Cyromazine detection method based on BOSS-PSO-BP neural network and fused with ultraviolet-fluorescence spectrum

Abstract

The invention belongs to the technical field of pesticide residue detection, and particularly discloses an ultraviolet-fluorescence spectrum fused cyromazine detection method based on a BOSS-PSO-BP neural network, and the method comprises the following steps: step 1, preparing a cyromazine sample with gradient concentration of 2.25 * 10 <-6 >-1.47 * 10 <-4 > mg / mL for training a neural network; 2, detecting the samples with different concentrations by using an LS55 fluorescence spectrophotometer and a UV-6300 ultraviolet spectrophotometer respectively, and completing data acquisition; 3, respectively carrying out maximum and minimum value normalization processing on the ultraviolet spectrum and the fluorescence spectrum to unify the data scale, and carrying out data preprocessing; step 4, screening key wavelengths by adopting a BOSS algorithm; step 5, optimizing the BP neural network by a PSO algorithm; and 6, after training is completed, evaluating indexes such as a decision coefficient, RMSE and a relative error, and after requirements are met, storing network parameters, and applying the network parameters to prediction of a to-be-tested sample. Compared with a single spectrum model, the precision of the detection method is improved by about 30%, and the detection accuracy of the trace pesticide in the complex water body is remarkably enhanced.

Description

Technical Field

[0001] This invention belongs to the field of pesticide residue detection technology, and relates to a method for detecting cyromazine using spectroscopy, and particularly to a method for detecting cyromazine based on a BOSS-PSO-BP neural network that integrates ultraviolet-fluorescence spectroscopy. Background Technology

[0002] The widespread use of chemical pesticides in modern agriculture leads to their residues in the environment, posing a potential threat to ecosystems and human health. Traditional detection methods mainly include high-performance liquid chromatography (HPLC), gas chromatography-mass spectrometry (GC-MS), and liquid chromatography-mass spectrometry (LC-MS). While these methods offer high sensitivity and accuracy, they generally require complex sample processing, expensive equipment, and relatively sophisticated operational skills, making them unsuitable for rapid, high-volume detection. Therefore, developing a simple, fast-responding, and highly sensitive detection method can significantly improve detection efficiency in environmental pollutant detection and has a wide range of applications.

[0003] Spectroscopic analysis is an analytical method that studies the physical or chemical properties of substances by measuring how they absorb, emit, or scatter light at different wavelengths. It is characterized by its speed, high sensitivity, non-destructive nature, and versatility. In recent years, it has been widely used in environmental monitoring, food and drug analysis, materials science, and other fields. Using ultraviolet, fluorescence, or infrared spectroscopy techniques to analyze different pollutants, it is possible to discover that the intensity of characteristic wavelength signals of various pollutants is highly correlated with their concentration. By combining feature extraction algorithms for spectral signals with machine learning modeling methods, accurate and rapid detection of pollutant concentrations can be achieved.

[0004] Current spectroscopic detection methods are mainly based on single spectra, but each single spectrum often has certain limitations. For example, ultraviolet spectroscopy can only obtain information on molecular conjugated structures and lacks sensitivity for non-absorbing pollutants or low-concentration samples. Fluorescence spectroscopy depends on specific excitation conditions, cannot capture non-fluorescent substances, and is easily affected by turbidity scattering in water.

[0005] Traditional modeling methods also have significant shortcomings. For example, full-spectrum modeling introduces a large number of irrelevant variables or even noise, leading to overfitting. Traditional feature extraction algorithms also have many flaws. For instance, the Competitive Adaptive System (CARS) algorithm, which selects variables based on the absolute value of regression coefficients, is susceptible to interference from high-noise regions. The iPLS algorithm relies on equally spaced spectral intervals, failing to consider the continuous distribution of feature peaks, resulting in the loss of effective information. The Random Frog Leaping Algorithm (SFLA) independently evaluates the importance of wavelengths, failing to consider the synergistic effects between wavelengths, leading to insufficient model fitting ability.

