Three-dimensional fluorescence spectrum scattering removal method based on physical information neural network

Abstract

The invention discloses a three-dimensional fluorescence spectrum scattering removal method based on a physical information neural network, and the method comprises the following steps: collecting three-dimensional fluorescence spectrum data of solutions with different turbidity gradients, and removing Raman scattering signals in the three-dimensional fluorescence spectrum data; expanding the three-dimensional fluorescence spectrum data, and labeling material scattering attributes, scattering peak values and peak width information of each group of three-dimensional fluorescence spectrum data to construct a three-dimensional fluorescence data set; constructing a de-scattering network model based on a physical information neural network and a generative adversarial network, and performing adversarial training by using the three-dimensional fluorescence data set; and generating three-dimensional fluorescence spectrum data after scattering removal by using the trained scattering removal model. According to the method, the scattering removing method based on scattering physical information is provided for the first time, the scattering data attributes are learned, meanwhile, the scattering physical attributes are considered, the material information can be well stored while scattering is removed, and the method has higher robustness and accuracy and is superior to an existing method.

Description

Technical Field

[0001] The present invention relates to the technical field of spectral analysis, specifically to the technical field of de-scattering processing of three-dimensional fluorescence spectral data, and particularly to a three-dimensional fluorescence spectral de-scattering method based on a physics-informed neural network. Background Art

[0002] In three-dimensional fluorescence spectral analysis, Rayleigh scattering and Raman scattering often interfere with the substance signal, seriously affecting the accuracy of spectral data. Therefore, it is crucial to perform appropriate preprocessing on the fluorescence data to reduce the influence of scattering. Common methods for Rayleigh scattering processing include blank subtraction, interpolation methods, and weighting methods, etc. Among them, the most common de-scattering idea is to set the fluorescence signal values within a certain width of the scattering region to zero, and then use the effective data obtained from the surrounding non-scattering regions to estimate the true signal values of these regions. In this way, the original fluorescence data can be corrected as necessary, thereby generating corrected fluorescence data and improving the usability and accuracy of the data.

[0003] However, in most three-dimensional fluorescence spectra, the scattering region will overlap with the substance peak region. If the interpolation method is used to directly set the scattering region to zero, it may lead to the annihilation of the substance signal or the incorrect retention of the scattering signal, thus affecting the accuracy of subsequent data analysis. At the same time, when the turbidity of the sample to be measured is high, the width of the region affected by Rayleigh scattering will change. After setting the three-dimensional fluorescence signals within a fixed width to zero using the traditional scattering removal method, the fluorescence signal may still be retained in the edge region of the scattering, and then the substance signal is interfered by the scattered light. Therefore, in high-turbidity samples, only using the scattering removal method with a specific width may not be able to effectively remove the interference caused by scattering, resulting in unstable results of spectral analysis. This requires considering more complex or more practical processing strategies to ensure more accurate measurement and characterization of the fluorescence signal.

[0004] For this reason, the present invention proposes an adaptive three-dimensional fluorescence de-scattering method to solve the above defects of the existing methods. Summary of the Invention

[0005] The purpose of the present invention is to provide a three-dimensional fluorescence spectral de-scattering method based on a physics-informed neural network in view of the deficiencies of the prior art. The present invention can improve the adaptive ability of the three-dimensional fluorescence de-scattering method in solutions with different turbidities, so that the three-dimensional fluorescence data after removing scattering can more restore the true substance signal.

[0006] The purpose of the present invention is achieved by the following technical solutions: A three-dimensional fluorescence spectral de-scattering method based on a physics-informed neural network, comprising the following steps:

[0007] (1) Collect the three-dimensional fluorescence spectral data of solutions with different turbidity gradients, and remove the Raman scattering signals therein; expand the three-dimensional fluorescence spectral data, and label the material scattering attributes, scattering peak values, and peak width information of each group of three-dimensional fluorescence spectral data to construct a three-dimensional fluorescence data set;

[0008] (2) Construct a de-scattering network model based on a physics-informed neural network and a generative adversarial network. The de-scattering network model includes a generator and a discriminator. The generator includes an encoder, a decoder, a physical reconstruction module, and an attribute classifier;

[0009] (3) Use the three-dimensional fluorescence data in the three-dimensional fluorescence data set to perform adversarial training on the de-scattering network model. During the training process, minimize the total loss function of the de-scattering network model as the optimization goal, and adjust the parameters of the de-scattering network model to obtain a trained de-scattering network model, and use the trained encoder and decoder as the final de-scattering model;

[0010] (4) Input the three-dimensional fluorescence data to be tested or de-scattered into the final de-scattering model, and successively pass through the trained encoder and decoder to output the three-dimensional fluorescence data with scattering removed.

[0011] Further, the step (1) includes the following sub-steps:

[0012] (1.1) Collect the three-dimensional fluorescence spectral data of solutions with different turbidity gradients. For each three-dimensional fluorescence spectral data, use the background subtraction method to remove the Raman scattering signals therein, specifically by subtracting the three-dimensional fluorescence spectral data of the blank solution sample from the three-dimensional fluorescence spectral data of the turbidity solution, where the blank solution sample is a pure water solution sample;

[0013] (1.2) For the three-dimensional fluorescence spectral data with Raman scattering signals removed obtained in step (1.1), expand it by extracting the Rayleigh scattering information therein, manually eliminating scattering, supplementing material information, and adjusting the scattering intensity. During the expansion process, simultaneously label the material scattering attributes of the peak values and peak widths of the material signals and scattering signals of each group of three-dimensional fluorescence spectral data to construct a three-dimensional fluorescence data set.

[0014] Further, the step (1.2) includes the following sub-steps:

[0015] (1.2.1) For each three-dimensional fluorescence spectral data data i , construct an attribute vector and generate the corresponding number of material signals N, material peak position vector according to the material scattering attributes of the three-dimensional fluorescence spectral data, and the peak ratio T of the scattering signal to the material signal i; wherein, the attribute vector indicates whether the three-dimensional fluorescence data contains a substance signal and a scattering signal, and its expression is:

[0016]

[0017] In the formula, represents the attribute vector of the i-th three-dimensional fluorescence spectrum data data i after removing the Raman scattering signal; a i indicates whether the i-th three-dimensional fluorescence data contains a substance signal, and its value range is {0, 1}, where 0 indicates the absence of a substance signal and 1 indicates the presence of a substance signal; b i indicates whether the i-th three-dimensional fluorescence data contains a first-order Rayleigh scattering signal, and its value range is {0, 1}, where 0 indicates the absence of a first-order Rayleigh scattering signal and 1 indicates the presence of a first-order Rayleigh scattering signal; c i indicates whether the i-th three-dimensional fluorescence data contains a second-order Rayleigh scattering signal, and its value range is {0, 1}, where 0 indicates the absence of a second-order Rayleigh scattering signal and 1 indicates the presence of a second-order Rayleigh scattering signal; the substance peak position vector is a combination of the position vectors of N substance peaks, and its expression is:

[0018]

