A rapid monitoring method for comprehensive spectral ecological parameters of seawater based on cross-neural networks

Abstract

This invention discloses a rapid monitoring method for comprehensive spectral ecological parameters of seawater based on a cross-neural network. The method includes: receiving the UV-Vis absorption spectrum and three-dimensional fluorescence spectrum of a standard solution sample as initial spectral data; preprocessing the initial spectral data to obtain a comprehensive spectral dataset; establishing a cross-neural network model, inputting the comprehensive spectral dataset into the model, and outputting predicted values ​​of multiple seawater ecological parameter components by fusing spectral feature matrices and cross-referencing information from the UV-Vis absorption spectrum and three-dimensional fluorescence spectrum, and adjusting the network weights by comparing these predicted values ​​with the actual values; stopping training once the model accuracy meets the target, fixing the network weights, and then inputting the preprocessed spectral data of the seawater sample to be tested into the trained cross-neural network model to output the content of the ecological parameter components. This method can simultaneously measure multiple ecological parameters, improving monitoring efficiency and accuracy, and providing a key technology for the development of multi-parameter sensors for seawater quality.

Description

Technical Field

[0001] This invention relates to the monitoring of multiple ecological parameters in seawater, specifically a rapid monitoring method for comprehensive spectral ecological parameters of seawater based on cross-neural networks. Background Technology

[0002] Ultraviolet-visible absorption spectroscopy is a water quality testing method that utilizes the absorption characteristics of substances in water to ultraviolet and visible light to infer the content of substances from the absorption spectrum. The quantitative law of light absorption by substances is based on Beer-Lambert's law. Generally, organic matter and some inorganic matter in water have a certain absorption capacity in the ultraviolet-visible region. Water quality parameters are determined by ultraviolet-visible absorption spectroscopy.

[0003] Fluorescence analysis, a type of direct spectroscopy, is a detection technique that utilizes the fluorescence properties of certain substances to identify their types and detect their content. By fixing the wavelength of excitation light, a water sample is irradiated with this light to induce fluorescence emission; the collected fluorescence spectrum is called the fluorescence emission spectrum. A three-dimensional fluorescence spectrum is a matrix spectrum characterized by three coordinates: excitation wavelength (y-axis), emission wavelength (x-axis), and fluorescence intensity (z-axis). A typical fluorescence spectrum is a planar plot obtained by scanning fluorescence intensity against emission wavelength.

[0004] Neural network methods are mathematical models based on artificial neurons that can be used to handle complex nonlinear problems, possessing advantages such as self-learning, self-organization, and adaptability. Neural network methods can handle complex system problems and simultaneously process quantitative and qualitative information, demonstrating significant advantages in multi-information fusion. In recent years, they have been widely applied in the field of marine ecological parameter monitoring.

[0005] In the field of multi-parameter monitoring of marine ecological parameters, in addition to instrument interference and interference from other substances, aliasing interference can easily occur between multiple parameters to be measured, affecting the accuracy of spectral concentration retrieval. For example, in the ultraviolet-visible absorption spectrum, NO3... - With NO2 - The absorption peaks overlap significantly, making it difficult to distinguish the effects of the two substances on the spectrum, thus hindering the simultaneous determination of NO3. - With NO2 - The difficulty lies in this. DOC (dissolved organic carbon) absorbs in both the ultraviolet and visible regions, and is similar to NO3. - With NO2 - The overlap also affects the subtraction of interferences such as turbidity, further increasing the difficulty of simultaneous measurement of multiple parameters. Summary of the Invention

[0006] To address the aliasing interference problem that easily occurs when monitoring multiple marine ecological parameters, this invention proposes a rapid monitoring method for comprehensive marine spectral ecological parameters based on cross-neural networks. This method can simultaneously measure multiple ecological parameters, improving monitoring efficiency and accuracy.

[0007] The present invention provides a rapid monitoring method for comprehensive spectral ecological parameters of seawater based on cross-neural networks, comprising the following steps:

[0008] 1) Receive the UV-Vis absorption spectrum and three-dimensional fluorescence spectrum of the standard solution sample as initial spectral data;

[0009] 2) Perform smoothing, denoising, differencing, and normalization preprocessing on the initial spectral data to obtain a comprehensive spectral dataset for constructing a cross-neural network model;

[0010] 3) Establish and train a cross-neural network model. Input the comprehensive spectral dataset into the model, fuse the spectral feature matrix, cross-reference the information of ultraviolet-visible absorption spectrum and fluorescence spectrum, output the predicted values ​​of the content of multiple marine ecological parameter components, and compare them with the true values ​​to adjust the network weights.