[0006] When using machine learning modeling, directly employing backpropagation neural networks (BPNNs) can lead to local optima due to the influence of initial parameters during the iterative process of adjusting weights and biases, resulting in slow convergence. Therefore, a method is needed to optimize the search for optimal solutions, ensuring the discovery of the most suitable parameters and improving modeling accuracy. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the present invention aims to provide a method for detecting cyromazine based on a BOSS-PSO-BP neural network that integrates ultraviolet-fluorescence spectroscopy. This invention employs ultraviolet-fluorescence spectroscopy fusion technology to effectively overcome the inherent defects of single spectra through a synergistic and complementary mechanism. By intelligently selecting key wavelengths using the BOSS algorithm, the information complementarity of ultraviolet and fluorescence spectra is finally achieved in the PSO-BP model, improving the accuracy by nearly 30% compared to single spectral models and significantly enhancing the detection accuracy of trace pesticides in complex water bodies.

[0008] This invention is achieved through the following technical solution:

[0009] A method for detecting cyromazine based on BOSS-PSO-BP neural network using ultraviolet-fluorescence spectroscopy includes the following steps:

[0010] Step 1, Sample Preparation: Add diluent to the still lake water in portions to ensure the diluent concentration in the sample reaches 2.25 × 10⁻⁶. -6 ~1.47×10 -4 The gradient range of mg / mL is used to train the neural network;

[0011] Step 2, Data Acquisition: After each addition of cyromazine dilution solution, mix thoroughly, inject the solution into a cuvette, and use an LS55 fluorescence spectrophotometer and a UV-6300 ultraviolet spectrophotometer to detect the data and obtain multiple sets of spectral data at different concentrations.

[0012] Step 3, Data Preprocessing: The UV and fluorescence spectra are normalized to unify the data scale by performing maximum and minimum value normalization, and the original spectral intensity values ​​are compressed to the [0,1] interval; in order to address instrument noise and matrix interference, the UV spectral data are preprocessed using standard normal transformation (SNV), and the fluorescence spectral data are preprocessed using SG filtering.

[0013] Step 4, Feature wavelength extraction: The BOSS algorithm is used to screen 35 key wavelengths, including 10 feature wavelengths of ultraviolet spectrum and 25 feature wavelengths of fluorescence spectrum.

[0014] Step 5: PSO Optimization of the BP Neural Network: Encode the weights and biases of the BP neural network (BPNN) into particle positions, constructing a particle swarm with a dimension equal to the sum of the number of nodes in the input, hidden, and output layers. Use the minimum root mean square error as the fitness function, and employ linearly decreasing inertial weights and a learning factor to drive particle position updates. In the first stage, use PSO global search to obtain approximately optimal parameters. In the second stage, use the PSO results as initial values ​​and combine them with the BP gradient descent algorithm for local fine-tuning. When the iteration reaches 100 times or RMSE < 10, the optimization continues. -5 Optimization will terminate at this time;

[0015] Step 6, Validation and Model Application: After training, evaluate metrics such as the coefficient of determination, RMSE, and relative error. Once the requirements are met, save the network parameters. When predicting new data, maintain the same data format as during training. Simply load the model to output results without retraining.

[0016] A further improvement to the present invention is as follows:

[0017] The concentration of the cyromazine diluent is 0.02-0.03 mg / mL, and its solvent is still lake water.

[0018] Furthermore, when acquiring data, baseline correction and full-scale correction are required before ultraviolet spectroscopy detection. The scanning range is set to 200-300 nm with a step size of 0.5 nm. For fluorescence spectroscopy, the emission wavelength is set to 200-600 nm with a step size of 0.5 nm and the excitation wavelength is set to 265 nm. Fluorescence peak data in the range of 320-520 nm are acquired.

[0019] Furthermore, the formula for normalizing the maximum and minimum values ​​is shown in equation (1):

[0020]

[0021] Where X is the intensity value of the original spectral data point, X max X represents the maximum intensity value of the spectral data. min X is the minimum intensity value of the spectral data. norm It is the normalized intensity value of the data point.

