A LIBS automatic quantitative analysis method based on self-supervised atlas fusion

Abstract

The application belongs to the field of laser spectrum analysis, and discloses a LIBS automatic quantitative analysis method based on self-supervised atlas fusion, which comprises the following steps: constructing a residual attention unit, and learning channel weights based on soft threshold filtering and channel attention mechanism; using the residual attention unit and a one-dimensional convolution layer in a deep network model to respectively extract the features of a plasma image and a spectrum of a sample to be measured, and performing feature fusion on the features of the plasma image and the spectrum to obtain a correction factor; inputting the average values of the plasma image and the spectrum into the deep network model to train, so as to obtain a correction model; and inputting the full-band spectrum signal and the plasma image of the sample to be measured into the correction model to obtain a calibration curve. The application can improve the accuracy and precision of LIBS automatic quantitative analysis.

Description

Technical Field

[0001] This invention belongs to the field of laser spectral analysis technology, and more specifically, relates to an automated quantitative analysis method for LIBS based on self-supervised spectral fusion. Background Technology

[0002] Laser-induced breakdown spectroscopy (LIBS) is an emerging elemental analysis technique based on atomic emission spectroscopy. While it has matured considerably in recent years, its quantitative analysis stability and accuracy are inferior to traditional high-precision elemental analysis techniques such as ICP-MS. This is primarily because LIBS is susceptible to interference from spectral fluctuations, matrix effects, and self-absorption effects, which disrupt the linear mapping between characteristic spectral intensities and elemental content. The observed spectral line intensities fail to accurately reflect the true elemental composition of the sample, resulting in a relative standard deviation (RSD) approximately one order of magnitude larger than other traditional analytical methods. Current methods for improving the quantitative analysis accuracy of LIBS can be categorized into two types: instrument improvement and data processing. However, the former is cumbersome, time-consuming, and lacks universality, while the latter typically corrects one-dimensional LIBS spectra, limiting the amount of information and thus only partially mitigating interference rather than eliminating it. Chinese patent document CN110987903B discloses a method and application for correcting LIBS matrix effect interference based on plasma images. It mainly involves normalizing the plasma image to extract two parameters: region area and average brightness. These parameters are used to estimate the ablation quality and plasma temperature to correct the quantitative results, thus obtaining more reliable measurement results. Professor Zheng Rong'er's team at Ocean University of China extracted five features from the plasma image—brightness, area, intensity, contrast, and flatness—and used the PLSR algorithm (Partial Least Squares Regression) to eliminate the self-absorption effect in underwater LIBS. However, the aforementioned existing technologies only focus on a single interference factor, using manual rules to extract a few individual features such as brightness, area, and intensity from the plasma image to compensate for the influence of the studied factor. This cannot avoid the limitations of manual rules, discarding a large amount of potential information contained in the image and the correlation between pixels. Therefore, the effect is limited when facing the combined influence of multiple interference factors. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the present invention aims to provide an automated quantitative analysis method for LIBS based on self-supervised spectral fusion, thereby solving the problems of high workload, poor accuracy, and low precision in existing LIBS quantitative analysis methods.

[0004] To achieve the above objectives, this invention provides an automated quantitative analysis method for LIBS based on self-supervised spectral fusion, comprising the following steps:

[0005] S1 constructs residual attention units and learns channel weights based on soft thresholding and channel attention mechanisms;

[0006] S2 uses the learned residual attention unit and the one-dimensional convolutional layer in the pre-built self-supervised plasma image-spectral feature fusion deep network model to extract the features of the plasma image and spectrum of the sample to be tested, respectively, and performs feature fusion on the features of the plasma image and the features of the spectrum to obtain a correction factor. The aforementioned plasma image and spectrum are acquired synchronously in advance.

[0007] S3 trains the deep network model based on the correction factor using the average values ​​of the plasma image and the spectrum to obtain a correction model;

[0008] S4 inputs the full-band spectral signal and plasma image of the sample to be tested into the calibration model to obtain the calibration curve in order to predict the element content.

