A raman spectrum decoupling method based on physical information neural network

Abstract

The application discloses a Raman spectrum decoupling method based on a physical information neural network and belongs to the technical field of spectral quantitative analysis. The application firstly collects pure substance Raman spectra, constructs spectrum generation sub-modules of each pure substance by taking a peak function as a neuron; then prepares binary standard samples, constructs a binary mixture spectrum neural network model, and calibrates the relative correction coefficients of each component relative to a reference substance; then, a multi-component mixture spectrum neural network model is constructed, which is composed of parallel pure substance spectrum generation sub-modules and offset correction modules, and the total loss function is the weighted sum of the spectrum reconstruction loss and the content loss based on the relative correction coefficient for joint optimization training; when applied, the input is the Raman spectrum of the sample to be measured, the weight parameters are optimized and mapped into the concentration of each component. In the complex multi-component system with serious peak overlap and nonlinear spectral behavior, the application can realize stable and interpretable Raman spectrum decoupling and quantitative analysis.

Description

Technical Field

[0001] This invention belongs to the field of quantitative spectral analysis and intelligent modeling technology, specifically relating to a Raman spectroscopy decoupling method based on a physical information neural network. Background Technology

[0002] Raman spectroscopy is widely used in chemical analysis and process monitoring due to its advantages such as speed, non-destructive nature, high sensitivity, and ability to provide molecular vibrational fingerprint information. It enables non-contact measurement through an optical window, making it suitable for online / in-situ monitoring of closed or continuous flow systems. However, in the quantitative analysis of complex mixed systems, Raman spectra often suffer from overlapping characteristic peaks and similar component structures, making direct separation of component signals difficult.

[0003] Existing Raman quantification methods are mostly based on the ideal assumption that the spectra of a mixture can be obtained by linear superposition of the spectra of pure components. Typical methods include chemometric models such as partial least squares regression (PLS) and linear regression models based on characteristic peaks. These methods can achieve good results when there are few components and the peaks are well separated; however, as the number of components in the mixture increases and peak overlap intensifies, the spectral decoupling problem tends to decompose into a strongly nonorthogonal structure, making it difficult for the model to stably distinguish the contributions of each component, and the regression results lack clear physical interpretability.

[0004] To improve interpretability and adaptability to overlapping peaks, Indirect Hard Modeling (IHM) represents the spectrum of a mixture as a linear combination of several Voigt peaks. It achieves spectral fitting by optimizing parameters such as peak position, peak width, and peak shape, and further infers component content using weighting coefficients. While IHM alleviates the difficulties caused by peak overlap to some extent, in multi-component organic mixtures such as aromatic nitration and ester nitration systems, as the number of components increases, interactions such as solvation, association, and medium polarization between components induce significant nonlinear spectral behavior, such as nonlinear scaling of peak intensity, peak broadening, and peak position shift, causing the mixture spectrum to deviate from the ideal linear superposition model. In this case, traditional fitting strategies relying solely on local peak parameter adjustments are prone to compensatory drift and non-physical convergence, leading to unstable quantitative results.

[0005] Therefore, there is an urgent need for a method that can achieve stable quantitative decoupling and maintain physical interpretability in Raman mixtures with strong overlap of multiple components and significant nonlinear effects, so as to learn the nonlinear spectral patterns of complex mixtures and thus provide a reliable data foundation for online / in-situ quantitative analysis and process monitoring. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a Raman spectroscopy decoupling method based on a physical information neural network. This method represents the Raman spectrum of a mixture as a differentiable neural network composed of several physically meaningful peak function generation units. Through backpropagation, it jointly optimizes peak parameters and component weights, learning the nonlinear spectral behavior of a multi-component system as the number of components increases under multi-sample constraints. This achieves robust decoupling and high-precision concentration inversion for strongly overlapping, strongly non-orthogonal, and nonlinearly perturbed spectra, while maintaining the physical interpretability at the peak parameter level.

