A method for rapid, automated differentiation of components in a mixture solution based on TLC-SERS technology

Abstract

The application relates to a method for rapidly and automatically distinguishing components in a mixed solution based on a TLC-SERS technology, and relates to the field of spectral detection; the application aims to solve the problems that TLC-SERS cannot accurately analyze substances with similar structures and is too dependent on subjective interpretation and ability of operators; the application assembles silver sol nanoparticles on the surface of a TLC plate, and a large-area TLC-SERS substrate with high sensitivity and repeatable detection is prepared; spectral collection is carried out, and TLC-SERS spectra of all target substances are generated; then, two machine learning algorithms are adopted to analyze the TLC-SERS spectral data; compared with a conventional manual SERS peak analysis method, the proposed machine learning algorithm depends on overall analysis of spatial information and spectral information of the TLC-SERS spectral data, instead of spectral information determined by main characteristic SERS peaks, so that the accuracy of SERS peak identification of different substances is greatly improved.

Description

Technical Field

[0001] This invention belongs to the field of chromatographic detection, specifically relating to a method for rapidly and automatically distinguishing components in a mixture by combining SERS technology, thin-layer chromatography technology, and machine learning algorithms. Background Technology

[0002] Surface-enhanced Raman scattering thin-layer chromatography (TLC-SERS) has shown great potential as a novel analytical chemistry tool and is gaining increasing research interest due to its high sensitivity and ease of implementation. TLC technology synergistically integrates multiple functions onto a latex plate, achieving the separation of mixtures by combining different eluents (mobile phase) and adsorption layers (stationary phase). SERS technology amplifies molecular vibrational spectra based on the plasmon resonance effect between nanostructures, thereby enabling trace-level chemical detection. Provided the mobile and stationary phases are properly matched, TLC-SERS can very successfully detect single target substances from complex chemical and biological samples. Therefore, TLC-SERS has been widely applied in fields such as food safety, agriculture, medical diagnostics, and industrial detection.

[0003] Traditional SERS detection is usually performed in a wet solution or dry film formation state. Dynamic SERS method is performed during the transition of nanosolution from wet to dry state to find the optimal substrate state to ensure maximum enhancement effect. At present, TLC-SERS faces several key challenges: (1) TLC-SERS technology usually involves placing gold nanoparticles and silver nanoparticles (NPs) onto a TLC plate containing the sample for dynamic SERS detection. Factors such as the concentration and volume of gold and silver sols and detection time will have a significant impact on the SERS detection results. In addition, this will also cause adverse interference to continuous SERS measurements. (2) The imperfection of TLC technology separation leads to greater challenges for TLC-SERS technology when detecting multiple target substances. When identifying multiple substances, the SERS signal intensity of different substances fluctuates greatly, resulting in a low signal-to-noise ratio. Manual analysis of SERS spectra can only analyze one or a few characteristic peaks of the target substance, making it difficult to complete the accurate analysis of mixtures. Especially when facing substances with similar structures, it is difficult to separate them by TLC and they have the same type of chemical bonds. (3) Traditional TLC-SERS technology, as a manual analysis technique, relies excessively on the operator's subjective interpretation and ability. As the number of spectra increases, this method will become unsustainable. Summary of the Invention

[0004] This invention addresses the problems mentioned above, such as the inability of TLC-SERS to accurately analyze structurally similar substances and its over-reliance on the subjective interpretation and ability of operators, by providing a rapid and automated method for distinguishing components in a mixed solution based on TLC-SERS technology.

[0005] The present invention provides a rapid and automated method for distinguishing components in a mixture solution based on TLC-SERS technology, which is carried out according to the following steps:

[0006] Step 1: Prepare silver sol nanoparticles (Ag NPs) using the sol-gel method;

[0007] Step 2: Prepare TLC-SERS substrates using a three-phase interface self-assembly method.

[0008] The prepared silver sol nanoparticles (Ag NPs), dichloromethane, and deionized water were mixed in a volume ratio of (0.01-0.1):(0.1-1):1 to obtain solution B; solution B was then mixed with n-hexane in a volume ratio of (2-100):1 and allowed to stand for 1-30 minutes; the silver sol nanoparticles (Ag NPs) were transferred to a TLC plate using the dip-coating method to obtain a TLC-SERS substrate.

[0009] Step 3: Separate the mixture using a TLC-SERS substrate and perform spectral acquisition.

