A Method and System for Quantitative Analysis of Micron-sized Single Particles Based on Domain Adaptive Transfer Learning

Abstract

This invention provides a method and system for quantitative analysis of micron-sized single particles based on domain adaptive transfer learning. The method first uses a dual-mode LIBS spectral acquisition device to acquire spectral data of compressed (source domain) and single-particle (target domain) carbon black samples. The spectral data is preprocessed and augmented. Then, feature variables are selected from the spectral data, and a quantitative correction model is constructed based on the TrAdaBoost transfer learning algorithm. Finally, the key hyperparameters of the model are refined using the Northern Eagle Optimization Algorithm (NGO). This establishes a transfer learning correction model that effectively utilizes source domain knowledge to assist in target domain prediction, effectively solving the problems of poor signal reproducibility and low quantitative accuracy, and providing a new paradigm for in-situ monitoring of complex aerosols.

Description

Technical Field

[0001] This invention belongs to the field of spectroscopic analysis and chemometrics technology, specifically relating to a method and system for quantitative analysis of micron-sized single particles based on domain adaptive transfer learning. More specifically, this invention relates to a rapid quantitative analysis method for heavy metal elements in suspended micron-sized particles that combines transfer learning and data augmentation techniques, particularly suitable for solving the problem of accurate in-situ detection of complex samples such as single-particle aerosols in the absence of a large amount of labeled data. Background Technology

[0002] LIBS technology, as an emerging elemental analysis method, has shown great application potential in environmental monitoring, industrial process control, and deep space exploration due to its advantages such as requiring no complex sample preparation and enabling in-situ, real-time, and simultaneous multi-element detection. However, when LIBS technology is applied to the in-situ analysis of single-particle micro-nano scale samples such as atmospheric aerosols, the accuracy and reliability of its quantitative analysis have long been limited by three core bottlenecks: First, there are signal instabilities and significant matrix effects. Due to the heterogeneity of the microstructure of single-particle samples (such as differences in size, morphology, and surface roughness), the interaction process between laser and matter is extremely unstable, resulting in large fluctuations in the generated plasma excitation state. Furthermore, the lack of a dense matrix support like that of solid bulk samples leads to severe matrix effects. This causes single-particle LIBS spectra to exhibit significant baseline drift, low signal-to-noise ratio, and drastic fluctuations in spectral line intensity, severely disrupting the linear mapping relationship between spectral intensity and elemental concentration.

[0003] Secondly, there is an extreme scarcity of high-quality labeled samples (the small sample problem). In practical environmental monitoring, obtaining a large number of single-particle standard reference materials (SRMs) with known and precise concentrations is extremely difficult and costly. Traditional machine learning models (such as random forests and partial least squares methods) typically require massive amounts of labeled data for training to cover complex chemical spaces. However, in the typical small-sample scenario of single-particle analysis, traditional models are prone to overfitting, resulting in poor generalization ability and failing to meet the needs of accurate quantification.

[0004] Finally, there is the difficulty of cross-domain knowledge transfer. Although macroscopic compressed standard samples (source domain) are easy to prepare in the laboratory, with stable spectral signals and readily available large amounts of labeled data, there are significant domain shifts between macroscopic compressed samples and microscopic single particles (target domain) in terms of physical morphology and spectral characteristics. Directly applying a calibration model based on compressed samples to single-particle predictions will result in significant prediction biases (i.e., negative transfer) due to matrix and particle size effects, leading to model failure.

[0005] The invention patent with authorization announcement number CN116256303B discloses a quantitative analysis method for micron-sized single particles based on nano-silver signal enhancement. This method uses a LIBS system to acquire and preprocess spectral data of carbon black particle target samples, employs variable importance measurement algorithms and continuous projection algorithms to screen target variables, and then inputs these variables into a model to obtain the metal element content. This method relies on nano-silver signal enhancement technology and is only applicable to carbon black particle samples doped with nano-silver and with unknown metal element content. It is not suitable for the quantitative analysis of ordinary micron-sized single particle samples without a carbon black matrix or without nano-silver doping.