[0007] To achieve the above objectives, the present invention proposes the following technical solution: The present invention provides a Raman spectroscopy decoupling method based on a physical information neural network, which includes the following steps:

[0008] S1: Collect the Raman spectra of each pure substance in the test system, and extract the characteristic parameter matrix of each pure substance by fitting the peak function; based on the characteristic parameter matrix, construct a spectrum generation submodule for each pure substance, the submodule consisting of several peak function neurons;

[0009] S2: Pair the reference substance with each other substance in the system to prepare a series of binary standard samples with different concentration gradients, and collect their Raman spectra; construct a binary mixture spectral neural network model consisting of a pure substance spectral generation submodule and concentration weight parameters, freeze the learnable parameters of the submodule, and fit to obtain the weight parameters corresponding to each binary standard sample; based on the Beer-Lambert law, use the weight parameters and the actual composition data of each binary standard sample to calibrate the relative correction coefficients of each component in the system relative to the reference substance.

[0010] S3: Construct a multi-component mixture spectral neural network model. This model consists of parallel pure substance spectrum generation sub-modules, concentration weight parameters, and a composition-dependent offset correction module. The composition-dependent offset correction module takes the predicted composition mapped by the concentration weight parameters as input and outputs the offset of the peak function parameters in each pure substance spectrum generation sub-module. It constructs a spectrum reconstruction loss function based on the difference between the generated spectrum and the measured spectrum of the mixture, and maps the concentration weight parameters to the predicted composition based on the relative correction coefficient. It constructs a content loss function based on the difference between the predicted composition and the true composition. Train the multi-component mixture spectral neural network model.

[0011] S4: When applying, input the Raman spectrum of the sample to be tested. The concentration weight parameters of the sample to be tested are optimized by the multi-component mixture spectral neural network model to obtain the optimal weight parameters. Then, based on the relative correction coefficient, the optimal weight parameters are mapped to the concentration of each component.

[0012] The Raman spectrum of a pure substance is represented as a combination of several peak functions; the eigenvectors of each spectral peak form the characteristic parameter matrix of the pure substance, and each pure substance's spectrum generation submodule is composed of several of the aforementioned peak functions connected in parallel as neurons.

[0013] It should be noted that the binary mixture spectral neural network model consists of two pure substances constructing a pure substance spectrum generation submodule connected in parallel. The output mixed spectrum of the binary mixture spectral neural network model is the sum of the generated spectra of the two pure substances multiplied by their respective concentration weight parameters. The loss function used in the fitting process is the mean square error between the output mixed spectrum and the measured binary mixture Raman spectrum.

[0014] According to a preferred embodiment of the present invention, the relative correction coefficient is used to describe the linear relationship between the mole fraction ratio of the component and the reference substance and the weighting parameter ratio, and is obtained through linear regression analysis.

[0015] According to a preferred embodiment of the present invention, the composition-dependent offset correction module in S3 corrects the peak parameters of each component in the mixed sample as follows: the predicted composition is calculated based on the concentration weight parameter and the relative correction coefficient, and the predicted composition is input into the composition-dependent offset correction module to obtain the offset term of each component peak parameter; the pure state reference parameter matrix is ​​added to the offset term to obtain the effective parameter matrix under mixed environment; the offset term acts on the peak intensity parameter, the full width at half maximum (FWHM) parameter and the peak position parameter to describe the peak intensity scaling, peak width broadening and peak position drift phenomena with composition change under mixed environment.

[0016] Based on the mapping of the effective parameter matrix, the corrected spectra of each substance after the offset correction module are obtained. The corrected spectra of each component are multiplied by the concentration weight parameter and then summed to obtain the generated spectrum of the corresponding mixture generated by the multi-component mixture spectral neural network model.

[0017] Furthermore, in a preferred embodiment of the present invention, in S3, a total loss function is constructed by combining the content loss function and the spectral reconstruction loss. During the training of the multi-component mixture spectral neural network model, the backpropagation algorithm is used to jointly optimize the network parameters in the offset correction module and the concentration weight parameters of each sample to minimize the total loss function.