[0019] In the formula, represents the substance peak position vector of the i-th three-dimensional fluorescence spectrum data data i after removing the Raman scattering signal; represents the position vector of the j-th substance peak in the i-th three-dimensional fluorescence data, em ij represents the emission wavelength corresponding to the j-th substance peak in the i-th three-dimensional fluorescence data, and ex ij represents the excitation wavelength corresponding to the j-th substance peak in the i-th three-dimensional fluorescence data;

[0020] (1.2.2) According to the attribute vector i of the three-dimensional fluorescence spectrum data data obtained in step (1.2.1), determine whether this data contains a substance signal. If this data contains a substance signal, then use the position vectors of the N substance peaks in the substance peak position vector as the center, and use the product of the substance signal peak value and the peak ratio as the maximum value of the scattering signal to randomly generate N elliptical substance signals that conform to the normal distribution; if this data does not contain a substance signal, then directly jump to step (1.2.3);

[0021] (1.2.3) The three-dimensional fluorescence spectrum data data obtained according to step (1.2.1) i 's attribute vector Determine whether there is a first-order Rayleigh scattering signal in this data. If this data contains a first-order Rayleigh scattering signal, traverse this data to obtain the peak p i1 and width L 1 ; if this data does not contain a first-order Rayleigh scattering signal, then set the first-order Rayleigh scattering signal in the three-dimensional fluorescence spectrum data data i to zero;

[0022] (1.2.4) The three-dimensional fluorescence spectrum data data obtained according to step (1.2.1) i 's attribute vector Determine whether there is a second-order Rayleigh scattering signal in this data. If this data contains a second-order Rayleigh scattering signal, traverse this data to obtain the peak p i2 and width L 2 ; if this data does not contain a second-order Rayleigh scattering signal, then set the second-order Rayleigh scattering signal in the three-dimensional fluorescence spectrum data data i to zero; finally, complete the construction of a new training data data_create i ;

[0023] (1.2.5) Repeat steps (1.2.2) - (1.2.4) to obtain three-dimensional fluorescence data with different scattering properties and material backgrounds, so as to construct a three-dimensional fluorescence data set.

[0024] Furthermore, the generator is used to edit the three-dimensional fluorescence data according to the input three-dimensional fluorescence data and its corresponding data attribute labels, and generate a three-dimensional fluorescence data image that conforms to the new attribute labels;

[0025] In the generator, the encoder is used to obtain the high-dimensional latent features of the input three-dimensional fluorescence data, specifically by gradually extracting and compressing the input three-dimensional fluorescence data through convolutional layers and activation functions, and converting it into a high-dimensional latent space representation feature; the decoder is used to obtain the spectral image corresponding to the input high-dimensional latent features, specifically by gradually expanding the high-dimensional latent feature vector output by the encoder through deconvolutional layers, restoring its spatial dimension, and restoring the latent space representation to a three-dimensional fluorescence spectrum image after removing scattering; the attribute classifier obtains the corresponding attribute classification result by comparing the attribute vector of the spectral image output by the decoder with the target attribute; the physical reconstruction module generates the scattering theoretical value according to the physical reconstruction formula of Rayleigh scattering and the peaks and widths of the first-order and second-order Rayleigh scattering signals in the three-dimensional fluorescence data, so as to guide the generator to generate the spectral image;

[0026] The discriminator is used to evaluate the authenticity of the three-dimensional fluorescence data image generated by the generator.

[0027] Further, the physical reconstruction module is constructed by the following method:

[0028] Construct a physical reconstruction formula according to the scattering physical model in the three-dimensional fluorescence spectral data. Based on the three-dimensional fluorescence data in the three-dimensional fluorescence dataset, generate the theoretical value of the three-dimensional fluorescence data that removes Rayleigh scattering and retains substance information according to the physical reconstruction formula. The specific implementation steps are as follows:

[0029] (2.1) Measure the minimum excitation wavelength of the three-dimensional fluorescence data on the first-order Rayleigh scattering The maximum fluorescence intensity value of the first-order Rayleigh scattering is and the minimum excitation wavelength At this time, the scattering width L when the first-order Rayleigh scattering drops to one-tenth of the peak 1 ;

[0030] (2.2) Measure the minimum excitation wavelength of the three-dimensional fluorescence data on the second-order Rayleigh scattering The maximum fluorescence intensity value of the second-order Rayleigh scattering is and the minimum excitation wavelength At this time, the scattering width L when the second-order Rayleigh scattering drops to one-tenth of the peak 2 ;

[0031] (2.3) According to the three-dimensional fluorescence physical information obtained in steps (2.1) and (2.2), perform data fitting on the first-order Rayleigh scattering and the second-order Rayleigh scattering. The fitting formulas for the first-order Rayleigh scattering and the second-order Rayleigh scattering are respectively:

[0032]

[0033] In the formula, data 1 represents the physical fitting intensity value of the first-order Rayleigh scattering, data 2 represents the physical fitting intensity value of the second-order Rayleigh scattering, em and ex respectively represent the emission wavelength and the excitation wavelength;

[0034] (2.4) According to the fitting information of the first-order Rayleigh scattering and the second-order Rayleigh scattering obtained in step (2.3), fit out the theoretical value x of the three-dimensional fluorescence data that conforms to the physical law, removes the Rayleigh scattering truth value, and retains the substance information according to the following physical reconstruction formula phys , where the expression of the physical reconstruction formula is:

[0035] x phys = data original - data 1 - data 2

[0036] In the formula, data original is a three-dimensional fluorescence data image without undergoing past scattering processing.

[0037] Further, the step (3) specifically includes:

[0038] Input the three-dimensional fluorescence data in the three-dimensional fluorescence data set into the de-scattering network model. First, it enters the generator. The three-dimensional fluorescence data passes through the encoder and decoder in sequence, and under the guidance of the physical reconstruction module, a de-scattered three-dimensional fluorescence sample x b is generated, and at the same time, a three-dimensional fluorescence sample containing scattering is generated Compare the generated three-dimensional fluorescence data containing scattering with the original input three-dimensional fluorescence data with scattering, and calculate the reconstruction loss function; input the de-scattered three-dimensional fluorescence sample x b and the three-dimensional fluorescence sample containing scattering into the attribute classifier respectively to identify their attributes and obtain the corresponding attribute classification values; compare the two attribute classification values with the expected attribute values respectively, and calculate the classification loss of the generator and the classification loss of the attribute classifier; input the de-scattered three-dimensional fluorescence sample x b and the three-dimensional fluorescence sample containing scattering into the discriminator respectively to identify whether the input three-dimensional fluorescence sample is real data and obtain the corresponding authenticity discrimination result; calculate the adversarial loss of the generator and the discriminator according to the authenticity discrimination result; at the same time, the three-dimensional fluorescence data in the three-dimensional fluorescence data set generates the corresponding three-dimensional fluorescence data theoretical value x phys According to the de-scattered three-dimensional fluorescence data output by the decoder and the three-dimensional fluorescence data theoretical value x phys calculate the physical information loss function; use the weighted sum of the reconstruction loss function, the classification loss of the generator and the attribute classifier, the adversarial loss of the generator and the discriminator, and the physical information loss function as the total loss function of the de-scattering network model;

[0039] During the training process, with minimizing the total loss function of the de-scattering network model as the optimization goal, adjust the parameters of the de-scattering network model until the preset number of training rounds is reached, obtain the trained de-scattering network model, and use the trained encoder and decoder as the final de-scattering model.