[0006] The methods described above generally suffer from problems such as difficulty in balancing model robustness and universality, reliance on complex physical calibration or traditional chemometrics, and difficulty in effectively utilizing readily available macroscopic standard sample knowledge to overcome domain distribution differences. Therefore, it is particularly important to find an innovative modeling strategy that can achieve in-situ rapid and accurate quantification of heavy metal elements in suspended micron-sized particulate samples. Summary of the Invention

[0007] This invention aims to address the technical challenges of existing LIBS technology in in-situ analysis of suspended micron-sized particulate samples, including poor signal reproducibility, significant matrix effects, and low quantitative accuracy due to a lack of a large number of labeled samples. It provides a method and system for quantitative analysis of micron-sized single particles based on domain adaptive transfer learning, achieving efficient knowledge transfer from macroscopic standard samples (source domain) to microscopic single particles (target domain), thereby enabling high-precision in-situ quantitative detection under small sample conditions.

[0008] To achieve the above objectives, this invention provides a method and system for quantitative analysis of micron-sized single particles based on domain adaptive transfer learning. Its core lies in constructing an integrated modeling framework that combines dual-morphological sample preparation, spectral preprocessing, data augmentation, feature selection, and improved transfer learning. The specific technical solution includes the following steps: Step 1, Dual-morphology sample preparation and spectral acquisition: Prepare carbon black-based standard samples with the same chemical gradient but different physical morphologies, including the target domain of a single-particle suspended sample simulating an atmospheric scenario and the source domain of a densely compressed sample. Then, use a dual-mode laser-induced breakdown spectroscopy (LIBS) acquisition device to acquire spectral data of the two types of samples respectively. Step 2, Spectral preprocessing and data augmentation: The acquired spectral data is preprocessed, and the source domain samples are augmented using the SMOTE data augmentation method; Step 3, Feature variable screening: Introduce a feature selection algorithm to reduce the dimensionality of the preprocessed spectral data and screen out feature variables related to heavy metal content; Step 4, Construct an improved TrAdaBoost calibration model: Construct a quantitative calibration model based on the improved TrAdaBoost algorithm, use the label information of the source domain tablet samples to assist in the modeling of the target domain single-particle samples, and achieve cross-domain knowledge transfer by dynamically adjusting the sample weights; Step 5, Heavy metal content prediction: Use the calibration model established in Step 4 to predict the content of heavy metal elements in a single particle sample.

[0009] In the above technical solution, the dual-morphological sample preparation step in step 1 is specifically as follows: using high-purity micron-sized carbon black as a matrix, the target heavy metal element is loaded by wet chemical adsorption to prepare a composite powder with a precise and controllable concentration gradient; wherein, the single-particle suspended sample is introduced through an aerosol generation system, and the tablet sample is formed by adding an adhesive and pressing it with a mold.

[0010] In the above technical solution, the acquisition range of LIBS spectral data in step 1 is 200-500nm.

[0011] In the above technical solution, the spectral preprocessing method in step 2 is further selected differently according to the different elements to be measured. The preprocessing methods include: standard normal transformation, SG smoothing and noise filtering, derivative method, wavelet transform (WT), and baseline correction (BL).

[0012] In the above technical solution, the SMOTE data augmentation method in step 2 is further described as follows: LIBS spectrum is regarded as a high-dimensional feature space vector. For spectral samples of scarce categories, its K nearest neighbor spectra are calculated, and linear interpolation is performed on the line connecting the original spectrum and the neighboring spectra to generate a synthetic spectrum with physical meaning. The enhancement ratio is optimized in the range of 0.5-2.5.

[0013] In the above technical solution, the feature selection algorithm in step 3 further includes mutual information (MI), average impurity reduction (MDI), and particle swarm optimization.

[0014] In the above technical solution, further, the construction of the improved TrAdaBoost calibration model in step 4 includes: replacing the base learner in the traditional TrAdaBoost with a single decision tree and a random forest (RF) model, and using the Northern Eagle Optimization Algorithm (NGO) to globally optimize the model hyperparameters.