[0022] Furthermore, the network structure of the BP neural network is nmk, which stands for input layer - hidden layer - output layer, and the particle swarm dimension is:

[0023] D=(n×m)+m+(m×k)+k (2)

[0024] The position X of each particle i = [W1,b1,W2,b2], representing a set of network parameters for a solution, with initial position and velocity randomly generated within [-1,1].

[0025] Furthermore, using the root mean square error (RMSE) as an evaluation metric, a fitness function is constructed to drive the particle swarm search direction by minimizing the error.

[0026]

[0027] Where N is the number of samples, this formula calculates the cumulative RMSE for each sample, which comprehensively reflects the performance of the parameters.

[0028] Furthermore, in the first stage, PSO global search is used to obtain approximately optimal parameters, while simultaneously accelerating the convergence speed. The particle velocity v in each iteration... id With position x id for:

[0029] v id (t+1)=ω·v id (t)+c1·r1·(p id -x id (t))+c2·r2·(p gd -x id (t)) (4)

[0030] x id (t+1)=x id (t)+v id (t+1) (5)

[0031] The inertia weight ω adopts a linear decreasing strategy, gradually decreasing from 0.9 to 0.3 to balance global and local exploration; id (t+1) and x id (t+1) represents the velocity and position of particle i at iteration (t+1), respectively; c1 and c2 are learning factors; p jd (t) is the historical optimal position (individual optimal) of particle j, p gd (t) is the global optimal position of all particles in the swarm; r1 and r2 are random numbers uniformly distributed in [0,1].

[0032] To overcome the limitations of single-spectrum analysis, this invention introduces a multispectral fusion method that combines ultraviolet and fluorescence spectroscopy. Multispectral fusion integrates the structural features of the ultraviolet spectrum with the emission characteristics of fluorescence to form a complementary information matrix. Compared to single-spectrum analysis, which is susceptible to environmental factors (such as turbidity and temperature), leading to baseline drift or signal distortion, multispectral fusion overcomes this limitation through a cross-spectral correction mechanism.

[0033] To improve the accuracy of feature extraction, this invention employs a soft threshold shrinkage algorithm (BOSS). The algorithm achieves efficient dimensionality reduction while preserving information from weakly correlated variables through a soft shrinkage strategy and population modeling analysis. The soft shrinkage strategy addresses the problems of traditional variable selection methods, such as over-eliminating weakly correlated variables and ignoring inter-variable synergies, thereby improving the model's prediction accuracy.

[0034] To improve the performance of BPNN neural networks, find the optimal biases and weights for modeling, and overcome their shortcomings of being prone to local optima and slow convergence, this invention uses Particle Swarm Optimization (PSO) to optimize the BPNN neural network. As a population-based global optimization algorithm, PSO simulates the foraging behavior of bird flocks and utilizes a collaborative search mechanism between individual and group historical optimal solutions to effectively find the optimal solution in the solution space. Combining PSO with BPNN can globally optimize the weights and thresholds of the BP network, improving its performance.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0036] The UV-fluorescence spectroscopy fusion technology provided by this invention effectively overcomes the inherent defects of single spectra through a synergistic and complementary mechanism: while UV spectroscopy can capture the characteristic absorption of cyromazine in the 200-300 nm range, its sensitivity is insufficient for low-concentration samples; and while fluorescence spectroscopy relies on the 320-520 nm emission spectrum excited at 265 nm, although it can identify molecular luminescence characteristics, it is easily affected by environmental interference. After fusion, the signal-to-noise ratio is first improved through differentiated preprocessing (UV SNV anti-scattering / fluorescence SG smoothing and noise reduction), and then 35 key wavelengths (10 UV absorption peaks + 25 fluorescence emission peaks) are intelligently screened using the BOSS algorithm, finally achieving information complementarity in the PSO-BP model. UV data compensates for the fluorescence quenching blind zone, and the high sensitivity of fluorescence compensates for the weak UV response, reducing the average relative error of the test set to 9.28%, an improvement of nearly 30% in accuracy compared to single-spectral models, significantly enhancing the detection accuracy of trace pesticides in complex water bodies. Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating the workflow of the model of the present invention;

[0038] Figure 2 A flowchart for feature fusion;

[0039] Figure 3 Optimize the BP neural network flowchart for PSO. Detailed Implementation

[0040] The present invention will now be described in detail with reference to specific embodiments.

