A method for correcting fluctuations in LIBS spectra

Abstract

This invention provides a LIBS spectral volatility correction method, belonging to the field of laser spectral analysis. The method involves: obtaining the full-band spectral signal and plasma image of the sample to be tested and establishing a training database; fitting the original spectral intensity of the element to be tested with its content to obtain a standard curve, which is then used as a training label to train a deep convolutional network model; when the loss function converges to a preset condition, optimizing the standard curve and using it as a new training label to continue training; when the loss function converges to the preset condition again, training is complete and a correction model is obtained; the corrected spectral intensity is linearly fitted with its content to obtain a calibration curve, thereby completing the LIBS spectral volatility correction. This invention offers simple operation, significant effect, and high reliability in correcting spectral volatility, greatly improving the quantitative detection accuracy of LIBS in harsh environments and effectively applying it to industrial scenarios.

Description

Technical Field

[0001] This invention belongs to the field of laser spectral analysis, and more specifically, relates to a method for correcting LIBS spectral fluctuations. Background Technology

[0002] Laser-induced breakdown spectroscopy (LIBS), also known as laser-induced plasma spectroscopy (LIPS) or laser probe, is an emerging elemental analysis technique based on atomic emission spectroscopy. Its main principle involves focusing a high-energy-density, short-pulse laser onto the surface of the sample, ablating the surface and generating plasma. The characteristic spectrum of this plasma is then collected and analyzed to qualitatively and quantitatively determine the elemental composition of the analyte. Compared to traditional elemental analysis techniques, LIBS offers unique advantages such as small size, portability, simple operation, short analysis time, and near-non-destructive testing. Therefore, it has become a research hotspot in recent years and is hailed as a "future star in chemical analysis," achieving initial applications. However, one of the main reasons limiting the further development of LIBS is spectral volatility. Unstable spectra lead to inaccurate quantification, with a relative standard deviation (RSD) approximately an order of magnitude larger than other traditional analytical methods such as ICP-MS and ICP-OES, severely hindering the large-scale application of LIBS in industrial testing.

[0003] To improve spectral stability, researchers both domestically and internationally have proposed various solutions. These methods can be categorized into three types: data processing methods, plasma manipulation methods, and the introduction of external reference signals. Plasma images, as a direct representation of plasma, can describe in detail the interaction process between laser and matter and the plasma evolution process, making them a highly promising spectral correction signal that has gained increasing attention in recent years. CN110987903B discloses a LIBS matrix effect correction method and its application. This method uses plasma images for LIBS matrix effect correction. By establishing the relationship between the plasma region area and ablation mass corresponding to the plasma image, and between image brightness and plasma temperature, the spectral intensity of the sample is corrected based on the aforementioned relationships between the plasma image and the sample type. The matrix effect correction is then completed by fitting calibration curves at the laser energy to be corrected based on the corrected spectra and their corresponding elemental contents. CN114894781A discloses a generalized method and system for underwater LIBS spectral normalization based on external signals. It provides a method for normalizing and correcting LIBS spectra based on external signals, including plasma images. This patent can extract plasma image signal features, then use the PCA algorithm to extract principal components of the image signal, and combine the spectral line intensity and relative deviation of the extracted and analyzed elements to establish a PLS regression model to normalize the spectrum. Deng Zhang et al. (Deng Zhang et al., Aplasma-image-assisted method for matrix effect correction in laser-induced breakdown spectroscopy[J], Analytica Chimica Acta, 2019.) proposed an IA-LIBS method, which quantifies the temperature and ablation quality of the plasma by extracting the average brightness and area of ​​the plasma image, and then corrects the spectral intensity to correct matrix effects. Yongquan Zhang et al. (Yongquan Zhang et al., Quantitation improvement of underwater laser-induced breakdown spectroscopy by using self-absorption correction based on plasma images[J], Analytica Chimica Acta, 2022.) successfully suppressed the self-absorption effect of underwater LIBS by calculating five key image features of simultaneously acquired plasma images: brightness, area, contrast, plasma intensity, and shape flatness.

