LIBS multi-component quantitative analysis method based on spectral mechanism constraint and feature decoupling

Abstract

The application relates to the field of intelligent analysis, and discloses a LIBS multi-component quantitative analysis method based on spectrum mechanism constraint and feature decoupling. Target spectrum data is acquired and pretreated. Spectrum mechanism sensing features are extracted according to the pretreated target spectrum data. Component-specific feature decoupling is performed according to the spectrum mechanism sensing features. Adaptive error regulation is performed. Finally, multi-component quantitative analysis results are output. The output multi-component quantitative analysis results are based on concentration range constraint and multi-component consistency constraint. The application realizes spectrum feature decoupling, avoids the problem of insufficient attention to low-content components, and improves the accuracy and stability of analysis.

Description

Technical Field

[0001] This invention relates to the field of intelligent analysis, and in particular to a LIBS multi-component quantitative analysis method based on spectral mechanism constraints and feature decoupling. Background Technology

[0002] Laser-induced breakdown spectroscopy (LIBS) has been widely applied in the multi-element composition analysis of solid, powder, and liquid samples due to its advantages such as no sample pretreatment required, fast analysis speed, and the ability to perform in-situ and online detection. However, in LIBS multi-component analysis, the characteristic spectral lines of different elements overlap in wavelength space, and the matrix effect leads to significant coupling between elemental signals, making it difficult for the model to distinguish the independent contributions of different components. Furthermore, for low-abundance elements, their emission signals are weak and easily dominated by high-abundance components during training, further reducing quantitative accuracy.

[0003] To address the aforementioned issues, existing technologies typically employ partial least squares, support vector machines, and random forest methods to model spectral data. However, these existing technologies generally suffer from the following shortcomings: 1. The feature extraction and modeling process did not distinguish between common component information and component-specific information; 2. The sharing of the same error feedback mechanism among different components leads to insufficient attention to low-content components during model training; 3. It is difficult to effectively improve the prediction performance of low-content components while simultaneously modeling multiple components.

[0004] Therefore, designing a multi-component quantitative analysis method to avoid the shortcomings of characteristic coupling and insufficient attention to low-content components, and to improve the accuracy and stability of the analysis, has become an urgent problem to be solved. Summary of the Invention

[0005] Based on this, the present invention proposes a LIBS multi-component quantitative analysis method based on spectral mechanism constraints and feature decoupling. It constructs a spectral mechanism-aware feature representation based on characteristic spectral line distribution to reduce interference from non-feature regions. Furthermore, it constructs component-specific feature modeling units and introduces feature correlation constraints to achieve decoupled modeling of multi-component spectral features. A two-factor adaptive error control mechanism based on component concentration distribution and prediction error is designed to optimize the error weight allocation of each component during model training. Concentration range constraints and multi-component consistency constraints are also applied during the model output stage to further improve the physical rationality of the prediction results. This effectively improves the quantitative analysis accuracy of low-content components and overcomes the problems of severe spectral coupling and uneven error distribution in existing technologies.

[0006] This invention proposes a LIBS multi-component quantitative analysis method based on spectral mechanism constraints and feature decoupling, comprising: Acquire the target spectral data and perform preprocessing; Spectral mechanism sensing features are extracted based on the preprocessed target spectral data, and the spectral mechanism sensing feature extraction is based on an adaptive mechanism constraint weight function. Component-specific features are decoupled based on spectral mechanism sensing characteristics, and the component-specific feature decoupling is used to construct a component-specific feature modeling unit based on spectral mechanism sensing characteristics; Adaptive error control is performed, which is based on concentration distribution factor and prediction error factor; The final output is a multi-component quantitative analysis result, which is based on concentration range constraints and multi-component consistency constraints.

[0007] Furthermore, the step of acquiring the target spectral data and performing preprocessing specifically includes: Collect target spectral data, which includes LIBS full-band spectral data; The target spectral data is subjected to background continuous spectrum subtraction based on the SNIP algorithm, followed by Savitzky-Golay smoothing. The smoothing process has a window length of 11, a polynomial order of 3, and maximum value normalization. A stratified sampling strategy was adopted to divide the target spectral data dataset. The concentrations of each component were divided into three intervals: low, medium, and high according to their distribution range. Within each interval, training set, validation set, and test set were extracted in a ratio of 60%:20%:20% respectively.

[0008] Furthermore, the step of extracting spectral mechanism sensing features based on the preprocessed target spectral data specifically includes: Mechanistic enhancement is performed on the preprocessed target spectral data based on an adaptive mechanism-constrained weighting function. The specific algorithm for mechanism enhancement is as follows: , , in, Indicates the mechanism weighting coefficient. The standard relative spectral intensity of an element Indicates wavelength. This indicates the probability of a transition to the corresponding spectral line. Indicates the center wavelength of the characteristic spectral lines of the component to be measured. Indicates the spectral line broadening influence factor. Indicating mechanism-enhanced spectral data, Represents the target spectral data. This represents the Hadamard product operation; Based on the center position and influence range of the characteristic spectral lines of each component to be measured in the wavelength space, the spectrum of the target spectral data is divided into multiple spectral sub-intervals; Local spectral features are extracted from each spectral sub-interval, and then global spectral features are extracted. The local and global spectral features are then fused to obtain spectral mechanism perception features.