[0011] 4) Once the model accuracy meets the standard, training is stopped, the network weights are fixed, and when applying the model, the spectral data of the seawater sample to be tested is preprocessed by smoothing, denoising, differencing and normalizing, and then input into the trained cross-neural network model to output the content of ecological parameter components.

[0012] In the rapid monitoring method for comprehensive spectral ecological parameters of seawater based on cross-neural networks of the present invention, step 2) involves preprocessing the initial spectral data, specifically as follows:

[0013] 201) Preprocess the UV-Vis absorption spectrum and convert the original light intensity into absorbance according to the following formula:

[0014]

[0015] Where A is absorbance, I0 is original light intensity, and I is emitted light intensity;

[0016] 202) The Savitzky-Golay convolution smoothing algorithm was used to smooth and filter the UV-Vis absorption spectrum and the three-dimensional fluorescence spectrum;

[0017] 203) Perform first-order and second-order difference analysis on the ultraviolet-visible absorption spectrum to obtain a total of three sets of data, including the ultraviolet-visible absorption spectrum and its first-order and second-order difference spectra.

[0018]

[0019]

[0020] Among them, Aλ λ represents the absorbance at wavelength λ. It is a first-order difference. This is a second-order difference, where Δ is the resolution of the absorption spectrum;

[0021] 204) Normalize the UV-Vis absorption spectrum and its first-order, second-order difference spectra and three-dimensional fluorescence spectra according to the following formula:

[0022]

[0023]

[0024]

[0025] Where, x i Here is the original data, and m is the number of samples. This is the final training set input data, where ∈ is the remainder to prevent overflow when dividing by zero, and is set to 10. -9 The values ​​of μ and σ need to be saved so that the input data can be normalized in practical applications.

[0026] 205) The preprocessed UV-Vis absorption spectra and their first and second order difference spectral vectors are concatenated to obtain the absorption spectral matrix. The preprocessed three-dimensional fluorescence spectra are directly used to obtain the fluorescence spectral matrix. The absorption spectral matrix, fluorescence spectral matrix and true values ​​are combined together as the modeling dataset and randomly divided into training set and test set in a 7:3 ratio.

[0027] In the rapid monitoring method for comprehensive spectral ecological parameters of seawater based on cross-neural networks of the present invention, step 3) specifically includes:

[0028] 301) The absorption spectrum matrix and fluorescence spectrum matrix are used as input data. After the absorption spectrum matrix is ​​convolved and pooled three times, the absorption spectrum feature matrix is ​​obtained, which characterizes the absorption spectrum characteristics of the analyte in the seawater sample. The fluorescence spectrum matrix is ​​convolved and pooled three times to obtain the fluorescence spectrum feature matrix, which characterizes the fluorescence spectrum characteristics of the analyte in the seawater sample.

[0029] 302) The absorption spectral feature matrix and the fluorescence spectral feature matrix are spliced ​​together to obtain the comprehensive spectral cross matrix;

[0030] 303) Perform two convolutional pooling operations on the comprehensive spectral cross matrix to cross-reference the absorption and fluorescence spectra, then flatten it into a one-dimensional vector, and output DOC and NO3 through a fully connected layer. - NO2 - Concentration; (Please find a suitable place in the DOC file to explain what it is)

[0031] 304) Compare the actual values ​​with the model's predicted values ​​to calculate the error, perform backpropagation to update the network weights, and record one epoch as the time it takes to use all training set data once. After 10 epochs, verify the model's generalization ability through the test set. Stop training when the model's accuracy on the test set reaches the target.

[0032] In the rapid monitoring method for comprehensive spectral ecological parameters of seawater based on cross-neural networks of the present invention, model training is divided into two stages:

[0033] The first stage uses single-substance standard solution data for training, to train an inversion model for a single substance; the second stage uses mixed solution data for training, to train an inversion model for multiple substances.