[0009] Furthermore, the residual attention unit includes a BN layer, a ReLU layer, a convolutional layer, an inter-channel attention layer, a soft thresholding filter layer, and a residual connection layer.

[0010] Furthermore, the implementation process of the inter-channel attention layer is as follows:

[0011] S101 performs pooling and ReLU activation on the output of the convolutional layer;

[0012] S102 calculates the cosine similarity matrix of the channel dimension and uses a fully connected layer to learn the influence coefficient of each channel on the final result in order to calculate the weight factor;

[0013] Preferably, the expression for the soft threshold filtering function used by the soft threshold filtering layer is:

[0014]

[0015] Where B is a given matrix and λ is the set threshold.

[0016] Furthermore, in step S2, the method for extracting the plasma image features and spectral features and performing feature fusion is as follows:

[0017] S201 uses a plasma image pipeline composed of several residual attention units to extract the image correction factor β(T,n) from the plasma image. s );

[0018] S202 uses a spectral pipeline composed of one-dimensional convolutional layers to extract the electron number density factor χ(n) from the spectrum. e );

[0019] S203 will use the image correction factor β(T,n) s ) and the electron number density factor χ(n e Feature fusion is performed.

[0020] Furthermore, the electron number density factor χ(n) is extracted. e The calculation formula used is:

[0021]

[0022] in, This represents the collected one-dimensional spectrum. Γ(n) represents the feature factor obtained after transforming the collected spectrum using a network model. e () indicates the electron number density influence factor;

[0023] Extract the image correction factor β(T, n) s The calculation formula used is:

[0024]

[0025] in, The acquired plasma images, F(n) represents the feature factors obtained after transforming the acquired image using a network model. s T) represents the particle number density n s The combined influence factors of plasma temperature T.

[0026] Furthermore, in step S3, the training process includes a pre-training phase and a fine-tuning phase. In the pre-training phase, training ends and fine-tuning begins when the loss function converges to a preset condition. In the fine-tuning phase, the loss function is optimized and converges to the preset condition, at which point fine-tuning ends. Preferably, the loss function is a discrete loss. With linearity loss The weighted sum, the formula for calculating the weighted sum is:

[0027]

[0028] Where α is the loss factor;

[0029] More preferably, the preset condition is that the fluctuation of the loss function is less than 1%.

[0030] Furthermore, for a given set of points containing n points... The discreteness loss The expression is:

[0031]

[0032] Among them, S t Let p be a point set S. t The point in the middle, For point set S t The two-dimensional linear space in which it resides;

[0033] Corresponding linearity loss The expression is:

[0034]

[0035] Where N is the number of samples, y i To normalize the original spectral intensity, The average spectral intensity, Predict spectral intensity for the model.

[0036] Furthermore, in step S3, during training, the parameters of each layer of the deep network model are iteratively updated using the gradient descent backpropagation method.

[0037] Furthermore, in step S4, the correction expression used in the correction model is:

[0038]

[0039] in, For the corrected spectral intensity, I i This represents the original spectral intensity.

[0040] Furthermore, prior to step S1, the method for synchronously acquiring the plasma image and the spectrum is as follows: the sample to be tested is ablated and excited using a laser, and the spectrum and image of the sample to be tested are acquired simultaneously using a spectrometer and an ICCD, and multiple acquisitions are performed at different positions of the sample to be tested. The delay of the spectrometer and the ICCD acquisition is adjusted to 2 μs, and the gate width is adjusted to 9 μs.

[0041] Compared with the prior art, the above technical solutions conceived by this invention have the following main advantages:

[0042] 1. This invention achieves accurate extraction and deep understanding of spectral and plasma image features through the proposed self-supervised deep learning algorithm and residual attention unit. It uses the embedded feature vectors for deep fusion and corrects the spectral quantitative results, which significantly improves the accuracy and precision of element quantification in LIBS technology, and the effect is significantly better than the original quantitative results.