[0018] The training process of S3 employs multi-sample joint training, allowing the same set of peak function neuron parameters to be shared among different component samples. The multi-component mixture spectral neural network model is physically constrained by the known component content to adaptively model the compensatory drift or non-physical convergence of the peak function neuron parameters.

[0019] According to a preferred embodiment of the present invention, in the application stage of S4, the parameters of the peak function neuron and the offset correction module are frozen, and only the concentration weight parameters are optimized or updated to obtain the component content of the unknown sample.

[0020] Through the above technical solution, the present invention has the following significant advantages:

[0021] 1. This invention uses an IHM-based spectral generation submodule to construct a spectral neural network, which enables the Raman spectra of each component to be represented by peak function parameters with clear physical meaning, thereby realizing a structured expression of the spectral generation process. Even under complex peak overlap conditions, it can still maintain good component separation ability and parameter physical consistency, avoiding the black box error and parameter drift problems that may be generated by pure statistical regression models.

[0022] 2. By establishing the calibration relationship between the weight ratio and the composition ratio in the binary system and obtaining the relative correction coefficient. This achieves a stable mapping between weighting parameters and actual concentrations, effectively reducing coupling errors caused by direct concentration fitting in multi-component systems and improving the accuracy and stability of quantitative analysis.

[0023] 3. In the process of modeling multi-component systems, spectral reconstruction loss and content loss are introduced simultaneously for joint optimization, so that the model can ensure the accuracy of spectral fitting while taking into account the accuracy of concentration prediction. This structurally suppresses the non-physical compensation phenomenon between peak parameters and weight parameters, thereby enhancing the model's generalization ability and robustness.

[0024] 4. This invention adopts a modular spectral generation structure, in which each pure molecular module can be independently constructed and expanded, making it easy to adapt to different reaction systems or new components without reconstructing the overall network architecture. It has good scalability and engineering applicability.

[0025] 5. After the model is trained, the component concentration can be inverted simply by quickly optimizing the weight parameters. The computational efficiency is high and can meet the real-time online monitoring and control requirements of continuous flow reaction processes. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the training process in one embodiment of the present invention;

[0027] Figure 2 This is a schematic diagram of a multi-component mixture spectral neural network model structure according to an embodiment of the present invention;

[0028] Figure 3 These are the prediction results for the test set and validation set in this embodiment of the invention. Detailed Implementation

[0029] The present invention will be further described and illustrated below with reference to specific embodiments. The embodiments described are merely examples of the content of this disclosure and do not limit the scope of the invention. The technical features of each embodiment in the present invention can be combined accordingly, provided that there is no mutual conflict.

[0030] This invention provides a Raman spectral decoupling method based on a physical information neural network. This method is designed for Raman spectral decoupling of complex multi-component mixed systems, and is particularly suitable for systems where there is significant peak overlap between components, nonlinear spectral shape changes, and peak position shifts, peak broadening, and peak intensity variations caused by changes in the medium environment.

[0031] The method of this invention mainly includes steps such as pure component feature parameter extraction, binary quantitative calibration, multi-component mixed sample modeling and training, and test sample concentration inversion, as detailed below:

[0032] S1: Extraction of characteristic parameters of pure components

[0033] Raman spectra of each pure substance in the test system are collected, and the characteristic parameter matrix of each pure substance is extracted by peak function fitting. Based on the characteristic parameter matrix, a spectrum generation submodule is constructed for each pure substance, and the submodule consists of several peak function neurons.

[0034] S2: Binary Quantitative Calibration (Ref.) Figure 1 (solid arrow in the middle)

[0035] This invention pairs a reference substance with each other substance in the system to prepare a series of binary standard samples with different concentration gradients and collects their Raman spectra. A binary mixture spectral neural network model is constructed, consisting of a pure substance spectral generation submodule and concentration weight parameters. The learnable parameters of the submodule are frozen, and the weight parameters corresponding to each binary standard sample are obtained by fitting. Based on the Beer-Lambert law, the relative correction coefficients of each component in the system relative to the reference substance are calibrated using the weight parameters and the actual composition data of each binary standard sample.

