A method for improving the classification accuracy of laser probes by utilizing spectral feature augmentation

Abstract

This invention belongs to the technical field of laser probe elemental analysis and discloses a method for improving the classification accuracy of laser probes by expanding spectral features. The method includes the following steps: S1, acquiring plasma spectra using a laser probe spectral acquisition device; S2, averaging the plasma spectra and selecting analytical lines and their corresponding start and end wavelengths from the resulting flat spectrum; S3, extracting spectral intensity, peak area, peak half-width at half-maximum, peak standard deviation, peak signal-to-background ratio, and peak signal-to-noise ratio features from the original spectrum; S4, expanding the input feature vector using the above features to obtain an expanded mixed spectral feature vector; S5, training a classification model based on the expanded mixed spectral features using a classification algorithm; S6, inputting the mixed spectral features of the test set into the classification model, and the classification model outputs the classification result, completing the classification. This invention effectively expands the traditional spectral feature vector, which is mainly based on spectral intensity, improving the representational ability and classification accuracy of the spectral feature vector.

Description

Technical Field

[0001] This invention belongs to the technical field of laser probe elemental analysis, and more specifically, relates to a method for improving the classification accuracy of laser probes by utilizing the expansion of spectral features. Background Technology

[0002] Laser probe microscopy, also known as laser-induced breakdown spectroscopy (LIBS), is a type of atomic emission spectroscopy. It typically uses a high-energy-density pulsed laser to ablate a sample, generating a plasma spectrum. The plasma spectrum is then analyzed to determine the qualitative and quantitative composition and content of elements in the sample. Currently, laser probe microscopy is widely used in Mars exploration, environmental pollution detection, metallurgical analysis, and rock and ore exploration.

[0003] Qualitative analysis using laser probe microanalysis typically combines spectral features with machine learning algorithms, such as principal component analysis, linear discriminant analysis, support vector machine, and K-nearest neighbor algorithms. Spectral intensity is directly positively correlated with the elemental content in a sample; generally, the higher the elemental content in the sample, the stronger the spectral intensity of its characteristic spectral lines. Therefore, spectral intensity features are most commonly used in qualitative and quantitative analysis using laser probe microanalysis. In basic calibration methods, the spectral intensity of a single analytical line is usually fitted to the elemental concentration. In multivariate regression analysis, the spectral intensities of multiple analytical lines are typically combined with a regression algorithm for fitting and modeling. In qualitative analysis based on laser probe microanalysis, traditional methods also utilize the spectral intensities of multiple analytical lines. However, spectral intensity is only one feature of the original spectrum. While it can indirectly reflect the elemental concentration in the sample, its characterization ability is easily weakened due to the limited number of characteristic spectral lines, interference from overlapping adjacent lines, matrix effects, and small differences among similar samples, thus affecting the overall analytical performance of laser probe microanalysis.

[0004] Currently, some research has been conducted by relevant personnel in this field. For example, the paper "In situ classification of rocks using stand-off laser-induced breakdown spectroscopy with a compact spectrometer" (J. Anal. At. Spectrom. (2020) 35: 1486-1486) discloses a classification method that uses the spectral intensity features of 49 analytical lines combined with the LDA algorithm. This method achieves a classification accuracy of 97.8% for 15 different rock categories. Since the 15 rock samples have large differences, the spectral intensity features alone can achieve a good classification accuracy. However, the paper "Application of laser-induced breakdown spectroscopy (LIBS) coupled with PCA for rapid classification of soil samples in geothermal areas" (Anal Bioanal Chem 411(13)(2019) 2855-2866) only achieved a classification accuracy of 80% when using spectral intensity features to classify soil samples. While the algorithm certainly has some impact on classification performance, the weak characterization ability of spectral intensity features is another important reason. In addition to spectral intensity characteristics, features such as spectral intensity ratios and peak areas are also increasingly being applied to laser probe technology. For example, patent CN104483292A, published on April 1, 2015, entitled "A Method for Improving the Accuracy of Laser Probe Analysis Using the Multi-Line Ratio Method," discloses a method for improving the quantitative analysis accuracy of laser probes using the multi-line ratio method. The paper "Research on Soil Physicochemical Information Detection Method Based on Laser-Induced Breakdown Spectroscopy" (Zhejiang University, 2016) discloses a quantitative analysis method for detecting soil physicochemical information using peak area combined with the PLSR algorithm. These methods have respectively transformed and replaced traditional spectral intensity characteristics, and all have improved the quantitative analysis performance based on laser probe technology to some extent. However, none of these studies have fundamentally changed the analytical model of using a single feature for qualitative or quantitative analysis. Furthermore, the insufficient characterization ability of a single feature due to the aforementioned reasons still has a significant impact on analytical performance. Correspondingly, in the field of laser probes, in addition to spectral intensity characteristics, laser probe spectra also contain many other physical quantities that can characterize specific spectral performance, such as peak half width at half maximum, peak standard deviation, peak signal-to-background ratio, and signal-to-noise ratio. Utilizing these characteristics to expand traditional spectral intensity characteristics has great potential to improve the characterization ability of spectral feature vectors. Summary of the Invention