[0043] 2. The plasma images and full-band spectral signals used in this invention are acquired synchronously and at the same frequency, thus obtaining both spatial and spectral information of the plasma. This significantly improves the information richness and preserves a large amount of effective information hidden in the images and the correlation between different features. This allows for better inversion of the interference of factors such as sample differences, energy differences, matrix effects, and self-absorption effects on the spectrum. Even when multiple interference factors coexist, it can still provide a basis for the extraction of correction information, demonstrating good selectivity and representativeness.

[0044] 3. The method proposed in this invention uses self-supervised learning, eliminating the need for manual labeling and other additional operations. Compared with other methods, it can save a lot of manpower and resources, significantly improve training efficiency, and has a short inference time, only in the millisecond range. It has good real-time performance, high reliability, is simple and convenient, and has low cost, making it suitable for promotion to industrial environments. Attached Figure Description

[0045] Figure 1 This is a flowchart of the LIBS self-supervised spectral fusion quantitative analysis method provided in the embodiments of the present invention;

[0046] Figure 2 This is a diagram of the experimental setup used for synchronous acquisition of full-band spectral signals and plasma images at the same frequency in an embodiment of the present invention;

[0047] Figure 3 This is a spectral intensity distribution map collected in a preferred embodiment of the present invention;

[0048] Figure 4 This is a structural diagram of the residual attention unit used in a preferred embodiment of the present invention;

[0049] Figure 5 This is a diagram of the architecture of the plasma image-spectral attention convolutional network used in a preferred embodiment of the present invention;

[0050] Figure 6 A comparison diagram of the original calibration curves and the corrected calibration curves of manganese and magnesium elements in different matrices provided for a preferred embodiment of the present invention;

[0051] Figure 7 This is a comparison chart of the original calibration curve and the corrected calibration curve of potassium element under the influence of self-absorption effect, provided as a preferred embodiment of the present invention.

[0052] Among them: 1-nanosecond laser, 2-spectrometer, 3-ICCD, 4-four-dimensional stage, 5-control unit. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0054] This invention provides an automated quantitative analysis method for LIBS based on self-supervised spectral fusion, which mainly includes the following steps:

[0055] S1 constructs residual attention units and learns channel weights based on the importance of each channel using soft thresholding and channel attention mechanisms.

[0056] S2 constructs a self-supervised plasma image-spectral feature fusion deep network model (hereinafter referred to as the deep network model), and uses the residual attention unit learned in step S1 and the one-dimensional convolutional layer in the self-supervised plasma image-spectral feature fusion deep network model to extract the features of the plasma image and spectrum of the sample to be tested, respectively. The features of the plasma image and the spectrum are fused to obtain the correction factor. The aforementioned plasma image and spectrum are acquired synchronously in advance.

[0057] Specifically, a timing controller is used to control the laser to emit laser light and focus the laser light on the surface of the sample to be tested for ablation excitation. After a certain delay, the enhanced charge-coupled device (ICCD) is controlled to aim at the plasma generation area to capture plasma images. The sample position is moved and the laser focus position is changed to re-capture images and collect spectra. The above steps are repeated several times. Then, the average value of the multiple acquired spectra and plasma images is used as the data signal.

[0058] S3 inputs the average values ​​of plasma images and spectra into a deep network model for training to obtain a network weight model, which is then used as a correction model for subsequent applications.

[0059] S4 inputs the full-band spectral signal and plasma image of the sample to be tested into the calibration model, and linearly fits the calibration spectral intensity of the characteristic spectral lines of the element to be tested with its content to obtain the calibration curve, so as to predict the element content and realize accurate quantitative analysis of LIBS.

[0060] Specifically, the analysis method flow in the embodiments of the present invention is as follows: Figure 1 As shown:

[0061] During training, spectral and plasma images of the sample to be tested are acquired simultaneously, and then the feature vectors of the plasma image and spectrum are determined. The features of the two are then pre-fused, and two corresponding correction coefficients are calculated based on the physical parameters of the plasma. The temperature, particle number density, and electron number density of the plasma are obtained sequentially from the feature vector of the plasma image, and latent influence factors are also obtained from the feature vectors of the plasma image and the spectrum. Based on the multiple influencing factors obtained in the previous step, the plasma image and spectrum are fused, and the correction coefficient β containing the latent influence factors is calculated. Then, the weighted sum of the dispersion loss and linearity loss is calculated as the loss function. Finally, training is started using the pre-acquired plasma image and spectral data, iterating repeatedly until the loss function converges to obtain the correction model. During interference, the plasma image and spectral data of the sample to be tested are input into the correction model, and the corrected spectral intensity is calculated based on the correction coefficient β. Finally, a calibration curve is established to predict the elemental content, thus completing the entire analysis process.

