Time-resolved LIBS (laser-induced breakdown spectroscopy)-based milk powder adulteration qualitative and quantitative collaborative analysis method

Abstract

The invention discloses a milk powder adulteration qualitative and quantitative collaborative analysis method based on time-resolved LIBS (laser-induced breakdown spectroscopy). Comprising the following steps: S1, acquiring LIBS spectral data of a milk powder sample at a plurality of delay times to form a time-resolved spectral matrix; s2, performing feature extraction on the spectrum matrix to obtain a multi-scale feature vector; s3, inputting the multi-scale features into a classification branch and a regression branch based on mixed experts, outputting the probability of each adulteration type by the classification branch, weighting the concentration prediction result of each expert sub-network by the regression branch according to the probability, and outputting the final adulteration concentration; s4, carrying out end-to-end training on the model by adopting a joint loss function; and S5, outputting the adulteration type and concentration of the milk powder sample to realize qualitative and quantitative collaborative analysis. According to the method, the limitation of weak generalization ability of traditional qualitative and quantitative splitting is broken through, and the stability, precision and industrial suitability of LIBS milk powder detection in small samples and high-dimensional complex scenes are remarkably improved. And the method has good engineering implementability and popularization value.

Description

Technical Field

[0001] This invention relates to the field of food safety testing, and in particular to a hypothetical quantitative collaborative analysis method for milk powder adulteration based on time-resolved LIBS and multi-task hybrid expert CNN. Background Technology

[0002] Laser-induced breakdown spectroscopy (LIBS) is a cutting-edge analytical technique that uses high-energy laser pulses to excite a sample and generate plasma, then analyzes its radiation spectrum to achieve rapid detection of elemental types and concentrations. With its unique advantages, including no need for complex sample pretreatment, fast analysis speed, in-situ operation, minimal damage, remote detection, and simultaneous analysis of all elements, LIBS technology has been widely applied in various industries such as food safety, materials composition, environmental monitoring, and drug screening.

[0003] However, in safety testing scenarios involving complex food matrices such as milk powder, LIBS spectroscopy often suffers from problems such as sparse spectral lines, strong matrix effects, severe self-absorption, and nonlinear distortion, which significantly affect the stability of the spectrum and the accuracy of the analysis. To address this challenge, academia and industry have proposed numerous improvement methods, mainly categorized into hardware improvement methods and data processing algorithms.

[0004] Hardware improvement methods include auxiliary airflow, spatial constraint, magnetic constraint, dual pulse, beam shaping, etc. While these methods have improved detection sensitivity and stability to some extent, they have also brought about practical bottlenecks such as increased equipment complexity, cumbersome operation, and increased costs, which seriously limit the widespread application of LIBS in industrial fields and online detection scenarios.

[0005] Data processing methods include traditional chemometrics (such as PCA and PLS), normalization correction, feature engineering, and multivariate calibration. Although these methods can alleviate spectral drift and differences between samples to some extent, the high dimensionality, sparse information, and limited number of samples in the context of milk powder adulteration lead to significant shortcomings in generalization ability, model stability, and complex tasks (such as simultaneous qualitative and quantitative collaborative identification).

[0006] Therefore, there is an urgent need to propose a novel LIBS data processing and intelligent analysis method for detecting adulteration in complex food products such as milk powder. This method should be able to effectively mine potential spectral information, achieve accurate and coordinated identification of adulteration type and concentration under high-dimensional, small-sample conditions, and possess practical usability that is both hardware and software friendly and easy to deploy in industrial applications. This is precisely the core technical problem that this invention aims to solve. Summary of the Invention

[0007] To address the core pain points of existing technologies in the detection of adulteration types and concentrations in complex food matrices such as milk powder—namely, their difficulty in simultaneously analyzing adulteration types and concentrations, poor model generalization ability, and high challenges in industrial deployment—this invention proposes a detection method based on a time-resolved LIBS multi-task-hybrid expert CNN model for the coordinated analysis and detection of adulteration in milk powder, combining qualitative and quantitative methods. This method uses time-resolved spectral data and fully mines the data information through modules such as TCN and Inception. Furthermore, it leverages a multi-task mechanism and a hybrid expert architecture to achieve high-precision coordinated analysis of adulteration types and concentrations, significantly improving the accuracy, robustness, and industrial applicability of the detection.