[0005] To address the shortcomings of existing analytical methods and meet practical application needs, this invention provides a method for improving the classification accuracy of laser probes by expanding spectral features. Based on the characteristics of existing laser probe classification methods, this method designs and proposes a method with good classification performance, strong generalization ability, and improved accuracy through effective expansion of spectral features. The expanded features used in this method include six features: spectral intensity, peak area, peak half-width at half-maximum, peak standard deviation, peak signal-to-background ratio, and peak signal-to-noise ratio. Therefore, it can comprehensively and stably characterize spectral properties, thereby improving the representational ability of feature vectors and thus enhancing the classification performance of the laser probe classification model.

[0006] To achieve the above objectives, the present invention provides a method for improving the classification accuracy of laser probes by utilizing spectral feature augmentation, characterized in that the method includes the following steps:

[0007] S1, using a laser probe spectral acquisition device to collect plasma spectra;

[0008] S2, average the plasma spectrum and select the analytical line and its corresponding start and end wavelengths from the obtained flat spectrum;

[0009] S3 extracts spectral intensity, peak area, peak half width at half maximum (FWHM), peak standard deviation, peak signal-to-background ratio (SNR), and peak signal-to-noise ratio (SNR) features from the original spectrum.

[0010] S4. Using the above features, the input feature vector is augmented to obtain the augmented mixed spectral feature vector.

[0011] S5, The expanded mixed spectral features are combined with a classification algorithm to train a classification model based on the mixed spectral features;

[0012] S6. Input the mixed spectral features of the test set into the classification model, and the classification model outputs the classification result to complete the classification.

[0013] Furthermore, step S3 specifically includes the following sub-steps:

[0014] S31. Based on the analysis line obtained in S2, query the wavelength position of the analysis line in the original spectral data, and retrieve the spectral intensity of the three adjacent points before and after it. Then, extract the maximum intensity in the neighborhood as the spectral intensity of the analysis line, thereby obtaining the spectral intensity feature F(SI).

[0015] S32. Based on the analysis line and the start and end wavelengths of the peak obtained in S2, extract the spectral wavelength data and spectral intensity data within the start and end bands. Then, with wavelength as the independent variable and intensity as the dependent variable, perform a three-segment fitting of the peak in the band. The first segment is from the starting wavelength to the analysis line wavelength, and the second segment is from the analysis line wavelength to the end wavelength. The fitting functions are shown in formula (1) and formula (2) respectively. Then, integrate the above formulas to find the area of ​​the peak in the band. Finally, sum the areas of the two segments to obtain the area of ​​the full peak F(SPA).

[0016] I1(W) = A1W 3 +B1W 2 +C1W+D1 (1)

[0017] I²(W) = A²W 3 +B2W 2 +C2W+D2 (2)

[0018] S33, based on the geometric shape of the spectral peak, its full width at half maximum (FWHM) characteristic F(FWHM) is obtained. First, the maximum value I(W0) and minimum value I(W3) of its spectral intensity are obtained within the start and end wavelength range. Then, I(W) is obtained according to formula (3). mid Finally, let I1(W) and I2(W) be equal to I(W) respectively. mid The real roots W1 and W2 can be obtained respectively. The absolute value of the difference between the two real roots ΔW is the half width at half maximum (WHM) of the spectral peak, as shown in formula (4). Thus, the half width at half maximum (WHM) characteristic F(FWHM) of the spectral peak can be obtained.