[0034] In the rapid monitoring method for comprehensive spectral ecological parameters of seawater based on cross-neural networks of the present invention, all activation functions adopt Leaky ReLU, as shown in the following formula:

[0035]

[0036] Where x is the forward propagation output value of a single neuron, and α is a hyperparameter, set to 0.01.

[0037] The rapid monitoring method for comprehensive spectral ecological parameters of seawater based on cross-neural networks of the present invention has the following beneficial effects and advantages:

[0038] 1. The rapid monitoring method for multiple components of ecological parameters of the present invention, based on a cross-neural network model and a comprehensive spectral method, can improve monitoring efficiency and accuracy, and provides key technologies for the development of multi-parameter sensors for seawater quality.

[0039] 2. This invention can solve the aliasing interference problem that is easily generated when monitoring multiple marine ecological parameters, has strong anti-interference ability, and can improve the stability of the integrated spectral multi-parameter monitoring system. Attached Figure Description

[0040] Figure 1 The flowchart shows a rapid monitoring method for integrated spectral ecological parameters of seawater based on cross-neural networks.

[0041] Figure 2 This is a schematic diagram of the network structure of a cross-neural network model;

[0042] Figure 3 This is a flowchart of the training process for a cross-neural network. Detailed Implementation

[0043] The present invention will now be further described with reference to the accompanying drawings.

[0044] like Figure 1As shown, the rapid monitoring method for comprehensive spectral ecological parameters of seawater based on cross-neural networks of the present invention includes the following steps:

[0045] 1) Receive the UV-Vis absorption spectrum and three-dimensional fluorescence spectrum of the standard solution sample as initial spectral data;

[0046] In practice, the UV-Vis absorption spectrum range is 200 nm to 700 nm with an interval of 0.625 nm, and the data volume is 800 × 1. The three-dimensional fluorescence spectral parameters are: emission spectrum range of 200 nm to 700 nm, excitation spectrum range of 200 nm to 700 nm with an interval of 5 nm, and the data volume is 100 × 100.

[0047] The concentration table for standard solution sample preparation is shown in Table 1, including DOC and NO3. - NO2 - The single-substance standard solution yielded 18 sets of data (6×3), while the multi-substance standard solution was prepared using an orthogonal experimental method, yielding 49 sets of data (5 factors, 7 levels).

[0048] Table 1. Concentration of Standard Solution Samples in Spectral Dataset

[0049]

[0050] 2) The initial spectral data is preprocessed by smoothing, denoising, differencing, and normalizing to obtain the comprehensive spectral dataset for constructing the cross-neural network model; the preprocessing of the initial spectral data is as follows:

[0051] 201) Preprocess the UV-Vis absorption spectrum and convert the original light intensity into absorbance according to the following formula:

[0052]

[0053] Where A is absorbance, I0 is original light intensity, and I is emitted light intensity;

[0054] 202) The Savitzky-Golay convolution smoothing algorithm was used to smooth and filter the UV-Vis absorption spectrum and the three-dimensional fluorescence spectrum;

[0055] 203) Perform first-order and second-order difference analysis on the UV-Vis absorption spectrum to obtain three sets of data: the UV-Vis absorption spectrum and its first-order and second-order difference spectra, with a total data volume of 800 × 3. The specific formula is as follows:

[0056]

[0057]

[0058] Among them, A λ λ represents the absorbance at wavelength λ. It is a first-order difference. The second-order difference is represented by Δ, which is the resolution of the absorption spectrum, and in this embodiment it is 0.625 nm.

[0059] 204) Normalize the UV-Vis absorption spectrum and its first-order, second-order difference spectra and three-dimensional fluorescence spectra according to the following formula:

[0060]

[0061]

[0062]

[0063] Where, x i Here is the original data, and m is the number of samples. This is the final training set input data, where ∈ is the remainder to prevent overflow when dividing by zero, and is set to 10. -9 The values ​​of μ and σ need to be saved so that the input data can be normalized in practical applications.

[0064] 205) The preprocessed UV-Vis absorption spectra and their first and second order difference spectral vectors are concatenated to obtain the absorption spectral matrix. The preprocessed three-dimensional fluorescence spectra are directly used to obtain the fluorescence spectral matrix. The absorption spectral matrix, fluorescence spectral matrix and true values ​​are combined together as the modeling dataset and randomly divided into training set and test set in a 7:3 ratio.