[0004] However, these methods have not fully extracted and utilized the deep information contained in plasma images. As a two-dimensional representation, plasma images have millions of pixels, making it difficult to describe them using only a few individual features such as brightness, area, and intensity. There is certainly much more information that cannot be easily quantified and remains to be discovered. For example, the brightness of plasma can be measured by total brightness or global average brightness, as well as feature information such as local brightness in different regions. Contrast, for instance, includes not only core region contrast but also edge contrast or contrast between adjacent pixels. In short, the final correction factor is affected by a considerable number of factors, and it is difficult to accurately quantify the correction coefficient using only a few simple feature parameters, ultimately leading to unsatisfactory quantitative results. If each feature parameter is expanded from a single value to a feature matrix, and multiple different feature matrices are concatenated into a unified correction feature matrix—that is, if the correction coefficient variable is expanded from a single digit to a larger range, such as 64×64—it can obviously include more information contained in the original image, thereby more accurately and robustly overcoming the influence of spectral fluctuations and improving the quantitative accuracy of LIBS.

[0005] Referring to the development of computer vision in RGB natural images, before the large-scale application of deep learning, tasks such as detection and segmentation were limited to recognition through manually designed features, with only minor improvements. However, with the introduction of deep learning, high-dimensional, large-scale deep feature layers were added, leading to numerous innovations and a dramatic improvement in detection accuracy, quickly reaching the level for commercial application. Therefore, methods for using deep learning technology and designing suitable learning networks based on the characteristics of plasma images to fully explore and utilize the deep information hidden in plasma images, mapping the learned high-dimensional features to large-scale feature maps, and applying them to the correction of LIBS spectral fluctuations have not yet been fully studied. Summary of the Invention

[0006] In view of the shortcomings of the prior art, the purpose of this invention is to provide a LIBS spectral fluctuation correction method, which aims to solve the problem of poor quantitative accuracy caused by spectral fluctuation in existing LIBS detection methods.

[0007] To achieve the above objectives, the present invention provides a LIBS spectral fluctuation correction method, which includes the following steps:

[0008] S1 ablation excitation of the sample under test is performed to obtain full-band spectral signals and plasma images and to establish a training database.

[0009] S2 For the element to be tested, the original spectral intensity of the characteristic spectral line of the element to be tested in the full-band spectral signal is fitted with its content to obtain a standard curve and use it as a training label. The training label is then input into the deep convolutional network model and trained using plasma images in the training database.

[0010] When the loss function of S3 training converges to the preset condition, the standard curve is optimized as a new training label and input into the deep convolutional network model, and the plasma images in the training database are used to continue training it.

[0011] When the loss function converges to the preset condition again during training, the training is complete and the trained deep convolutional network model is used as the correction model.

[0012] S5 inputs the full-band spectral signal and plasma image of the sample to be tested into the correction model, and linearly fits the corrected spectral intensity of the characteristic spectral lines of the element to be tested with its content to obtain the calibration curve, thereby completing the LIBS spectral fluctuation correction.

[0013] As a further preferred option, in step S1, lasers of different energies are used to ablate the sample under test a predetermined number of times in order to acquire full-band spectral signals and plasma images at different energies.

[0014] As a further preferred option, in step S2, partial least squares method is used for fitting, and the fitting relationship is a linear fitting relationship.

[0015] As a further preferred embodiment, in step S2, the deep convolutional network model comprises seven combined convolutional units and four fully connected layers in a linear fashion. The number of neurons in the first fully connected layer is equal to the number of channels in the output feature map of the last convolutional unit. The number of neurons in the second, third, and fourth convolutional layers are 1 / 4, 1 / 16, and 1 of the number of neurons in the first convolutional layer, respectively.

[0016] As a further preferred embodiment, in step S2, each convolutional unit in the deep convolutional network model includes a two-dimensional convolutional layer, a max pooling layer, a BatchNorm layer, and a LeakyReLU activation layer.

[0017] As a further preferred embodiment, in steps S2 and S4, the loss function for:

[0018]

[0019] in, These are the spectral intensity values ​​predicted by the model during iterative training. is the spectral intensity value corresponding to the standard fitted curve, B is the training batch size, and n is the total number of training sets.

[0020] As a further preferred option, the process of optimizing the standard curve in step S3 is as follows:

[0021] (1) Interrupt training and save the weight model with the best performance;

[0022] (2) Input the original spectral intensity of the feature spectral lines in the training database into the optimal weight model to obtain the initial calibration standard curve;

[0023] (3) Use the initial calibration standard curve as the new training label.

[0024] As a further preferred embodiment, in steps S3 and S4, the preset condition is: the fluctuation of the loss function is less than 1%.