[0015] In the above technical solution, further, the model hyperparameters include the number of base learners n_estimators, learning rate learning_rate, random forest sub-parameters rf_n_estimators, maximum number of features max_features, maximum tree depth max_depth, minimum number of split samples min_samples_split, and minimum number of leaf samples min_samples_leaf; wherein, the optimization range for the number of base learners n_estimators is 3-50, the optimization range for the learning rate learning_rate is 0.001-1, the optimization range for the random forest sub-parameters rf_n_estimators is 3-50, the optimization range for the maximum number of features max_features is 0.001-1, the optimization range for the maximum tree depth max_depth is 2-50, the optimization range for the minimum number of split samples min_samples_split is 2-30, and the optimization range for the minimum number of leaf samples min_samples_leaf is 1-30.

[0016] This invention also provides a domain-adaptive transfer learning-based quantitative analysis system for micron-sized single particles, comprising a dual-mode LIBS spectral acquisition device, a spectral data preprocessing module for implementing the method described above, a data augmentation and feature selection module for implementing the method described above, and an improved TrAdaBoost modeling module for implementing the method described above; wherein, the dual-mode LIBS spectral acquisition device includes a laser, an optical path system, a spectrometer, an optical tweezers system for capturing single particles, and a sample stage for fixing the pellet.

[0017] Compared with existing technologies, this invention has the following advantages: First, by introducing SMOTE data augmentation technology, the training set size is effectively expanded, solving the problem of model overfitting caused by the difficulty in obtaining single-particle standard reference materials and the scarcity of labeled samples, and significantly improving the robustness of the model under small sample conditions. Second, an innovative transfer learning strategy of pellet (source domain) - single particle (target domain) is proposed. The improved TrAdaBoost algorithm is used to transfer the rich label information of readily available macroscopic pellet samples from the laboratory to microscopic single-particle analysis, overcoming the domain distribution differences between the two and significantly reducing prediction bias caused by matrix effects and particle size effects. Third, RF is used as the base learner of TrAdaBoost, and combined with the NGO algorithm for hyperparameter optimization, enabling the model to maintain stability while having a stronger ability to extract nonlinear spectral features and significantly improving quantitative accuracy. Fourth, by combining average impurity reduction and mutual information feature screening methods, the feature dimension is significantly compressed while ensuring accuracy, improving the model's computational speed and real-time performance, and providing a highly robust technical means for in-situ monitoring of trace elements in complex atmospheric aerosols. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the modeling process in an embodiment of the present invention; Figure 2 The image shows a comparison of the LIBS spectra of sample #14 in this embodiment of the invention. (a) shows the original spectrum of the tablet sample, (b) shows the original spectrum of the single particle sample, and (c) shows the distribution of the two spectral data in the PCA three-dimensional principal component space and the compactness of the confidence interval. Figure 3 This is a comparison of the effects of different preprocessing methods on the prediction performance of the TrAdaBoost correction model in the embodiments of the present invention. (a) and (b) show the effect of WT on Cu, (c) and (d) show the effect of BL on Zn, and (e) and (f) show the effect of WT on Ni. Figure 4 This is a comparison diagram of the original spectrum and the spectrum enhanced by SMOTE in an embodiment of the present invention, wherein (a) is the distribution diagram after PCA dimensionality reduction, used to visualize the feature space filling effect, and (b) is a comparison of the LIBS spectral morphology before and after enhancement. Figure 5 This is a distribution diagram of characteristic variables in an embodiment of the present invention, where (a) corresponds to Cu element, (b) corresponds to Zn element, and (c) corresponds to Ni element; Figure 6 This is a scatter plot of the prediction performance of the optimized model in this embodiment of the invention, where (a) corresponds to Cu, (b) corresponds to Zn, and (c) corresponds to Ni. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] This invention provides a method for quantitative analysis of micron-sized single particles based on domain adaptive transfer learning. The specific technical solution includes the following steps: Step 1, Dual-morphology sample preparation and spectral acquisition: Prepare carbon black-based standard samples with the same chemical gradient but different physical morphologies, including the target domain of a single-particle suspended sample simulating an atmospheric scenario and the source domain of a densely compressed sample. Then, use a dual-mode laser-induced breakdown spectroscopy (LIBS) acquisition device to acquire spectral data of the two types of samples respectively. Step 2, Spectral preprocessing and data augmentation: The acquired spectral data is preprocessed, and the source domain samples are augmented using the SMOTE data augmentation method; Step 3, Feature variable screening: Introduce a feature selection algorithm to reduce the dimensionality of the preprocessed spectral data and screen out feature variables related to heavy metal content; Step 4, Construct an improved TrAdaBoost calibration model: Construct a quantitative calibration model based on the improved TrAdaBoost algorithm, use the label information of the source domain tablet samples to assist in the modeling of the target domain single-particle samples, and achieve cross-domain knowledge transfer by dynamically adjusting the sample weights; Step 5, Heavy metal content prediction: Use the calibration model established in Step 4 to predict the content of heavy metal elements in a single particle sample.

