Libs cross-instrument spectral library transfer method and system based on spectral calibration and deep learning
By combining spectral calibration with deep learning, the nonlinear differences in LIBS across instrument spectral libraries are solved, enabling high-precision spectral library migration and cross-instrument application, reducing detection costs, improving material classification and quantification accuracy, and making it suitable for LIBS analysis in multiple fields.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HEFEI INSTITUTE OF PHYSICAL SCIENCE CHINESE ACADEMY OF SCIENCES
- Filing Date
- 2026-06-26
- Publication Date
- 2026-07-24
AI Technical Summary
Existing LIBS cross-instrument calibration methods only support linear correction, cannot fit nonlinear spectral differences between instruments, have insufficient spectral reconstruction accuracy, poor reusability of cross-instrument spectral libraries, high detection costs, and low material classification/quantification accuracy.
A method based on spectral calibration and deep learning is adopted. Baseline correction and wavelength alignment are performed by asymmetric least squares method. Combined with sliding window local regression and tunable multilayer perceptron model, nonlinear mapping and intensity calibration of the spectrum are achieved to eliminate instrument differences.
It enables high-precision cross-instrument transfer of spectral libraries, reduces detection costs, supports rapid deployment of portable devices, and improves the accuracy of material classification and quantification. It is applicable to fields such as geological exploration, industrial testing, and planetary exploration.
Smart Images
Figure CN122448826A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of laser-induced breakdown spectroscopy (LIBS) analysis technology, and in particular to a method and system for LIBS cross-instrument spectral library transfer based on spectral calibration and deep learning. Background Technology
[0002] Laser-induced breakdown spectroscopy (LIBS) is a rapid, in-situ, and minimally destructive elemental analysis technique. With its advantages of simple sample pretreatment, near real-time detection, and applicability to multiphase samples (solid, liquid, and gas), it has been widely applied in fields such as geological exploration, agricultural testing, industrial process control, biomedical diagnostics, and planetary exploration.
[0003] In practical applications, LIBS instruments are divided into high-end laboratory equipment and portable field devices. These two types of instruments differ significantly in optical configuration, detector response, spectral resolution, laser excitation parameters, and light collection methods. This leads to issues such as baseline shift, wavelength misalignment, intensity distortion, and peak shape distortion in the spectra acquired by different instruments, resulting in cross-instrument domain differences. These differences prevent the direct reuse of spectral libraries and machine learning models built for a single instrument. Data must be reacquired and a dedicated spectral library built for each instrument, which is time-consuming, labor-intensive, and costly, severely limiting the large-scale deployment and cross-device application of LIBS technology.
[0004] Existing LIBS cross-instrument calibration transfer methods, represented by direct normalization (DS) and piecewise direct normalization (PDS), are based on linear / piecewise linear assumptions. They can only correct linear spectral differences between instruments and cannot fit nonlinear spectral distortions caused by plasma excitation and detector response. They suffer from low spectral reconstruction accuracy and severe performance degradation of downstream classification / quantification models.
[0005] In summary, the existing technical problems are as follows: 1. Existing LIBS cross-instrument calibration methods only support linear correction and cannot fit nonlinear spectral differences between instruments, resulting in insufficient spectral reconstruction accuracy. 2. LIBS spectral libraries have poor cross-instrument reusability, requiring repeated library construction, resulting in high detection costs and low efficiency; 3. The accuracy of material classification / quantification is low after cross-instrument migration, which cannot meet the accuracy requirements of actual testing.
[0006] Currently, there is no LIBS cross-instrument solution that combines physical-guided spectral calibration with deep learning nonlinear transfer, making it difficult to achieve high-precision and robust cross-instrument spectral conversion and model deployment. Summary of the Invention
[0007] To overcome the shortcomings of the prior art, the present invention provides a LIBS cross-instrument spectral library transfer method and system based on spectral calibration and deep learning.
