dripped onto the filter membrane to allow AuNPs to adhere to the surface of the microplastic particles. The SERS signal enhancement of microplastics is achieved by utilizing the "hot spot" effect between gold nanoparticles. It should be noted that gold nanoparticles will fail when encountering ions (such as salt in seawater and acidic rainwater). Therefore, between filtering the microplastic solution and the gold colloid solution, it is necessary to wash twice with pure water to ensure the enhancement effect. Then, put the filter membrane into a microwave oven to dry. After drying, the SERS spectrum of the microplastics can be obtained by detecting the surface of the filter membrane with a micro-Raman system.
[0078] Figure 2 Scanning electron microscope (SEM) images of blank filter paper and gold nanoparticle-doped filter paper containing microplastic particles are shown. It can be clearly seen that the microplastic particles are captured by the fiber pores of the filter paper. The pictures show the SEM images of the filter paper, microplastic particles on the filter paper substrate, and enlarged AuNPs-doped filter paper substrate with microplastic particles.
[0079] The enhancement factor is one of the most important parameters to characterize the performance of SERS substrates and can be calculated by the following formula:
[0080]
[0081] Among them, I blank and I SERS are the integrated intensities of the normal Raman and SERS signals of the analyte solution at the characteristic peaks. blank is the concentration of the analyte solution under normal Raman signal, C SERS is the concentration of the analyte solution used in SERS.
[0082] 4. Standard spectral preprocessing
[0083] After Raman spectroscopy data is collected, it is necessary to perform baseline correction, smoothing filtering and other processing. Figure 3It shows the data acquisition and database establishment process. The preparation of test samples and the parameter settings of the Raman spectroscopy measurement system will affect the quality of the spectrum. Processing the spectral data is very helpful for feature extraction in subsequent algorithms. It mainly includes three steps: baseline correction, smoothing filtering, and normalization.
[0084] (a) Baseline correction
[0085] There is always fluorescence background interference in Raman measurement, which will cause the baseline drift of the obtained Raman spectrum and seriously affect the effectiveness and accuracy of spectral analysis. The present invention uses the adaptive iteratively reweighted penalized least squares (airPLS) model for baseline correction. airPLS is an iterative weighted method based on error. The weights of each point are updated according to the difference between the baseline value and the difference between the original signal fitted in the previous loop, expressed as:
[0086]
[0087] In the formula, x is the original signal, z is the fitted baseline, and d t is the sum of the absolute values less than the variable x i -z i in the t-th iteration, x i represents the eigenvalue of the i-th sample, represents the reference value of the i-th sample in the previous round (round t-1);
[0088] (b) Smoothing filtering
[0089] There are many high-frequency noise signals in the initial Raman spectrum, which will interfere with the algorithm to extract the main features of the spectrum. The present invention uses the Savitzky-Golay (SG) filter to smooth the noise spectrum. The SG filter performs least squares fitting on each window of the data through polynomials to obtain the smoothing coefficient h i for each point in the window, expressed as:
[0090]
[0091] In the formula, w is the window width, x k represents the smoothed value at position k, is the normalization factor.
[0092] (c) Normalization
[0093] In order to perform a linear transformation on the original data, the present invention normalizes the data to obtain scaled data in the range of (0, 1). The conversion formula is as follows:
[0094]
[0095] where x, x max , and x min are the original spectral data, the maximum value in the spectral data, and the minimum value in the spectral data, respectively.
[0096] The min-max normalization process performs a central shift transformation and dimensionless compression on the Raman spectral data, which can effectively eliminate the influence of laser intensity, focusing effect, and exposure time on the Raman signal intensity.
[0097] Figure 4 shows the SERS spectral results of PET under different treatment methods as an example.
[0098] 5. Creation of the sediment microplastic Raman spectral database: The CNN model is constructed and trained by establishing a standardized Raman spectral database for inspecting microplastics in sediments. To ensure authenticity and diversity, before establishing the standardized Raman spectral database, the present invention considered the possible interferences during the sample detection process. The main interference factors are the organic matters and other impurities extracted from lake sediments. There are mainly three interferences in the Raman spectra of microplastics in sediments, namely fluorescence, high-frequency noise signals, and the overlap of characteristic peaks of different microplastics, as Figure 5 shown.
[0099] The present invention innovatively proposes a solution to the difference in the intensity of Raman characteristic peaks in the microplastic solution:
[0100] Taking PE as an example, PE particles and AuNPs are randomly aggregated and distributed on the surface of the glass fiber filter membrane, resulting in different Raman characteristic peak values collected at different points on the sample. For example, when filtering 100 mL of a PE solution with a concentration of 100 ppm, 10×10 points are selected at different positions on the substrate for SERS spectral measurement.