[0008] The specific technical solutions are as follows:

[0009] A time-resolved LIBS-based hypothetical quantitative synergistic analysis method for milk powder adulteration includes the following steps:

[0010] Step 1: Acquire LIBS spectral signals of the milk powder and adulterated samples at multiple delay times to form a time-resolved spectral matrix S = [s ij ]∈R C×L Where C is the number of time delay channels and L is the spectral length per channel, comprehensively covering the spectral information of the sample at different plasma evolution stages.

[0011] Step 2: Automatically perform multi-stage data augmentation on the acquired time-resolved spectra. The automatic data augmentation strategy aims to minimize the joint loss of the validation set. The TPE Bayesian optimization algorithm is used for automatic hyperparameter search and tuning. Each search sequentially applies Gaussian noise, frequency masking, and normalization perturbation operations (Gaussian noise probability [0,1], standard deviation

[10] ). -6 [0, 0.15]; frequency masking probability [0, 1], width [100, 2000]; normalized perturbation mean [-0.1, 0.1], variance [10, 0.15] -4 [0.05]), the maximum number of iterations of the algorithm is set to 100 rounds. In each round of iteration, the joint loss performance of the current parameter combination is evaluated through the validation set, and finally the enhancement parameter combination corresponding to the optimal validation loss is selected from all iteration results.

[0012] Step 3: After channel weighting, the enhanced time-resolved spectrum is input into a feature extraction network that fuses one-dimensional convolution, TCN and Inception structures (with 3 to 7 parallel Inception branches) to achieve in-depth mining of the dynamic spectral features of complex milk powder matrix at multiple scales and time series.

[0013] Step 4: Input the extracted multi-scale feature vectors into both the classification branch and the hybrid expert regression branch simultaneously.

[0014] The classification branch outputs the probability distribution of each doping type, which is used to characterize the possible doping category of the sample;

[0015] The regression branch consists of N expert subnetworks that correspond one-to-one with the adulteration type, with each expert focusing on the concentration prediction of the corresponding category;

[0016] The final concentration output is weighted by the probability output from the classification branch, and the regression results of each expert are weighted and fused. That is, based on the distribution of possible sample types, the predictions of each expert are dynamically integrated to achieve accurate quantitative discrimination guided by information on adulteration type. This mechanism completely breaks through the inherent defects of traditional "serial" models, such as type-concentration separation, weak generalization ability, and large quantitative errors under confused categories.

[0017] Step 5: End-to-end collaborative training is performed using a classification-regression multi-task joint loss to ensure that qualitative identification and quantitative regression mutually optimize each other. All model structures and hyperparameters (including data augmentation, network width, loss weights, etc.) are automatically adjusted by the TPE Bayesian optimization framework to achieve optimal performance throughout the entire process. The joint loss function used in this invention for end-to-end multi-task training is:

[0018]

[0019] Where, α cls β reg The values ​​of are all set within the range of [0.1, 10]. The granularity of TPE Bayesian optimization is 0.1, and the maximum number of optimization iterations is set to 100 rounds. Finally, the optimal combination of hyperparameters corresponding to the minimum validation set joint loss is selected.

[0020] Step 6: Based on the outputs of the classification branch and the weighted regression branch, the adulteration type and concentration of milk powder samples are automatically and collaboratively identified, which greatly improves the accuracy and efficiency of qualitative and quantitative integrated detection.

[0021] Step 7: To address different industrial deployment scenarios and computing power requirements, the feature network channel width is variably compressed, and dynamic INT8 quantization and other technologies are adopted to significantly reduce the model's computational load and storage requirements. This enables rapid deployment and efficient inference of the model on multiple platforms such as MCU, FPGA, and ARM-SoC, while also meeting the real-time high-throughput detection needs of industrial sites.

[0022] A time-resolved LIBS-based hypothetical quantitative collaborative analysis system for milk powder adulteration includes a time-resolved LIBS spectral acquisition module, an automatic data augmentation module, a multi-scale feature extraction module, a classification-regression collaborative analysis module, a processor, and a memory. The time-resolved LIBS spectral acquisition module acquires LIBS spectral data of milk powder samples at multiple time delays to form a time-resolved spectral matrix. The automatic data augmentation module performs multi-stage data augmentation on the time-resolved spectra of the dataset. The multi-scale feature extraction module extracts features from the augmented spectral data. The classification-regression collaborative analysis module simultaneously inputs the extracted multi-scale feature vectors into a classification branch and a mixed expert regression branch to obtain the final adulteration concentration. The memory stores one or more programs, and the steps of the method are executed by the processor.