[0008] To achieve the above objectives, the present invention adopts the following technical solution, including: The LIBS cross-instrument spectral library transfer method based on spectral calibration and deep learning includes the following steps: S1. Physically guided spectral preprocessing is performed on the original LIBS spectra of the source domain and the original LIBS spectra of the target domain. The baseline correction is completed using the asymmetric least squares method. The common wavelength range of the source domain and the target domain is calculated and truncated. The wavelength alignment between the target domain spectrum and the source domain spectrum is achieved through linear interpolation. S2, local intensity calibration is performed on the target domain spectrum and source domain spectrum after wavelength alignment using the sliding window local regression method to eliminate the difference in local spectral response of the instrument; S3. Construct a deep spectral transfer model based on an tunable multilayer perceptron. With the source domain spectrum as input and the target domain spectrum as output, learn the nonlinear mapping relationship from the source domain to the target domain and generate the reconstructed target domain spectrum.
[0009] Preferably, in step S1, the formula for calculating the common wavelength range is: ; ; in, This is the lower limit of the common wavelength. The upper limit of the common wavelength, For the target domain wavelength, The wavelength is the source domain wavelength.
[0010] Preferably, in step S2, the calculation formula for the local intensity calibration is: ; in, For the calibrated wavelength Spectral intensity at that location Before calibration wavelength Spectral intensity at that location , These are the local regression coefficients within the sliding window.
[0011] Preferably, in step S3, the deep spectral transfer model based on an adjustable multilayer perceptron includes: an input layer, two fully connected hidden layers, and an output layer; the number of neurons in the hidden layer is 128~512, the activation function is LeakyReLU, L2 regularization is added to the model to prevent overfitting, and the Adam optimizer is used for training, with mean squared error, mean absolute error, and coefficient of determination as optimization indicators.
[0012] This invention also provides a LIBS cross-instrument spectral library transfer system based on spectral calibration and deep learning, applicable to the aforementioned LIBS cross-instrument spectral library transfer method based on spectral calibration and deep learning, comprising: The source domain LIBS acquisition module is used to acquire the raw LIBS spectrum in the source domain; The target domain LIBS acquisition module is used to acquire the raw LIBS spectrum of the target domain. The preprocessing module is used to perform baseline correction, common wavelength interception, and wavelength alignment on the original LIBS spectra of the source domain and the target domain. The local intensity calibration module is used to perform spectral intensity calibration through local regression via a sliding window. The deep spectral transfer module is used to load a deep spectral transfer model based on an tunable multilayer perceptron to achieve a nonlinear conversion from the source domain spectrum to the target domain spectrum.
[0013] Preferably, the source domain LIBS acquisition module is a laboratory LIBS system, employing a 1064nm nanosecond Q-switched Nd:YAG laser, a five-channel light collection structure, and a spectral detection range of 230~750nm.
[0014] Preferably, the target domain LIBS acquisition module is a portable LIBS system, employing a 1064nm diode-pumped solid-state laser, a three-channel optical collection structure, and a spectral detection range of 280~816nm.
[0015] The present invention also provides a computer program product comprising a computer program / instruction that, when executed by a processor, implements the LIBS cross-instrument spectral library transfer method based on spectral calibration and deep learning.
[0016] The present invention also provides an electronic device, which includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the LIBS cross-instrument spectral library transfer method based on spectral calibration and deep learning.
[0017] The present invention also provides a readable storage medium having a computer program stored thereon, which, when executed, implements the LIBS cross-instrument spectral library migration method based on spectral calibration and deep learning.
[0018] The advantages of this invention are: (1) This invention proposes an integrated framework of hybrid spectral calibration-deep spectral transfer (HSH-DSM), which combines physical-guided spectral preprocessing with data-driven deep learning nonlinear mapping to eliminate instrument differences step by step. First, a unified spectral benchmark is achieved through physical calibration, and then a tunable multilayer perceptron (MLP) is used to learn the nonlinear mapping from the source domain to the target domain, ultimately realizing high-precision transfer of the LIBS spectral database across instruments.
[0019] (2) The present invention uses the asymmetric least squares method (ALS) to remove low-frequency baseline noise introduced by plasma and instrument response; calculates the spectral overlap range between the source domain and the target domain to determine the common wavelength range; and resamples the target domain spectrum to the source domain wavelength grid through linear interpolation to achieve accurate wavelength alignment.