[0040] Further, the calculation formula of the reconstruction loss function is:

[0041]

[0042] In the formula, represents the reconstruction loss function, x ais the three-dimensional fluorescence data of the original input containing Rayleigh scattering, is the generated three-dimensional fluorescence data containing scattering, G enc represents the encoder, G dec represents the decoder, represents taking the expected value of the data and attribute labels under the joint distribution p data and p attr x represents taking the expected value of the data and attribute labels under the joint distribution p a ~p data b~p represents sampling an image from the three-dimensional fluorescence data set distribution attr represents randomly selecting an attribute vector from the attribute vector distribution;

[0043] The calculation formulas for the classification loss of the generator and the classification loss of the attribute classifier are respectively:

[0044]

[0045]

[0046]

[0047]

[0048] In the formula, is the classification loss of the generator, l g (x a , b) is the binary cross-entropy loss for classification, x a represents the three-dimensional fluorescence data of the original input containing Rayleigh scattering, b represents the target attribute label, is the probability value output by the attribute classifier for the i-th attribute, b i is the true label of the i-th attribute; is the classification loss of the classifier, represents taking the expected value after sampling the input data x a according to the data distribution p data l r (x a , a) is the classification loss for the reconstructed data, x a represents the three-dimensional fluorescence data of the original input containing Rayleigh scattering, a represents the attribute label of the original input data, C i (x a ) is the probability value output by the classifier for the i-th attribute, a i is the true label of the i-th attribute;

[0049] The calculation formulas for the adversarial loss of the generator and the discriminator are respectively:

[0050]

[0051]

[0052] In the formula, represents the adversarial loss of the discriminator, represents the adversarial loss of the generator, represents the input data x a sampled according to the data distribution p data and taking the expected value, D(x a ) represents the authenticity discrimination result of the discriminator for the real data x a ; represents the authenticity discrimination result of the discriminator for the generated data ; represents taking the expected value of the data and attribute labels under the joint distributions p data and p attr ;

[0053] The calculation formula of the physical information loss function is:

[0054]

[0055] In the formula, represents the physical information loss function, represents the three-dimensional fluorescence data with scattering removed output by the decoder, x phys represents the theoretical value of the three-dimensional fluorescence data fitted by the physical reconstruction module, and M represents the total number of three-dimensional fluorescence data;

[0056] The calculation formula of the total loss function of the de-scattering network model is:

[0057]

[0058] In the formula, represents the total loss function of the de-scattering network model, λ 1 , λ 2 , λ 3 and λ 4 are respectively and weight parameters.

[0059] Compared with the prior art, the beneficial effects of the present invention are:

[0060] (1) The present invention first proposes a de-scattering method based on scattering physical information. While learning the data attributes of scattering, considering the physical attributes of scattering, it can better preserve the material information while removing scattering, and is more robust and accurate.

[0061] (2) The present invention adopts a generative adversarial network architecture, combines a generator and a discriminator, and realizes the removal of scattered signals and the retention of substance signals. During the training process, the generator simulates and generates three-dimensional fluorescence data without scattering, and the discriminator evaluates the generated data to ensure the authenticity and integrity of the generation result; it significantly improves the robustness of the network when processing high-turbidity samples and overcomes the common problems of overfitting or signal loss in traditional interpolation methods.

[0062] (3) The present invention uses the theoretical scattering value generated by the physical reconstruction module as a constraint condition for the generative adversarial network. During network training, guided by physical information, the de-scattered data output by the generator not only satisfies the distribution characteristics of experimental data but also is highly consistent with the model corresponding to the physical information. This design ensures that the finally generated three-dimensional fluorescence data can be consistent with Rayleigh scattering theory during the scattering removal process, enabling the data after scattering removal to truly reflect substance information, effectively solving the problem of inaccurate reduction of substance signals when traditional methods have overlapping scattered and substance signals, and enhancing the reliability and physical consistency of the scattering removal method. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 is a flowchart of the three-dimensional fluorescence spectrum scattering removal method based on a physical information neural network of the present invention;

[0064] Figure 2 is a flowchart of the expanded three-dimensional fluorescence data set of the present invention;

[0065] Figure 3 is an architecture flowchart of the scattering removal network model of the present invention;

[0066] Figure 4 is an architecture flowchart of the scattering removal model of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0067] Here, exemplary embodiments will be described in detail, and the examples are shown in the drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims. It should be understood that the above general description and the following detailed description are merely exemplary and explanatory and should not limit the present application.

[0068] The terms used in this application are for the purpose of describing specific embodiments only and are not intended to limit this application. The singular forms "a", "the", and "said" used in this application and the appended claims are also intended to include the plural forms unless the context clearly dictates otherwise. It should also be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0069] It should be understood that although the terms first, second, third, etc. may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from each other. For example, without departing from the scope of this application, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "when" or "while" or "in response to determining". Moreover, the terms "comprise", "include" or any other variant thereof are intended to cover non-exclusive inclusion, such that a process, method, article or apparatus that comprises a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or apparatus. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or apparatus that comprises the element.

[0070] The present invention will be described in detail below with reference to the accompanying drawings. Without conflict, the features in the following embodiments and implementation manners can be combined with each other.

[0071] The three-dimensional fluorescence spectrum de-scattering method based on the physics-informed neural network of the present invention combines the physics-informed neural network (PINN) with the generative adversarial network (GAN), that is, combines the physical model with the deep learning method. The physical formula of Rayleigh scattering is used to reconstruct the scattering signal, and a generator and a discriminator are constructed for adversarial training in the generative adversarial network. The physics-informed neural network calculates the theoretical scattering intensity that conforms to physical laws by inputting the scattering peak and width information, thereby providing precise constraints for the generative adversarial network. Compared with the traditional interpolation method and blank deduction method, the present invention can retain the integrity and authenticity of the substance signal while removing the scattering signal. Thanks to the combination of the physical model and deep learning, the present invention has higher generalization ability, can effectively process samples with different concentration gradients, and provides a more reliable solution for spectral analysis of various complex samples.

[0072] See Figure 1 , the three-dimensional fluorescence spectrum de-scattering method based on the physics-informed neural network of the present invention specifically includes the following steps:

[0073] (1)Collect the three-dimensional fluorescence spectral data of solutions with different turbidity gradients, and remove the Raman scattering signals therein; expand the three-dimensional fluorescence spectral data, and label the material scattering attributes, scattering peak values, and peak width information of each group of three-dimensional fluorescence spectral data to construct a three-dimensional fluorescence data set.

[0074] It should be noted that in this embodiment, the three-dimensional fluorescence spectral data is collected through a three-dimensional fluorescence experiment to obtain a batch of basic data (the three-dimensional fluorescence data of the solution with turbidity, which only has scattering and no material signals), and then this batch of basic data is further processed. For example, artificial simulated material signals are added to the basic data, or one of the Raman scattering signals is eliminated, etc. After processing, attribute values are labeled for this data. Repeating the above steps can expand this batch of three-dimensional fluorescence spectral data obtained from the experiment into a larger three-dimensional fluorescence data set that is rich and covers different material backgrounds and scattering bands.