[0019]

[0020] ΔW=|W2-W1| (4)

[0021] S34, extract the peak intensity data within the start and end wavelength range, then calculate the standard deviation of the peak intensity to obtain the standard deviation characteristic F(SD) of the peak.

[0022] In step S35, extract the spectral intensities of the starting wavelength and the two points preceding it, then extract the spectral intensities of the ending wavelength and the two points following it. Combine the spectral intensities of these six points into a single sequence, and calculate the mean and standard deviation of this sequence. The resulting values ​​represent the background and noise of the spectral peak. Next, divide the spectral intensities obtained in step S31 by the background and noise levels respectively to obtain the signal-to-background ratio (SBR) and signal-to-noise ratio (SNR) characteristics of the spectrum, denoted as F(SBR) and F(SNR), respectively.

[0023] Furthermore, in step S3, the spectral features include, but are not limited to, spectral intensity, peak area, peak half width at half maximum, peak standard deviation, peak signal-to-background ratio, and peak signal-to-noise ratio. Other features extracted from the spectrum or features that have undergone certain mathematical transformations, such as spectral intensity after internal standardization and spectral intensity ratio, can also be used for feature expansion.

[0024] Furthermore, in step S32, the peak area is obtained by integrating the segmented peaks three times.

[0025] Furthermore, in step S35, the background and noise of the spectrum are estimated by estimating the band formed by the three points before the starting point of the spectral peak and the three points after the ending point.

[0026] Furthermore, in step S4, the feature vectors extracted in S3 are vertically concatenated to achieve effective expansion of the spectral feature vectors based on multiple mixed spectral features.

[0027] Furthermore, in step S5, the classification algorithm is any one of the following algorithms: linear discriminant analysis algorithm, support vector machine algorithm, neural network algorithm, and K-nearest neighbor algorithm.

[0028] Furthermore, the spectral acquisition device includes a laser, a spectrometer, a computer, a lithium battery, a controller, a acquisition head, and a focusing lens. The laser is connected to the controller, which is connected to the computer, the lithium battery, and the spectrometer. The spectrometer is connected to the acquisition head via an optical fiber and is located behind it. The focusing lens is located in front of the laser.

[0029] Furthermore, the lithium battery is used to power the entire laser probe system, the computer is used to implement the hardware and software control and data analysis of the laser probe system, the laser is used to generate high-energy-density pulsed laser at a fixed frequency, the laser is focused on the sample surface by the focusing lens, the focused laser ablates the sample to generate plasma emission light, which is collected by the acquisition head and transmitted to the spectrometer via optical fiber, the spectrometer converts the optical signal into an electrical signal and outputs spectral data to the computer.

[0030] In summary, compared with the prior art, the method for improving laser probe classification accuracy by utilizing spectral feature augmentation provided by the present invention has the following beneficial effects:

[0031] 1. The expanded hybrid spectral features describe and characterize the spectrum from six dimensions: spectral intensity, peak area, peak half width at half maximum, peak standard deviation, peak signal-to-background ratio, and peak signal-to-noise ratio. This feature construction method is more comprehensive, richer, and more robust than traditional single features such as intensity features or peak area features, which can make the feature vector more representative and thus improve the classification performance.

[0032] 2. By utilizing the extended features, more effective features can be extracted in classification applications with fewer samples and fewer feature spectral lines, which appropriately increases the feature dimension and thus avoids overfitting and underfitting of the classification model.

[0033] 3. The spectral feature expansion is formed by vertically splicing multiple single features, and the feature expansion process is simple and convenient. In addition, there are many types of features that can be used for expansion, including but not limited to spectral intensity, peak area, peak half width at half maximum, peak standard deviation, peak signal-to-background ratio, and peak signal-to-noise ratio. In actual classification operations, each feature can be deleted or added depending on its independent classification performance and complementary classification performance, and the expansion method is flexible and diverse.