[0009] Furthermore, the step of decoupling component-specific features based on spectral mechanism sensing features specifically includes: Based on spectral mechanism sensing features, multiple independent component-specific feature modeling units are constructed. Each component-specific feature modeling unit has a unique corresponding component to be tested. The component-specific feature modeling unit includes a one-dimensional convolutional layer, a batch normalization layer, a nonlinear activation function layer, and a regression output layer. The one-dimensional convolutional layer has a kernel size of 5, a stride of 1, and a padding method of "same". The regression output layer has an output dimension of 1. Component-specific feature decoupling is performed based on a feature decoupling loss function, which is based on a feature decoupling regularization constraint term. The specific algorithm for the feature decoupling loss function is as follows: , in, The loss represents the feature decoupling loss, where N represents the number of component-specific feature modeling units, and i and j represent two distinct component-specific feature modeling units. and These represent the extracted feature vectors of two different component-specific feature modeling units. It represents a very small number.

[0010] Furthermore, the step of adaptive error control specifically includes: Adaptive error control is performed based on a two-factor dynamic error control mechanism, which is based on a concentration distribution factor and a prediction error factor. The specific algorithm of the two-factor dynamic error control mechanism is as follows: , , , in, Indicates the concentration distribution factor. This represents the average concentration value of the component corresponding to the i-th component-specific feature modeling unit in the entire training set. Represents the smoothing constant. Indicates the prediction error factor. and Let represent the root mean square error of the component corresponding to the i-th and j-th component-specific feature modeling units on the validation set in the previous evaluation period, respectively; N represents the number of component-specific feature modeling units; and i and j represent two distinct component-specific feature modeling units. Represents a very small number. Indicates the weight of the overall error feedback; The objective is optimized based on the total loss function, the specific algorithm for which the total loss function is defined is as follows: , in, Indicates the total loss. This represents the mean squared error loss of the component corresponding to the i-th component-specific feature modeling unit. The hyperparameters representing the decoupling constraint terms, This represents the feature decoupling loss; An update cycle constraint is applied to the comprehensive error feedback weights, and the update cycle constraint is based on the training rounds.

[0011] Furthermore, the step of finally outputting the multi-component quantitative analysis results specifically includes: Based on the modeling unit of each component's specific characteristics after adaptive error control, the quantitative analysis results corresponding to each component to be tested are output, and the quantitative analysis results are subject to concentration range constraints and multi-component consistency constraints. The specific algorithm for the concentration range constraint is as follows: , in, This represents the constrained prediction value output by the i-th component-specific feature modeling unit. This represents the initial predicted value output by the i-th component-specific feature modeling unit. and They represent the first i The minimum and maximum concentration values ​​of the component corresponding to each component-specific feature modeling unit in the training set samples; The specific algorithm for the multi-component consistency constraint is as follows: , in, This represents a multi-component consistency penalty term, where K represents the number of stoichiometric relationships requiring consistency constraints, and k represents the ordinal number of the stoichiometric relationships requiring consistency constraints. This represents the k-th stoichiometric relationship obtained from the preliminary prediction calculation, where the stoichiometric relationship is a component ratio relationship. Theoretical stoichiometric values ​​representing the proportions of components; The total loss function is expanded based on a multi-component consistency penalty term. The specific algorithm for the expanded total loss function is as follows: , in, Indicates the total loss. This represents the mean squared error loss of the component corresponding to the i-th component-specific feature modeling unit. The hyperparameters representing the decoupling constraint terms, This represents the feature decoupling loss, where N represents the number of component-specific feature modeling units. Hyperparameters representing consistency constraints; Based on the predicted values ​​constrained by concentration range and multi-component consistency, the results of multi-component quantitative analysis are obtained.

[0012] This invention proposes a LIBS multi-component quantitative analysis system based on spectral mechanism constraints and feature decoupling, comprising: The preprocessing module is used to acquire and preprocess the target spectral data; The feature extraction module is used to extract spectral mechanism-aware features based on the preprocessed target spectral data. The spectral mechanism-aware feature extraction is based on an adaptive mechanism constraint weighting function. The feature decoupling module is used to decouple component-specific features based on spectral mechanism sensing features. The component-specific feature decoupling is used to construct a component-specific feature modeling unit based on spectral mechanism sensing features. An error control module is used to perform adaptive error control, which is based on a concentration distribution factor and a prediction error factor. The results output module is used to output the final multi-component quantitative analysis results, which are based on concentration range constraints and multi-component consistency constraints.

[0013] The present invention also provides a storage medium that stores one or more programs, which, when executed by a processor, implement the LIBS multi-component quantitative analysis method based on spectral mechanism constraints and feature decoupling as described above.