[0075] (1.1)Collect the three-dimensional fluorescence spectral data of solutions with different turbidity gradients. For each three-dimensional fluorescence spectral data, the Raman scattering signal therein is removed by using a background subtraction method, which is specifically realized by subtracting the three-dimensional fluorescence spectral data of the blank solution sample from the three-dimensional fluorescence spectral data of the turbidity solution, where the blank solution sample is a pure water solution sample.

[0076] Specifically, during the collection process of the three-dimensional fluorescence spectral data, the three-dimensional fluorescence spectral data with an excitation wavelength range of 240 nm to 800 nm is obtained through a fluorescence spectrometer. First, measure the three-dimensional fluorescence spectral data of a pure water solution, which is used as the three-dimensional fluorescence spectral data of the blank solution sample; then measure the three-dimensional fluorescence spectral data of solutions with different turbidity gradients. For each three-dimensional fluorescence spectral data, the Raman scattering signal therein is removed by using a background subtraction method, that is, subtracting the three-dimensional fluorescence spectral data of the blank solution sample from the three-dimensional fluorescence spectral data of the turbidity solution, and what is obtained is the three-dimensional fluorescence spectral data with the Raman scattering signal removed.

[0077] (1.2)For the three-dimensional fluorescence spectral data with the Raman scattering signal removed obtained in step (1.1), it is expanded by extracting the Rayleigh scattering information therein, artificially eliminating scattering, supplementing material information, and adjusting the scattering intensity, etc. During the expansion process, the peak values, peak widths, and other material scattering attributes of the material signals and scattering signals of each group of three-dimensional fluorescence spectral data are accurately labeled to construct a larger-scale three-dimensional fluorescence data set, specifically as Figure 2 shown, which is convenient to provide comprehensive data support for the training of the model in subsequent steps.

[0078] (1.2.1)For each three-dimensional fluorescence spectral data data i , construct an attribute vector for it And generate the corresponding number of substance signals N and the substance peak position vector according to the scattering attributes of the substances in the three-dimensional fluorescence spectrum data The peak ratio T of the scattering signal to the substance signal i . Among them, the attribute vector Indicates whether the three-dimensional fluorescence data contains substance signals and scattering signals, and its expression is:

[0079]

[0080] In the formula, Represents the attribute vector of the i-th three-dimensional fluorescence spectrum data data after removing the Raman scattering signal i ; a i Indicates whether the i-th three-dimensional fluorescence data contains substance signals, and its value range is {0, 1}, where 0 indicates the absence of substance signals and 1 indicates the presence of substance signals; b i Indicates whether the i-th three-dimensional fluorescence data contains first-order Rayleigh scattering signals, and its value range is {0, 1}, where 0 indicates the absence of first-order Rayleigh scattering signals and 1 indicates the presence of first-order Rayleigh scattering signals; c i Indicates whether the i-th three-dimensional fluorescence data contains second-order Rayleigh scattering signals, and its value range is {0, 1}, where 0 indicates the absence of second-order Rayleigh scattering signals and 1 indicates the presence of second-order Rayleigh scattering signals. The substance peak position vector Is a combination of the position vectors of N substance peaks, and its expression is:

[0081]

[0082] In the formula, Represents the substance peak position vector of the i-th three-dimensional fluorescence spectrum data data after removing the Raman scattering signal i ; Represents the position vector of the j-th substance peak in the i-th three-dimensional fluorescence data, em ij Represents the emission wavelength (x-axis coordinate) corresponding to the j-th substance peak in the i-th three-dimensional fluorescence data, ex ij Represents the excitation wavelength (y-axis coordinate) corresponding to the j-th substance peak in the i-th three-dimensional fluorescence data.

[0083] It should be understood that the three-dimensional fluorescence spectrum data is a matrix spectrum characterized by the three-dimensional coordinates of excitation wavelength (y-axis) - emission wavelength (x-axis) - fluorescence intensity (z-axis). When artificially generating three-dimensional fluorescence data, it is necessary to artificially add the fluorescence intensity of the substance signal. The peak ratio refers to the ratio of the strongest fluorescence intensity of the substance signal to the strongest fluorescence intensity of the scattering.

[0084] (1.2.2) According to the three-dimensional fluorescence spectral data data obtained in step (1.2.1) i 's attribute vector Determine whether this data contains a substance signal. If this data contains a substance signal, then use the position vectors of N substance peaks in the substance peak position vector as the center, and use the product of the substance signal peak value and the peak ratio as the maximum value of the scattering signal, and randomly generate N elliptical substance signals that conform to the normal distribution, as shown in ; if this data does not contain a substance signal, directly jump to step (1.2.3). Figure 2 shown; if this data does not contain a substance signal, directly jump to step (1.2.3).

[0085] (1.2.3) According to the three-dimensional fluorescence spectral data data obtained in step (1.2.1) i 's attribute vector Determine whether this data has a first-order Rayleigh scattering signal. If this data contains a first-order Rayleigh scattering signal, traverse this data to obtain the peak p of the first-order Rayleigh scattering signal i1 and width L 1 ; if this data does not contain a first-order Rayleigh scattering signal, then set the first-order Rayleigh scattering signal in the three-dimensional fluorescence spectral data data i to zero.

[0086] It should be noted that the peak value of the Rayleigh scattering signal is obtained by using the traversal method. Specifically, starting from the central axis (em = ex) of the three-dimensional fluorescence spectral data, scan step by step to the left and right sides to find the maximum fluorescence intensity, which is the peak signal corresponding to the excitation wavelength. The width of the Rayleigh scattering signal is obtained by using the traversal method. Specifically, starting from the determined fluorescence intensity peak, scan to the left until the fluorescence intensity drops to one-tenth of the peak value. The distance between this point and the central axis (em = ex) of the three-dimensional fluorescence spectrum is the width of the Rayleigh scattering, that is, the full width at one-tenth of the peak (FWTM).

[0087] (1.2.4) According to the three-dimensional fluorescence spectral data data obtained in step (1.2.1) i 's attribute vector Determine whether this data has a second-order Rayleigh scattering signal. If this data contains a second-order Rayleigh scattering signal, traverse this data to obtain the peak p of the second-order Rayleigh scattering signal i2 and width L 2 ; if this data does not contain a second-order Rayleigh scattering signal, then set the second-order Rayleigh scattering signal in the three-dimensional fluorescence spectral data data i to zero. Finally, complete the construction of a new training data data_create i .

[0088] (1.2.5) Repeat steps (1.2.2) - (1.2.4) to obtain three-dimensional fluorescence data with different scattering properties and material backgrounds, so as to construct a larger-scale three-dimensional fluorescence data set.

[0089] It should be understood that by repeating each real experimental data n times, n three-dimensional fluorescence data set data with different properties and different material backgrounds can be obtained from one experimental data.