[0023] The classification-guided regression multi-task-hybrid expert CNN architecture of this invention has groundbreaking significance in the dynamic fusion and quantitative regression of doping type information. It significantly improves the recognition accuracy, generalization ability and deployment flexibility under small sample high-dimensional conditions, and provides a new technical path for the industrialization of LIBS milk powder adulteration detection.

[0024] Beneficial effects

[0025] The beneficial effects of this invention are:

[0026] 1. Sufficient feature information fusion and high accuracy: This invention utilizes LIBS spectral information under multiple delay times and the multi-scale feature extraction capability of deep learning. The channel weighting mechanism can highlight the feature contribution of high-delay channels in qualitative and quantitative analysis. Through the fusion of TCN and Inception structure, the discrimination and quantitative limits of trace adulteration and complex matrices are significantly improved.

[0027] 2. Multi-task collaboration and hybrid expert mechanism: The multi-task-hybrid expert architecture enables qualitative and quantitative collaborative optimization, significantly enhancing the model's generalization ability and stability under small sample and high noise conditions, which is significantly higher than the traditional "classification first and then regression" concatenation scheme.

[0028] 3. Enhanced generalization ability and industrial adaptability: Employing multi-stage automatic data augmentation and global Bayesian hyperparameter optimization, it overcomes the overfitting bottleneck of small-sample, high-dimensional data. Simultaneously, it supports on-demand compression and quantization of model structures, exhibits strong hardware adaptability, and facilitates deployment and large-scale promotion in industrial settings.

[0029] 4. Simple operation and flexible deployment: The model can be deployed on the basis of the conventional LIBS acquisition process without complicated sample pretreatment; model compression and quantization do not require manual intervention, and the degree of automation is high, making it suitable for rapid application in various industrial sites such as edge devices and mobile devices.

[0030] In summary, this invention can significantly improve the accuracy and quantitative precision of LIBS detection for complex food matrices such as milk powder, and has extremely high industrial application value. It can provide high-throughput, automated, robust and reliable detection methods for high-risk areas such as food safety, and promote the further development and practical application of LIBS spectral analysis technology. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 A flowchart of a multi-task-hybrid expert CNN model based on temporally resolved LIBS is provided for an embodiment of the present invention;

[0033] Figure 2 This is a calibration curve showing the difference between the adulterant concentration predicted by the regression branch of the MTMoE-CNN model and the actual concentration, provided in an embodiment of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. 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.

[0035] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0036] This invention addresses the problems of poor generalization, high qualitative classification error, and low quantitative accuracy in LIBS detection of milk powder adulteration due to factors such as complex sample matrices, variable types and amounts of adulterants, and small sample high-dimensional data. By collecting LIBS spectral data at different delay times to enrich the information content of the original data, this invention further proposes a multi-task-hybrid expert convolutional neural network (MTMoE-CNN) for the collaborative analysis of milk powder adulteration qualitative and quantitative aspects. This effectively achieves accurate extraction of time-resolved spectral features and collaborative prediction of adulteration types and their concentrations.

[0037] After the method of this invention begins, it first acquires the LIBS time-resolved spectra of milk powder and adulterated samples using a LIBS system. Then, it performs data augmentation and deep learning model training, finally achieving collaborative detection and analysis of milk powder adulteration. The specific steps are as follows:

[0038] Step 1: Sample preparation and LIBS time-series spectral acquisition

[0039] Pure milk powder (MP) was mixed with five common adulterant proteins (soy protein isolate (SPI), rice hydrolyzed protein (RHP), wheat hydrolyzed protein (WHP), peanut protein (PP), and egg white protein (ALB)) at weight ratios of 1%, 3%, 5%, 8%, 10%, 15%, 20%, and 25%, respectively. After homogenization, the mixture was pressed into discs with a diameter of 40 mm and a thickness of 3 mm. A Nd:YAG laser (532 nm, 80 mJ) was then used to excite the samples, generating plasma after focusing through a lens assembly. LIBS spectra (range 200–950 nm) were then collected at 21 different delay times (from 1.0 μs to 5.0 μs, with a step size of 0.2 μs) using a spectrometer, constructing time-resolved spectral data.

[0040] The time-resolved spectrum of each sample after acquisition can be formalized as a matrix:

[0041] S = [s ij ]∈R C×L

[0042] Where C = 21 is the number of time delay channels, L is the spectral length of each channel, and s ij =I(t) i ,λ j ) represents the i-th time delay and wavelength λ j The light intensity at that location.