[0010] The mixture to be analyzed was dropped onto the bottom of the prepared TLC-SERS substrate, dried for 3-10 min, placed in the developing solvent, developed for 10-40 min, dried for 10-30 min, and then the SERS spectrum was acquired by one-dimensional scanning along the development path of the mixture on the TLC-SERS substrate using a Raman spectrometer. The scanning step size was 50-200 nm, the scanning time was 5-30 seconds, and the laser power was 10-100 mW.

[0011] Step 4: Automated analysis of the spectra acquired on the TLC-SERS substrate using machine learning algorithms.

[0012] Based on the convolutional neural network algorithm and spectral angle algorithm in Python, a spectral classification model and a spectral similarity model were written respectively. The collected SERS spectra were input into the spectral classification model and the spectral similarity model for automated analysis: the spectral classification model is used to predict the substance type information corresponding to the SERS spectrum. The spectral classification model uses a convolutional neural network. The input of the convolutional neural network is the two-dimensional SERS spectrum data, and the output is the classification curve corresponding to the substance type information of the SERS spectrum.

[0013] The spectral similarity model is used to calculate the similarity between the standard spectrum and the acquired SERS spectrum; the similarity is calculated using the spectral angles of the standard spectrum and the acquired SERS spectrum to obtain the similarity curve;

[0014] Simultaneously, based on the positions of the classification model curve and the similarity model curve on the TLC-SERS substrate, the range of different components in the mixture to be analyzed is accurately identified.

[0015] The range for identifying different components in the mixture to be analyzed refers to identifying the position of the mixture on the TLC plate. This invention involves continuously scanning and acquiring spectra along a straight line on the TLC plate, that is, acquiring a spectrum at fixed intervals.

[0016] For example, ten spectra are collected at 200µm intervals on a TLC plate, representing the spectra of substances A, B, and C. Starting from the fifth spectrum, there are three consecutive positions corresponding to substance B. Since this invention continuously scans and collects spectra along a straight line on the TLC plate, i.e., one spectrum is collected every 200µm, this means that substance B is located on the spectral plate from 5 * 200 = 1000µm to 1400µm. This invention uses continuous spectral acquisition to imbue the spectra with spatial location attributes. Randomly acquired spectra do not possess this attribute. With continuous acquisition and controlled intervals, the 100th spectrum, i.e., the location of the acquisition, is the interval distance × 100. Therefore, by identifying the spectral sequence number of substance B from beginning to end, its position on the TLC plate can be determined.

[0017] The standard spectrum mentioned is used to train the classification model.

[0018] Furthermore, the silver sol nanoparticles described in step one are prepared as follows:

[0019] At room temperature, ascorbic acid and sodium citrate were dispersed in ultrapure water and stirred to dissolve to obtain solution A. Then, solution A was transferred to a water bath, and silver nitrate solution was added to solution A and boiled to react. After reacting for 1-3 hours, silver sol nanoparticles were obtained. Finally, after the reaction was completed and cooled to room temperature, Ag NPs were collected, washed twice by centrifugation with aqueous solution, and the collected centrifuged product was dispersed in water to complete the preparation of silver sol nanoparticles.

[0020] Furthermore, the mass-to-volume ratio of ascorbic acid, sodium citrate, and ultrapure water is (1-100) mg: (100-500) mg: 1 mL.

[0021] Furthermore, the mass-to-volume ratio of silver nitrate solution to ultrapure water is 1:1 to 20.

[0022] Furthermore, the centrifugation speed is 5000-10000 rpm, and the centrifugation time is 10-40 min.

[0023] Furthermore, the amount of the mixture to be analyzed added is 1-10 μL.

[0024] Furthermore, in step 4, the automated analysis automatically determines the species information in the experimental spectrum through a classification model, and the similarity model compares the similarity between the experimental spectrum and the standard spectrum.

[0025] Furthermore, the automated analysis of the collected SERS spectra by inputting them into the spectral classification model is as follows:

[0026] The SERS spectrum is predicted using a convolutional neural network classification algorithm, and the prediction result is the information on the types of substances corresponding to the SERS spectrum.

[0027] Furthermore, the standard spectrum refers to the Raman scattering spectrum of a pure substance. It is obtained by acquiring the spectrum of the pure substance using a Raman spectrometer. The pure substance refers to the pure substance corresponding to the analyte.