[0014] The present invention also provides a computer device, the computer device including a memory and a processor, wherein: The memory is used to store computer programs; When the processor executes the computer program stored in the memory, it implements the LIBS multi-component quantitative analysis method based on spectral mechanism constraints and feature decoupling as described above.

[0015] The present invention has at least the following beneficial effects: 1. The spectrum is divided into sub-intervals based on the distribution of elemental characteristic spectral lines, and a spectral mechanism perception feature representation is constructed on this basis. At the same time, a weight function based on the physical properties of spectral lines (standard relative spectral line intensity, transition probability and spectral line broadening factor) is introduced, which can enhance the model's response to key spectral line regions, effectively suppress interference from non-feature regions, and thus improve the accuracy and robustness of spectral feature extraction. 2. By constructing component-specific feature modeling units and introducing feature correlation constraints, different components can form independent representations in the feature space. Compared with traditional unified modeling methods, this can effectively reduce the multi-component spectral coupling effect and improve the model's ability to distinguish between components. 3. By introducing a two-factor adaptive error control mechanism based on component concentration distribution and prediction error, the dynamic allocation of loss weights for different components is achieved, so that low-content components can obtain higher optimization weights during model training, thereby significantly improving the quantitative analysis accuracy of low-content components. 4. The error control mechanism can adaptively adjust the training strategy according to the component concentration level and prediction error, and update the weights according to a preset period (3 to 10 training rounds) to avoid training oscillations. Without increasing the complexity of the model, it can achieve a balanced improvement in the learning ability of each component in the multi-component modeling process. 5. By applying concentration range constraints and multi-component consistency constraints in the model output stage, abnormal predicted values ​​can be effectively suppressed, and the predicted results of multiple components can meet known stoichiometric relationships, thereby further improving the accuracy and physical rationality of quantitative analysis. 6. The method of the present invention is applicable to the simultaneous quantitative analysis of multiple components in various forms of samples such as solids, powders and liquids, and has good versatility, stability and engineering application value. Attached Figure Description

[0016] Figure 1 This is a flowchart of the LIBS multi-component quantitative analysis method based on spectral mechanism constraints and feature decoupling proposed in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the LIBS multi-component quantitative analysis system based on spectral mechanism constraints and feature decoupling proposed in the second embodiment of the present invention; Figure 3 This is a schematic diagram of the multi-branch spectral deep learning model structure proposed in the first embodiment of the present invention; Figure 4 This is a schematic diagram of the adaptive error control mechanism proposed in the first embodiment of the present invention; Figure 5 This is a comparison chart of the root mean square error of the comparative experiment in the first embodiment of the present invention; The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0017] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0018] It should be noted that when a component is said to be "fixed to" another component, it can be directly on the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0020] Please see Figure 1 The diagram shows a flowchart of the LIBS multi-component quantitative analysis method based on spectral mechanism constraints and feature decoupling proposed in the first embodiment of the present invention. This LIBS multi-component quantitative analysis method based on spectral mechanism constraints and feature decoupling includes steps S01 to S05, wherein: Step S01: Acquire target spectral data and perform preprocessing; It should be noted that in this embodiment, target spectral data is collected, which includes LIBS full-band spectral data; The target spectral data is subjected to background continuous spectrum subtraction based on the SNIP algorithm, followed by Savitzky-Golay smoothing. The smoothing process has a window length of 11, a polynomial order of 3, and maximum value normalization. A stratified sampling strategy was adopted to divide the target spectral data dataset. The concentrations of each component were divided into three intervals: low, medium, and high according to their distribution range. Within each interval, training set, validation set, and test set were extracted in a ratio of 60%:20%:20% respectively.

[0021] In a preferred embodiment, to enhance model robustness, random noise perturbation is applied to the standardized spectral data during the model training phase. The random noise perturbation is additive Gaussian noise with a mean of zero and a standard deviation of 1% to 5% of the standard deviation of the spectral signal, preferably 2%. This processing is performed only during the training phase; no noise perturbation is applied during the verification and testing phases.

[0022] Step S02: Extract spectral mechanism sensing features based on the preprocessed target spectral data; It should be noted that in this embodiment, the spectral mechanism sensing feature extraction is based on an adaptive mechanism constraint weight function, and the preprocessed target spectral data is enhanced based on the adaptive mechanism constraint weight function. The specific algorithm for mechanism enhancement is as follows: , , in, Indicates the mechanism weighting coefficient. The standard relative spectral intensity of an element Indicates wavelength. This indicates the probability of a transition to the corresponding spectral line. Indicates the center wavelength of the characteristic spectral lines of the component to be measured. Indicates the spectral line broadening influence factor. Indicating mechanism-enhanced spectral data, Represents the target spectral data. This represents the Hadamard product operation; Based on the center position and influence range of the characteristic spectral lines of each component to be measured in the wavelength space, the spectrum of the target spectral data is divided into multiple spectral sub-intervals; Local spectral features are extracted from each spectral sub-interval, and then global spectral features are extracted. The local and global spectral features are then fused to obtain spectral mechanism perception features.