[0043] Step 2: Automatic Data Augmentation Strategy

[0044] Gaussian noise, frequency masking, and normalization perturbation are applied sequentially to the original spectral data. The enhancement operations are defined as follows:

[0045] 1. Gaussian noise:

[0046]

[0047] in This represents the input data after perturbation, s ij This represents the normalized disturbance term, where σ is the standard deviation of Gaussian noise, and its value ranges from

[10] . -6 ,0.15].

[0048] 2. Frequency masking:

[0049]

[0050] Where w is the mask width, with a value range of [100, 2000], and j0 is the starting position, which is randomly selected.

[0051] 3. Normalized perturbation:

[0052]

[0053] Where μ and σ 2 Let μ and σ be the mean and variance of the disturbance, respectively, where μ ∈ [-0.1, 0.1], and σ are the mean and variance of the disturbance. 2 ∈[10 -4 ,0.05].

[0054] All enhancement methods and parameters are jointly determined through TPE Bayesian optimization, enabling automatic hyperparameter search:

[0055]

[0056] in L represents the optimal combination of hyperparameters obtained after TPE Bayesian optimization. val (x,y;θ,p) represents the loss function on the validation set.

[0057] Step 3: Construction of the feature extraction layer of the MTMoE-CNN model

[0058] like Figure 1 As shown, the enhanced LIBS time-series spectrum Further weighting factors are applied sequentially to each time-delay channel, specifically decreasing from 1 to 0.1 according to the time channel order from the first channel to the last. This differentiates the importance of information from different plasma evolution stages, significantly improving the expressive power of high-delay channel features. The weighted spectrum is then input into the MTMoE-CNN network, whose main structure includes:

[0059] • Initial one-dimensional convolution extracts local spectral features;

[0060] • Temporal Convolutional Networks (TCNs) extract plasma temporal dynamics information:

[0061]

[0062] Where D is the expansion coefficient, Δ is the expansion step size, and x t-d·4 The input feature at time td·Δ represents the input back to the past, W. d This represents the weight of the d-th dilated convolution kernel, where d represents the index of the current convolution kernel.

[0063] • The Inception module extracts multi-scale features in parallel:

[0064]

[0065] in This indicates that the kernel size is k. M The one-dimensional convolution branch.

[0066] Dropout and pooling operations reduce overfitting and computational cost.

[0067] Step 4: Construction of the classification and regression branches of the MTMoE-CNN model

[0068] like Figure 1 As shown, the obtained multi-scale spectral features h are input into the classification branch and the regression branch consisting of N expert sub-networks, respectively. The specific calculations are as follows:

[0069] 1. Classification branch output (qualitative recognition):

[0070]

[0071] ●where p i b represents the predicted probability for each doping type. cls The bias term representing the classification branch has dimension N, p i This represents the probability of predicting the i-th doping type.

[0072] 2. Regression branch composed of multiple expert subnetworks (quantitative prediction):

[0073] r i =g i (h)=W reg,i h+b reg,i

[0074] ·where r i W represents the concentration regression prediction value for the i-th doping type. reg,i This represents the weight matrix of the classification branches.

[0075] 3. Concentration output of the final regression branch (probability-weighted fusion):

[0076]

[0077] Step 5: End-to-end training and Bayesian optimization of multi-task joint loss

[0078] Training employs multi-task joint loss:

[0079]

[0080] in:

[0081]

[0082]

[0083] In the formula α clsβ represents the classification loss weight coefficient. reg L represents the regression loss weight coefficient. CE (p,y cls ) represents the cross-entropy loss for a classification task. This represents the mean absolute error loss of the regression task. Where α... cls β reg The values ​​are all set between [0.1, 10], and all hyperparameters (data augmentation, loss weights, network width, etc.) are automatically adjusted through TPE Bayesian optimization.

[0084] Step 6: Collaborative Qualitative and Quantitative Detection of the Model

[0085] For unknown milk powder samples, time-resolved LIBS spectra are acquired and the above-mentioned data enhancement, feature extraction and classification, and regression prediction are performed to output the sample doping type and concentration, achieving rapid, automated, and collaborative detection.

[0086] Step 7: Model Deployment and Quantization Compression Adaptation

[0087] To meet the needs of actual industrial deployments, the channel widths of the model's convolutional and fully connected layers are compressed proportionally by alpha (α can be 0.1, 0.25, 0.5, etc.), and further dynamic INT8 quantization is implemented layer by layer. Under different compression ratios, the model can achieve efficient inference capabilities on edge devices such as MCUs, FPGAs, or ARM-SoCs, enabling high-precision online applications of the model.