[0062] In this embodiment, the constructed residual attention unit is as follows: Figure 4 As shown, it consists of a BN layer, a ReLU activation layer, a 3×3 convolutional layer with a stride of 2, an inter-channel attention layer, a soft thresholding filter layer, and a residual connection layer; the structure of the self-supervised plasma image-spectral feature fusion network model in step S4 is as follows. Figure 5 As shown.

[0063] In a preferred embodiment, the aforementioned convolutional layer is a 3×3 convolutional layer with a stride of 2.

[0064] In a more preferred embodiment, the implementation process of the aforementioned inter-channel attention layer is as follows:

[0065] S101 performs pooling and ReLU activation on the output of the convolutional layer;

[0066] S102 calculates the cosine similarity matrix of the channel dimension and uses a fully connected layer to learn the influence coefficient of each channel on the final result in order to calculate the weight factor.

[0067] In a preferred embodiment, the expression for the soft threshold filtering function used in the aforementioned soft threshold filtering layer is:

[0068]

[0069] Where B is a given matrix and λ is the set threshold.

[0070] In a preferred embodiment, the method steps for extracting plasma image features and spectral features and performing feature fusion in step S2 specifically include:

[0071] S201 uses a plasma image pipeline composed of several residual attention units to extract the image correction factor β(T,n) from the plasma image. s );

[0072] S202 utilizes a spectral pipeline composed of one-dimensional convolutional layers to extract the electron number density factor χ(n) from the spectrum. e );

[0073] S203 will use the image correction factor β(T,n) s ) and electron number density factor χ(n e Feature fusion is performed by concatenating the feature vectors and then fitting them using a fully connected network.

[0074] In a more preferred embodiment, the aforementioned electron number density factor χ(n) is extracted. e The calculation formula used is:

[0075]

[0076] in, This represents the collected one-dimensional spectrum. Γ(n) represents the feature factor obtained after transforming the collected spectrum using a deep network model. e () indicates the electron number density influence factor;

[0077] Extracting image correction factor β(T, n) s The calculation formula used is:

[0078]

[0079] in, The acquired plasma images, F(n) represents the feature factors obtained after transforming the acquired image using a deep network model. s ,T) represents the combined influence factor of particle number density and plasma temperature.

[0080] In a preferred embodiment, step S3 includes a pre-training phase and a fine-tuning phase. In the pre-training phase, training ends and fine-tuning begins when the loss function converges to a preset condition. In the fine-tuning phase, the loss function is optimized and converges to the preset condition, and then fine-tuning ends.

[0081] In a preferred embodiment, the aforementioned loss function is the discreteness loss. With linearity loss The weighted sum, and the specific formula for calculating the weighted sum is:

[0082]

[0083] Where α is the loss factor, which is selected as 0.6 in the pre-training stage and 0.2 in the fine-tuning stage;

[0084] In a more preferred embodiment, the aforementioned preset condition is that the fluctuation of the loss function is less than 1%.

[0085] In a more preferred embodiment, for a given set of points containing n points Dispersion loss The expression is:

[0086]

[0087] Among them, S t Let p be a point set S. t The point in the middle, For point set S t The two-dimensional linear space in which it resides;

[0088] Corresponding linearity loss The expression is:

[0089]

[0090] Where N is the number of samples, y i To normalize the original spectral intensity, The average spectral intensity, Predict spectral intensity for the model.

[0091] In a preferred embodiment, during step S3, the parameters of each layer of the deep network model are iteratively updated using the gradient descent backpropagation method during training.