[0020] (3) The present invention uses the sliding window local regression method to perform intensity correction on each wavelength point, thereby eliminating the local intensity differences caused by the sensitivity and optical response of different instrument detectors.
[0021] (4) The present invention constructs a supervised adjustable MLP model, which uses the calibrated source domain spectrum as input and the target domain spectrum as label to learn cross-instrument nonlinear mapping; the model adopts LeakyReLU activation function, L2 regularization and Adam optimizer to improve nonlinear fitting ability and generalization, and outputs high-precision reconstructed target domain spectrum.
[0022] (5) This invention breaks through the limitations of traditional linear calibration methods and accurately fits the nonlinear spectral differences between instruments; it enables cross-instrument reuse of LIBS spectral libraries without the need for repeated library construction, reduces detection costs, supports rapid deployment of portable devices, and has strong engineering practicality; it can be used for cross-device LIBS analysis in multiple fields such as geological exploration, industrial testing, environmental monitoring, and planetary exploration, and has a wide range of applicable scenarios. Attached Figure Description
[0023] Figure 1 This is a schematic diagram of the experimental setup for laser-induced breakdown spectroscopy (LIBS) in the source and target domains.
[0024] Figure 2 This is a flowchart of a LIBS cross-instrument spectral library transfer method based on spectral calibration and deep learning.
[0025] Figure 3 This is a comparison of the source and target domain spectra before and after preprocessing.
[0026] Figure 4 This is a comparison chart of the spectral migration effects of the traditional segmented direct normalization (PDS) method and the tunable multilayer perceptron (MLP) method of this invention.
[0027] Figure 5 This is the confusion matrix for KNN classification based on PDS transfer learning.
[0028] Figure 6 This is the KNN classification confusion matrix based on the MLP transfer model of this invention. Detailed Implementation
[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0030] The LIBS cross-instrument spectral library transfer system based on spectral calibration and deep learning of the present invention includes: The source domain LIBS acquisition module is used to acquire the raw LIBS spectrum in the source domain; The target domain LIBS acquisition module is used to acquire the raw LIBS spectrum of the target domain. The preprocessing module is used to perform baseline correction, common wavelength interception, and wavelength alignment on the original LIBS spectra of the source domain and the target domain. The local intensity calibration module is used to perform spectral intensity calibration through local regression via a sliding window. The deep spectral transfer module is used to load a deep spectral transfer model based on an tunable multilayer perceptron to achieve a nonlinear conversion from the source domain spectrum to the target domain spectrum.
[0031] In this embodiment, as Figure 1 As shown in (a), the source domain LIBS acquisition module is a laboratory high-resolution LIBS system, which uses a nanosecond Q-switched Nd:YAG laser with a wavelength of 1064nm, a pulse energy of 50mJ, and a repetition frequency of 1Hz; a five-channel optical collection structure with an optical fiber core diameter of 200μm; a four-channel spectrometer with a spectral range of 230~750nm, an acquisition delay of 1.28μs, and an integration time of 1.05ms.
[0032] In this embodiment, as Figure 1 As shown in (b), the target domain LIBS acquisition module is a portable LIBS system, which uses a diode-pumped solid-state Nd:YAG laser with a wavelength of 1064nm and a pulse energy of 6mJ; a three-channel optical collection structure; a NEXOS spectrometer with a spectral range of 280~816nm and a spectral resolution of 0.3~0.36nm; and a 4096-pixel CMOS linear array detector.
[0033] like Figure 2 As shown, the present invention provides a LIBS cross-instrument spectral library transfer method based on spectral calibration and deep learning, comprising: S1, Physically guided spectral preprocessing is performed on the original LIBS spectra of the source domain and the original LIBS spectra of the target domain. The baseline correction is completed using the asymmetric least squares method (ALS). The common wavelength range of the source domain and the target domain is calculated and truncated. The wavelength alignment between the target domain spectrum and the source domain spectrum is achieved through linear interpolation. S2, The local intensity calibration of the wavelength-aligned spectrum is performed using the sliding window local regression method to eliminate the local spectral response differences of the instrument; S3. Construct a deep spectral transfer model based on a tunable multilayer perceptron (MLP), taking the source domain spectrum as input and the target domain spectrum as output, to learn the nonlinear mapping relationship from the source domain to the target domain and generate the reconstructed target domain spectrum.