[0041] This invention uses lake water from Yangguang Lake Park in Huai'an City, Jiangsu Province as the experimental sample;

[0042] Cypermethrin, with an effective content of 75%, was purchased from Jiangxi Heyi Chemical Co., Ltd.

[0043] LS55 fluorescence spectrophotometer, PerkinElmer, USA;

[0044] UV-6300 UV-Vis spectrophotometer, Shanghai Meipuda Instrument Co., Ltd.

[0045] Example 1

[0046] This embodiment provides a method for detecting cyromazine based on a BOSS-PSO-BP neural network that integrates ultraviolet-fluorescence spectroscopy. The following is a detailed description with reference to the accompanying drawings.

[0047] Step 1, Sample Preparation: The experiment used actual water samples. 5L of water from Yangguang Lake Park in Huai'an City was taken. Under laboratory conditions, the water sample was allowed to stand for 1 hour, and the supernatant was extracted for processing. The cyromazine used had an active ingredient content of 75% and was produced by Jiangxi Heyi Chemical Co., Ltd. 30mg of the 75% cyromazine active ingredient was dissolved in 100ml of the settled lake water, resulting in a concentration of 0.225mg / mL. 10ml of this solution was then diluted 10 times with 90ml of lake water, resulting in a concentration of 0.0225mg / mL. 0.1ml of the 0.0225mg / mL solution was added to the prepared 100ml of settled lake water each time, for a total of 70 additions.

[0048] Step 2, Data Acquisition: After each addition of solution, the solution was stirred thoroughly and then dropped into cuvettes until the solution level was 2 / 3 full. The cuvettes were then placed in a fluorescence spectrophotometer and a UV spectrophotometer for detection. The instruments used in the experiment were an LS55 fluorescence spectrophotometer and a UV-6300 UV-Vis spectrophotometer. For UV data acquisition, baseline and full-scale calibration were performed first. The acquisition range was then set to 200-300 nm, the scan step size to 0.5 nm, and the scan speed to normal. For fluorescence data acquisition, the emission wavelength scan range was set to 200 nm–600 nm, with scans every 0.5 nm at a scan speed of 1500 nm / min. The excitation wavelength was set to 265 nm (the optimal excitation wavelength for the sample), and the slit width to 5 nm. Fluorescence peak data were analyzed, with a wavelength range of 320 nm–520 nm. Each sample contained 400 variables. The experiment yielded data from 70 different samples at varying concentrations, with the concentration range of cyromazine being 2.25 × 10⁻⁶. -6 mg / mL~1.47×10 -4 mg / mL.

[0049] Step 3: Data Preprocessing: The unit of measurement for ultraviolet spectra is absorbance, while the unit of measurement for fluorescence spectra is fluorescence intensity. The orders of magnitude of these two results differ significantly. Therefore, during modeling, it is necessary to ensure that both spectra can be compared using a unified standard input model. This requires data normalization. Data normalization eliminates intensity differences between different spectra and standardizes the data scale.

[0050] This invention performs maximum-minimum normalization on the two spectra separately, and then performs joint modeling. Maximum-minimum normalization is a simple and effective spectral data preprocessing method, as shown in equation (1):

[0051]

[0052] Where X is the intensity value of the original spectral data point, X max X represents the maximum intensity value of the spectral data. min X is the minimum intensity value of the spectral data. norm It is the normalized intensity value of the data point.

[0053] The intensity values ​​of all data in the spectrum are scaled to the range of [0,1] according to their maximum and minimum values. Max-min normalization can eliminate signal differences between different samples and different spectral types while preserving the shape and characteristics of the spectrum.