[0028] Figure 8 This paper presents a flowchart illustrating the automated analysis process of spectra acquired on a TLC-SERS substrate using a machine learning algorithm. The CNN consists of multiple functional layers, including convolutional layers, pooling layers, and fully connected layers. Pooling layers, located after the convolutional layers, reduce data dimensionality while preserving spectral features. The functional units of the convolutional and pooling layers explore and extract the most critical features from the input data, thereby optimizing model performance. Finally, the fully connected layers map the extracted key feature representations to the sample label space. In this work, the feature extraction part of the CNN algorithm comprises two pooling layers and three convolutional layers. The CNN employs an adaptive moment estimation optimization algorithm to train on raw Raman spectral data with class labels. The training process is accelerated by CUDA, using the ReLU function to implement a nonlinear transformation of the feature map as the activation function, with a kernel size of 1*3 and a stride of 2. The learning rate is 0.01. 75% of the total Raman spectral data is used for model training, and the remaining data is used for testing. The selection of training and testing data is randomized, which further improves the model's generalization ability. This invention also uses an iterative algorithm to ensure the generalization ability of the CNN model. After each iteration, the selection of the CNN model for the training data is completely random, which can effectively prevent the model from overfitting.

[0029] The convolutional neural network used in this invention is: a first convolutional layer + a first pooling layer + a second convolutional layer + a second pooling layer + a third convolutional layer + a first fully connected layer + a second fully connected layer + a third fully connected layer;

[0030] First convolutional layer: kernel size is 1*3, number of kernels is 10, stride is 2; after ReLU activation, a 650*10 feature matrix is ​​obtained;

[0031] First pooling layer: The pooling method is selected as max pooling, the pooling kernel size is 1*3, the stride is 2, and the feature matrix output by convolutional layer 1 is 325*10 after the pooling operation.

[0032] The second convolutional layer has a kernel size of 1*3, a number of kernels of 20, and a stride of 2. After ReLU activation, a feature matrix of 163*20 is obtained.

[0033] Second pooling layer: The pooling method is selected as max pooling, the pooling kernel size is 1*3, the stride is 2, and the output matrix size is 82*20.

[0034] The third convolutional layer has a kernel size of 1*3, a kernel count of 40, and a stride of 1. After ReLU activation, it yields an 82*40 feature matrix.

[0035] First fully connected layer: The 3280 extracted and flattened features are fully connected into 1000 features;

[0036] The second fully connected layer connects 1000 features into 300 features;

[0037] The third fully connected layer connects the 300 features to form the required number of categories. In the concentration recognition model, the final number of categories after full connection is 1, which is the category to be recognized.

[0038] Furthermore, the standard spectrum and the acquired SERS spectrum are input into the spectral similarity algorithm for secondary discrimination. By calculating the spectral angle between the standard spectrum and the acquired SERS spectrum, and then performing a cosine operation, a value between 0 and 1 is obtained. If the output value is close to 1, the spectral similarity is high, that is, the convolutional neural network classification model result is correct. In this way, the degree of similarity between the standard spectrum and the acquired SERS spectrum is determined, and the secondary discrimination is completed.

[0039] Furthermore, the calculation of the spectral angle is based on the following formula:

[0040]

[0041] Where θ is the spectral angle, arccos is the arccosine, and x i y is the value at the i-th Raman shift in the standard spectrum. i The value is for the i-th Raman shift in the test spectrum.

[0042] This invention develops an objective, accurate, and efficient TLC-SERS sensing approach capable of rapidly and automatically detecting multiple substances in complex mixtures. First, silver sol nanoparticles (Ag NPs) are assembled on the surface of a TLC plate using an interface self-assembly method, preparing a large-area TLC-SERS substrate with high sensitivity and repeatable detection. One-dimensional spectral acquisition is performed along the TLC path to generate TLC-SERS spectra for all target substances. Then, two machine learning algorithms (a convolutional neural network classification model and a spectral angle similarity model) are used to analyze the TLC-SERS spectral data. Compared to traditional manual SERS peak analysis methods, the proposed machine learning algorithm relies on a holistic analysis of the spatial and spectral information of the TLC-SERS spectral data, rather than relying on the spectral information determined by the main characteristic SERS peaks, which significantly improves the accuracy of SERS peak identification for different substances. Furthermore, this invention only requires acquiring the SERS spectrum of a pure substance once as a standard spectrum, which can be applied to the automatic identification of various mixtures, greatly improving the scalability of the proposed method. Most importantly, the convolutional neural network (CNN) model works in conjunction with the spectral angle (SA) model, achieving high accuracy even when target substances overlap along the TLC path. Therefore, this automated and objective TLC-SERS spectral detection and analysis method has great potential for rapid mixture detection. Attached Figure Description