[0092] In a preferred embodiment, in step S4, the correction model is adjusted for the optimized image correction factor β(T, n) s ) and electron number density factor χ(n e The correction expression used is:

[0093]

[0094] in, For the corrected spectral intensity, I i This represents the original spectral intensity.

[0095] In a preferred embodiment, before step S1, the synchronous acquisition method for plasma images and spectra is as follows: the sample to be tested is ablated and excited; after plasma is generated, the delay of the spectrometer and ICCD acquisition is adjusted to 2 μs, and the gate width is adjusted to 9 μs; then, images are taken at different positions of the sample to be tested to obtain multiple images. Specifically: during acquisition, the following method is used... Figure 2The experimental setup shown acquires plasma images and spectra of the sample under test. The setup mainly includes a nanosecond laser 1, a spectrometer 2, an ICCD 3, a four-dimensional stage 4, and a control unit 5. The four-dimensional stage 4 is used to place the sample under test. The nanosecond laser 1, spectrometer 2, ICCD 3, and control unit 5 are connected to each other. The control unit 5 controls the nanosecond laser 1 to emit lasers of different energies onto the sample under test. Specifically, the control unit 5 is selected as a timing controller. The timing controller controls the nanosecond laser 1 to emit lasers onto the sample under test for ablation excitation. The ICCD is then used to perform the ablation excitation. The laser beam is placed on the sample surface, with the sample surface forming a 90-degree angle with the laser incident angle. The probe of spectrometer 2 is placed at a 60-degree angle with the sample surface. Two μs after laser emission, ICCD 3 and spectrometer 2 are controlled to synchronously acquire images and spectra in a synchronized mode, with a threshold width of 9 μs. Specifically, images and spectra of the sample are synchronously acquired a preset number of times with a delay of 2 μs and a threshold width of 9 μs. Multiple continuously acquired spectra and plasma images (e.g., 10 images) are averaged to reduce plasma fluctuations. The acquired spectra are shown below. Figure 3 As shown.

[0096] To better illustrate the implementation details of the present invention, the following embodiments are provided to further illustrate the present invention. It should be understood that the following embodiments are only preferred implementation methods and are not intended to limit the scope of protection of the present invention in any way.

[0097] Example 1

[0098] This embodiment provides a method for accurate quantitative detection of matrix effect interference in LIBS by using the calibration model obtained in the present invention. This embodiment mainly uses the calibration model to quantitatively analyze manganese (Mn) in different metal standard samples such as aluminum alloy and microalloyed steel, and magnesium (Mg) in different powder pressed samples such as soil pressed samples and potassium feldspar pressed samples. The specific steps are as follows:

[0099] Multiple samples were selected as the test samples, and the contents of Mn and Mg in different samples are shown in Table 1. The serial numbers in the table represent the sample numbers of different types.

[0100] Table 1. Mn and Mg content in aluminum alloy and microalloyed steel standard samples

[0101]

[0102] Simultaneously, spectral and plasma images of the sample to be tested are acquired. Specifically, after routine preprocessing, the sample is placed... Figure 2In the experimental setup shown, after the sample to be tested is excited by lasers of different energies emitted by the nanosecond laser 1, the full-band spectral signal and plasma image of the sample to be tested are acquired by the spectrometer 2 and ICCD 3, respectively. After repeating the above steps multiple times, all the obtained spectral signal and plasma image signal data are stored in the training database.

[0103] The spectral signals and plasma image signal data from the training database are input into the calibration model obtained by the aforementioned method of this invention to obtain the corrected spectral intensities of the MnI 293.931nm and MgI 285.213nm spectral lines and construct new calibration curves. The original calibration curves of Mn and Mg elements and the calibration curves after correction by the self-supervised spectral fusion algorithm are shown below. Figure 6 As shown in Table 2, the evaluation indicators for the quantitative analysis results are as follows.