[0034] The specific experimental process of this embodiment is as follows.
[0035] The experimental samples consisted of 26 types of rock samples, with 5 samples from each type. Five measurement sites were selected for each sample, and the corresponding raw spectra were acquired through the source domain LIBS acquisition module and the target domain LIBS acquisition module, respectively.
[0036] The data was divided into three stratified groups: 30% for training (21% for calibration and 9% for validation) and 70% for testing. The training set was used for model training, and the testing set was used to evaluate the classification performance of unknown samples.
[0037] The specific process of physical-guided spectral preprocessing is as follows: (1) Baseline correction: The ALS algorithm is used to remove low-frequency background noise from the original spectrum; (2) Common wavelength cutoff: The calculated common wavelength range is 280~750nm, and the overlapping spectral band is retained; (3) Wavelength alignment: Linear interpolation is performed on the target domain spectrum to correspond one-to-one with the source domain wavelength grid.
[0038] like Figure 3 The image shown is a comparison of the source domain spectrum and the target domain spectrum before and after preprocessing.
[0039] The specific process of local strength calibration is as follows: using sliding window local regression, according to the formula... Correct the intensity at each wavelength point to eliminate local response differences in the instrument.
[0040] The construction and training process of the deep spectral transfer model is as follows: (1) Model structure: It includes an input layer, two fully connected hidden layers and an output layer; the input layer dimension is the source domain spectral dimension, the number of neurons in the hidden layer is adjustable from 128 to 512, the activation function is LeakyReLU (negative slope α=0.01), L2 regularization is added to the model to prevent overfitting, and the output layer dimension is the target domain spectral dimension. (2) Training parameters: The Adam optimizer was used for training, with a batch size of 64, a maximum training epoch of 300, and an early stopping strategy to prevent overfitting; (3) Optimization indicators: The mean square error (MSE), mean absolute error (MAE), and coefficient of determination (R²) are used as the core optimization indicators to select the optimal model parameters.
[0041] Next, spectral fusion and classification are performed, the specific process of which is as follows: (1) Data fusion: The reconstructed target domain spectrum generated by the deep spectral transfer model is fused with the real target domain spectrum to construct an enhanced dataset; (2) Feature selection: The ANOVA F-score univariate percentage method was used to retain 10%~65% of discriminative features; (3) Classification and recognition: KNN classifier, Euclidean distance metric, cross-validation to select the optimal k value, and output the classification result.
[0042] The experimental results are as follows: The spectral reconstruction performance is shown in Table 1 below: Table 1 Optimization metrics for each dataset
[0043] The classification performance is as follows: Classification of 26 rock samples: precision 97.8%, recall 97.5%, F1 score 97.5%, overall classification accuracy 97.65%.
[0044] like Figure 4 The figure shown is a comparison of the spectral migration effects of the traditional segmented direct normalization (PDS) method and the tunable multilayer perceptron (MLP) method of this invention. Figure 5 This is the confusion matrix for KNN classification based on PDS transfer learning. Figure 6 This is the KNN classification confusion matrix based on MLP transfer learning in this invention. Figure 5 (a) in the figure corresponds to the KNN classification confusion matrix based on PDS transfer learning in the training set. Figure 5 (b) corresponds to the KNN classification confusion matrix based on PDS transfer in the test set. Figure 6 (a) in the figure corresponds to the KNN classification confusion matrix based on the MLP transfer of this invention under the training set. Figure 6 In the middle (b), the KNN classification confusion matrix based on the MLP transfer of this invention is shown in the test set.
[0045] Compared with the traditional PDS linear calibration method, the MLP deep migration method of this invention improves the test set R² by 4.86% and the classification accuracy by 10.26%, effectively solves the problem of nonlinear spectral differences, and has significantly better cross-instrument migration performance than the traditional method.