[0101] To reduce the difference in the intensity of Raman characteristic peaks of the same concentration PE solution, the strategy adopted by the present invention is to collect multiple SERS spectra of a specific area of the sample, preprocess these spectra first, and finally take the average value of each Raman shift of these spectra to obtain AMS. AMS can reduce the SERS peak variation caused by factors such as uneven sample filtration.
[0102] Mix 6 types of microplastics and organic matters into different combinations (a total of 2 7 -1 = 127), and collect 10 groups of Raman spectra for each combination (a total of 1280 groups of Raman spectra are collected). Mix 6 types of microplastics with lake sediments (microplastic: sediment = 1:1000) in various combinations (2 6Mix (equal to 64) to create a test set. Separate microplastics from the mixture and collect 100 sets of Raman spectra from the surface of the filter membrane. Finally, standardize 6,400 Raman spectra to complete the creation of the test set.
[0103] 6. Establish a sediment microplastic analysis model based on a convolutional neural network:
[0104] The feedforward neural network is the basic foundation of deep learning algorithms and is usually composed of an input layer, a hidden layer, and an output layer. The input layer receives raw data as input, the hidden layer is used to extract high-level features of the data, and the output layer generates the final output result of the network. Through repeated iterations, the neural network can gradually optimize the weights to improve the fitting ability and generalization ability for the data.
[0105] This invention uses a one-dimensional CNN model, including two convolutional layers (C1 and C2) and three fully connected layers (F1, F2, and F3). Figure 6 is the structural diagram of the CNN model. The content in the brackets represents the number of convolutional kernels, kernel size, and convolutional stride respectively. F1 flattens 16 features into a one-dimensional vector. F3 contains 6 neurons, corresponding to 6 types of microplastics. This invention makes improvements to the used CNN model. Remove the max pooling layer, increase the convolutional kernel and convolutional stride, so that there are fewer operations and the training time is reduced. 6 types of microplastics use independent fully connected layers, reducing the interference between data and improving the accuracy.
[0106] The output of the output layer is classified using the Sigmoid function, and the activation function of other layers uses the rectified linear unit. The binary cross-entropy loss function is used as the objective function. To prevent the model from overfitting, an L2 regularization term is added. Therefore, the objective equation of the convolutional neural network can be expressed as:
[0107]
[0108] In the formula, n is the total number of sample trainings, yi is the actual label of the sample, is the predicted label of the sample, λ is the regularization coefficient, and ω is the weight to be regularized.
[0109] 7. CNN model training: Use the obtained database to train the CNN. Divide the database into 1,024 training sets and 256 validation sets at a ratio of 8:2. The learning rate, batch size, maximum number of iterations, and L2 regularization weight are set to 0.002, 126, 1,000, and 0.006 respectively. To obtain the best state of the model and prevent overfitting, this invention uses early stopping and takes the accuracy of the validation set of the validation set as the termination judgment condition. After about 200 iterations, the loss function converges. After 619 iterations, the accuracy of the validation set reaches 99.65%.
[0110] 8. Comparison of CNN model performance: Microplastics in sediments were extracted onto a filter membrane, and 100 sets of Raman spectra were collected on the surface of the filter membrane. These 100 sets of Raman spectra were simultaneously input into the trained model to obtain 100 predicted labels. An OR operation was performed on the 100 labels to obtain the predicted label for the distribution of microplastics in the sediment. This method can effectively prevent omissions caused by the overly sparse distribution of microplastics on the filter membrane surface. When the test set was input into the trained CNN model, the accuracy, precision, recall, F1 value, and Kappa coefficient were 0.9427, 0.9680, 0.9167, 0.9404, and 0.8854, respectively. Figure 7 The output comparison results of different algorithms are shown. Two types of microplastics, PVC and PET, have high densities, resulting in poor recycling efficiency and low identification accuracy. PS has fluorescence and low identification accuracy. Different machine learning methods, such as random forest, logistic regression, support vector machine, K-nearest neighbor algorithm, decision tree, etc., were used for comparison. All machine learning models were trained using the same database as CNN. The results show that the accuracy, precision, recall, F1 value, and Kappa coefficient of CNN are relatively higher than those of other machine learning models, confirming that CNN is more effective in identifying microplastics in lake sediments.
[0111] 9. Microplastics and detection: Identification of microplastics at different sampling points. The numbers correspond to the number of spectra of a certain type of microplastic identified from 100 spectra collected on the surface of each filter membrane after sediment treatment, and the spectra correspond to each sampling point.
[0112] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the technical principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention.