[0034] 4. The expanded spectral features can be combined with various classification algorithms, including but not limited to linear discriminant analysis, support vector machine, neural network, and K-nearest neighbor algorithms, which can avoid the non-positive definite problem of traditional full-spectrum intensity feature matrices and improve the flexibility and practicality of the features. Attached Figure Description

[0035] Figure 1 This is a flowchart illustrating a method for improving the classification accuracy of laser probes by utilizing spectral feature augmentation, as provided by the present invention.

[0036] Figure 2 This is a schematic diagram of the spectral acquisition system involved in the method for improving the classification accuracy of laser probes by utilizing spectral feature augmentation provided by the present invention;

[0037] Figure 3 It is to utilize Figure 2 The average spectrum acquired by the spectral acquisition system shown;

[0038] Figure 4 This is a schematic diagram showing the extraction of the spectral peaks, start points, end points, and peak areas of the iron analysis lines in the average spectrum.

[0039] Figure 5 This is a schematic diagram of the spectral feature extraction and expansion involved in a method for improving the classification accuracy of laser probes by utilizing spectral feature expansion provided by the present invention.

[0040] Figure 6 The diagram shows the classification results obtained by using the traditional strength method for rock classification.

[0041] Figure 7 The diagram shows the classification results obtained by classifying sedimentary rock samples using the 15 methods provided by this invention.

[0042] Table 1 is a detailed table of classification results obtained by classifying 15 sedimentary rock samples using the method provided in this invention.

[0043] In all the accompanying drawings, the same reference numerals are used to denote the same elements or structures, wherein 1-sample, 2-computer, 3-controller, 4-lithium battery, 5-laser, 6-wire, 7-spectrometer, 8-optical fiber, 9-focusing lens, 10-collection head. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail 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. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0045] Please see Figure 1 , Figure 2 , Figure 3 , Figure 4 and Figure 5 This invention provides a method for improving the classification accuracy of laser probes by expanding spectral features. The laser probe classification method first requires the use of a laser probe system to collect plasma spectra, and then extracts spectral intensity, peak area, peak half width at half maximum, peak standard deviation, peak signal-to-background ratio, and peak signal-to-noise ratio features one by one from the spectrum. Then, the above individual features are vertically stitched together to achieve expansion. Finally, the expanded spectral features are combined with a classification algorithm to achieve sample identification and classification.

[0046] The laser probe classification method mainly includes the following steps:

[0047] S1, using a laser probe spectral acquisition device to collect plasma spectra;

[0048] Specifically, the sample 1 to be tested is placed in the laser probe spectral acquisition system, and the plasma spectrum of sample 1 is acquired using the system. The laser probe system involved in the method for improving the classification accuracy of laser probes by utilizing spectral feature expansion provided by this invention includes sample 1, computer 2, controller 3, lithium battery 4, laser 5, wire 6, spectrometer 7, optical fiber 8, focusing lens 9, and acquisition head 10.

[0049] The lithium battery 4 is used to power the entire laser probe system. The controller 3 is used to stabilize and modulate the voltage. The computer 2 is used for the hardware and software control and data analysis of the laser probe system. The laser 5 is used to generate high-energy-density pulsed laser at a fixed frequency. The spectrometer 7 is used to separate the polychromatic light signal of the plasma into spectral data of each wavelength. The focusing lens 9 is used to focus the laser generated by the laser 5 to improve the energy density of the laser. The acquisition head 10 is used to collect the plasma emitted light. The optical fiber 8 is used for the transmission of the plasma emitted light.

[0050] The lithium battery 4 is connected to the controller 3 via the wire 6. The controller 3 is connected to the computer 2, the laser 5, and the spectrometer 7. The spectrometer is connected to the acquisition head 10 via the optical fiber 8. The acquisition head 10 is located at the front end of the spectrometer 7 and is placed at a certain angle to the front plane of the sample 1. The focusing lens 9 is located at the horizontal front end of the laser 5. The front plane of the sample 1 is located at the focal point of the focusing lens 9. The focal point of the acquisition head 10 and the focal point of the focusing lens 9 coincide on the surface of the sample 1.