[0054] In addition, actual instruments exhibit phenomena such as thermal noise, light source fluctuations, and vibrations. In solutions, sample inhomogeneity and scattering effects also contribute to noise that affects experimental results. As can be seen from actual spectra, the measured data often deviates from the true data due to noise, impacting the final results. Therefore, various preprocessing methods are necessary. Experimental verification shows that for cyromazine samples, Standard Normal Variation (SNV) preprocessing for ultraviolet spectral data and Savitzky-Golay (SG) filtering preprocessing for fluorescence spectral data are the optimal preprocessing methods. SG smoothing effectively removes noise while preserving signal peaks, valleys, trends, and other characteristic information. This paper selects an SG filter with a window size of 11 and a polynomial of 3. After SG filtering, the noise in the original spectrum is significantly reduced, but the overall shape of the data remains unchanged. SNV assumes that the spectral curve follows a certain statistical distribution, which can be used to correct outliers in the spectral curve.

[0055] Step 4, Feature Wavelength Extraction: Use the BOSS algorithm to extract features from the collected data.

[0056] Among different feature selection methods, the BOSS (Bootstrap Soft Shrinkage) algorithm is clearly the best. This algorithm generates a PLS sub-model through weighted bootstrap sampling, assesses variable importance based on the absolute value of regression coefficients, and updates weights accordingly. It gradually optimizes the variable size through an iterative soft shrinkage strategy, retaining low-weight variables to avoid erroneously deleting important features, and introduces a collinearity compensation mechanism to balance the effects of multicollinearity. Finally, it selects the optimal wavelength subset with the smallest cross-validation error. Its relative error on single-spectral data is within 30%, while the maximum relative error of CARS, SFLA, SPA, and iPLS algorithms exceeds 40% under the same conditions. In UV-fluorescence spectral fusion data, the BOSS feature extraction algorithm's error is within 20%, while although the relative error values ​​of CARS, SFLA, SPA, and iPLS algorithms are more stable than those on single-spectral data, their fluctuation range is still around 40%, which is inferior to the performance of the BOSS algorithm. Therefore, the BOSS algorithm can extract more effective information to improve model accuracy. The BOSS algorithm is used to select key wavelengths.

[0057] The BOSS Feature Wavelength Extraction Algorithm is a variable selection method based on weighted bootstrap sampling and model population analysis. It optimizes the variable space through an iterative soft contraction strategy: first, it uses bootstrap sampling to generate a subset of variables to construct a PLS sub-model, and evaluates the importance of variables and updates their weights using the absolute values ​​of regression coefficients; then, it employs weighted sampling to gradually shrink the variable size, retaining low-weight variables for subsequent iterations to avoid erroneously removing potentially important variables, while simultaneously balancing the impact of multicollinearity through a collinearity compensation mechanism. In each iteration, the algorithm retains the model with the minimum cross-validation error, ultimately selecting the globally optimal subset of variables.

[0058] Step 5: PSO Optimization of BP Neural Network: The specific steps of optimizing the BP neural network using the Particle Swarm Optimization (PSO) algorithm are as follows:

[0059] Parameter encoding. The BPNN network structure is nmk (input layer - hidden layer - output layer). The weights and biases in the BP neural network are encoded as particle positions, constructing the position space of individuals in the particle swarm with the following dimensions:

[0060] D=(n×m)+m+(m×k)+k (2)

[0061] The position X of each particle i = [W1,b1,W2,b2], representing a set of network parameters for a solution, with initial position and velocity randomly generated within [-1,1].

[0062] Fitness function construction. Root mean square error (RMSE) is used as the evaluation metric, and minimizing the error drives the particle swarm search.

[0063]

[0064] Where N is the number of samples, this formula calculates the cumulative RMSE for each sample, which comprehensively reflects the performance of the parameters.

[0065] Hybrid training method: In the first stage, PSO global search is used to obtain approximately optimal parameters, while simultaneously accelerating convergence. The particle velocity (v) in each iteration... id ) and position (x id )for:

[0066] v id (t+1)=ω·v id (t)+c1·r1·(p id -x id (t))+c2·r2·(p gd -x id (t)) (4)

[0067] x id (t+1)=x id (t)+v id (t+1) (5)

[0068] The inertia weight ω adopts a linear decreasing strategy, gradually decreasing from 0.9 to 0.3 to balance global and local exploration; id (t+1) and x id (t+1) represents the velocity and position of particle i at iteration (t+1), respectively; c1 and c2 are learning factors; p jd (t) is the historical optimal position (individual optimal) of particle j, p gd (t) represents the global optimal position of all particles in the swarm; r1 and r2 are random numbers uniformly distributed in [0,1]. After obtaining the optimal parameters of the PSO search, the BPNN uses the optimal parameters of the PSO as input and combines the gradient descent algorithm of the BPNN for local fine-tuning, balancing efficiency and accuracy.