[0046] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of 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 LIBS cross-instrument spectral library transfer method based on spectral calibration and deep learning, characterized in that, Includes the following steps: S1. Physically guided spectral preprocessing is performed on the original LIBS spectra of the source domain and the original LIBS spectra of the target domain. The baseline correction is completed using the asymmetric least squares method. The common wavelength range of the source domain and the target domain is calculated and truncated. The wavelength alignment between the target domain spectrum and the source domain spectrum is achieved through linear interpolation. S2, local intensity calibration is performed on the target domain spectrum and source domain spectrum after wavelength alignment using the sliding window local regression method to eliminate the difference in local spectral response of the instrument; S3. Construct a deep spectral transfer model based on an tunable multilayer perceptron. With the source domain spectrum as input and the target domain spectrum as output, learn the nonlinear mapping relationship from the source domain to the target domain and generate the reconstructed target domain spectrum.
2. The LIBS cross-instrument spectral library transfer method based on spectral calibration and deep learning according to claim 1, characterized in that, In step S1, the formula for calculating the common wavelength range is: in, This is the lower limit of the common wavelength. The upper limit of the common wavelength, For the target domain wavelength, The wavelength is the source domain wavelength.
3. The LIBS cross-instrument spectral library transfer method based on spectral calibration and deep learning according to claim 1, characterized in that, In step S2, the calculation formula for the local intensity calibration is: in, For the calibrated wavelength Spectral intensity at that location Before calibration wavelength Spectral intensity at that location , These are the local regression coefficients within the sliding window.
4. The LIBS cross-instrument spectral library transfer method based on spectral calibration and deep learning according to claim 1, characterized in that, In step S3, the deep spectral transfer model based on tunable multilayer perceptron includes: an input layer, two fully connected hidden layers, and an output layer; the number of neurons in the hidden layer is 128~512, the activation function is LeakyReLU, L2 regularization is added to the model to prevent overfitting, and the Adam optimizer is used for training, with mean squared error, mean absolute error, and coefficient of determination as optimization indicators.
5. A LIBS cross-instrument spectral library transfer system based on spectral calibration and deep learning, characterized in that, The LIBS cross-instrument spectral library transfer method based on spectral calibration and deep learning, applicable to any one of claims 1-4, includes: The source domain LIBS acquisition module is used to acquire the raw LIBS spectrum in the source domain; The target domain LIBS acquisition module is used to acquire the raw LIBS spectrum of the target domain. The preprocessing module is used to perform baseline correction, common wavelength interception, and wavelength alignment on the original LIBS spectra of the source domain and the target domain. The local intensity calibration module is used to perform spectral intensity calibration through local regression via a sliding window. The deep spectral transfer module is used to load a deep spectral transfer model based on an tunable multilayer perceptron to achieve a nonlinear conversion from the source domain spectrum to the target domain spectrum.
6. The LIBS cross-instrument spectral library transfer system based on spectral calibration and deep learning according to claim 5, characterized in that, The source domain LIBS acquisition module is a laboratory LIBS system, which uses a 1064nm nanosecond Q-switched Nd:YAG laser, a five-channel light collection structure, and a spectral detection range of 230~750nm.
7. The LIBS cross-instrument spectral library transfer system based on spectral calibration and deep learning according to claim 6, characterized in that, The target domain LIBS acquisition module is a portable LIBS system that uses a 1064nm diode-pumped solid-state laser, a three-channel optical collection structure, and a spectral detection range of 280~816nm.
8. A computer program product, characterized in that, It includes a computer program / instruction that, when executed by a processor, implements the LIBS cross-instrument spectral library transfer method based on spectral calibration and deep learning as described in any one of claims 1 to 4.
9. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the LIBS cross-instrument spectral library transfer method based on spectral calibration and deep learning as described in any one of claims 1 to 4.
10. A readable storage medium, characterized in that, It stores a computer program, which, when executed, implements the LIBS cross-instrument spectral library migration method based on spectral calibration and deep learning as described in any one of claims 1 to 4.