[0088] Construction of control methods and comparison of results:

[0089] In this embodiment of the invention, in order to objectively evaluate the comprehensive performance of the MTMoE-CNN model in small sample, high-dimensional, and complex LIBS spectral scenarios, and to further verify the innovation and superiority of the proposed method, several commonly used machine learning methods and deep learning networks were specifically selected as comparison schemes. In strict accordance with the unified experimental standards, a systematic performance comparison of the classification and regression tasks was carried out respectively.

[0090] The construction and parameter selection of the control methods strictly adhered to existing technologies and academic paper standards, including but not limited to mainstream machine learning classification algorithms such as Decision Tree (DT), Support Vector Machine (SVC), Random Forest (RF), Logistic Regression (LR), K-Nearest Neighbors (KNN), and XGBoost (XGB), as well as typical regression methods such as Support Vector Regression (SVR), Multilayer Perceptron (MLP), K-Nearest Neighbors Regression (KNNR), and Random Forest Regression (RFR). For deep learning methods, Inception CNN (SCNN) based on single-time-series spectroscopy and Inception CNN (MCNN) based on time-resolved spectroscopy were selected as control networks. In the specific experiments, all control models underwent systematic hyperparameter optimization, and PCA dimensionality reduction and multiple normalization methods were introduced in the data preprocessing process to eliminate data distribution bias and the influence of high-dimensional noise, ensuring a fair and scientific comparison.

[0091] The dataset was randomly stratified into training, validation, and test sets, with proportions of 70%, 20%, and 10%, respectively. All models were trained and optimized on the same dataset and under the same data augmentation environment. All evaluation metrics were based on results from the independent test set. For classification tasks, accuracy, precision, and recall were the primary performance indicators, while for regression tasks, the coefficient of determination (R²) was used. 2 The core evaluation criteria are root mean square error (RMSE) and mean absolute error (MAE).

[0092] During the specific experiment, follow these steps:

[0093] S1. Prepare samples and collect their time-resolved LIBS spectra to complete dataset construction and labeling.

[0094] S2. Perform unified preprocessing on all control models and MTMoE-CNN, including PCA dimensionality reduction and standardization.

[0095] S3. For different models, grid search and cross-validation are used to determine the optimal hyperparameters.

[0096] S4. Complete model training and tuning on the training and validation sets, and use methods such as early termination to prevent overfitting.

[0097] S5. Finally, evaluate the various evaluation metrics of each classification model and quantitative model on the independent test set.

[0098] Comparative analysis shows that the MTMoE-CNN model proposed in this invention significantly outperforms traditional machine learning methods and classic deep learning networks in all major performance metrics, including classification accuracy, precision, recall, regression determination coefficient, RMSE, and MAE. Taking classification tasks as an example, in small-sample, high-dimensional scenarios, the MTMoE-CNN test set accuracy reaches 0.950, while traditional decision trees and support vector machines generally fall below 0.800. Although deep networks SCNN-C and MCNN-C show some improvement, they still lag significantly behind the method proposed in this invention. In regression tasks, such as... Figure 2 As shown, the test set determination coefficient (R²) of MTMoE-CNN 2 The value was as high as 0.864, and both RMSE and MAE were significantly lower than the control method. For detailed comparison data, please refer to Appendix 1 and Appendix 2.

[0099] Appendix 1

[0100]

[0101]

[0102] Appendix 2

[0103]

[0104] Ablation experiments and results analysis:

[0105] To further verify the specific contributions of the main technical modules and core data augmentation strategies of this invention to the final detection performance, a systematic ablation experiment was conducted on the MTMoE-CNN network structure proposed in this invention. This ablation experiment targeted key components of the model architecture (such as the temporal convolution module, the Inception multi-scale extraction module, the MoE expert mechanism, and data augmentation strategies), removing or replacing one of each component, retaining only the others, and examining the changes in recognition accuracy, regression ability, and robustness of the model under different simplified structures.

[0106] The specific procedure for an ablation experiment includes the following steps:

[0107] S1. Using the complete structure of MTMoe-CNN as the baseline model, key network modules such as channel weighting (w / o LCW), temporal convolution (w / o TCN), Inception structure (w / o Inc), regression branch (w / o Reg), and classification branch (w / o Cla) are removed in sequence.