[0025] In summary, compared with the prior art, the above-described technical solutions conceived by this invention have the following advantages:

[0026] Beneficial effects:

[0027] 1. This invention combines laser-induced full-band spectral signals with plasma images for composite acquisition. A deep convolutional network model is then used to extract large-scale feature matrices such as brightness, area, and contrast from the plasma images, which are then stitched together to form a higher-dimensional, larger-scale deep feature information map. This yields a more accurate and robust correction model, resulting in a unified normalized spectral intensity at different energy levels. The normalized spectrum corrected by this model significantly reduces spectral fluctuations caused by sample differences, energy differences, and other accidental factors, significantly improving the elemental quantification accuracy of LIBS technology. The effect is significantly better than the original spectral quantification results. Furthermore, the method provided by this invention is simple to operate and easy to implement, requiring minimal modification to existing equipment. The model inference time is extremely short, only on the order of milliseconds, without affecting detection efficiency, fully meeting real-time requirements. It also boasts high reliability and is suitable for application in industrial environments.

[0028] 2. At the same time, the present invention uses lasers of different energies to ablate and excite the sample under test a preset number of times in order to acquire full-band spectral signals and plasma images at different energies, which can establish a comprehensive and accurate training database with good selectivity and representativeness.

[0029] 3. Furthermore, by optimizing the deep convolutional network model and loss function, this invention can obtain the required information more accurately and quickly for correction, and can better avoid overfitting problems, eliminate the influence of noise in the training data, and has good generalization ability for new samples. Attached Figure Description

[0030] Figure 1 This is a flowchart of the LIBS spectral fluctuation correction method provided in the embodiments of the present invention;

[0031] Figure 2 This is a diagram of the experimental setup used to acquire full-band spectral signals and plasma images in an embodiment of the present invention;

[0032] Figure 3 This is a schematic diagram of the fitting between the original spectral intensity of the characteristic spectral line of the element to be measured and its content in a preferred embodiment of the present invention;

[0033] Figure 4 This is a structural diagram of the deep convolutional network model used in a preferred embodiment of the present invention;

[0034] Figure 5 This is a comparison chart of the original standard curves and the calibrated calibration curves of the spectral lines of aluminum alloy Fe II 239.563nm and microalloyed steel MnII 293.931nm provided in the preferred embodiment of the present invention.

[0035] Figure 6 This is a comparison diagram of the distribution of relative spectral intensity before and after correction obtained by multiple excitations with single energy and multiple energies, provided by a preferred embodiment of the present invention;

[0036] Figure 7 These are box plots of relative spectral intensity before and after correction obtained from multiple excitations with single energy and multiple energies, provided in a preferred embodiment of the present invention.

[0037] Figure 8 This is a comparison chart of the evaluation indicators before and after correction of the relative spectral intensity obtained by multiple excitations with single energy and multiple energies using a preferred embodiment of the present invention.

[0038] In all the accompanying drawings, the same reference numerals are used to denote the same elements or structures, wherein:

[0039] 1-Nanosecond laser, 2-Spectrometer, 3-ICCD, 4-Four-dimensional stage, 5-Control unit. Detailed Implementation

[0040] 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.

[0041] like Figure 1As shown, this invention provides a LIBS spectral volatility correction method. This method addresses the problem of poor LIBS spectral stability leading to poor calibration results. By acquiring plasma images, a correction model of the plasma image on the LIBS spectral intensity is obtained, thereby achieving the purpose of correcting the LIBS spectrum. The specific steps are as follows:

[0042] S1 ablation excitation of the sample under test is performed to obtain full-band spectral signals and plasma images and to establish a training database.

[0043] S2 fits the original spectral intensity and content of the characteristic spectral line of the element to be tested in the full-band spectral signal to obtain a standard curve and use it as a training label. The training label is then input into the deep convolutional network model and trained using plasma images in the training database.

[0044] When the S3 training reaches the point where the loss function converges to a fluctuation of less than 1%, the standard curve is optimized as a new training label and input into the deep convolutional network model, and the model continues to be trained using plasma images from the training database.

[0045] S4 training is complete when the loss function converges again to a fluctuation of less than 1%, and the resulting trained deep convolutional network model is used as the calibration model.

[0046] S5 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, thereby completing the LIBS spectral fluctuation correction.