[0069] Termination condition: When the maximum number of iterations T = 100 is reached or the fitness ΔRMSE < 10. -5 The process terminates at this point, and gbest at this time represents the optimal parameters for the BPNN network.

[0070] Through PSO optimization, BPNN can quickly approximate its global optimum.

[0071] Step 6: Verify the training results and save the network data.

[0072] Once the BPNN network is trained, it outputs a prediction error image. If the coefficient of determination, root mean square error, and relative error all meet the detection requirements, then the output result is the accurate prediction value. At this point, the trained network data is saved. When new data needs to be predicted, ensuring the input data format is consistent with the training data format allows for direct loading and use without retraining.

[0073] Comparative Example 1

[0074] In this comparative example, a UV spectrophotometer was used for data acquisition. After acquisition, the data was preprocessed using SNV. Other operations were largely the same as in Example 1, and will not be repeated here.

[0075] Comparative Example 2

[0076] In this comparative example, a fluorescence spectrophotometer was used for data acquisition. After acquisition, the data was preprocessed using SG filtering. Other operations were largely the same as in Example 1, and will not be repeated here.

[0077] Tables 1 to 3 record the mean relative error (MRE) and coefficient of determination (R²) predicted by different feature extraction methods before and after fusion. 2 The values ​​for RMSE (root mean square error) and root mean square error (RMSE) are as follows:

[0078] Table 1. Average relative errors predicted by different feature extraction methods before and after fusion.

[0079]

[0080] Table 2. R-values ​​predicted by different feature extraction methods before and after fusion. 2 value

[0081]

[0082] Table 3. Root Mean Square Error (RMSE) values ​​predicted by different feature extraction methods before and after fusion.

[0083]

[0084] As shown in Tables 1 to 3, feature wavelength fusion is a highly effective strategy. By integrating the feature wavelengths extracted after UV and fluorescence spectral preprocessing, the complementarity of different spectral information sources is fully utilized, significantly reducing prediction errors (MRE, RMSE) and improving the model's interpretability (R). 2 This approach enhances the model's robustness to different feature extraction methods. The fused model outperforms the model built using only a single preprocessed spectral data source in overall prediction performance, with particularly significant improvements for methods that previously performed poorly on a single data source.

[0085] The above description of the embodiments is only for illustrating the technical concept and features of the present invention. Its purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. Those skilled in the art can obviously easily make various modifications to these embodiments and apply the general principles described herein to other embodiments without creative effort. Therefore, the above embodiments should not be used to limit the scope of protection of the present invention. All improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for detecting cyromazine based on a BOSS-PSO-BP neural network using ultraviolet-fluorescence spectroscopy, characterized in that, Includes the following steps: Step 1, Sample Preparation: Add diluent to the still lake water in portions to ensure the diluent concentration in the sample reaches 2.25 × 10⁻⁶. -6 ~1.47×10 -4 The gradient range of mg / mL is used to train the neural network; Step 2, Data Acquisition: After each addition of cyromazine dilution solution, mix thoroughly, inject the solution into a cuvette, and use an LS55 fluorescence spectrophotometer and a UV-6300 ultraviolet spectrophotometer to detect the data and obtain multiple sets of spectral data at different concentrations. Step 3, Data Preprocessing: Perform maximum and minimum value normalization on both the ultraviolet and fluorescence spectra to unify the data scale, and compress the original spectral intensity values ​​to the [0,1] interval; To address instrument noise and matrix interference, UV spectral data were preprocessed using standard normal transformation (SNV), and fluorescence spectral data were preprocessed using SG filtering. Step 4, Feature wavelength extraction: A total of 35 key wavelengths were selected using the BOSS algorithm, including 10 feature wavelengths from ultraviolet spectroscopy and 25 feature wavelengths from fluorescence spectroscopy. Step 5: PSO Optimization of the BP Neural Network: Encode the weights and biases of the BP neural network into particle positions, constructing a particle swarm with a dimension equal to the sum of the number of nodes in the input, hidden, and output layers; use the minimum root mean square error as the fitness function, and employ linearly decreasing inertial weights and a learning factor to drive particle position updates; initially, use PSO global search to obtain approximately optimal parameters; in the second stage, use the PSO results as initial values ​​and combine them with the BP gradient descent algorithm for local fine-tuning; when the iteration reaches 100 times or RMSE < 10... -5 Optimization will terminate at this time; Step 6, Validation and Model Application: After training, evaluate metrics such as the coefficient of determination, RMSE, and relative error. Once the requirements are met, save the network parameters. When predicting new data, maintain the same data format as during training. Simply load the model to output results without retraining.