[0023] The mechanistic constraint weights are not randomly initialized, but are calculated based on prior physical parameters in the spectral database, so that the model has a bias towards the characteristic spectral region in the initial state.

[0024] Unlike conventional deep learning models that automatically learn features directly from raw spectral data, the “spectral mechanism perception” feature extraction described in this invention has the following distinctive features: (1) The sub-interval division is not uniform but is non-uniformly divided based on the known feature spectral line positions of each element in the NIST atomic spectral database, so that each sub-interval physically corresponds to the feature response region of a specific element; (2) The mechanism constraint weight function introduces the physical parameter of transition probability, so that the model gives innate higher attention to spectral line transitions with higher probability of occurrence, rather than relying entirely on data-driven methods; (3) The mechanism weight is applied to the input spectrum through the Hadamard product operation, realizing the guided constraint of physical prior knowledge on the data-driven model, which is the core feature that distinguishes this invention from the pure data-driven method.

[0025] Step S03: Decouple component-specific features based on spectral mechanism sensing characteristics; It should be noted that in this embodiment, the component-specific feature decoupling is used to construct a component-specific feature modeling unit based on the spectral mechanism sensing features. Multiple independent component-specific feature modeling units are constructed based on the spectral mechanism sensing features. Each component-specific feature modeling unit has a unique corresponding component to be measured. The component-specific feature modeling unit includes a one-dimensional convolutional layer, a batch normalization layer, a nonlinear activation function layer, and a regression output layer. The kernel size of the one-dimensional convolutional layer is 5, the stride is 1, and the padding method is "same". The output dimension of the regression output layer is 1. Component-specific feature decoupling is performed based on a feature decoupling loss function, which is based on a feature decoupling regularization constraint term. The specific algorithm for the feature decoupling loss function is as follows: , in, The loss represents the feature decoupling loss, where N represents the number of component-specific feature modeling units, and i and j represent two distinct component-specific feature modeling units. and These represent the extracted feature vectors of two different component-specific feature modeling units. It represents a very small number.

[0026] This invention minimizes this constraint term through backpropagation, forcing the network to ensure that the hidden layer feature vectors learned by different component-specific branches tend to be orthogonal during parameter updates. Physically, even if the input LIBS spectrum overlaps at the cobalt (Co) and iron (Fe) spectral lines, after decoupling training, the feature vector of the cobalt branch is only sensitive to Co concentration, and the feature vector of the iron branch is only sensitive to Fe concentration, thus achieving component-level feature decoupling.

[0027] The "feature decoupling" described in this invention differs from the hard or soft parameter sharing mechanisms in conventional multi-task learning. Conventional multi-task learning achieves multi-component prediction only through independent branch output layers, but the features of different components in the shared feature extraction layer remain highly coupled. This invention explicitly introduces a feature correlation constraint term, Ldecouple, into the loss function, forcing the hidden layer feature vectors of different branches to tend towards orthogonality, thereby achieving active separation of information between components at the feature space level. This mechanism allows even if the feature spectral lines of two components severely overlap in the original wavelength space (e.g., the 345.35nm spectral line of Co and the 345.28nm spectral line of Fe), after decoupling training, the two branches can still learn mutually independent discriminative features.

[0028] Step S04: Perform adaptive error control; It should be noted that in this embodiment, the adaptive error control is based on the concentration distribution factor and the prediction error factor, and is performed based on a two-factor dynamic error control mechanism. The specific algorithm of the two-factor dynamic error control mechanism is as follows: , , , in, Indicates the concentration distribution factor. This represents the average concentration value of the component corresponding to the i-th component-specific feature modeling unit in the entire training set. Represents the smoothing constant. Indicates the prediction error factor. and Let represent the root mean square error of the component corresponding to the i-th and j-th component-specific feature modeling units on the validation set in the previous evaluation period, respectively; N represents the number of component-specific feature modeling units; and i and j represent two distinct component-specific feature modeling units. Represents a very small number. Indicates the weight of the overall error feedback; The objective is optimized based on the total loss function, the specific algorithm for which the total loss function is defined is as follows: , in, Indicates the total loss. This represents the mean squared error loss of the component corresponding to the i-th component-specific feature modeling unit. The hyperparameters representing the decoupling constraint terms, This represents the feature decoupling loss; An update cycle constraint is applied to the comprehensive error feedback weights, and the update cycle constraint is based on the training rounds.

[0029] To avoid frequent weight oscillations that could lead to training instability, the error feedback weights are not updated at every iteration, but rather according to a preset training epoch period K. As a preferred implementation, K ranges from 3 to 10 training epochs, with K=5 being the preferred value. This means that after every 5 complete dataset traversals (Epochs), βi(t) is recalculated and the loss weights of each branch are updated.