[0021] This embodiment will describe the details of each step in the order of the above method.

[0022] To achieve the above method, the present invention also provides a micron-sized single-particle quantitative analysis system based on domain adaptive transfer learning, including a dual-mode LIBS spectral acquisition device, a spectral data preprocessing module for implementing the method described above, a data augmentation and feature selection module for implementing the method described above, and an improved TrAdaBoost modeling module for implementing the method described above; wherein, the dual-mode LIBS spectral acquisition device includes a laser, an optical path system, a spectrometer, an optical tweezers system for capturing single particles, and a sample stage for fixing the pellet. Example 1

[0023] This embodiment provides a single-particle heavy metal detection method based on LIBS domain adaptive modeling, specifically including the following steps: Step 1: Preparation of carbon black-based standard samples and acquisition of spectral data: Sample Preparation: High-purity micron-sized carbon black (average particle size approximately 1.86 μm) was selected as the matrix. Analytical-pure metal salts (CuSO4·5H2O, ZnSO4·7H2O, NiCl2·6H2O) were weighed according to the concentration values ​​listed in Table 1 and dissolved in deionized water to prepare a series of concentration gradient solutions. 2.0 g of carbon black powder was weighed and added to each solution, magnetically stirred at room temperature for 2 h, then ultrasonically dispersed, dried at 80°C for 8 h, and finely ground to prepare a series of composite powders (numbered 1#-15#) with precisely controllable heavy metal mass fractions. Source Domain Sample (Dense Compacted State): The above composite powder was mixed with polyvinyl alcohol binder at a 5:1 mass ratio, ground and placed in a stainless steel mold, and pressurized at 20 tons for 2 min to form uniform discs with a diameter of 13 mm and a thickness of approximately 2 mm. Target Domain Sample (Single Particle Suspension State): The above composite powder was introduced into an aerosol generation system, and single particles were captured in the air using optical tweezers technology.

[0024] Table 1. Elemental Content Information (wt.%)

[0025] Spectral Acquisition: Data acquisition was performed using a dual-mode LIBS spectral acquisition system. For the source domain (pressed state): a lateral collection geometry was employed, with the laser focused on the sample surface. Ten points were randomly selected from each sample, and five pulses were accumulated at each point, with the average taken to obtain a stable source domain spectrum. For the target domain (single-particle suspended state): a long working distance microscope objective was used to achieve stable optical tweezers capture and LIBS excitation focusing of aerosol particles. A back-collection mode was used to record the microplasma emission light, with each spectrum corresponding to the single-particle signal under a single laser pulse.

[0026] Step 2: Differentiated spectral preprocessing, performing denoising and background correction on the spectral data acquired in Step 1. To address the issues of large signal fluctuations and severe baseline drift in single particles, this invention employs a differentiated preprocessing strategy. For Cu and Ni elements: WT is used for denoising. After optimization through five-fold cross-validation, the db1 wavelet basis was determined to be used, with a decomposition layer of 1. For Zn element: BL method is used to eliminate background interference. Through the above processing, spectral intensity fluctuations and distribution shifts caused by differences in sample morphology (compacted tablets versus single particles) are effectively eliminated.

[0027] Step 3: SMOTE data augmentation, which augments the target domain (single particle) spectral data after preprocessing in Step 2. Due to the difficulty and scarcity of single-particle acquisition, this invention introduces the SMOTE algorithm, and its LIBS data augmentation process is as follows: Each LIBS spectrum is considered as a vector in a high-dimensional feature space.

[0028] Calculate the K nearest neighbors of the scarce class sample in the feature space.