[0043] Figure 1 Flowchart of TLC-SERS substrate preparation;

[0044] Figure 2 A physical image of a large-area silver sol nanoparticle assembly;

[0045] Figure 3 Physical image of a TLC-SERS substrate;

[0046] Figure 4 Scanning electron microscope image of TLC-SERS substrate;

[0047] Figure 5 Schematic diagram of TLC-SERS substrate for separating mixture solutions;

[0048] Figure 6 TLC-SERS spectrum;

[0049] Figure 7 Curve graphs of convolutional neural network (CNN) classification model and spectral angle (SA) similarity model;

[0050] Figure 8 A flowchart of the SERS spectrum collection process for automated analysis of spectral classification and spectral similarity models. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the spirit of the contents disclosed in the present invention will be described in detail below. After understanding the embodiments of the present invention, any person skilled in the art can make changes and modifications based on the technology taught in the present invention without departing from the spirit and scope of the present invention.

[0052] The illustrative embodiments and descriptions of the present invention are used to explain the present invention, but are not intended to limit the present invention.

[0053] Example 1

[0054] This embodiment presents a rapid and automated method for distinguishing components in a mixture solution based on TLC-SERS technology, as detailed below:

[0055] (1) At room temperature of 27°C, ascorbic acid (26.45 mg) and sodium citrate (220.6 mg) were dispersed and dissolved in 250 mL of ultrapure water;

[0056] (2) After stirring for 30 minutes, transfer the solution to a 100°C water bath for heating;

[0057] (3) Dissolve silver nitrate (42.5 mg) in ultrapure water and add it to the solution in step (2) and react for 3 hours;

[0058] (4) Wash the sample twice with ultrapure water, centrifuge at 5000-8000 rpm for 20 min, disperse the collected centrifuged product in ultrapure water to obtain high-purity silver sol nanoparticles (Ag NPs).

[0059] (5) Mix dichloromethane, deionized water, and the Ag NPs solution obtained in step (4) at a volume ratio of 12:30:1, then mix with n-hexane at a volume ratio of 4:1. After sonication for 2 minutes, allow to stand for 10 minutes. During this standing process, the Ag NPs are driven to the water-n-hexane interface. Then, using a dip-coating method, transfer the Au@Ag NRs to a thin-layer chromatography (TLC) plate, thus completing the TLC-SERS substrate preparation. A schematic diagram of the TLC-SERS substrate preparation process is shown below. Figure 1 As shown in the image. A photograph of a large-area silver sol nanoparticle is shown below. Figure 2 As shown. The physical image and scanning electron microscope image of the TLC-SERS substrate are shown below. Figure 3 and 4 As shown, the results indicate that this technique can produce large-area, flat TLC-SERS substrates.

[0060] (6) Take a mixed solution containing Sudan I, malachite green and rhodamine 6G as the research object, drop it onto a TLC-SERS substrate, and after drying, put it into a developing solution of methanol, ethanol and water (V / V = 8:2:2) for separation of the mixture;

[0061] (7) After separation is complete, remove the TLC-SERS substrate and dry it at 30℃ for 1 hour before performing spectral acquisition. The flowchart is shown below. Figure 5 As shown;

[0062] (8) A Raman spectrometer was used to acquire spectra along the one-dimensional path of the TLC, and the generated SERS spectrum is shown below. Figure 6 As shown;

[0063] (9) The above SERS spectral dataset was input into a convolutional neural network classification algorithm and a spectral angle similarity model for automated classification and recognition. The results are as follows: Figure 7 As shown, the Raman spectrum is first assigned using two algorithms, and then the position of the substance on the TLC plate is automatically determined based on the intersection point of the curves obtained by the two algorithms.

[0064] The convolutional neural network used in this embodiment is: first convolutional layer + first pooling layer + second convolutional layer + second pooling layer + third convolutional layer + first fully connected layer + second fully connected layer + third fully connected layer;

[0065] First convolutional layer: kernel size is 1*3, number of kernels is 10, stride is 2; after ReLU activation, a 650*10 feature matrix is ​​obtained;

[0066] First pooling layer: The pooling method is selected as max pooling, the pooling kernel size is 1*3, the stride is 2, and the feature matrix output by convolutional layer 1 is 325*10 after the pooling operation.