[0030] In the described dual-factor mechanism, the concentration distribution factor is a static prior factor, determined statistically based on the concentration labels of the dataset before training begins, reflecting the inherent sparsity of each component in the sample space. The prediction error factor is a dynamic feedback factor, updated periodically as training progresses, reflecting the model's learning progress for each component. This product-like combination ensures that for a low-concentration component, even with a large initial prediction error, it can receive a higher weight in the early stages of training to accelerate learning. As prediction accuracy improves, the weight gradually decreases to a steady-state level dominated by the concentration factor, avoiding overfitting. This dual mechanism of "static guidance + dynamic adjustment" is the key feature that distinguishes the adaptive error control of this invention from fixed-weighted or purely error-driven methods.

[0031] Step S05: Finally output the multi-component quantitative analysis results; It should be noted that in this embodiment, the output of multi-component quantitative analysis results is based on concentration range constraints and multi-component consistency constraints. The quantitative analysis results corresponding to each analyte are output by the modeling unit of each component specific characteristics after adaptive error adjustment, and the quantitative analysis results are subject to concentration range constraints and multi-component consistency constraints. The specific algorithm for the concentration range constraint is as follows: , in, This represents the constrained prediction value output by the i-th component-specific feature modeling unit. This represents the initial predicted value output by the i-th component-specific feature modeling unit. and They represent the first i The minimum and maximum concentration values ​​of the component corresponding to each component-specific feature modeling unit in the training set samples; The specific algorithm for the multi-component consistency constraint is as follows: , in, This represents a multi-component consistency penalty term, where K represents the number of stoichiometric relationships requiring consistency constraints, and k represents the ordinal number of the stoichiometric relationships requiring consistency constraints. This represents the k-th stoichiometric relationship obtained from the preliminary prediction calculation, where the stoichiometric relationship is a component ratio relationship. Theoretical stoichiometric values ​​representing the proportions of components; The total loss function is expanded based on a multi-component consistency penalty term. The specific algorithm for the expanded total loss function is as follows: , in, Indicates the total loss. This represents the mean squared error loss of the component corresponding to the i-th component-specific feature modeling unit. The hyperparameters representing the decoupling constraint terms, This represents the feature decoupling loss, where N represents the number of component-specific feature modeling units. Hyperparameters representing consistency constraints; Based on the predicted values ​​constrained by concentration range and multi-component consistency, the results of multi-component quantitative analysis are obtained.

[0032] In the model output stage, in order to further ensure that the prediction results conform to the actual physical meaning and eliminate abnormal prediction values, the present invention applies two physical constraints to the preliminary prediction results output by the model. To avoid outliers in the model output that exceed the reasonable concentration range, a concentration range constraint is imposed on the initial predicted values ​​of the components, cropping them to between the minimum and maximum concentrations of the component in the training set. This results in constrained predicted values. The concentration range constraint ensures that all predicted concentrations fall within the reasonable concentration range observed in the training set, effectively suppressing outliers caused by spectral noise or model uncertainty.

[0033] For component systems with known stoichiometric relationships (such as the specific molar ratio between lithium (Li) and transition metals like nickel (Ni), cobalt (Co), and manganese (Mn) in lithium-ion battery cathode material black powder), this invention further introduces a multi-component consistency penalty term into the loss function during model training to constrain the proportional relationship between predicted values ​​to conform to the theoretical stoichiometric ratio. By introducing the aforementioned concentration range constraint and multi-component consistency constraint, this invention can effectively suppress abnormal predicted values ​​and ensure that the predicted results of multiple components satisfy the known stoichiometric relationship, thereby further improving the accuracy and physical rationality of quantitative analysis.

[0034] Comparative experiments were conducted based on the multi-branch spectral deep learning model of this embodiment. For the specific structure of the multi-branch spectral deep learning model, please refer to [link / reference needed]. Figure 3 The specific experimental setup is as follows: In the specific implementation, the input spectrum is divided into several sub-intervals based on the wavelength range of the characteristic spectral lines of the target element. Each sub-interval corresponds to the feature response region of at least one element, and feature extraction is performed through local convolution or weighting mechanisms. The feature decoupling constraint is implemented by introducing a feature correlation penalty term or an orthogonality constraint term into the loss function to reduce the similarity between the feature representations of different components.

[0035] In this embodiment: spectral wavelength range: 200–900 nm; spectral dimensions: 7902 wavelength variables; number of samples: 600 groups; number of convolutional layer channels: 32 / 64 / 128; optimizer: Adam; initial learning rate: 0.001; batch size: 32; number of training epochs: 800 epochs.

[0036] Different types of lithium battery black powder were selected, with the main components including six elements: lithium (Li), nickel (Ni), cobalt (Co), manganese (Mn), iron (Fe), and aluminum (Al). Full-band spectral data were acquired using a LIBS device. Background continuum subtraction and intensity normalization were performed on the acquired spectral data. During the training phase, additive Gaussian noise perturbation was applied to the standardized spectral data, with the noise standard deviation being 2% of the spectral signal standard deviation, to enhance model robustness. A multi-branch deep learning model was constructed, comprising a shared spectral feature extraction module and multiple component-specific feature modeling units, and a joint training method was used for model training.