[0090] (2) Construct a de-scattering network model based on the physics-informed neural network (PINN) and the generative adversarial network (GAN), combining the physical model with deep learning to improve the accuracy of spectral data and the adaptive ability of the model. The de-scattering network model includes a generator G and a discriminator D. The generator G includes an encoder G enc 、a decoder G dec 、a physical reconstruction module R phy and an attribute classifier C, as Figure 3 shown.

[0091] It should be noted that the combination of the physical model and deep learning is mainly manifested in two parts: the physical information reconstruction module and the physical information loss function in the loss function. The physical reconstruction module generates the "theoretical three-dimensional fluorescence data x phys that conforms to physical laws and removes the true value of Rayleigh scattering" through the formula "theoretical three-dimensional fluorescence data x phys that conforms to physical laws and removes the true value of Rayleigh scattering" and the width and peak information of Rayleigh scattering. This part of the x phys data participates in the loss function to guide the network to generate de-scattering values close to physical laws. After using the physical reconstruction module to reconstruct the scattering value x phys that conforms to physical laws, after the generator generates the de-scattering spectral value, randomly select n points from the generated values and compare them with this scattering value x phys . The difference is the physical information loss function. By minimizing the physical information loss function, the finally generated spectral data can be closer to the theoretically de-scattered data, that is, more in line with physical laws.

[0092] Specifically, input the artificially simulated three-dimensional fluorescence data into the de-scattering network model. The generator generates de-scattered three-dimensional fluorescence samples to deceive the discriminator under the guidance of the physical information reconstruction module, and at the same time generates three-dimensional fluorescence samples containing scattering to ensure the accuracy of the generator; the discriminator needs to identify whether the de-scattered three-dimensional fluorescence samples are real data; after several rounds of adversarial training, the generator generates de-scattered three-dimensional fluorescence samples.

[0093] In this embodiment, as Figure 3As shown, the generator is used to edit the three-dimensional fluorescence data according to the input three-dimensional fluorescence data and its corresponding data attribute tags, and generate a three-dimensional fluorescence data image that conforms to the new attribute tags. In the generator, the encoder is used to obtain the high-dimensional latent features of the input three-dimensional fluorescence data. Specifically, through a series of convolutional layers and activation functions shown in Table 1, the input three-dimensional fluorescence data is gradually extracted and compressed, and converted into a high-dimensional latent space representation feature. This encoder can capture the important features in the spectral data and perform high-dimensional feature representation on them for subsequent decoding and generation processes; the decoder is used to obtain the spectral image corresponding to the input high-dimensional latent features. Specifically, through a series of transposed convolutional layers shown in Table 2, the high-dimensional latent feature vector output by the encoder is gradually expanded to restore its spatial dimension, and the latent space representation is restored to obtain a three-dimensional fluorescence spectrum image after removing scattering. This decoder can re-decode the information in the latent space into a de-scattered image and maintain the material signal of the original data; the attribute classifier obtains the corresponding attribute classification result by comparing the attribute vector of the spectral image output by the decoder with the target attribute. By comparing the attributes of the generated image with the target attributes, the attribute classifier can reduce the classification error, ensure that the generated spectral image contains the correct material attributes and scattering characteristics, and minimize the classification error during subsequent training to ensure that the generated image has the correct attributes; the physical reconstruction module generates the theoretical value of scattering according to the physical reconstruction formula of Rayleigh scattering and the peak and width of the first-order and second-order Rayleigh scattering signals in the three-dimensional fluorescence data, that is, according to the physical quantity p i1 , p i2 , L 1 , L 2 Generate the theoretical physical image The theoretical physical image is used to guide the generator to generate the spectral image to ensure that the generated spectral image is not only visually reasonable but also conforms to physical laws. The theoretical physical image is used to guide the generation result to ensure that the de-scattering result is consistent with the experimental observation. The discriminator is used to evaluate the authenticity of the three-dimensional fluorescence data image generated by the generator to ensure the authenticity of the newly generated three-dimensional fluorescence data image. Its network hierarchy is shown in Table 3. The discriminator compares the generated spectral data with the real experimental data and gives feedback to guide the generator to optimize. Through adversarial learning, the discriminator continuously improves the quality of the three-dimensional fluorescence data image generated by the generator to ensure that the generated three-dimensional fluorescence data is visually reasonable and conforms to the distribution of the experimental data. Through the collaborative action of each module of the generator and the discriminator, the present invention can generate three-dimensional fluorescence spectral data that conforms to physical laws and has been de-scattered, effectively improving the spectral de-scattering effect under high turbidity conditions, and ensuring the integrity of the material signal and the consistency of the experimental data.

[0094] Table 1: Encoder network hierarchy

[0095]

[0096] Table 2: Network Layers of the Decoder

[0097]

[0098] Table 3: Network Layers of the Discriminator and the Attribute Classifier

[0099]

[0100] It should be understood that the present invention proposes a physical-information-based de-scattering network model. Specifically, a Rayleigh scattering model is constructed through a physics-informed neural network, and physical fitting is performed based on the known scattering peak and width data of the three-dimensional fluorescence spectrum, so as to achieve an accurate simulation of the actual scattering phenomenon and ensure that the de-scattering result can truly reflect the physical law of the scattering signal. In addition, combined with the generative adversarial network, an adaptive network for generating de-scattered data is designed to perform adversarial training on the de-scattering network model. Among them, the generator and the discriminator are continuously optimized in the adversarial training. The generator is responsible for removing the scattering and retaining the substance signal, while the discriminator ensures the authenticity and integrity of the output data, enhancing the robustness of the de-scattering model in removing the scattering.

[0101] Furthermore, the physical reconstruction module is specifically constructed by the following method: a physical reconstruction formula is constructed according to the scattering physical model in the three-dimensional fluorescence spectrum data. Based on the three-dimensional fluorescence data in the three-dimensional fluorescence dataset, the theoretical value of the three-dimensional fluorescence data that removes Rayleigh scattering and retains the substance information is generated through the following steps:

[0102] It should be noted that in the three-dimensional fluorescence spectrum detection, the physical model of Rayleigh scattering can be described by classical electromagnetic theory. According to the scattering theory, the interaction between the incident light and the molecules will cause the polarization and vibration of the molecules, thereby generating radiation. In Rayleigh scattering, the frequency and wavelength of the scattered light are the same as those of the incident light, so the frequency of the incident light will not change, but only the propagation direction of the light is changed. According to the amplitude and phase changes of the scattered light, the intensity formula of Rayleigh scattering can be obtained, which is the scattering physical model, and its expression is:

[0103]

[0104] In the formula, I represents the intensity of Rayleigh scattering, n represents the number of particles per unit volume, v represents the size of the particles, E 0 represents the amplitude of the incident light, λ represents the wavelength of the incident light, n 1 represents the refractive index of the dispersed phase, n 0 represents the refractive index of the dispersion medium.

[0105] (2.1) Measure the minimum excitation wavelength of the three-dimensional fluorescence data at the first-order Rayleigh scattering The maximum fluorescence intensity value of the first-order Rayleigh scattering is and the minimum excitation wavelength At this time, the scattering width L when the first-order Rayleigh scattering drops to one-tenth of the peak 1 .