Claims
1. A method for separating, identifying and detecting microplastics in sediments, characterized in that The steps include: (1) Sediment collection (2) Sediment treatment and microplastic separation The collected sediment is dried, sieved, subjected to two density flotation and organic matter digestion, and the upper clarified liquid is taken and filtered onto a glass fiber filter membrane using vacuum-assisted filtration; (3) Acquisition of surface-enhanced Raman spectroscopy of microplastics By adding gold nanoparticles, stronger Raman characteristic peaks of microplastics can be obtained; (4) Spectral preprocessing Collect Raman spectrum data and process it through three steps: baseline correction, smoothing filtering and normalization; (5) Creation of Raman spectroscopy database of sediment microplastics The convolutional neural network model was constructed and trained by establishing a standardized Raman spectral database for examining microplastics in test samples; Since there are three main interferences in the Raman spectrum of microplastics in sediments, namely fluorescence, high-frequency noise signals and overlapping characteristic peaks of different microplastics, the strategy adopted is to collect multiple surface-enhanced Raman scattering (SERS) spectra of specific areas of the sample, pre-process these spectra first, and finally take the average of each Raman shift of these spectra to obtain the average mapping spectrum (AMS). AMS can reduce the peak variation of the SERS spectrum; (6) Creation of sediment microplastic analysis model based on convolutional neural network; (7) Convolutional neural network model training Train a convolutional neural network using the obtained database; (8) Evaluation of Convolutional Neural Network Model Effect The Raman spectrum of the microplastics on the glass fiber membrane in step (2) is collected and input into the convolutional neural network model in step (6) to obtain a prediction label, and the prediction label is calculated to obtain a prediction label for the distribution of microplastics in the sediment.
2. A method for separating, identifying and detecting microplastics in sediments according to claim 1, characterized in that The specific steps of step (2) are as follows: (2-1) After the sediment is dried, it is sieved using a wire mesh with a pore size of 5 mm; (2-2) Performing density flotation and organic matter digestion twice; For the first density flotation and organic matter digestion, 30 mL of 35% hydrogen peroxide (H2O2) and 30 mL of 0.11 g / mL sodium chloride (NaCl) solution were added to the sediment, stirred for 1 hour, allowed to settle for 3 hours, and then the upper clarified liquid was taken; the electrostatic adsorption between microplastics and other particles (inorganic mineral particles, organic particles, metal oxide particles, etc.) was broken by stirring; For the second density flotation and organic matter digestion, 30 mL of 35% H2O2 and excess NaCl solution were added to the supernatant, stirred for 1 hour, and allowed to settle for 3 hours. The upper clarified liquid was then taken and filtered onto a glass fiber filter membrane using vacuum-assisted filtration.
3. A method for separating, identifying and detecting microplastics in sediments according to claim 1, characterized in that The specific steps of the three steps of baseline correction, smoothing filtering and normalization in step (3) are as follows: (3-1) Baseline Correction The adaptive iterative reweighted penalized least squares model is used for baseline correction. The weight of each point is updated according to the difference between the baseline value and the original signal of the previous loop fitting, which is expressed as: Among them, x is the original signal, z is the fitted baseline, and d t is the absolute value of the variable that is smaller than the value in the tth iteration The sum of x i represents the eigenvalue of the i-th sample, z i t-1 Represents the reference value of the i-th sample in the previous round (t-1 round). (3-2) Smoothing Filter The invention uses a Savitzky-Golay filter to smooth the noise spectrum. The Savitzky-Golay filter performs least squares fitting on each window of the data through a polynomial to obtain a smoothing coefficient h corresponding to each point in the window. i , expressed as: Among them, w is the window width, x k represents the smoothed value at position k, is the normalization factor. (3-3) Normalization In order to perform linear transformation on the original data, the data is normalized to obtain scaled data in the range of (0,1). The conversion formula is as follows: Among them, x, x max 、x min are the original spectral data, the maximum value in the spectral data, and the minimum value in the spectral data respectively.
4. The method for separating, identifying and detecting microplastics in sediment according to claim 1, characterized in that The specific steps of step (5) are as follows: A one-dimensional convolutional neural network model is used, including two convolutional layers and three fully connected layers. The convolutional neural network model used is improved by removing the maximum pooling layer, increasing the convolution kernel and convolution step size, so that fewer operations are performed and the training time is reduced. The output of the output layer is classified using the Sigmoid function, the activation function of other layers uses the rectified linear unit, and the binary cross entropy loss function is used as the objective function. In order to prevent the model from overfitting, the L2 regularization term is added. Therefore, the objective equation of the convolutional neural network is expressed as: Among them, n is the total number of sample training, y i is the actual label of the sample, is the predicted label of the sample, λ is the regularization coefficient, and ω is the weight to be regularized.
5. The method for separating, identifying and detecting microplastics in sediment according to claim 1, characterized in that The specific steps of step (6) are as follows: The database was divided into 1024 training sets and 256 validation sets in a ratio of 8:2; the learning rate, batch size, maximum number of iterations and L2 regularization weight were set to 0.002, 126, 1000 and 0.006 respectively; in order to obtain the best state of the model and prevent overfitting, early stopping was adopted, and the validation set accuracy of the validation set was used as the termination judgment criterion.