[0051] The high-energy-density pulsed laser generated by the laser 5 is focused onto the sample 1 by the focusing lens 9. The plasma emitted by the laser ablation of the sample 1 is collected by the acquisition head 10 and transmitted to the spectrometer 7 via the optical fiber 8. The spectrometer 7 converts the polychromatic plasma light signal into a separated spectrum and outputs the spectral data to the computer 2.

[0052] S2, average the plasma spectrum, and select the analytical line and its corresponding start and end wavelengths from the resulting flat spectrum. (See [link to relevant documentation]). Figure 3 Specifically, it includes the following steps:

[0053] S21, Read all spectral data. Let the total number of spectra obtained in the experiment be m. The spectral resolution of the spectrometer 7 is n pixels. Within its effective band, each pixel corresponds to a wavelength.

[0054] S22, by averaging the spectral intensities of the same wavelength in all spectra, an average spectrum can be obtained.

[0055] S23. Analytical lines are selected using the average spectrum. N peaks with strong spectral intensity, large signal-to-background ratio, and no obvious self-erosion are chosen, and their peak wavelengths are recorded as analytical lines W. c At the same time, the starting wavelength W of the spectral peak is specified. s and termination wavelength W e , with W s1 To W e1 For example, the total number of pixels within its band is M.

[0056] S3. Extract spectral intensity, peak area, peak half-width at half-maximum, peak standard deviation, peak signal-to-background ratio, and peak signal-to-noise ratio features from the original spectrum. Let these features be F(SI), F(SPA), F(FWHM), F(SD), F(SBR), and F(SNR), respectively. Please refer to [link to relevant documentation]. Figure 5 With the first analysis line W c1 Taking the spectral peak as an example, the specific steps include:

[0057] S31, In the raw spectral data, query the wavelength W of the analysis line. c1 The position is determined, and the spectral intensities of the three points before and three points after it are retrieved. Then, the maximum intensity value in the neighborhood is extracted as the spectral intensity of the analysis line, thus obtaining the spectral intensity feature F1(SI) of a single spectral peak. The intensity feature dimension of a single spectrum is N×1.

[0058] S32, Extract start and end bands [W] s1 W e1 Spectral wavelength data W within ] M×1 and spectral intensity data I M×1 Then, using wavelength W as the independent variable and intensity I as the dependent variable, a segmented cubic fitting is performed on the spectral peak of this band. The initial wavelength W... s1 To the analysis line wavelength W c1 The first segment is within the range, and the analysis line wavelength W c1 To the termination wavelength W e1 The range is the second segment, and the fitting functions are shown in formula (1) and formula (2) respectively. Then, the fitting functions are integrated according to the above formulas to find the peak area in the range of each spectrum segment. Then, the areas of the two segments are summed to obtain the total area F1 (SPA) of a single peak. The characteristic dimension of the peak area of ​​a single spectrum is N×1.

[0059] I1(W) = A1W 3 +B1W 2 +C1W+D1 (1)

[0060] I²(W) = A²W 3 +B2W 2 +C2W+D2 (2)

[0061] S33, based on the geometric shape of the spectral peak, its full width at half maximum (FWHM) characteristic F(FWHM) is obtained. First, the maximum value I(W0) and minimum value I(W3) of its spectral intensity are obtained within the start and end wavelength range. Then, I(W) is obtained according to formula (3). mid Finally, let I1(W) and I2(W) be equal to I(W) respectively. midThe real roots W1 and W2 of the cubic fitting function in their respective band ranges can be obtained respectively. The absolute value of the difference between the two real roots is the half-width at half maximum (FWHM) of the spectral peak, as shown in formula (4). Thus, the FWHM feature of a single spectral peak F1 (FWHM) can be obtained. The FWHM feature dimension of a single spectrum peak is N×1.

[0062]

[0063] ΔW=|W2-W1| (4)

[0064] S34, Extract the start and end wavelength range [W] s1 W e1 All spectral peak intensity data within ] M×1 Then calculate its standard deviation to obtain the standard deviation feature F1(SD) of the spectral peak. The standard deviation feature dimension of a single spectrum is N×1.