[0106] (2.2) Measure the minimum excitation wavelength of the three-dimensional fluorescence data in the second-order Rayleigh scattering The maximum fluorescence intensity value of the second-order Rayleigh scattering is and the minimum excitation wavelength At this time, the scattering width L when the second-order Rayleigh scattering drops to one-tenth of the peak 2 .

[0107] (2.3) According to the three-dimensional fluorescence physical information obtained in steps (2.1) and (2.2), perform data fitting on the first-order Rayleigh scattering and the second-order Rayleigh scattering. The fitting formulas for the first-order Rayleigh scattering and the second-order Rayleigh scattering are respectively:

[0108]

[0109]

[0110] In the formula, data 1 represents the physical fitting intensity value of the first-order Rayleigh scattering, data 2 represents the physical fitting intensity value of the second-order Rayleigh scattering, and em and ex represent the emission wavelength and the excitation wavelength respectively.

[0111] (2.4) According to the fitting information of the first-order Rayleigh scattering and the second-order Rayleigh scattering obtained in step (2.3), fit the theoretical value x of the three-dimensional fluorescence data that conforms to the physical law and removes the Rayleigh scattering truth value and retains the substance information according to the following physical reconstruction formula phys , where the expression of the physical reconstruction formula is:

[0112] x phys = data original - data 1 - data 2

[0113] In the formula, data original is the three-dimensional fluorescence data image without de-scattering processing.

[0114] It should be noted that the theoretical value x of the three-dimensional fluorescence data fitted according to the above physical reconstruction formula phys is completely dependent on the scattering physical model, so x phys is the de-scattering value that conforms to the physical law.

[0115] (3) Use the three-dimensional fluorescence data in the three-dimensional fluorescence dataset to perform adversarial training on the de-scattering network model. During the training process, minimize the total loss function of the de-scattering network model as the optimization objective, and adjust the parameters of the de-scattering network model to obtain a trained de-scattering network model. Then, use the trained encoder and decoder as the final de-scattering model.

[0116] Specifically, as Figure 3 shown, input the three-dimensional fluorescence data in the three-dimensional fluorescence dataset into the de-scattering network model. First, it enters the generator. The three-dimensional fluorescence data passes through the encoder and decoder in sequence, and under the guidance of the physical reconstruction module, a de-scattered three-dimensional fluorescence sample x b is generated to deceive the discriminator, and at the same time, a three-dimensional fluorescence sample containing scattering is generated to ensure the accuracy of the generator. Compare the generated three-dimensional fluorescence data containing scattering with the original input three-dimensional fluorescence data with scattering, and calculate the reconstruction loss function. Input the de-scattered three-dimensional fluorescence sample x b and the three-dimensional fluorescence sample containing scattering into the attribute classifier respectively to identify their attributes and obtain the corresponding attribute classification values. Compare the two attribute classification values with the expected attribute values respectively, and calculate the classification loss of the generator and the classification loss of the attribute classifier. Input the de-scattered three-dimensional fluorescence sample x b and the three-dimensional fluorescence sample containing scattering into the discriminator respectively to identify whether the input three-dimensional fluorescence sample is real data and obtain the corresponding authenticity discrimination result. Calculate the adversarial loss of the generator and the discriminator according to the authenticity discrimination result. At the same time, the three-dimensional fluorescence data in the three-dimensional fluorescence dataset generates the corresponding three-dimensional fluorescence data theoretical value x phys through the physical reconstruction module. According to the de-scattered three-dimensional fluorescence data output by the decoder and the three-dimensional fluorescence data theoretical value x phys calculate the physical information loss function. Use the weighted sum of the reconstruction loss function, the classification loss of the generator and the attribute classifier, the adversarial loss of the generator and the discriminator, and the physical information loss function as the total loss function of the de-scattering network model. Among them, the theoretical value generated by the physical reconstruction module will be used as part of the loss function. Through the method of adversarial learning, it guides the training of the generator and the discriminator. The corresponding loss function is specifically expressed as: randomly select N points from the de-scattered three-dimensional fluorescence data generated by the generator, and make the values of these points as close as possible to the three-dimensional fluorescence data theoretical value x phys. During the training process, with the goal of minimizing the total loss function of the de-scattering network model, the parameters of the de-scattering network model are adjusted until the preset number of training epochs is reached, obtaining a trained de-scattering network model, and the trained encoder and decoder are used as the final de-scattering model for subsequent application stages.

[0117] It should be understood that during the training process of the de-scattering network model, the theoretical scattering intensity is generated through the physical reconstruction module, and the theoretical scattering intensity is used as the constraint of the generative adversarial network. The encoder and decoder generate scatter-free fluorescence data, which is evaluated by the discriminator, and the performance of the generator and discriminator is optimized through the adversarial loss function to ensure that the generated data gradually approaches reality and conforms to the physical model.

[0118] Furthermore, the calculation formula of the reconstruction loss function is:

[0119]

[0120] In the formula, represents the reconstruction loss function, x a is the original input three-dimensional fluorescence data containing Rayleigh scattering, is the generated three-dimensional fluorescence data containing scattering, G enc represents the encoder, G dec represents the decoder, represents taking the expected value of the data and attribute labels under the joint distribution p data and p attr , x a ~p data means sampling an image from the three-dimensional fluorescence data set distribution, b~p attr means randomly selecting an attribute vector from the attribute vector distribution.

[0121] Furthermore, the calculation formulas of the classification loss of the generator and the classification loss of the attribute classifier are respectively:

[0122]

[0123]

[0124]

[0125]

[0126] In the formula, is the classification loss of the generator, l g (x a , b) is the binary cross-entropy loss for classification, x a represents the original input three-dimensional fluorescence data containing Rayleigh scattering, and b represents the target attribute label. is the probability value output by the attribute classifier for the i-th attribute, b i is the true label of the i-th attribute; is the classification loss of the classifier, denotes taking the expected value after sampling the input data x a according to the data distribution p data and l r (x a , a) is the classification loss for the reconstructed data, x a represents the three-dimensional fluorescence data with Rayleigh scattering in the original input, a represents the attribute label of the original input data, C i (x a ) is the probability value output by the classifier for the i-th attribute, a i is the true label of the i-th attribute.

[0127] Furthermore, the calculation formulas for the adversarial losses of the generator and the discriminator are respectively:

[0128]

[0129]

[0130] In the formula, represents the adversarial loss of the discriminator, represents the adversarial loss of the generator, denotes taking the expected value after sampling the input data x a according to the data distribution p data and D(x a ) represents the authenticity discrimination result of the discriminator for the real data x a ; represents the authenticity discrimination result of the discriminator for the generated data ; denotes taking the expected value of the data and attribute labels under the joint distributions p data and p attr .

[0131] Furthermore, the calculation formula for the physical information loss function is:

[0132]

[0133] In the formula, represents the physical information loss function, represents the three-dimensional fluorescence data with scattering removed output by the decoder, x phys represents the theoretical value of the three-dimensional fluorescence data fitted by the physical reconstruction module, and M represents the total number of three-dimensional fluorescence data.