[0037] During training, the error feedback weights of branches corresponding to low-content components are dynamically increased based on the content distribution of each component and the statistical information of prediction errors. Specifically, the error adjustment weight update mechanism is as follows: during model training, the mean prediction error of each component is calculated every 5 training epochs, and combined with the average concentration of each component in the dataset, the loss weight coefficient of the corresponding component is updated according to the above formula.

[0038] During the model output phase, a concentration range constraint is imposed on the initial predicted values, cropping them to between the minimum and maximum concentrations of each component in the training set. Simultaneously, a multi-component consistency penalty term is introduced into the loss function to constrain the molar ratio of Li to the total amount of transition metals (Ni+Co+Mn) to be close to the theoretical value of 1:1, with the hyperparameter η set to 0.01.

[0039] To verify the technical effectiveness of the adaptive error control mechanism proposed in this invention, please refer to the details of the adaptive error control mechanism. Figure 4 Under the same dataset, network structure parameters, training epochs, and optimization algorithms, the following two models were constructed for comparison: (1) Multi-branch feature decoupling model without error control mechanism (referred to as "model without error control mechanism"); (2) Introduce the multi-branch feature decoupling model of the adaptive error control mechanism of the present invention (denoted as "the model with error control mechanism").

[0040] The two models differ only in the way the loss function is constructed; the rest—network structure, parameter size, training strategy, and data partitioning method—are completely identical to ensure that the experimental results have a strict basis of controlled variables.

[0041] Calculate the root mean square error (RMSE) for each component on the test set. For comparison results, please refer to [link to relevant documentation]. Figure 5 The details are shown in Table 1.

[0042] Table 1 Comparison of RMSE of each component with and without error control mechanism. As shown in Table 1, after adopting the adaptive error control mechanism of the present invention, the RMSE of low-content components (such as Al, Co, Ni, Mn) decreased significantly more than that of high-content components (such as Fe, Li), which verifies the targeted improvement effect of the method of the present invention on the quantitative accuracy of low-content components.

[0043] The experimental results above show that by constructing a two-factor weighting mechanism based on concentration distribution characteristics and prediction error statistical characteristics, this invention redistributes the error feedback intensity of different branches during multi-component synchronous modeling, enabling the model to give higher optimization weights to low-content components during parameter updates, thereby effectively alleviating the problem of low-content components being "masked" by high-content components in multi-component coupled modeling.

[0044] Existing multi-component modeling methods typically employ a uniform loss function or a fixed weight structure, making it difficult to dynamically adjust based on the concentration distribution characteristics of different components. This invention, by incorporating concentration distribution information into the construction of the loss function weights, achieves an adaptive error control mechanism based on physical content distribution. This mechanism is not a simple multi-task learning or conventional weighted averaging, but rather dynamically adjusts the branch optimization intensity during training, demonstrating a clear technical focus.

[0045] Therefore, the above comparative experimental results verify that there is a direct causal relationship between the technical features described in this invention and its technical effects. It can significantly improve the quantitative accuracy of low-content components without increasing the complexity of the network structure, demonstrating outstanding substantive features and significant progress.

[0046] Please see Figure 2 The figure shows a schematic diagram of the LIBS multi-component quantitative analysis system based on spectral mechanism constraints and feature decoupling proposed in the second embodiment of the present invention. The system includes: Preprocessing module 10 is used to acquire target spectral data and perform preprocessing; Feature extraction module 20 is used to extract spectral mechanism sensing features based on preprocessed target spectral data, wherein the spectral mechanism sensing feature extraction is based on an adaptive mechanism constraint weight function. The feature decoupling module 30 is used to decouple component-specific features based on spectral mechanism sensing features. The component-specific feature decoupling is used to construct a component-specific feature modeling unit based on spectral mechanism sensing features. Error control module 40 is used to perform adaptive error control, which is based on concentration distribution factor and prediction error factor. The result output module 50 is used to finally output the multi-component quantitative analysis results, which are based on concentration range constraints and multi-group consistency constraints.

[0047] The present invention also proposes a computer storage medium storing one or more programs that, when executed by a processor, implement the above-described LIBS multi-component quantitative analysis method based on spectral mechanism constraints and feature decoupling.

[0048] The present invention also proposes a computer device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to realize the above-mentioned LIBS multi-component quantitative analysis method based on spectral mechanism constraints and feature decoupling.