[0036] S3: Multi-component mixed sample modeling training (refer to...) Figure 1 (dashed arrow in the middle)

[0037] Constructing a spectral neural network model for multi-component mixtures, such as Figure 2As shown, the model consists of parallel pure substance spectrum generation submodules, concentration weight parameters, and a composition-dependent offset correction module. The composition-dependent offset correction module takes the predicted composition mapped from the weight parameters as input and outputs the offset of the peak function parameters in each pure substance spectrum generation submodule. It constructs a spectrum reconstruction loss function based on the difference between the generated spectrum and the measured spectrum of the mixture, and maps the weight parameters to the predicted composition based on the relative correction coefficient. A content loss function is then constructed based on the difference between the predicted composition and the true composition. This model is used to train the multi-component mixture spectral neural network model.

[0038] S4: Test Sample Concentration Inversion

[0039] When applying the method, the Raman spectrum of the sample to be tested is input. The concentration weight parameters of the sample to be tested are optimized by the multi-component mixture spectral neural network model to obtain the optimal weight parameters. Then, the optimal weight parameters are mapped to the concentration of each component based on the relative correction coefficient.

[0040] To implement the method of the present invention, in one embodiment, a physical information neural network is further provided, comprising a pure substance spectrum generation submodule, a composition-dependent shift correction module, a mixture spectrum reconstruction module, and a concentration inversion module. The pure substance spectrum generation submodule represents the reference Raman spectra of each component in the system; the composition-dependent shift correction module describes the shift of the spectral parameters of each component relative to the pure state parameters under a multi-component mixing environment; the mixture spectrum reconstruction module reconstructs the Raman spectrum of the mixture based on the generated spectra of each component and their weights; and the concentration inversion module calculates the mole fraction of each component based on its weight and a pre-calibrated relative quantification coefficient.

[0041] The following detailed explanation of each step of the method of this invention is based on a specific case. This embodiment uses a chlorobenzene nitration system as the verification object. The main components involved in this system include chlorobenzene, o-nitrochlorobenzene, m-nitrochlorobenzene, p-nitrochlorobenzene, and dichloromethane. It is a complex multi-component system, exhibiting significant peak overlap, nonlinear spectral changes, and peak position shifts, peak broadening, and peak intensity variations caused by changes in the medium environment. In this embodiment, dichloromethane is used as the solvent, and chlorobenzene is used as the reference substance to establish the relative quantitative relationship between each component and the reference substance.

[0042] The following provides a detailed description of this embodiment. In this embodiment, the Raman spectroscopy decoupling method based on physical information neural networks includes the following steps:

[0043] 1) Establish an initial multi-component prediction model

[0044] The initial multi-component prediction model was obtained using the following method:

[0045] 1-1) Collect Raman spectra of pure substances and extract characteristic parameters

[0046] The Raman spectra of each pure substance in the system are collected. In this embodiment, the pure substances include chlorobenzene, o-nitrochlorobenzene, m-nitrochlorobenzene, p-nitrochlorobenzene, and dichloromethane. The spectra are preprocessed, preferably including Savitzky-Golay filtering and smoothing, as well as baseline removal using the Morphological method.

[0047] Peak shape fitting was performed on the Raman spectra of each pretreated pure substance. The spectrum of each pure substance is composed of the superposition of several physically meaningful peak functions. In this embodiment, the peak function is a simplified Voigt function, with the following form:

[0048]

[0049] in, For Raman shift, This represents the eigenvector of a single spectral peak. For peak intensity parameters, These are Gaussian weighting coefficients. The half-width and height parameters, This refers to the peak position parameter.