[0134] Furthermore, the calculation formula for the total loss function of the de-scattering network model is as follows:

[0135]

[0136] In the formula, represents the total loss function of the de-scattering network model, and λ 1 , λ 2 , λ 3 and λ 4 are the weight parameters of and respectively.

[0137] (4) Input the three-dimensional fluorescence data to be tested or de-scattered into the final de-scattering model, and successively pass through the trained encoder and decoder to output the de-scattered three-dimensional fluorescence data, as shown in Figure 4 .

[0138] Specifically, apply the final de-scattering model to the three-dimensional fluorescence test data of different turbidity samples, and evaluate its performance through comparative experimental results. The experimental results show that the method described in the present invention exhibits excellent robustness and accuracy in high-turbidity samples.

[0139] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A three-dimensional fluorescence spectrum descattering method based on physical information neural network, characterized in that: The following steps are involved: (1) collecting three-dimensional fluorescence spectral data of solutions with different turbidity gradients and removing Raman scattering signals therefrom; expanding the three-dimensional fluorescence spectral data and annotating the material scattering properties, scattering peak values ​​and peak width information of each set of three-dimensional fluorescence spectral data to construct a three-dimensional fluorescence data set; (2) constructing a descattering network model based on a physical information neural network and a generative adversarial network, wherein the descattering network model includes a generator and a discriminator, and the generator includes an encoder, a decoder, a physical reconstruction module, and an attribute classifier; (3) Using the three-dimensional fluorescence data in the three-dimensional fluorescence data set to conduct adversarial training on the descattering network model, minimizing the total loss function of the descattering network model as the optimization goal during the training process, adjusting the parameters of the descattering network model to obtain a trained descattering network model, and using the trained encoder and decoder as the final descattering model; (4) The three-dimensional fluorescence data to be tested or descattered is input into the final descattering model, and is sequentially passed through the trained encoder and decoder to output the three-dimensional fluorescence data with the scattering removed.

2. The three-dimensional fluorescence spectrum descattering method based on physical information neural network according to claim 1 is characterized in that: The step (1) comprises the following sub-steps: (1.1) collecting three-dimensional fluorescence spectrum data of solutions with different turbidity gradients, and using a background subtraction method to remove the Raman scattering signal from each set of three-dimensional fluorescence spectrum data, specifically by subtracting the three-dimensional fluorescence spectrum data of a blank solution sample from the three-dimensional fluorescence spectrum data of the turbidity solution, wherein the blank solution sample is a pure water solution sample; (1.2) The three-dimensional fluorescence spectral data obtained in step (1.1) with the Raman scattering signal removed are expanded by extracting the Rayleigh scattering information, manually eliminating the scattering, supplementing the material information and adjusting the scattering intensity. During the expansion process, the material scattering properties of the peak value and peak width of the material signal and scattering signal of each set of three-dimensional fluorescence spectral data are annotated to construct a three-dimensional fluorescence data set.

3. The three-dimensional fluorescence spectrum descattering method based on physical information neural network according to claim 2 is characterized in that: The step (1.2) includes the following sub-steps: (1.2.1) For each piece of three-dimensional fluorescence spectrum data with Raman scattering signal removed i , construct the attribute vector for it And according to the material scattering properties of the three-dimensional fluorescence spectrum data, the corresponding material signal quantity N and material peak position vector are generated The peak ratio of the scattered signal to the material signal is T i ; Among them, the attribute vector It is marked whether the three-dimensional fluorescence data contains material signals and scattering signals. The expression is: In the formula, represents the ith three-dimensional fluorescence spectrum data with Raman scattering signal removed i The attribute vector of i Indicates whether the i-th three-dimensional fluorescence data contains material signals, and its value range is {0,1}, 0 means no material signal, and 1 means there is a material signal; b i Indicates whether the i-th three-dimensional fluorescence data contains a first-order Rayleigh scattering signal, and its value range is {0,1}, 0 means there is no first-order Rayleigh scattering signal, and 1 means there is a first-order Rayleigh scattering signal; c i Indicates whether the i-th three-dimensional fluorescence data contains a second-order Rayleigh scattering signal, and its value range is {0,1}, 0 means there is no second-order Rayleigh scattering signal, and 1 means there is a second-order Rayleigh scattering signal; the material peak position vector is the combination of the position vectors of N material peaks, and its expression is: In the formula, represents the ith three-dimensional fluorescence spectrum data with Raman scattering signal removed i The material peak position vector of represents the position vector of the jth substance peak in the i-th three-dimensional fluorescence data, em ij represents the emission wavelength corresponding to the jth substance peak in the i-th three-dimensional fluorescence data, ex ij Indicates the excitation wavelength corresponding to the jth substance peak in the i-th three-dimensional fluorescence data; (1.2.2) The three-dimensional fluorescence spectrum data obtained according to step (1.2.1) i The attribute vector Determine whether the data contains material signals. If the data contains material signals, the material peak position vector The position vectors of the N material peaks in As the center, take the product of the material signal peak value and the peak ratio as the maximum value of the scattering signal, and randomly generate N elliptical material signals that conform to the normal distribution; if the data does not contain material signals, jump directly to step (1.2.3); (1.2.3) The three-dimensional fluorescence spectrum data obtained according to step (1.2.1) i The attribute vector Determine whether there is a first-order Rayleigh scattering signal in the data. If the data contains a first-order Rayleigh scattering signal, traverse the data to obtain the peak value p of the first-order Rayleigh scattering signal. i1 and width L1; if the data does not contain a first-order Rayleigh scattering signal, the three-dimensional fluorescence spectrum data data i The first-order Rayleigh scattering signal in is set to zero; (1.2.4) The three-dimensional fluorescence spectrum data obtained according to step (1.2.1) i The attribute vector Determine whether the data contains a second-order Rayleigh scattering signal. If the data contains a second-order Rayleigh scattering signal, traverse the data to obtain the peak value p of the second-order Rayleigh scattering signal. i2 and width L2; if the data does not contain second-order Rayleigh scattering signal, the three-dimensional fluorescence spectrum data data i The second-order Rayleigh scattering signal in is set to zero; finally a new training data data_create is completed i The construction of (1.2.5) Repeat steps (1.2.2) to (1.2.4) to obtain three-dimensional fluorescence data with different scattering properties and material backgrounds to construct a three-dimensional fluorescence data set.

4. The three-dimensional fluorescence spectrum descattering method based on physical information neural network according to claim 1 is characterized in that: The generator is used to edit the three-dimensional fluorescence data according to the input three-dimensional fluorescence data and its corresponding data attribute label, and generate a three-dimensional fluorescence data image that conforms to the new attribute label; In the generator, the encoder is used to obtain high-dimensional potential features of input three-dimensional fluorescence data, specifically, by gradually extracting and compressing the input three-dimensional fluorescence data through convolution layers and activation functions, and converting it into high-dimensional potential space representation features; the decoder is used to obtain the spectral image corresponding to the input high-dimensional potential features, specifically, by gradually expanding the high-dimensional potential feature vector output by the encoder through deconvolution layers, restoring its spatial dimension, and restoring the potential space representation to a three-dimensional fluorescence spectral image after removing scattering; the attribute classifier obtains the corresponding attribute classification result by comparing the attribute vector of the spectral image output by the decoder with the target attribute; the physical reconstruction module generates a scattering theoretical value according to the physical reconstruction formula of Rayleigh scattering and the peak value and width of the first-order and second-order Rayleigh scattering signals in the three-dimensional fluorescence data, so as to guide the generator to generate a spectral image; The discriminator is used to evaluate the authenticity of the three-dimensional fluorescence data image generated by the generator.