[0049] Those skilled in the art will understand that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can mean any means that can contain stored, communicated, propagated, or transmitted programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0050] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0051] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0052] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0053] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A LIBS multi-component quantitative analysis method based on spectral mechanism constraints and feature decoupling, characterized in that, include: Acquire the target spectral data and perform preprocessing; Spectral mechanism sensing features are extracted based on the preprocessed target spectral data, and the spectral mechanism sensing feature extraction is based on an adaptive mechanism constraint weight function. The step of extracting spectral mechanism sensing features based on the preprocessed target spectral data specifically includes: Mechanistic enhancement is performed on the preprocessed target spectral data based on an adaptive mechanism-constrained weighting function. The specific algorithm for mechanism enhancement is as follows: ， ， in, Indicates the mechanism weighting coefficient. The standard relative spectral intensity of an element Indicates wavelength. This indicates the probability of a transition to the corresponding spectral line. Indicates the center wavelength of the characteristic spectral lines of the component to be measured. Indicates the spectral line broadening influence factor. Indicating mechanism-enhanced spectral data, Represents the target spectral data. This represents the Hadamard product operation; Based on the center position and influence range of the characteristic spectral lines of each component to be measured in the wavelength space, the spectrum of the target spectral data is divided into multiple spectral sub-intervals; Local spectral features are extracted from each spectral sub-interval, and then global spectral features are extracted. The local and global spectral features are then fused to obtain spectral mechanism perception features. Component-specific features are decoupled based on spectral mechanism sensing characteristics, and the component-specific feature decoupling is used to construct a component-specific feature modeling unit based on spectral mechanism sensing characteristics; The step of decoupling component-specific features based on spectral mechanism sensing features specifically includes: Based on spectral mechanism sensing features, multiple independent component-specific feature modeling units are constructed. Each component-specific feature modeling unit has a unique corresponding component to be tested. The component-specific feature modeling unit includes a one-dimensional convolutional layer, a batch normalization layer, a nonlinear activation function layer, and a regression output layer. The one-dimensional convolutional layer has a kernel size of 5, a stride of 1, and a padding method of "same". The regression output layer has an output dimension of 1. Component-specific feature decoupling is performed based on a feature decoupling loss function, which is based on a feature decoupling regularization constraint term. The specific algorithm for the feature decoupling loss function is as follows: ， in, The loss represents the feature decoupling loss, where N represents the number of component-specific feature modeling units, and i and j represent two distinct component-specific feature modeling units. and These represent the extracted feature vectors of two different component-specific feature modeling units. Represents a very small number; Adaptive error control is performed, which is based on concentration distribution factor and prediction error factor; The steps for adaptive error control specifically include: Adaptive error control is performed based on a two-factor dynamic error control mechanism, which is based on a concentration distribution factor and a prediction error factor. The specific algorithm of the two-factor dynamic error control mechanism is as follows: ， ， ， in, Indicates the concentration distribution factor. This represents the average concentration value of the component corresponding to the i-th component-specific feature modeling unit in the entire training set. Represents the smoothing constant. Indicates the prediction error factor. and Let represent the root mean square error of the component corresponding to the i-th and j-th component-specific feature modeling units on the validation set in the previous evaluation period, respectively. Let N represent the number of component-specific feature modeling units, and i and j represent two different component-specific feature modeling units. Represents a very small number. Indicates the weight of the overall error feedback; The objective is optimized based on the total loss function, the specific algorithm for which the total loss function is defined is as follows: ， in, Indicates the total loss. This represents the mean squared error loss of the component corresponding to the i-th component-specific feature modeling unit. The hyperparameters representing the decoupling constraint terms, This represents the feature decoupling loss; The update cycle constraint is applied to the comprehensive error feedback weights, and the update cycle constraint is based on the training rounds. The final output is a multi-component quantitative analysis result, which is based on concentration range constraints and multi-component consistency constraints.

2. The LIBS multi-component quantitative analysis method based on spectral mechanism constraints and feature decoupling as described in claim 1, characterized in that, The steps of acquiring target spectral data and preprocessing it specifically include: Collect target spectral data, which includes LIBS full-band spectral data; The target spectral data is subjected to background continuous spectrum subtraction based on the SNIP algorithm, followed by Savitzky-Golay smoothing. The smoothing process has a window length of 11, a polynomial order of 3, and maximum value normalization. A stratified sampling strategy was adopted to divide the target spectral data dataset. The concentrations of each component were divided into three intervals: low, medium, and high according to their distribution range. Within each interval, training set, validation set, and test set were extracted in a ratio of 60%:20%:20% respectively.

3. The LIBS multi-component quantitative analysis method based on spectral mechanism constraints and feature decoupling as described in claim 1, characterized in that, The steps for finally outputting multi-component quantitative analysis results specifically include: Based on the modeling unit of each component's specific characteristics after adaptive error control, the quantitative analysis results corresponding to each component to be tested are output, and the quantitative analysis results are subject to concentration range constraints and multi-component consistency constraints. The specific algorithm for the concentration range constraint is as follows: ， in, This represents the constrained prediction value output by the i-th component-specific feature modeling unit. This represents the initial predicted value output by the i-th component-specific feature modeling unit. and They represent the first i The minimum and maximum concentration values ​​of the component corresponding to each component-specific feature modeling unit in the training set samples; The specific algorithm for the multi-component consistency constraint is as follows: ， in, This represents a multi-component consistency penalty term, where K represents the number of stoichiometric relationships requiring consistency constraints, and k represents the ordinal number of the stoichiometric relationships requiring consistency constraints. This represents the k-th stoichiometric relationship obtained from the preliminary prediction calculation, where the stoichiometric relationship is a component ratio relationship. Theoretical stoichiometric values ​​representing the proportions of components; The total loss function is expanded based on a multi-component consistency penalty term. The specific algorithm for the expanded total loss function is as follows: ， in, Indicates the total loss. This represents the mean squared error loss of the component corresponding to the i-th component-specific feature modeling unit. The hyperparameters representing the decoupling constraint terms, This represents the feature decoupling loss, where N represents the number of component-specific feature modeling units. The hyperparameters representing the consistency constraints; Based on the predicted values ​​constrained by concentration range and multi-component consistency, the results of multi-component quantitative analysis are obtained.