[0067] The second convolutional layer has a kernel size of 1*3, a number of kernels of 20, and a stride of 2. After ReLU activation, a feature matrix of 163*20 is obtained.

[0068] Second pooling layer: The pooling method is selected as max pooling, the pooling kernel size is 1*3, the stride is 2, and the output matrix size is 82*20.

[0069] The third convolutional layer has a kernel size of 1*3, a kernel count of 40, and a stride of 1. After ReLU activation, it yields an 82*40 feature matrix.

[0070] First fully connected layer: The 3280 extracted and flattened features are fully connected into 1000 features;

[0071] The second fully connected layer connects 1000 features into 300 features;

[0072] The third fully connected layer connects the 300 features to form the required number of categories. In the concentration recognition model, the final number of categories after full connection is 1, which is the category to be recognized.

[0073] The standard spectrum and the acquired SERS spectrum are input into a spectral similarity algorithm for secondary discrimination. By calculating the spectral angle between the standard spectrum and the acquired SERS spectrum, and then performing a cosine operation, a value between 0 and 1 is obtained. If the output value is close to 1, the spectral similarity is high, that is, the convolutional neural network classification model result is correct. This determines the degree of similarity between the standard spectrum and the acquired SERS spectrum, thus completing the secondary discrimination.

[0074] The calculation of the spectral angle is based on the following formula:

[0075]

[0076] Where θ is the spectral angle, arccos is the arccosine, and x i y is the value at the i-th Raman shift in the standard spectrum. i The value is for the i-th Raman shift in the test spectrum.

[0077] This mode is not limited to silver sol nanoparticles; it can assemble any metal nanoparticles to prepare TLC-SERS substrates. This mode is not limited to the two models mentioned above; any machine learning spectral analysis algorithm can be used in this mode to perform automated spectral analysis on the spectra of TLC-SERS substrates.

Claims

1. A rapid and automated method for distinguishing components in a mixture solution based on TLC-SERS technology, characterized in that... It is done in the following steps: Step 1: Prepare silver sol nanoparticles (Ag NPs) using the sol-gel method; Step 2: Prepare TLC-SERS substrates using a three-phase interface self-assembly method. The prepared silver sol nanoparticles (Ag NPs), dichloromethane, and deionized water were mixed in a volume ratio of (0.01-0.1):(0.1-1):1 to obtain solution B; solution B was then mixed with n-hexane in a volume ratio of (2-100):1 and allowed to stand for 1-30 min; the silver sol nanoparticles (Ag NPs) were transferred to a TLC plate using the dip-coating method to obtain a TLC-SERS substrate. Step 3: Separate the mixture using a TLC-SERS substrate and perform spectral acquisition. The mixture to be analyzed was dropped onto the bottom of the prepared TLC-SERS substrate, dried for 3-10 min, placed in the developing solvent, developed for 10-40 min, dried for 10-30 min, and then the SERS spectrum was acquired by one-dimensional scanning along the development path of the mixture on the TLC-SERS substrate using a Raman spectrometer. The scanning step size was 50-200 nm, the scanning time was 5-30 seconds, and the laser power was 10-100 mW. Step 4: Automated analysis of the spectra acquired on the TLC-SERS substrate using machine learning algorithms. Based on convolutional neural network algorithms and spectral angle algorithms in Python, spectral classification models and spectral similarity models were written respectively. The collected SERS spectra were then input into the spectral classification models and spectral similarity models for automated analysis. The spectral classification model is used to predict the substance type information corresponding to the SERS spectrum. The spectral classification model adopts a convolutional neural network. The input of the convolutional neural network is the two-dimensional SERS spectrum data, and the output is the classification curve corresponding to the substance type information of the SERS spectrum. The spectral similarity model is used to calculate the similarity between the standard spectrum and the acquired SERS spectrum; the similarity is calculated using the spectral angles of the standard spectrum and the acquired SERS spectrum to obtain the similarity curve; Simultaneously, based on the positions of the classification model curve and the similarity model curve on the TLC-SERS substrate, the range of different components in the mixture to be analyzed is accurately identified.