[0029] Linear interpolation is performed on the line connecting the original sample and randomly selected neighboring samples to synthesize a new spectral sample with physical meaning.

[0030] In this embodiment, the enhancement ratio is set to 1, which means that the number of original target domain samples is increased to twice the original number, in order to balance the data distribution and improve the robustness of the model.

[0031] Step 4: Feature Variable Screening. Feature selection is performed on the enhanced spectral data from Step 3 to reduce dimensionality and eliminate redundant information. This invention employs different screening methods for the spectral characteristics of different elements: Cu and Ni: The MDI method is used. This method effectively captures multiple sensitive lines for Cu (e.g., 324.7 nm, 327.4 nm) and Ni (e.g., 341.5 nm, 352.4 nm). After screening, 87 variables are retained for Cu and 46 variables for Ni. Zn: The MI method is used. This method exhibits stronger correlation capture ability when screening Zn characteristic wavelengths (e.g., 213.8 nm). After screening, 116 variables are retained for Zn.

[0032] Step 5: Construct and optimize the improved TrAdaBoost correction model. An improved TrAdaBoost transfer learning model is constructed, replacing the single decision tree base learner in the traditional algorithm with Random Forest (RF) to enhance the ability to express complex spectral features. The NGO algorithm is used to globally optimize the model hyperparameters. The hyperparameters optimized include: number of base learners (n_estimators), learning rate (learning_rate), random forest sub-parameters (rf_n_estimators), maximum number of features (max_features), maximum tree depth (rf_max_depth), minimum number of split samples (min_samples_split), and minimum number of leaf samples (min_samples_leaf). The optimal parameter combinations for each heavy metal element obtained through the NGO algorithm are shown in Table 2 below. Table 2 Model parameters optimized based on NGO

[0033] Step 6: Heavy Metal Content Prediction and Model Validation. Using the optimized and improved TrAdaBoost calibration model established in Step 5, the contents of Cu, Zn, and Ni heavy metals in single-particle samples are predicted. Model performance is evaluated using R... 2 Evaluations were conducted using RMSE and MRE. Implementation effect

[0034] Tests showed that the method in this embodiment performed excellently on the prediction set: Cu element: R 2 P The value reached 0.9610, and the MRE was 4.98%; Zn element: R 2 P The concentration reached 0.9818, and the MRE was 3.23%; Ni element: R 2 P The RMSE reached 0.9731, and the MRE was 4.02%. These data indicate that the single-particle heavy metal detection method based on LIBS domain adaptive modeling proposed in this invention exhibits excellent quantitative prediction performance for Cu, Zn, and Ni heavy metals in a carbon black matrix. Compared to the traditional Raw-RF benchmark model, the optimized model constructed in this invention, while maintaining a high coefficient of determination, significantly reduced the prediction set RMSE by 84.02% (Cu), 85.93% (Zn), and 75.24% (Ni), respectively, and the MRE was reduced to below 5% for all three.

[0035] In summary, this method, by introducing SMOTE data augmentation, differential feature selection, and an improved TrAdaBoost transfer learning algorithm, successfully solves the common problems of the curse of small samples and domain distribution drift in micro- and nano-scale spectral analysis, achieving efficient cross-domain knowledge transfer from macroscopic standard samples to microscopic single-particle detection. This method possesses advantages such as strong anti-interference capability, high quantitative accuracy, and the ability to directly perform in-situ analysis of single particles without complex physical sample preparation (such as pelleting), making it suitable for rapid detection and on-site monitoring of trace heavy metals in atmospheric aerosols in complex environments.

Claims

1. A method for quantitative analysis of micron-sized single particles based on domain adaptive transfer learning, characterized in that, Includes the following steps: Step 1, Dual-morphology sample preparation and spectral acquisition: Prepare carbon black-based standard samples with the same chemical gradient but different physical morphologies, including the target domain of a single-particle suspended sample simulating an atmospheric scenario and the source domain of a densely compressed sample. Then, use a dual-mode laser-induced breakdown spectroscopy (LIBS) acquisition device to acquire spectral data of the two types of samples respectively. Step 2, Spectral preprocessing and data augmentation: The acquired spectral data is preprocessed, and the source domain samples are augmented using the SMOTE data augmentation method; Step 3, Feature variable screening: Introduce a feature selection algorithm to reduce the dimensionality of the preprocessed spectral data and screen out feature variables related to heavy metal content; Step 4, Construct an improved TrAdaBoost calibration model: Construct a quantitative calibration model based on the improved TrAdaBoost algorithm, use the label information of the source domain tablet samples to assist in the modeling of the target domain single-particle samples, and achieve cross-domain knowledge transfer by dynamically adjusting the sample weights; Step 5, Heavy metal content prediction: Use the calibration model established in Step 4 to predict the content of heavy metal elements in a single particle sample.