[0047] Furthermore, adopting, such as Figure 2 The experimental setup shown acquires full-band spectral signals and plasma images. The setup includes a four-dimensional stage 4, a nanosecond laser 1, a spectrometer 2, an ICCD 3, and a control unit 5. The four-dimensional stage 4 is used to place the sample to be tested. The nanosecond laser 1, spectrometer 2, and ICCD 3 are connected to the control unit 5 to emit lasers of different energies to the sample under the control of the control unit 5, and to ablate and excite it for a preset number of times. The spectrometer 2 and ICCD 3 are used to acquire full-band spectral signals and plasma images at different energies to establish a training database. For example, if aluminum alloy is selected as the sample to be tested, it is excited by pulsed lasers with energies of 40, 45, 50, 55, and 60 mJ, and full-band spectral signals and plasma images are acquired.

[0048] Furthermore, in step S2, partial least squares method is used for fitting, and the fitting relationship is a linear relationship, such as... Figure 3 As shown, this method can obtain a linear fit relationship with minimal difference from the data distribution.

[0049] Furthermore, in step S2, the structure of the deep convolutional network model is as follows: Figure 4 As shown, it consists of seven combined convolutional units and four fully connected layers in a linear fashion. Each convolutional unit includes a two-dimensional convolutional layer, a max-pooling layer, a BatchNorm layer, and a LeakyReLU activation layer. The number of neurons in the first fully connected layer is the same as the number of channels in the output feature map of the last convolutional unit. The number of neurons in the second, third, and fourth convolutional layers are 1 / 4, 1 / 16, and 1 of the number of neurons in the first convolutional layer, respectively. This allows for the full extraction of feature information contained in the image and the quantification of loss relationships, thus optimizing the model's performance.

[0050] Furthermore, in steps S2 and S4, the loss function for:

[0051]

[0052] in, These are the spectral intensity values ​​predicted by the model during iterative training. is the spectral intensity value corresponding to the standard fitted curve, B is the training batch size, and n is the total number of training sets.

[0053] Furthermore, in step S3, the process of optimizing the standard curve is as follows:

[0054] (1) Interrupt training and save the weight model with the best performance;

[0055] (2) Input the original spectral intensity of the feature spectral lines in the training database into the optimal weight model to obtain the initial calibration standard curve;

[0056] (3) Use the initial calibration standard curve as the new training label.

[0057] The solution provided by the present invention will be further described below with reference to specific embodiments.

[0058] Example 1

[0059] This embodiment describes the correction of spectral fluctuations in aluminum alloy and microalloyed steel standard samples under pulsed laser excitation at energies of 40, 45, 50, 55, and 60 mJ. The specific steps are as follows:

[0060] The contents of Fe and Mn in the selected aluminum alloy and microalloyed steel samples are shown in Table 1. The serial number represents the sample number in each type of sample.

[0061] Table 1. Fe and Mn elemental contents in aluminum alloy and microalloyed steel standard samples.

[0062]

[0063] After pretreatment of the sample, place it into Figure 2 In the device shown, after the nanosecond laser 1 excites the sample to be tested with different energies, the full-band spectral signal and plasma sub-image of the sample are acquired by the spectrometer 2 and ICCD 3 and stored in the training database.

[0064] The data is input into the calibration model obtained by the aforementioned method to construct a new calibration curve. The original standard curves for Fe II (239.563 nm) and Mn II (293.931 nm) and the calibrated calibration curves obtained after model training are shown below. Figure 5 As shown in Table 2, the evaluation indices for the calibration curves are as follows. It can be seen from the table that for the 239.563 nm spectral line of aluminum alloy Fe II, the coefficient of determination R... 2 The coefficient of determination (R²) increased from 0.0653 to 0.9801, the root mean square error (RMSE) decreased from 0.3931 to 0.0204, the mean relative error (MRE) decreased from 1.6359 to 0.0791, and the relative standard deviation (RSD) decreased from 1.7154 to 0.0552; the spectral line of the microalloyed steel MnII at 293.931 nm, and the coefficient of determination (R²) increased. 2 The metric improved from 0.8376 to 0.9940, the root mean square error (RMSE) decreased from 0.3131 to 0.0479, the mean relative error (MRE) decreased from 0.6408 to 0.1242, and the relative standard deviation (RSD) decreased from 1.5096 to 0.1408. Through calibration of the calibration model, the evaluation indicators of the calibration curve were significantly improved, spectral fluctuations were suppressed, and quantitative accuracy was greatly enhanced.