[0065] S35, Extract the initial wavelength W s1 The spectral intensities of the first two points are used to extract the termination wavelength W. s1 The spectral intensities of the next two points are then calculated, and the spectral intensities of the above six points are combined into a single series. The mean and standard deviation of this series are then calculated, and the resulting values ​​represent the background and noise of the spectral peak. Subsequently, by dividing the spectral intensity by the background and noise respectively, the signal-to-background ratio (SBR) and signal-to-noise ratio (SNR) characteristics of the spectrum can be obtained, which are F1(SBR) and F1(SNR), respectively. The dimensions of the SBR and SNR characteristics of a single spectrum peak are both N×1.

[0066] S4. Using the above features, the input feature vector is augmented to obtain the augmented mixed spectral feature vector.

[0067] Specifically, please refer to Figure 5 The extracted feature vectors F(SI), F(SPA), F(FWHM), F(SD), F(SBR), and F(SNR) are vertically concatenated to effectively expand the spectral feature vector based on multiple mixed features. Taking a single spectrum as an example, if the number of analysis lines is N, then the dimension of the expanded spectral feature vector is 6N×1.

[0068] S5, The expanded mixed spectral features are combined with a classification algorithm to train a classification model based on the mixed spectral features;

[0069] Specifically, assuming there are m spectra in total, the dimension of the expanded feature matrix of all spectra is 6N×m, and the ratio of training set to test set data size is 3:1. The classification algorithm can be any of the following: Linear Discriminant Analysis (LDA), Neural Network Algorithm (CNN), Support Vector Machine (SVM), and K-Nearest Neighbors (KNN). The expanded spectral features are then incorporated into the classification algorithm, and after training, a laser probe classification model based on the expanded spectral features can be obtained.

[0070] S6. Input the mixed spectral features of the test set into the classification model, and the classification model outputs the classification result to complete the classification.

[0071] Example 1

[0072] The first embodiment of the present invention provides a method for improving the classification accuracy of laser probes by utilizing spectral feature augmentation, which mainly includes the following steps:

[0073] Step 1: Sample Preparation and Plasma Spectroscopy Acquisition. This example uses 15 sedimentary rock samples, with 5 samples of each type. To overcome potential spectral instability caused by uneven sample surfaces, each rock was polished before spectral acquisition to ensure at least one flat surface. In the experiment, two points were collected for each sample, and 20 plasma spectra were acquired from each point. Therefore, a total of 200 spectra were acquired for each type of sedimentary rock sample, and a total of 3000 spectra were acquired in the entire classification experiment. In this example, the laser's single-pulse energy was approximately 6.5 mJ, the frequency was 10 Hz, and the wavelength was 1064 nm.

[0074] Step two: Spectral averaging and analysis line selection. The spectrometer used in this embodiment has a wavelength range of approximately 260–430 nm and a detector pixel count of 4096 pixels, with each pixel or wavelength corresponding to a spectral intensity. In this step, the 3000 spectra acquired in Step one are averaged according to wavelength to obtain an average spectrum. This average spectrum is then used to select analysis lines, start wavelengths, and end wavelengths. In this embodiment, a total of 60 analysis lines are selected, each corresponding to a start wavelength and an end wavelength.

[0075] Step 3: Extraction of spectral intensity, peak area, peak half-width at half-maximum, peak standard deviation, peak signal-to-background ratio, and peak signal-to-noise ratio features. First, based on the position of the analytical line, the spectral intensity of its neighboring range is retrieved, and the maximum value within its neighborhood is taken as the spectral intensity of that analytical line. In this embodiment, there are 3000 spectra and 60 analytical lines, therefore the spectral intensity feature can be obtained as F. 60×3000(SI); then, based on the analytical line wavelength and start and end bands of each spectral peak, the area of ​​each spectral peak and the total area of ​​the spectral peaks are extracted using piecewise cubic fitting and integration methods to obtain the spectral peak area characteristics, which in this embodiment can be expressed as F. 60×3000 (SA); then, based on the geometry of the spectral peak, the wavelength of the analytical line, and the start and end wavelengths, the full width at half maximum (FWHM) value of each analytical line is calculated to obtain the FWHM characteristic of the spectral peak. In this embodiment, it can be expressed as F 60×3000 (FWHM); then, based on the start and end wavelengths of the spectral peak, all spectral intensities within that band are extracted and their standard deviations are calculated. This standard deviation value is used as the standard deviation characteristic of the spectral peak. In this embodiment, the standard deviation characteristic is F... 60×3000 (SD); Finally, extract the spectral intensity of the starting wavelength and the two points before it, then extract the spectral intensity of the ending wavelength and the two points after it. Combine the spectral intensities of these six points into a single series, and calculate the mean and standard deviation of this series. The resulting values ​​represent the background and noise of the spectral peak, respectively. Then, by dividing the spectral intensity by the background and noise, the signal-to-background ratio and signal-to-noise ratio characteristics of the spectrum can be obtained. In this embodiment, they are respectively F... 60×3000 (SBR), F 60×3000 (SNR).