5. The three-dimensional fluorescence spectrum descattering method based on physical information neural network according to claim 1 or 4, characterized in that: The physical reconstruction module is constructed by the following method: According to the scattering physical model in the three-dimensional fluorescence spectrum data, a physical reconstruction formula is constructed. Based on the three-dimensional fluorescence data in the three-dimensional fluorescence data set, the theoretical value of three-dimensional fluorescence data with Rayleigh scattering removed and material information retained is generated according to the physical reconstruction formula. This is achieved through the following steps: (2.1) Measure the minimum excitation wavelength of the first-order Rayleigh scattering to obtain three-dimensional fluorescence data The maximum fluorescence intensity of the first-order Rayleigh scattering is and the minimum excitation wavelength Under the condition of , the scattering width L1 when the first-order Rayleigh scattering drops to one-tenth of the peak value; (2.2) Measure the minimum excitation wavelength of three-dimensional fluorescence data on second-order Rayleigh scattering The maximum fluorescence intensity of the second-order Rayleigh scattering is and the minimum excitation wavelength Under the condition of , the scattering width L2 when the second-order Rayleigh scattering drops to one tenth of the peak value; (2.3) According to the three-dimensional fluorescence physical information obtained in step (2.1) and step (2.2), data fitting is performed on the first-order Rayleigh scattering and the second-order Rayleigh scattering, wherein the fitting formulas of the first-order Rayleigh scattering and the second-order Rayleigh scattering are: Where data1 represents the physical fitting intensity value of the first-order Rayleigh scattering, data2 represents the physical fitting intensity value of the second-order Rayleigh scattering, em and ex represent the emission wavelength and excitation wavelength respectively; (2.4) Based on the fitting information of the first-order Rayleigh scattering and the second-order Rayleigh scattering obtained in step (2.3), the theoretical value x of the three-dimensional fluorescence data that removes the true value of Rayleigh scattering and retains the material information is fitted according to the following physical reconstruction formula: phys , where the expression of the physical reconstruction formula is: x phys =data original -data1-data2 In the formula, data original This is a three-dimensional fluorescence data image without past scattering processing.

6. The three-dimensional fluorescence spectrum descattering method based on physical information neural network according to claim 1 is characterized in that: The step (3) specifically comprises: The three-dimensional fluorescence data in the three-dimensional fluorescence data set is input into the descattering network model. It first enters the generator, and the three-dimensional fluorescence data passes through the encoder and decoder in turn. Under the guidance of the physical reconstruction module, the descattered three-dimensional fluorescence sample x is generated. b , while generating a three-dimensional fluorescent sample including scattering The generated three-dimensional fluorescence data containing scattering is compared with the original input three-dimensional fluorescence data with scattering, and the reconstruction loss function is calculated; the de-scattered three-dimensional fluorescence sample x output by the decoder is b and three-dimensional fluorescent samples containing scattering are respectively input into the attribute classifier, and the attributes are identified to obtain the corresponding attribute classification values; the two attribute classification values ​​are compared with the expected attribute values, and the classification loss of the generator and the classification loss of the attribute classifier are calculated; the descattered three-dimensional fluorescent sample x output by the decoder is b and three-dimensional fluorescent samples containing scattering The three-dimensional fluorescence data in the three-dimensional fluorescence data set are respectively input into the discriminator to identify whether the input three-dimensional fluorescence sample is real data, and the corresponding authenticity discrimination result is obtained; the adversarial loss of the generator and the discriminator is calculated according to the authenticity discrimination result; at the same time, the three-dimensional fluorescence data in the three-dimensional fluorescence data set is used to generate the corresponding three-dimensional fluorescence data theoretical value x through the physical reconstruction module phys , according to the three-dimensional fluorescence data without scattering output by the decoder and the theoretical value of the three-dimensional fluorescence data x phys Calculate the physical information loss function; use the reconstruction loss function, the classification loss of the generator and the classification loss of the attribute classifier, the adversarial loss of the generator and the discriminator, and the weighted sum of the physical information loss function as the total loss function of the descattering network model; During the training process, the total loss function of the de-scattering network model is minimized as the optimization goal, and the parameters of the de-scattering network model are adjusted until the preset training rounds are reached to obtain the trained de-scattering network model, and the trained encoder and decoder are used as the final de-scattering model.

7. The three-dimensional fluorescence spectrum descattering method based on physical information neural network according to claim 6 is characterized in that: The calculation formula of the reconstruction loss function is: In the formula, represents the reconstruction loss function, x a is the original input three-dimensional fluorescence data containing Rayleigh scattering, To generate three-dimensional fluorescence data including scattering, G cnc represents the encoder, G dec Represents a decoder, Denotes the joint distribution p data and p attr The data and attribute labels under take the expected value, x a ~p data Indicates that the image is sampled from the distribution of the three-dimensional fluorescence data set, b~p attr Indicates randomly selecting an attribute vector from the attribute vector distribution; The calculation formulas for the classification loss of the generator and the classification loss of the attribute classifier are: In the formula, is the classification loss of the generator, l f (x a ,b) is the binary cross entropy loss for classification, x a represents the original input three-dimensional fluorescence data containing Rayleigh scattering, b represents the target attribute label, is the probability value output by the attribute classifier for the i-th attribute, b i is the true label of the i-th attribute; is the classification loss of the classifier, Represents the input data x a According to the data distribution p data After sampling, take the expected value, l r (x a ,a) is the classification loss for the reconstructed data, x a represents the original input three-dimensional fluorescence data containing Rayleigh scattering, a represents the attribute label of the original input data, C i (x a ) is the probability value output by the classifier for the i-th attribute, a i is the true label of the i-th attribute; The calculation formulas of the adversarial loss of the generator and the discriminator are: In the formula, represents the adversarial loss of the discriminator, represents the adversarial loss of the generator, Represents the input data x a According to the data distribution p data After sampling, take the expected value, D(x a ) represents the discriminator for the real data x a The authenticity judgment result of Represents the discriminator to generate data The authenticity judgment result of Denotes the joint distribution p data and p attr The data and attribute labels below take expected values; The calculation formula of the physical information loss function is: In the formula, represents the physical information loss function, represents the 3D fluorescence data without scattering output by the decoder, x phys represents the theoretical value of three-dimensional fluorescence data fitted by the physical reconstruction module, and M represents the total number of three-dimensional fluorescence data; The calculation formula of the total loss function of the descattering network model is: In the formula, represents the total loss function of the descattering network model, λ1, λ2, λ3 and λ4 are and The weight parameter of .

Images