[0050] For any pure substance i, its pure state Raman spectrum is represented as:

[0051]

[0052] in, The peak number of substance i (i.e. the number of peak functions, one peak function corresponds to one spectral peak). Let be the eigenvector of the m-th spectral peak of substance i.

[0053] By minimizing the mean square error between the true spectrum and the fitted spectrum, the peak parameters are fitted to obtain the characteristic parameter matrix for each pure substance:

[0054]

[0055] This feature parameter matrix serves as the initial parameters for subsequent training of the multi-component model.

[0056] 1-2) Construct a binary quantitative model and collect the spectra of standard samples.

[0057] Using chlorobenzene as the reference substance, all substances in the system except dichloromethane (the solvent is not involved in quantification) are paired with dichloromethane to form quantitative groups. For each quantitative group, multiple binary standard solution samples with different concentration gradients are prepared, and the Raman spectra of each sample are acquired. The samples are then filtered, baseline removed, and standardized according to the method in step 1-1).

[0058] 1-3) Establish a spectral neural network model for a binary mixture and obtain the weight parameters.

[0059] Based on the feature parameter matrix obtained in step 1-1), a corresponding pure substance spectrum generation submodule is constructed for each pure substance (the input of the submodule is the wavenumber range vector of the Raman spectrum). The model explicitly calculates the Raman spectrum of the corresponding pure substance. For any binary quantitative group, the two pure substance spectrum generation submodules are combined with concentration weight parameters to form a binary mixture spectrum neural network model. The output mixture spectrum of this model is:

[0060]

[0061] in, and These are the weighting parameters for substance i and reference substance q, respectively. and The generated spectra are those of substance i and reference substance q, respectively.

[0062] In this step, the peak parameters in the frozen pure substance spectrum generation submodule are optimized only for the weighting parameters of each standard sample. and The mean square error between the output mixture spectrum and the measured binary mixture Raman spectrum is used as the loss function:

[0063]

[0064] in, The image shows the measured spectrum, where N is the number of wavenumber sampling points.

[0065] This invention is based on the measured Raman spectra of binary mixtures, and combines the aforementioned methods to obtain the weighting parameters corresponding to each standard sample.

[0066] 1-4) Establish a standard curve and obtain the relative quantitative coefficients.

[0067] Based on Beer-Lambert's law, using the weighting parameters obtained in steps 1-3) and the actual composition of the standard sample, a quantitative relationship between the composition ratio and the weighting ratio in a binary system is established. Preferably, its expression is:

[0068]

[0069] in, and Let i and q be the mole fractions of substance i and reference substance q, respectively. Let be the relative quantification coefficient of substance i relative to reference substance q.

[0070] Linear regression analysis was performed on each binary system sample to obtain the relative quantitation coefficients of each component relative to the reference material. And use it as a fixed parameter in subsequent multi-component models.

[0071] 2) Collect spectral and concentration data of multi-component system samples.

[0072] A series of multi-component mixed samples were designed, allowing for different molar fraction combinations of chlorobenzene, o-nitrochlorobenzene, m-nitrochlorobenzene, p-nitrochlorobenzene, and dichloromethane in different samples. Preferably, the samples cover the possible compositional range of the target system to improve the generalization ability of the model.

[0073] Acquire the true Raman spectra of each multi-component sample and perform the same preprocessing as in step 1-1). Calculate the true mole fraction of each component in each sample according to the sample preparation ratio, and form a sample set by combining the corresponding true Raman spectra and mole fraction data. Divide the sample set into a training set and a test set; preferably, the training set accounts for 70% and the test set accounts for 30%.

[0074] 3) Train and save the multi-component model based on offset.

[0075] 3-1) Constructing a spectral neural network model for multi-component mixtures

[0076] For any sample, the pure-state baseline parameter matrix of each component is first defined as follows: Considering that a multi-component mixed environment can cause a systematic shift in spectral peak parameters relative to the pure state, this embodiment introduces a composition-dependent shift correction module to correct the peak parameters of each component.