[0076] Step four, expansion of spectral features. The individual features F obtained in step three are then expanded. 60×3000 (SI), F 60×3000 (SA), F 60×3000 (FWHM), F 60×3000 (SD), F 60×3000 (SBR) and F 60×3000 (SNR) is vertically concatenated, that is, the feature vectors are arranged row by row to form a new, higher-dimensional, and more feature-rich mixed spectral feature vector. In this embodiment, the resulting expanded spectral feature vector is F. 360×3000 .

[0077] Step 5: Combine the expanded mixed spectral features with the Linear Discriminant Analysis (LDA) algorithm to establish a classification model. This implementation combines the expanded spectral features with the LDA algorithm to achieve laser probe classification. Specifically, the mixed feature data is divided according to a training set to test set ratio (i.e., 3:1). Then, the mixed spectral feature matrix of the first 150 spectra of each sample is input into the LDA algorithm for training to generate a multi-classification model of unknown samples based on the mixed spectral features.

[0078] Step 6: Input the expanded spectral features of the above 3000 spectra of the test set into the classification model, and the classification model outputs the classification result.

[0079] The classification results of the laser probe classification method based on expanded mixed spectral features (spectral feature expansion method) and the traditional classification method based on a single feature such as spectral intensity features (intensity method) provided by this invention are as follows: Figure 6 and Figure 7 As shown in the figure above, the average classification accuracy of the 15 sedimentary rock samples using the traditional intensity method is 85%, which is relatively low. However, when the spectral feature augmentation method provided by this invention is used, the average classification accuracy is improved to 94.83%. Specifically, the classification accuracy of individual samples, such as limestone and conglomerate, is improved from 40% and 50% to 90% and 95%, respectively, representing increases of 50% and 45%, indicating a significant improvement in classification performance.