[0104] Table 2 Comparison of Evaluation Parameters for Fe and Mn Spectral Calibration Curves

[0105]

[0106] The table shows that for the 293.931 nm spectral line of manganese (MnII), the coefficient of determination R... 2 The coefficient of determination (R0.05) increased from 0.4151 to 0.9984, the root mean square error (RMSE) decreased from 0.4326 to 0.0042, and the mean relative error (MRE) decreased from 0.4457 to 0.0098; the magnesium element MgI 285.213 nm spectral line, the coefficient of determination (R0.05) increased from 0.4151 to 0.9984. 2 The root mean square error (RMSE) increased from 0.0244 to 0.9998, while the mean square error (MRE) decreased from 0.1382 to 0.0013 and from 0.6330 to 0.0064. Combined with... Figure 6 As shown in Table 2, the LIBS method based on self-supervised spectral fusion has significantly improved the evaluation index of the calibration curve compared with the original LIBS spectral quantitative detection. The spectral fluctuation is suppressed, and the quantitative accuracy and precision are greatly improved.

[0107] Example 2

[0108] This embodiment provides a method for correcting the interference of self-absorption effect in LIBS using the aforementioned analytical method to achieve accurate quantitative detection. It mainly focuses on the quantitative analysis of potassium (K) in soil-potassium carbonate mixed pressed samples. The specific steps are as follows:

[0109] Multiple soil-potassium carbonate mixed pressing samples were selected. The K element content in the soil-potassium carbonate mixed pressing samples is shown in Table 3. The serial number represents the sample number with different K element contents.

[0110] After performing routine pretreatment on the aforementioned groups of samples, they were placed... Figure 2 In the experimental setup shown, the nanosecond laser 1 emits lasers of different energies to excite the sample under test, and then the full-band spectral signal and plasma image of the sample under test are acquired by the spectrometer 2 and ICCD 3 and stored in the training database as signal data for subsequent applications.

[0111] Table 3. K content in soil-potassium carbonate mixed pressed samples

[0112] K content (wt.%) 2.8292 3.5365 4.2438 4.9512 5.6585 Serial Number 6 7 8 9 10 K content (wt.%) 8.4877 11.3169 14.1462 16.9754 19.8046 Serial Number 11 12 13 14 15 K content (wt.%) 22.6339 25.4631 28.2923 31.1216 33.9508

[0113] The aforementioned signal data is input into the correction model obtained by the aforementioned method of the present invention to obtain the corrected spectral intensities of the KI766.490nm and KI769.896nm spectral lines and to construct a new calibration curve.

[0114] It is worth noting that the severe self-absorption effect significantly disrupts the linear relationship between spectral feature intensity and elemental content, making linear fitting impossible. Only an exponential function can be used for fitting. Therefore, in this embodiment, all indices of the original spectral calibration curve are calculated based on exponential function fitting. The original calibration curve for element K and the calibration curve corrected by the self-supervised spectral fusion algorithm provided in this invention are shown below. Figure 7 As shown in Table 4, the evaluation indicators for the corresponding quantitative analysis results are as follows.

[0115] Table 4: Comparison of Evaluation Parameters for K-Line Calibration Curves

[0116]

[0117] As can be seen from the table, for the KI 766.490 nm spectral line, the coefficient of determination R... 2 The linear fit improved from 0.5% to 0.9991, the root mean square error (RMSE) decreased from 0.4266 to 0.0043, and the mean relative error (MRE) decreased from 0.2138 to 0.0022; the KI 769.896 nm spectral line had a coefficient of determination R0. 2 The linear fit improved from 0.5% to 0.9989, the root mean square error (RMSE) decreased from 0.3669 to 0.0053, and the mean relative error (MRE) decreased from 0.3852 to 0.0046. Combined with... Figure 7 As shown in Table 4, compared with the original LIBS spectral quantitative detection method, the LIBS method based on self-supervised spectral fusion has a significant improvement effect on the interference of self-absorption effect on the quantitative results. The accuracy and precision of the results obtained by the LIBS quantitative analysis method in this embodiment are greatly improved.