[0065] Table 2 Comparison of Evaluation Parameters for Fe and Mn Spectral Calibration Curves of Aluminum Alloys and Microalloyed Steels

[0066]

[0067] Example 2

[0068] This embodiment describes the correction of spectral fluctuations in a microalloyed steel standard sample under different energy pulsed laser excitation. The specific steps are as follows:

[0069] The Mn content in the selected microalloyed steel samples is shown in Table 1, where the serial number represents the sample number. After pretreatment, the samples were placed... Figure 2 In the device shown, after the nanosecond laser 1 excites the sample under test with different energies, the spectrometer 2 and ICCD 3 acquire the full-band spectral signal and plasma image of the sample under test and store them in the training database.

[0070] The data was input into the correction model obtained by the aforementioned method to obtain the corrected Mn II 293.931nm spectral intensity. The original and corrected spectral intensities collected after laser excitation with different pulses at the same and multiple energies are shown below. Figure 6 As shown, the box plot of the relative intensity of the spectrum is as follows: Figure 7 As shown in the figure, the evaluation index of spectral fluctuation is as follows: Figure 8 As shown.

[0071] As can be seen from the figure, regardless of whether the energy is the same or there are energy fluctuations, the spectral distribution is more concentrated, extreme values ​​are effectively removed, the relative standard deviation (RSD) and the mean standard deviation (SE mean) decrease significantly, the spectral fluctuations are well corrected, and the accuracy of LIBS detection is significantly improved.

[0072] Those skilled in the art will readily understand that the above description is merely 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 scope of protection of the present invention.

Claims

1. A method for correcting the spectral volatility of LIBS, characterized in that, The method includes the following steps: S1 ablation excitation of the sample under test is performed to obtain full-band spectral signals and plasma images and to establish a training database. S2 For the element to be tested, the original spectral intensity of the characteristic spectral line of the element to be tested in the full-band spectral signal is fitted with its content to obtain a standard curve and use it as a training label. The training label is then input into the deep convolutional network model and trained using plasma images in the training database. When the loss function of S3 training converges to the preset condition, the standard curve is optimized as a new training label and input into the deep convolutional network model, and the plasma images in the training database are used to continue training it. When the loss function converges to the preset condition again during training, the training is complete and the trained deep convolutional network model is used as the correction model. S5 inputs the full-band spectral signal and plasma image of the sample to be tested into the correction model, and linearly fits the corrected spectral intensity of the characteristic spectral lines of the element to be tested with its content to obtain the calibration curve, thereby completing the LIBS spectral fluctuation correction.

2. The LIBS spectral fluctuation correction method as described in claim 1, characterized in that, In step S1, lasers of different energies are used to ablate the sample under test a preset number of times in order to acquire full-band spectral signals and plasma images at different energies.

3. The LIBS spectral fluctuation correction method as described in claim 1, characterized in that, In step S2, partial least squares method is used for fitting, and the fitting relationship is a linear fitting relationship.

4. The LIBS spectral fluctuation correction method as described in claim 1, characterized in that, In step S2, the deep convolutional network model consists of seven combined convolutional units and four fully connected layers in a linear fashion. The number of neurons in the first fully connected layer is the number of channels in the output feature map of the last convolutional unit. The number of neurons in the second, third, and fourth convolutional layers are 1 / 4, 1 / 16, and 1 of the number of neurons in the first convolutional layer, respectively.

5. The LIBS spectral fluctuation correction method as described in claim 4, characterized in that, In step S2, each convolutional unit in the deep convolutional network model includes a two-dimensional convolutional layer, a max pooling layer, a BatchNorm layer, and a LeakyReLU activation layer.

6. The LIBS spectral fluctuation correction method as described in claim 1, characterized in that, In steps S2 and S4, the loss function for: in, These are the spectral intensity values ​​predicted by the model during iterative training. is the spectral intensity value corresponding to the standard fitted curve, B is the training batch size, and n is the total number of training sets.

7. The LIBS spectral fluctuation correction method as described in claim 1, characterized in that, In step S3, the process of optimizing the standard curve is as follows: (1) Interrupt training and save the weight model with the best performance; (2) Input the original spectral intensity of the feature spectral lines in the training database into the optimal weight model to obtain the initial calibration standard curve; (3) Use the initial calibration standard curve as the new training label.

8. The LIBS spectral fluctuation correction method according to any one of claims 1 to 7, characterized in that, In steps S3 and S4, the preset condition is: the fluctuation of the loss function is less than 1%.

Images