[0065] 3) such as Figure 2 As shown, a cross-neural network model is established and trained. The comprehensive spectral dataset is input into the model, and by fusing the spectral feature matrix and cross-referencing information from UV-Vis absorption and fluorescence spectra, predicted values ​​of multiple marine ecological parameters are output. These predicted values ​​are then compared with the actual values ​​to adjust the network weights. The specific process is as follows: Figure 3 As shown, it includes the following steps:

[0066] 301) The absorption spectrum matrix and fluorescence spectrum matrix are used as input data. After the absorption spectrum matrix is ​​convolved and pooled three times, the absorption spectrum feature matrix is ​​obtained, which characterizes the absorption spectrum characteristics of the analyte in the seawater sample. The fluorescence spectrum matrix is ​​convolved and pooled three times to obtain the fluorescence spectrum feature matrix, which characterizes the fluorescence spectrum characteristics of the analyte in the seawater sample.

[0067] 302) The absorption spectral feature matrix and the fluorescence spectral feature matrix are spliced ​​together to obtain the comprehensive spectral cross matrix;

[0068] 303) Perform two convolutional pooling operations on the comprehensive spectral cross matrix to cross-reference the absorption and fluorescence spectra, then flatten it into a one-dimensional vector, and output DOC and NO3 through a fully connected layer. - NO2- concentration;

[0069] In this embodiment, H 11 The convolutional layer size is 3×3×1×3, the pooling layer size is 2×2, and max pooling is used; H 12 The convolutional layer size is 3×3×3×3, the pooling layer size is 2×2, and max pooling is used; H 13 The convolutional layer size is 3×3×3×3, the pooling layer size is 2×2, and max pooling is used; after convolution and pooling, the feature layer size of the three-dimensional fluorescence spectral matrix is ​​12×12×3. H 21 The convolutional layer size is 3×3×1×3, the pooling layer size is 4×4, and max pooling is used; H 22 The convolutional layer size is 3×1×3×3, the pooling layer size is 4×4, and max pooling is used; H 23 The convolutional layer size is 3×1×3×3, and the pooling layer size is 4×4, using max pooling. After convolution and pooling, the feature layer size of the UV-Vis absorption spectrum and its difference matrix is ​​12×1×3. Concatenating the feature matrices yields a comprehensive spectral cross matrix of size 12×13×3, H. 31 The convolutional layer size is 3×3×3×3, the pooling layer size is 2×2, and max pooling is used; H 32 The convolutional layer has a size of 3×3×3×1, and the pooling layer has a size of 2×2, using max pooling. After two convolutional pooling operations, a feature matrix of size 3×3×1 is obtained, which is then expanded into a one-dimensional vector. The fully connected layer has a size of 9×3, and the final output value is a one-dimensional vector of size 3×1.

[0070] In practice, all activation functions use Leaky ReLU, as shown in the following formula:

[0071]

[0072] Where x is the forward propagation output value of a single neuron, and α is a hyperparameter, typically taken as 0.01.

[0073] 304) Compare the actual values ​​with the model's predicted values ​​to calculate the error, perform backpropagation to update the network weights, and record one epoch as the time it takes to use all training set data once. After 10 epochs, verify the model's generalization ability through the test set. Stop training when the model's accuracy on the test set reaches the target.

[0074] In practice, model training is divided into two stages:

[0075] The first stage uses single-substance standard solution data for training, to train an inversion model for a single substance; the second stage uses mixed solution data for training, to train an inversion model for multiple substances.

[0076] 4) Once the model accuracy meets the target, training is stopped and the network weights are fixed. In application, the spectral data of the seawater sample to be tested undergoes smoothing, denoising, differencing, and normalization preprocessing, and is then input into the trained cross-neural network model to output the content of ecological parameter components.