[0118] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An automated quantitative analysis method for LIBS based on self-supervised spectral fusion, characterized in that, Includes the following steps: S1 constructs residual attention units and learns channel weights based on soft thresholding and channel attention mechanisms; S2 uses the learned residual attention unit and the one-dimensional convolutional layer in the pre-built self-supervised plasma image-spectral feature fusion deep network model to extract the features of the plasma image and spectrum of the sample to be tested, respectively, and performs feature fusion on the features of the plasma image and the features of the spectrum to obtain a correction factor. The aforementioned plasma image and spectrum are acquired synchronously in advance. S3 trains the deep network model based on the correction factor using the average values ​​of the plasma image and the spectrum to obtain a correction model; S4 inputs the full-band spectral signal and plasma image of the sample to be tested into the calibration model to obtain the calibration curve in order to predict the element content; In step S2, the method for extracting the plasma image features and spectral features and performing feature fusion is as follows: S201 uses a plasma image pipeline composed of several residual attention units to extract the image correction factor β(T,n) from the plasma image. s ); S202 uses a spectral pipeline composed of one-dimensional convolutional layers to extract the electron number density factor χ(n) from the spectrum. e ); S203 will use the image correction factor β(T,n) s ) and the electron number density factor χ(n e Feature fusion is performed. Extract the electron number density factor χ(n) e The calculation formula used is: in, This represents the collected one-dimensional spectrum. Γ(n) represents the feature factor obtained after transforming the collected spectrum using a deep network model. e () indicates the electron number density influence factor; Extract the image correction factor β(T, n) s The calculation formula used is: in, The acquired plasma images, F(n) represents the feature factors obtained after transforming the acquired image using a deep network model. s T) represents the particle number density n s The combined influencing factors of plasma temperature T; In step S4, the correction expression used in the correction model is: in, For the corrected spectral intensity, I i This represents the original spectral intensity.

2. The LIBS automated quantitative analysis method based on self-supervised graph fusion as described in claim 1, characterized in that, The residual attention unit includes a BN layer, a ReLU layer, a convolutional layer, an inter-channel attention layer, a soft thresholding filter layer, and a residual connection layer.

3. The LIBS automated quantitative analysis method based on self-supervised spectral fusion as described in claim 2, characterized in that, The implementation process of the inter-channel attention layer is as follows: S101 performs pooling and ReLU activation on the output of the convolutional layer; S102 calculates the cosine similarity matrix of the channel dimension and uses a fully connected layer to learn the influence coefficient of each channel on the final result in order to calculate the weight factor; The expression for the soft threshold filtering function used by the soft threshold filtering layer is: Where B is a given matrix and λ is a preset threshold.

4. The LIBS automated quantitative analysis method based on self-supervised spectral fusion as described in claim 1, characterized in that, In step S3, the training process includes a pre-training phase and a fine-tuning phase. During the pre-training phase, training ends and fine-tuning begins when the loss function converges to a preset condition. During the fine-tuning phase, the loss function is optimized and converges to the preset condition, at which point fine-tuning ends. The loss function is the discrete loss. With linearity loss The weighted sum, the formula for calculating the weighted sum is: Where α is the loss factor; The preset condition is that the fluctuation of the loss function is less than 1%.

5. The LIBS automated quantitative analysis method based on self-supervised spectral fusion as described in claim 4, characterized in that, Given a set of n points The discreteness loss The expression is: Among them, S t Let p be a point set S. t The point in the middle, For point set S t The two-dimensional linear space in which it resides; Corresponding linearity loss The expression is: Where N is the number of samples, y i To normalize the original spectral intensity, The average spectral intensity, Predict spectral intensity for the model.

6. The LIBS automated quantitative analysis method based on self-supervised graph fusion as described in claim 1, characterized in that, In step S3, during training, the parameters of each layer of the deep network model are iteratively updated using the gradient descent backpropagation method.

7. The LIBS automated quantitative analysis method based on self-supervised spectral fusion as described in claim 1, characterized in that, The method for synchronously acquiring the plasma image and the spectrum is as follows: the sample to be tested is ablated and excited by laser, and the spectrum and image of the sample to be tested are acquired simultaneously by a spectrometer and an ICCD. Multiple acquisitions are performed at different positions of the sample to be tested. The delay of the spectrometer and the ICCD acquisition are both adjusted to 2 μs, and the gate width is adjusted to 9 μs.

Images