4. A LIBS multi-component quantitative analysis system based on spectral mechanism constraints and feature decoupling, characterized in that, include: The preprocessing module is used to acquire and preprocess the target spectral data; The feature extraction module is used to extract spectral mechanism-aware features based on the preprocessed target spectral data. The spectral mechanism-aware feature extraction is based on an adaptive mechanism constraint weighting function. The step of extracting spectral mechanism sensing features based on the preprocessed target spectral data specifically includes: Mechanistic enhancement is performed on the preprocessed target spectral data based on an adaptive mechanism-constrained weighting function. The specific algorithm for mechanism enhancement is as follows: ， ， in, Indicates the mechanism weighting coefficient. The standard relative spectral intensity of an element Indicates wavelength. This indicates the probability of a transition to the corresponding spectral line. Indicates the center wavelength of the characteristic spectral lines of the component to be measured. Indicates the spectral line broadening influence factor. Indicating mechanism-enhanced spectral data, Represents the target spectral data. This represents the Hadamard product operation; Based on the center position and influence range of the characteristic spectral lines of each component to be measured in the wavelength space, the spectrum of the target spectral data is divided into multiple spectral sub-intervals; Local spectral features are extracted from each spectral sub-interval, and then global spectral features are extracted. The local and global spectral features are then fused to obtain spectral mechanism perception features. The feature decoupling module is used to decouple component-specific features based on spectral mechanism sensing features. The component-specific feature decoupling is used to construct a component-specific feature modeling unit based on spectral mechanism sensing features. The step of decoupling component-specific features based on spectral mechanism sensing features specifically includes: Based on spectral mechanism sensing features, multiple independent component-specific feature modeling units are constructed. Each component-specific feature modeling unit has a unique corresponding component to be tested. The component-specific feature modeling unit includes a one-dimensional convolutional layer, a batch normalization layer, a nonlinear activation function layer, and a regression output layer. The one-dimensional convolutional layer has a kernel size of 5, a stride of 1, and a padding method of "same". The regression output layer has an output dimension of 1. Component-specific feature decoupling is performed based on a feature decoupling loss function, which is based on a feature decoupling regularization constraint term. The specific algorithm for the feature decoupling loss function is as follows: ， in, The loss represents the feature decoupling loss, where N represents the number of component-specific feature modeling units, and i and j represent two distinct component-specific feature modeling units. and These represent the extracted feature vectors of two different component-specific feature modeling units. Represents a very small number; An error control module is used to perform adaptive error control, which is based on a concentration distribution factor and a prediction error factor. The steps for adaptive error control specifically include: Adaptive error control is performed based on a two-factor dynamic error control mechanism, which is based on a concentration distribution factor and a prediction error factor. The specific algorithm of the two-factor dynamic error control mechanism is as follows: ， ， ， in, Indicates the concentration distribution factor. This represents the average concentration value of the component corresponding to the i-th component-specific feature modeling unit in the entire training set. Represents the smoothing constant. Indicates the prediction error factor. and Let represent the root mean square error of the component corresponding to the i-th and j-th component-specific feature modeling units on the validation set in the previous evaluation period, respectively. Let N represent the number of component-specific feature modeling units, and i and j represent two different component-specific feature modeling units. Represents a very small number. Indicates the weight of the overall error feedback; The objective is optimized based on the total loss function, the specific algorithm for which the total loss function is defined is as follows: ， in, Indicates the total loss. This represents the mean squared error loss of the component corresponding to the i-th component-specific feature modeling unit. The hyperparameters representing the decoupling constraint terms, This represents the feature decoupling loss; The update cycle constraint is applied to the comprehensive error feedback weights, and the update cycle constraint is based on the training rounds. The results output module is used to output the final multi-component quantitative analysis results, which are based on concentration range constraints and multi-component consistency constraints.

5. A storage medium, characterized in that, The storage medium stores one or more programs that, when executed by a processor, implement the LIBS multi-component quantitative analysis method based on spectral mechanism constraints and feature decoupling as described in any one of claims 1-3.

6. A computer device, characterized in that, The computer device includes a memory and a processor, wherein: The memory is used to store computer programs; When the processor executes the computer program stored in the memory, it implements the LIBS multi-component quantitative analysis method based on spectral mechanism constraints and feature decoupling as described in any one of claims 1-3.

Images