[0077] This invention connects the spectral generation submodules corresponding to each pure component in parallel and assigns weight coefficients to each submodule. It also introduces a composition-dependent offset correction module to form a multi-component mixture spectral neural network model. The composition-dependent offset correction module takes the predicted composition obtained by mapping the concentration weight parameters as input and outputs the offset of the peak function parameters in each pure substance's spectral generation submodule.

[0078] For substance i, its effective parameter matrix in the mixed sample is represented as:

[0079]

[0080] in, The offset of substance i is used to characterize the changes in peak intensity, peak width, and peak position caused by the mixture environment. Preferably, the offset mainly affects the peak intensity parameter. Half-width and height parameters and peak position parameters This describes peak intensity scaling, peak width broadening, and peak position shift phenomena under mixed environments. Specifically, x is the predicted composition vector calculated from the current concentration weighting parameters and relative correction coefficients. Since x affects the peak parameter shift... This further affects the mixture formation spectrum, thus requiring iterative updates of x during training and prediction. After correction, the formation spectrum of substance i is represented as:

[0081]

[0082] in, This represents a function mapping that generates a spectrum from an effective parameter matrix.

[0083] Therefore, the generated spectrum of a multi-component mixed sample can be represented as:

[0084]

[0085] Where C represents the total number of components. Let be the concentration weighting parameter for substance i.

[0086] 3-2) Construction of spectral loss and content loss

[0087] To ensure both the accuracy of mixed spectrum reconstruction and the accuracy of concentration inversion, this embodiment employs joint training using spectrum reconstruction loss and concentration loss.

[0088] The spectral reconstruction loss is defined as:

[0089]

[0090] in, This is the measured Raman spectrum.

[0091] Based on the relative quantification coefficients obtained in steps 1-4) The weight parameters of each component The mapping is to normalized mole fractions. Using the reference substance q as a reference, the relative composition values ​​of each component are expressed as:

[0092]

[0093] After further normalization, the predicted mole fractions of each component can be obtained:

[0094]

[0095] Based on this, a content loss function is constructed:

[0096]

[0097] in, This represents the true mole fraction of the sample.

[0098] Therefore, the total loss function is expressed as:

[0099]

[0100] in, The loss weighting coefficient is used to balance the accuracy of spectrum fitting and concentration prediction.

[0101] 3-3) Joint optimization model parameters

[0102] During training, the backpropagation algorithm is used to jointly optimize the network parameters in the concentration-dependent offset correction module and the concentration weight parameters of each sample to minimize the total loss function.

[0103] Preferably, to prevent non-physical divergence of parameters, the following constraints are applied to the peak parameters: the peak intensity parameter is non-negative, the peak width parameter is greater than zero, the peak position offset is limited to a set range, and the Gaussian weighting coefficient is limited to between 0 and 1. These constraints can be implemented using methods such as softplus, sigmoid, or clamp.

[0104] After training, the optimized multi-component mixture spectral neural network model is saved for subsequent spectral decoupling and concentration inversion of unknown samples.

[0105] 4) Prediction of unknown sample concentration

[0106] In step 4), when predicting unknown samples in this mixed system, the parameters of the pure component spectrum generation submodule in step 3) and the parameters of the offset correction module after training are fixed, and only the concentration weight parameters of the sample to be tested are optimized.

[0107] The specific process is as follows:

[0108] For any test sample, input the measured Raman spectrum and initialize the weighting parameters for each component. Calculate the sample concentration vector based on the current weighting parameters, then input the peak parameter offset of each component into the offset correction module to generate the corrected spectra of each component and reconstruct the mixture spectrum. Using the mean square error between the reconstructed spectrum and the measured spectrum as the optimization objective, perform inner-layer iterative optimization of the weighting parameters until the spectral loss converges. After obtaining the optimal weighting parameters, based on the relative quantification coefficients from steps 1-4),... The predicted mole fraction x of each component in the test sample is calculated and compared with the x input by the offset correction module. The process is iterated until the difference between the two is less than the set threshold. The final result is the model prediction value.