[0080] 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 improving the classification accuracy of a laser probe using spectral feature augmentation, characterized in that, The method comprises the following steps: S1, collecting a plasma spectrum by using a laser probe spectrum acquisition device; S2, averaging the plasma spectrum, and selecting an analysis line and its corresponding start and end wavelengths in the obtained flat spectrum; S3, extracting spectral intensity, spectral peak area, spectral peak half width, spectral peak standard deviation, spectral peak signal-to-background ratio, and spectral peak signal-to-noise ratio characteristics in the original spectrum; The step S3 specifically comprises the following sub-steps: S31, according to the analysis line obtained in S2, querying the wavelength position of the analysis line in the original spectrum data, retrieving the spectral intensities of the three adjacent points before and after the analysis line, and then extracting the maximum intensity in the neighborhood as the spectral intensity of the analysis line, so as to obtain the spectral intensity characteristic F(SI); S32, according to the analysis line obtained in S2 and the start and end wavelengths of the spectral peak where the analysis line is located, extracting the spectral wavelength data and spectral intensity data in the start and end wavelength ranges, and then performing piecewise cubic fitting on the wavelength range spectral peak with wavelength as the independent variable and intensity as the dependent variable, wherein the first segment is in the range from the start wavelength to the analysis line wavelength, and the second segment is in the range from the analysis line wavelength to the end wavelength, the fitting functions are shown in formulas (1) and (2) respectively, then the areas of the above formulas are integrated respectively, and then the sum of the areas of the two segments is obtained to obtain the full spectral peak area characteristic F(SPA); S33, the half width characteristic F(FWHM) of the spectral peak is obtained according to the geometric shape of the spectral peak, first the maximum value I(W0) and the minimum value I(W3) of the spectral intensity in the start and end wavelength ranges are obtained, then I(Wmid) is obtained according to formula (3); then I1(W) and I2(W) are respectively equal to I(Wmid), the real roots W1 and W2 can be obtained respectively, and the absolute value ΔW of the difference between the two real roots is the half width of the spectral peak, as shown in formula (4), and thus the half width characteristic F(FWHM) of the spectral peak is obtained; I1(W) = A1W 3 + B1W 2 + C1W + D1(1) I2(W) = A2W 3 + B2W 2 + C2W + D2(2) ΔW = |W2-W1| (4) (3) S34, the spectral intensity data in the start and end wavelength ranges are extracted, and then the standard deviation of the spectral peak intensity is calculated to obtain the standard deviation characteristic F(SD) of the spectral peak; S35, the spectral intensities of the start wavelength and the two points before it are extracted, and then the spectral intensities of the end wavelength and the two points after it are extracted, then the spectral intensities of the above six points are combined into a number sequence, and the mean and standard deviation of the number sequence are calculated, and the obtained values are respectively the background and noise of the spectral peak; then, the spectral intensity obtained by S31 is divided by the background and noise respectively to obtain the signal-to-background ratio and signal-to-noise ratio characteristics of the spectrum, which are F(SBR) and F(SNR) respectively; S4, the above characteristics are used to perform feature expansion on the input feature vector to obtain an expanded mixed spectrum feature vector; S5, combining the mixed spectrum feature obtained after expansion with a classification algorithm to train a classification model based on the mixed spectrum feature; S6, inputting the mixed spectrum feature of the test set into the classification model, and outputting a classification result by the classification model, and completing the classification. In step S3, the spectral characteristics further include the spectral intensity or spectral intensity ratio after internal standard correction.

2. The method of claim 1, wherein the spectral feature is augmented to improve the classification accuracy of the laser probe. In step S32, the spectral peak area is obtained by piecewise cubic fitting of the spectral peak.

3. The method of claim 1, wherein the spectral feature is augmented to improve the classification accuracy of the laser probe. ​ 4. The method of claim 1, wherein the spectral feature is augmented to improve the classification accuracy of the laser probe. In step S35, the background and noise of the spectrum are obtained by estimating the wave band composed of the three points before the starting point of the spectral peak and the three points after the ending point.

5. The method of claim 1, wherein the spectral feature is augmented to improve the classification accuracy of the laser probe. In step S4, the feature vectors extracted in S3 are longitudinally spliced to achieve effective expansion of the spectral feature vectors based on multiple mixed spectral features.

6. The method of claim 1, wherein the spectral feature is augmented to improve the classification accuracy of the laser probe. In step S5, the classification algorithm is any one of the following algorithms: linear discriminant analysis algorithm, support vector machine algorithm, neural network algorithm, K nearest neighbor algorithm.

7. The method of claim 1-6, wherein the method is used to improve the classification accuracy of laser probe by augmenting spectral features. The spectrum acquisition device comprises a laser, a spectrometer, a computer, a lithium battery, a controller, a collection head and a focusing mirror; the laser is connected to the controller; the controller is connected to the computer, the lithium battery and the spectrometer; the spectrometer is connected to the collection head through an optical fiber and is located behind the collection head; and the focusing mirror is located at the front end of the laser.

8. The method of claim 7, wherein the spectral features are used to improve the classification accuracy of the laser probe. The lithium battery is used to power the entire laser probe system; the computer is used to realize software and hardware control and data analysis of the laser probe system; the laser is used to generate high-energy density pulsed laser with fixed frequency; the high-energy density pulsed laser is focused on the sample through the focusing mirror; the plasma emission light generated by the focused laser ablation sample is collected by the collection head and transmitted to the spectrometer through the optical fiber; the spectrometer converts the optical signal into an electrical signal and outputs the spectral data to the computer.

Images