[0108] S2. For the model after removing the specified modules, adopt the same dataset partitioning and preprocessing process as the main model, while keeping other experimental conditions unchanged.

[0109] S3. Complete model training and evaluation on the same training set, validation set, and test set respectively. Record core indicators such as classification accuracy, regression determination coefficient, RMSE, and MAE on the test set.

[0110] S4. Compare the experimental results of all ablation models with those of the complete model.

[0111] Experimental results show that removing any core data augmentation strategy or key model structure significantly degrades the model's performance on the test set. Detailed experimental data can be found in Appendix Table 3.

[0112] Appendix 3

[0113]

[0114] The above ablation experiment results further demonstrate the necessity and synergistic gain of the various technical innovations of this invention, which play a key role in improving the detection generalization ability of complex high-dimensional spectral data, suppressing overfitting, and extracting intrinsic features.

[0115] Model quantization capability

[0116] To meet the diverse needs of industrial deployments, the MTMoE-CNN model established according to this invention was tested for its actual inference performance under different model compression ratios. Specific performance details are shown in Appendix Table 4 below.

[0117] Appendix 4

[0118]

[0119] In summary, the MTMoE-CNN network proposed in this invention, along with its automatic data augmentation strategy, hybrid expert weighted regression, and compressed quantization deployment mechanism, significantly improves the classification and quantitative performance of LIBS adulteration detection in milk powder under high-dimensional, small-sample conditions. It is also suitable for high-throughput automated detection in industrial settings, providing a complete and feasible technical solution for the practical industrial application of LIBS technology.

Claims

1. A hypothetical quantitative synergistic analysis method for adulteration of milk powder based on time-resolved LIBS, characterized in that, include: S1. Obtain LIBS spectral data of milk powder samples at multiple delay times to form a time-resolved spectral matrix; S2. Extract features from the spectral matrix using a feature extraction network to obtain multi-scale feature vectors; S3. Input the multi-scale features into the classification branch and the regression branch based on the hybrid experts. The classification branch outputs the probability of each adulteration type. The regression branch weights the concentration prediction results of each expert sub-network according to the probability and outputs the final adulteration concentration. S4. Use a joint loss function to train the model end-to-end; S5. Output the adulteration type and concentration of milk powder samples to achieve a combined qualitative and quantitative analysis.

2. The method as described in claim 1, characterized in that, The resolution spectral matrix in step S1 is: S=[s ij ]∈R C×L Where C is the number of time delay channels, L is the spectral length of each channel, and s ij =I(t) i ,λ j ) represents the i-th time delay and wavelength λ j The light intensity at that location.

3. The method as described in claim 2, characterized in that, Before the time-resolved spectral data in step S2 is input into the feature extraction network, data augmentation operations are applied, including one or more of Gaussian noise injection, frequency masking, and normalization perturbation. The data augmentation parameters are adaptively determined by a Bayesian optimization algorithm.

4. The method as described in claim 3, characterized in that, Gaussian noise probability [0,1], standard deviation [10] -6 [0, 0.15]; frequency masking probability [0, 1], width [100, 2000]; normalized perturbation mean [-0.1, 0.1], variance [10, 0.15] -4 ,0.05].

5. The method as described in claim 1, characterized in that, The feature extraction network in step S2 integrates one-dimensional convolution, TCN and Inception structures, wherein the number of parallel Inception branches is 3 to 7.

6. The method as described in claim 1, characterized in that, The regression branch in step S3 includes an expert subnetwork equal to the number of adulteration types. The probability output by the classification branch is used as a weighting coefficient to weight and fuse the regression outputs of each expert subnetwork.

7. A time-resolved LIBS-based hypothetical quantitative synergistic analysis system for milk powder adulteration, characterized in that, It includes a time-resolved LIBS spectral acquisition module, an automatic data augmentation module, a multi-scale feature extraction module, a classification-regression co-analysis module, a processor, and a memory; The time-resolved LIBS spectral acquisition module is used to acquire LIBS spectral data of milk powder samples at multiple delay times to form a time-resolved spectral matrix. The automatic data augmentation module is used to perform multi-stage data augmentation on the time-resolved spectrum of the set; The multi-scale feature extraction module is used to extract features from the enhanced spectral data; The classification-regression co-analysis module is used to simultaneously input the extracted multi-scale feature vectors into the classification branch and the hybrid expert regression branch to obtain the final adulteration concentration. The memory is used to store one or more programs, which, when executed by a processor, implement the steps of the method according to any one of claims 1-6.

Images