Claims

1. A rapid monitoring method for comprehensive spectral ecological parameters of seawater based on cross-neural networks, characterized in that, Includes the following steps: 1) Receive the UV-Vis absorption spectrum and three-dimensional fluorescence spectrum of the standard solution sample as initial spectral data; 2) Perform smoothing, denoising, differencing, and normalization preprocessing on the initial spectral data to obtain a comprehensive spectral dataset for constructing a cross-neural network model; 3) Establish and train a cross-neural network model. Input the comprehensive spectral dataset into the model, fuse the spectral feature matrix, cross-reference the information of ultraviolet-visible absorption spectrum and fluorescence spectrum, output the predicted values ​​of the content of multiple marine ecological parameter components, and compare them with the true values ​​to adjust the network weights. 4) Once the model accuracy meets the standard, training is stopped, the network weights are fixed, and when applying the model, the spectral data of the seawater sample to be tested is preprocessed by smoothing, denoising, differencing, and normalizing, and then input into the trained cross-neural network model to output the content of ecological parameter components.

2. The rapid monitoring method for comprehensive spectral ecological parameters of seawater based on cross-neural networks according to claim 1, characterized in that, Step 2) involves preprocessing the initial spectral data, specifically as follows: 201) Preprocess the ultraviolet-visible absorption spectrum, and convert the original light intensity into absorbance according to the following formula: Where A is absorbance, I0 is original light intensity, and I is emitted light intensity; 202) The Savitzky-Golay convolution smoothing algorithm was used to smooth and filter the UV-Vis absorption spectrum and the three-dimensional fluorescence spectrum; 203) Perform first-order and second-order difference analysis on the ultraviolet-visible absorption spectrum to obtain a total of three sets of data, including the ultraviolet-visible absorption spectrum and its first-order and second-order difference spectra; Among them, A λ λ represents the absorbance at wavelength λ. It is a first-order difference. This is a second-order difference, where Δ is the resolution of the absorption spectrum; 204) Normalize the UV-Vis absorption spectrum and its first-order, second-order difference spectra and three-dimensional fluorescence spectra according to the following formula: Where, x i Here is the original data, and m is the number of samples. This is the final training set input data, where ∈ is the remainder to prevent overflow when dividing by zero, and is set to 10. -9 The values ​​of μ and σ need to be saved so that the input data can be normalized in practical applications. 205) The preprocessed UV-Vis absorption spectra and their first and second order difference spectral vectors are concatenated to obtain the absorption spectral matrix. The preprocessed three-dimensional fluorescence spectra are directly used to obtain the fluorescence spectral matrix. The absorption spectral matrix, fluorescence spectral matrix and true values ​​are combined together as the modeling dataset and randomly divided into training set and test set in a 7:3 ratio.

3. The rapid monitoring method for comprehensive spectral ecological parameters of seawater based on cross-neural networks according to claim 2, characterized in that, Step 3) specifically involves: 301) The absorption spectrum matrix and fluorescence spectrum matrix are used as input data. After the absorption spectrum matrix is ​​convolved and pooled three times, the absorption spectrum feature matrix is ​​obtained, which characterizes the absorption spectrum features of the analyte in the seawater sample. The fluorescence spectral matrix is ​​subjected to three convolutional pooling operations to obtain the fluorescence spectral feature matrix, which characterizes the fluorescence spectral features of the analytes in the seawater sample. 302) The absorption spectral feature matrix and the fluorescence spectral feature matrix are spliced ​​together to obtain the comprehensive spectral cross matrix; 303) Perform two convolutional pooling operations on the comprehensive spectral cross matrix to cross-reference the absorption and fluorescence spectra, then flatten it into a one-dimensional vector, and output DOC and NO3 through a fully connected layer. - NO2 - concentration; 304) Compare the actual values ​​with the model's predicted values ​​to calculate the error, perform backpropagation to update the network weights, and record one epoch as the time it takes to use all training set data once. After 10 epochs, verify the model's generalization ability through the test set. Stop training when the model's accuracy on the test set reaches the target.

4. The rapid monitoring method for comprehensive spectral ecological parameters of seawater based on cross-neural networks according to claim 3, characterized in that, Model training is divided into two stages: The first stage uses single-substance standard solution data for training, to train an inversion model for a single substance; the second stage uses mixed solution data for training, to train an inversion model for multiple substances.

5. The rapid monitoring method for comprehensive spectral ecological parameters of seawater based on cross-neural networks according to claim 3, characterized in that, All activation functions use Leaky ReLU, as shown in the following formula: Where x is the forward propagation output value of a single neuron, and α is a hyperparameter, set to 0.01.

Images