2. The method for rapid and automated differentiation of components in a mixture solution based on TLC-SERS technology according to claim 1, characterized in that... The silver sol nanoparticles described in step one are prepared as follows: At room temperature, ascorbic acid and sodium citrate were dispersed in ultrapure water and stirred to dissolve to obtain solution A. Then, solution A was transferred to a water bath, and silver nitrate solution was added to solution A and boiled to react. After reacting for 1-3 hours, silver sol nanoparticles were obtained. Finally, after the reaction was completed and cooled to room temperature, Ag NPs were collected, washed twice by centrifugation with aqueous solution, and the collected centrifuged product was dispersed in water to complete the preparation of silver sol nanoparticles.

3. The method for rapid and automated differentiation of components in a mixture solution based on TLC-SERS technology according to claim 2, characterized in that... The mass-to-volume ratio of ascorbic acid, sodium citrate and ultrapure water is (1~100) mg: (100~500) mg: 1 mL.

4. The method for rapid and automated differentiation of components in a mixture solution based on TLC-SERS technology according to claim 2, characterized in that... The mass-to-volume ratio of silver nitrate solution to ultrapure water is 1:1~20.

5. The method for rapid and automated differentiation of components in a mixture solution based on TLC-SERS technology according to claim 2, characterized in that... The centrifugation speed is 5000-10000 rpm, and the centrifugation time is 10-40 min.

6. The method for rapid and automated differentiation of components in a mixture solution based on TLC-SERS technology according to claim 1, characterized in that... The amount of the mixture to be analyzed added is 1-10 μL.

7. The method for rapid and automated differentiation of components in a mixture solution based on TLC-SERS technology according to claim 1, characterized in that... In step 4, the automated analysis uses a classification model to automatically determine the species information in the experimental spectrum, and a similarity model to compare the similarity between the experimental spectrum and the standard spectrum.

8. A method for rapid and automated differentiation of components in a mixture solution based on TLC-SERS technology according to claim 1 or 7, characterized in that... The process of inputting the collected SERS spectra into the spectral classification model for automated analysis is as follows: The SERS spectrum is predicted using a convolutional neural network classification algorithm. The prediction result is the information of the substance type corresponding to the SERS spectrum. The convolutional neural network is: first convolutional layer + first pooling layer + second convolutional layer + second pooling layer + third convolutional layer + first fully connected layer + second fully connected layer + third fully connected layer. First convolutional layer: kernel size is 1*3, number of kernels is 10, stride is 2; after ReLU activation, a 650*10 feature matrix is ​​obtained; First pooling layer: The pooling method is selected as max pooling, the pooling kernel size is 1*3, the stride is 2, and the feature matrix output by convolutional layer 1 is 325*10 after the pooling operation. The second convolutional layer has a kernel size of 1*3, a number of kernels of 20, and a stride of 2. After ReLU activation, a feature matrix of 163*20 is obtained. Second pooling layer: The pooling method is selected as max pooling, the pooling kernel size is 1*3, the stride is 2, and the output matrix size is 82*20. The third convolutional layer has a kernel size of 1*3, a kernel count of 40, and a stride of 1. After ReLU activation, it yields an 82*40 feature matrix. First fully connected layer: The 3280 extracted and flattened features are fully connected into 1000 features; The second fully connected layer connects 1000 features into 300 features; The third fully connected layer connects the 300 features to the required number of categories. In the spectral classification model, the final number of fully connected layers is the number of categories to be identified.

9. A method for rapid and automated differentiation of components in a mixture solution based on TLC-SERS technology according to claim 1, characterized in that, The standard spectrum and the acquired SERS spectrum are input into the spectral similarity algorithm for secondary discrimination. By calculating the spectral angle between the standard spectrum and the acquired SERS spectrum, and then performing a cosine operation, a value between 0 and 1 is obtained. If the output value is close to 1, the spectral similarity is high, that is, the convolutional neural network classification model result is correct. In this way, the degree of similarity between the standard spectrum and the acquired SERS spectrum is determined, and the secondary discrimination is completed.

10. A method for rapid and automated differentiation of components in a mixture solution based on TLC-SERS technology according to claim 9, characterized in that... The calculation of the spectral angle is based on the following formula: in, is the spectral angle, and arccos is the arccosine. This is the value at the i-th Raman shift in the standard spectrum. The value is for the i-th Raman shift in the test spectrum.

Images