[0109] Example: Multi-component Raman decoupling experiment of chlorobenzene nitration system based on offset-corrected PIHM

[0110] The effectiveness of the method of the present invention will be verified by a specific experimental case below.

[0111] This embodiment uses the chlorobenzene nitration system as the research object. The system contains chlorobenzene, o-nitrochlorobenzene, m-nitrochlorobenzene, p-nitrochlorobenzene, and dichloromethane. Referring to the specific process of the Raman spectroscopy decoupling method based on physical information neural networks described above, this embodiment first acquires the pure Raman spectra of each of the above pure components, performs Savitzky-Golay filtering and smoothing, and then uses the Morphological method to remove the baseline. Afterwards, peak shape fitting is performed on the spectra of each pure substance based on the simplified Voigt peak function to extract their respective peak characteristic parameters. , , and This forms the corresponding feature parameter matrix.

[0112] Subsequently, using chlorobenzene as a reference material, binary quantitative groups of o-nitrochlorobenzene / chlorobenzene, m-nitrochlorobenzene / chlorobenzene, and p-nitrochlorobenzene / chlorobenzene were constructed. Multiple concentration gradient standard samples were prepared for each quantitative group, Raman spectra were acquired, and binary spectrum fitting was performed to extract the concentration weight parameters for each sample. Through linear regression analysis of the composition ratios and weight ratios, the relative quantitative coefficients of each component with respect to dichloromethane were obtained.

[0113] Building upon this foundation, multi-component mixed samples containing all components were further formulated to construct training and testing sets. The spectral generation submodules for each pure component were connected in parallel to form a multi-component model. A shift correction module was introduced into the model to characterize the overall shift of the spectral parameters of each component under different mixed compositions, resulting in the shifted PIHM model. During model training, spectral reconstruction loss and content loss were simultaneously used for joint optimization.

[0114] After training, the model was validated using a test set. The results show that the offset PIHM model described in this invention can accurately reconstruct the multi-component mixed Raman spectra of the chlorobenzene nitration system and can stably predict the mole fractions of chlorobenzene, o-nitrochlorobenzene, m-nitrochlorobenzene, and p-nitrochlorobenzene. Figure 3This diagram illustrates the prediction results for the test set samples in this embodiment. As can be seen, the test set sample points are mainly distributed near the diagonal where the true mole fraction equals the predicted mole fraction. The test set R² is 0.9641, the mean square error is 0.024, and the average error is 0.018, indicating that the model still has high prediction accuracy on samples not used in training and can be used for online quantitative analysis of Raman spectra in complex multi-component systems. Compared to the traditional IHM model without shift correction, the method of this invention improves both spectral reconstruction accuracy and concentration inversion accuracy. In this embodiment, R²... 2 =0.9461, compared to the traditional IHM model R 2 =0.3210, a three-fold increase, indicating that it still maintains good robustness and interpretability even under conditions of severe overlap of multi-component peaks, nonlinear peak position shifts, and peak width variations.

[0115] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of protection of this invention. For those skilled in the art, various modifications and improvements can be made without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of this invention.