2. The method for quantitative analysis of micron-sized single particles based on domain adaptive transfer learning according to claim 1, characterized in that, The specific steps for preparing the dual-morphology sample in step 1 are as follows: using high-purity micron-sized carbon black as a matrix, the target heavy metal elements are loaded by wet chemical adsorption to prepare a composite powder with a precise and controllable concentration gradient; wherein, the single-particle suspended sample is introduced through an aerosol generation system, and the tablet sample is formed by adding an adhesive and pressing it into shape with a mold.

3. The method for quantitative analysis of micron-sized single particles based on domain adaptive transfer learning according to claim 1, characterized in that, The range of LIBS spectral data acquisition in step 1 is 200-500 nm.

4. The method for quantitative analysis of micron-sized single particles based on domain adaptive transfer learning according to claim 1, characterized in that, The spectral preprocessing method in step 2 is selected differently depending on the element to be measured. The preprocessing methods include: standard normal transformation, SG smoothing and noise filtering, derivative method, wavelet transform (WT), and baseline correction (BL).

5. The method for quantitative analysis of micron-sized single particles based on domain adaptive transfer learning according to claim 1, characterized in that, The SMOTE data augmentation method in step 2 is as follows: LIBS spectra are regarded as high-dimensional feature space vectors. For spectral samples of scarce categories, their K nearest neighbor spectra are calculated, and linear interpolation is performed on the line connecting the original spectrum and the neighboring spectra to generate a synthetic spectrum with physical meaning. The enhancement ratio is optimized in the range of 0.5-2.

5.

6. A method for quantitative analysis of micron-sized single particles based on domain adaptive transfer learning according to claim 1, characterized in that, The feature selection algorithm in step 3 includes mutual information (MI), average impurity reduction (MDI), and particle swarm optimization.

7. The method for quantitative analysis of micron-sized single particles based on domain adaptive transfer learning according to claim 1, characterized in that, The construction of the improved TrAdaBoost calibration model in step 4 includes: replacing the base learner in the traditional TrAdaBoost with a single decision tree and a random forest (RF) model, and using the Northern Eagle Optimization Algorithm (NGO) to globally optimize the model hyperparameters.

8. A method for quantitative analysis of micron-sized single particles based on domain adaptive transfer learning according to claim 7, characterized in that, The model hyperparameters include the number of base learners (n_estimators), learning rate (learning_rate), random forest sub-parameters (rf_n_estimators), maximum number of features (max_features), maximum tree depth (max_depth), minimum number of split samples (min_samples_split), and minimum number of leaf samples (min_samples_leaf). The optimization ranges for the number of base learners (n_estimators) are 3-50, the learning rate (learning_rate) is 0.001-1, the random forest sub-parameters (rf_n_estimators) are 3-50, the maximum number of features (max_features) is 0.001-1, the maximum tree depth (max_depth) is 2-50, the minimum number of split samples (min_samples_split) is 2-30, and the minimum number of leaf samples (min_samples_leaf) is 1-30.

9. A quantitative analysis system for micron-sized single particles based on domain adaptive transfer learning, characterized in that, The invention includes a dual-mode LIBS spectral acquisition device, a spectral data preprocessing module for implementing the method of any one of claims 1-8, a data augmentation and feature selection module for implementing the method of any one of claims 1-8, and an improved TrAdaBoost modeling module for implementing the method of any one of claims 1-8; wherein the dual-mode LIBS spectral acquisition device includes a laser, an optical path system, a spectrometer, an optical tweezers system for capturing single particles, and a sample stage for fixing the pressed tablet.

Images