2. The method for detecting cyromazine based on a BOSS-PSO-BP neural network using fused ultraviolet-fluorescence spectroscopy according to claim 1, characterized in that: The concentration of the cyromazine diluent is 0.02-0.03 mg / mL, and its solvent is still lake water.

3. The method for detecting cyromazine based on a BOSS-PSO-BP neural network using integrated ultraviolet-fluorescence spectroscopy according to claim 1, characterized in that: When collecting data, baseline correction and full-scale correction are required before ultraviolet spectroscopy detection. Set the scanning range to 200-300nm and the step size to 0.5nm. For fluorescence spectroscopy, set the emission wavelength to 200-600nm, the step size to 0.5nm, and the excitation wavelength to 265nm. Collect fluorescence peak data in the range of 320-520nm.

4. The method for detecting cyromazine based on a BOSS-PSO-BP neural network using fused ultraviolet-fluorescence spectroscopy according to claim 1, characterized in that: The formula for normalizing the maximum and minimum values ​​is shown in equation (1): Where X is the intensity value of the original spectral data point, X max X represents the maximum intensity value of the spectral data. min X is the minimum intensity value of the spectral data. norm It is the normalized intensity value of the data point.

5. The method for detecting cyromazine based on BOSS-PSO-BP neural network using integrated ultraviolet-fluorescence spectroscopy according to claim 1, characterized in that: The network structure of a BP neural network is nmk, which stands for input layer - hidden layer - output layer. The particle swarm dimension is: D=(n×m)+m+(m×k)+k (2) The position X of each particle i = [W1,b1,W2,b2], representing a set of network parameters for a solution, with initial position and velocity randomly generated within [-1,1].

6. The method for detecting cyromazine based on a BOSS-PSO-BP neural network using fused ultraviolet-fluorescence spectroscopy according to claim 1, characterized in that: Using the root mean square error (RMSE) as an evaluation metric, a fitness function is constructed to drive the particle swarm search direction by minimizing the error. Where N is the number of samples, this formula calculates the cumulative RMSE for each sample, which comprehensively reflects the performance of the parameters.

7. The method for detecting cyromazine based on a BOSS-PSO-BP neural network using integrated ultraviolet-fluorescence spectroscopy according to claim 1, characterized in that: In the first stage, PSO global search is used to obtain approximately optimal parameters, while simultaneously accelerating the convergence speed. The particle velocity v in each iteration... id With position x id for: v id (t+1)=ω·v id (t)+c1·r1·(p id -x id (t))+c2·r2·(p gd -x id (t)) (4) x id (t+1)=x id (t)+v id (t+1) (5) The inertia weight ω adopts a linear decreasing strategy, gradually decreasing from 0.9 to 0.3 to balance global and local exploration; id (t+1) and x id (t+1) represents the velocity and position of particle i at iteration (t+1), respectively; c1 and c2 are learning factors; p jd (t) is the historical optimal position (individual optimal) of particle j, p gd (t) is the global optimal position of all particles in the swarm; r1 and r2 are random numbers uniformly distributed in [0,1].

Images