Claims

1. A Raman spectroscopy decoupling method based on a physical information neural network, characterized in that, The steps include the following: S1: Collect the Raman spectra of each pure substance in the test system, and extract the characteristic parameter matrix of each pure substance by fitting the peak function; Based on the feature parameter matrix, a spectrum generation submodule is constructed for each pure substance, and the submodule consists of several peak function neurons. S2: Pair the reference substance with each other substance in the system to prepare a series of binary standard samples with different concentration gradients, and collect their Raman spectra; A binary mixture spectral neural network model consisting of a pure substance spectral generation submodule and concentration weight parameters is constructed. The learnable parameters of the submodule are frozen, and the weight parameters corresponding to each binary standard sample are obtained by fitting. Based on the Beer-Lambert law, the relative correction coefficients of each component in the system relative to the reference substance are determined using the weight parameters and the actual composition data of each binary standard sample. S3: Construct a multi-component mixture spectral neural network model. This model consists of parallel pure substance spectrum generation sub-modules, concentration weight parameters, and a composition-dependent offset correction module. The composition-dependent offset correction module takes the predicted composition mapped by the concentration weight parameters as input and outputs the offset of the peak function parameters in each pure substance spectrum generation sub-module. It constructs a spectrum reconstruction loss function based on the difference between the generated spectrum and the measured spectrum of the mixture, and maps the concentration weight parameters to the predicted composition based on the relative correction coefficient. It constructs a content loss function based on the difference between the predicted composition and the true composition. Train the multi-component mixture spectral neural network model. S4: When applying, input the Raman spectrum of the sample to be tested. The concentration weight parameters of the sample to be tested are optimized by the multi-component mixture spectral neural network model to obtain the optimal weight parameters. Then, based on the relative correction coefficient, the optimal weight parameters are mapped to the concentration of each component.

2. The method according to claim 1, characterized in that, The peak function described in step S1 has the following form: ； in, For Raman shift, This represents the eigenvector of a single spectral peak. For peak intensity parameters, These are Gaussian weighting coefficients. The half-width and height parameters, For peak position parameters; The Raman spectrum of a pure substance is represented as a combination of several peak functions; the eigenvectors of each spectral peak form the characteristic parameter matrix of the pure substance, and each pure substance's spectrum generation submodule is composed of several of the aforementioned peak functions connected in parallel as neurons.

3. The method according to claim 1, characterized in that, In S2, the binary mixture spectral neural network model is constructed by two pure substances and connected in parallel to generate pure substance spectra. The output mixed spectrum of the binary mixture spectral neural network model is the sum of the generated spectra of the two pure substances multiplied by their respective concentration weight parameters. The loss function used in the fitting process is the mean square error between the output mixed spectrum and the measured binary mixture Raman spectrum.

4. The method according to claim 1, characterized in that, In S2, the relative correction coefficient is used to describe the linear relationship between the mole fraction ratio of the component and the reference substance and the weighting parameter ratio, and is obtained through linear regression analysis.

5. The method according to claim 1, characterized in that, The composition-dependent offset correction module described in S3 corrects the peak parameters of each component in the mixed sample as follows: Based on the concentration weight parameter and the relative correction coefficient, the predicted composition is calculated, and the predicted composition is input into the composition-dependent offset correction module to obtain the offset term for each component peak parameter; the pure-state reference parameter matrix is ​​added to the offset term to obtain the effective parameter matrix under mixed conditions; the offset term acts on the peak intensity parameter, the full width at half maximum (FWHM) parameter, and the peak position parameter to describe the peak intensity scaling, peak width broadening, and peak position drift phenomena that occur with composition changes under mixed conditions.

6. The method according to claim 5, characterized in that, Based on the mapping of the effective parameter matrix, the corrected spectra of each substance after the offset correction module are obtained. The corrected spectra of each component are multiplied by the concentration weight parameter and then summed to obtain the generated spectrum of the corresponding mixture generated by the multi-component mixture spectral neural network model.

7. The method according to claim 1, characterized in that, In S3, the total loss function is further constructed from the content loss function and the spectrum reconstruction loss. During the training of the multi-component mixture spectral neural network model, the backpropagation algorithm is used to jointly optimize the network parameters in the offset correction module and the concentration weight parameters of each sample to minimize the total loss function.

8. The method according to claim 2, characterized in that, The training process of S3 employs multi-sample joint training, allowing the same set of peak function neuron parameters to be shared among different component samples. The multi-component mixture spectral neural network model is physically constrained by the known component content to adaptively model the compensatory drift or non-physical convergence of the peak function neuron parameters.

9. The method according to claim 2, characterized in that, In the application phase of S4, the parameters of the peak function neuron and the offset correction module are frozen, and only the concentration weight parameters are optimized or updated to obtain the component content of the unknown sample.

Images