Water pollution source analysis method based on adaptive EEM and ACNN-MS

Abstract

The invention belongs to the technical field of environmental monitoring, and particularly relates to a water pollution source identification method based on adaptive three-dimensional fluorescence spectrum (3D-EEM) analysis and a multi-scale convolutional neural network (ACNN-MS). Specifically, the invention provides a method combining a self-adaptive three-dimensional fluorescence spectrum analysis technology and a multi-scale convolutional neural network (MS-CNN), 3D-EEM data are obtained through an improved high-sensitivity three-dimensional fluorescence spectrometer, residual connection is introduced in model construction to avoid the gradient disappearance problem, and the method has the advantages that the method is simple and easy to implement, and the real-time performance of the 3D-EEM data is improved. And a new adaptive wavelength selection algorithm (AWSSA) is utilized to screen the most representative wavelength interval for high-precision calculation, so that the defects of high detection limit, low precision, need of opening a black box and difficulty in traceability positioning of a traditional spectrum and algorithm are overcome. According to the method, the pollution source can be efficiently positioned, the early warning information can be provided, and the quick response capability of the water body pollution event is remarkably improved.

Description

1. Technical Field

[0001] The present invention belongs to the technical field of environmental pollution monitoring, and in particular, relates to a method for identifying water pollution sources based on adaptive three-dimensional fluorescence spectrum (3D-EEM) analysis and multi-scale convolutional neural network (ACNN-MS). The present invention solves the problems of weak background differentiation ability, poor identification ability, insufficient traceability accuracy, high positioning difficulty, etc. of the existing micro-pollutant source tracing technology. The method combines adaptive three-dimensional fluorescence spectrum analysis technology and multi-scale convolutional neural network (MS-CNN) to efficiently locate pollution sources and provide early warning information, which significantly improves the rapid response capability of water pollution events. 2. Background Technology

[0002] Water pollution is an important issue in global environmental governance. In recent years, with the rapid development of industrialization and urbanization, water pollution has become increasingly serious, especially the widespread presence of nitrogen and phosphorus pollution, heavy metal pollution, and emerging pollutants (such as microplastics, pharmaceuticals, and personal care products), which pose a serious threat to ecosystems and human health. In terms of water pollution source analysis technology, research at home and abroad has made significant progress. For example, the use of isotope tracing technology, multi-isotope joint analysis, microbial fingerprint technology and other methods can identify the source and migration path of pollutants to a certain extent. In addition, source analysis technology based on characteristic pollution factors is also constantly developing. By combining big data analysis and diffusion models, it can achieve preliminary positioning of pollution sources and contribution rate assessment. However, although these technologies have achieved certain results at the macro level, there are still many challenges at the micro level, especially in the accurate identification and tracing of micropollutants. These challenges not only limit the level of refinement of water pollution control, but also affect the scientific decision-making of environmental management.

[0003] The main problems faced by existing water pollution source analysis technologies in practical applications are insufficient identification of micropollutants, limited traceability accuracy, and high difficulty in positioning. First, the concentration of micropollutants (such as antibiotic resistance genes, nanomaterials, etc.) in the environment is usually extremely low (nanogram or even picogram level). Although existing detection technologies such as liquid chromatography-mass spectrometry (LC-MS / MS) and gas chromatography-mass spectrometry (GC-MS / MS) have high sensitivity, they are still difficult to meet the detection needs of trace pollutants in complex environmental matrices. Secondly, the sources of micropollutants are diverse and the migration paths are complex. They may undergo multiple media conversions and long-distance migration, making it difficult to accurately restore the traceability path. In addition, the existing technology has a weak ability to distinguish between environmental background and pollutants. For example, in the effluent samples of sewage treatment plants, complex organic matter matrices may mask or interfere with the detection of micropollutants, increasing analytical uncertainty. Finally, although isotope tracing technology and microbial fingerprint technology have improved the traceability accuracy to a certain extent, the application of these methods in micropollutant traceability is still limited by insufficient data accumulation and high technical costs.

[0004] In recent years, domestic and foreign scholars have made significant progress in improving the traceability technology of micropollutants in water. For example, source analysis technology based on characteristic pollution factors can achieve preliminary identification and traceability of micropollutants by constructing a pollution source DNA library and diffusion model (CN201910341774.3). In addition, the application of isotope tracing technology has also been further expanded. For example, the ratio of hydrogen isotopes to carbon isotopes can distinguish micropollutants from human activities and natural sources (Journal of Environmental Engineering Technology, 2020, 10(2):242). In the field of traceability of microplastic pollution, research progress is particularly outstanding. For example, separation and extraction technology and qualitative and quantitative detection methods can identify the source of microplastics and their ecotoxicological effects (Journal of Environmental Engineering Technology, 2023, 13(5):1883). At the same time, traceability technology based on big data and artificial intelligence is also developing continuously. For example, deep mining of complex environmental data through machine learning algorithms can improve the traceability accuracy and efficiency of micropollutants (Xiong Wen et al., Science Press, 2024). The advancement of these technologies provides important support for the accurate tracing of micropollutants.

[0005] However, current research and technical solutions still have significant defects, as follows:

[0006] 1) Low detection accuracy and insufficient traceability. For example, liquid chromatography-mass spectrometry and other equipment can only detect water-soluble organic pollutants of a certain concentration in water, and the detected species, molecular weight and response value are all limited;

[0007] 2) Species identification cannot replace the tracing of pollution sources. For example, pollutants can be accurately characterized by mass spectrometry, spectroscopy, and nuclear magnetic resonance, but it is actually difficult to accurately determine where the pollutants come from;

[0008] 3) Existing equipment is bulky, expensive, has slow detection speed, and long response cycle, making it impossible to achieve timely detection and difficult to provide accurate early warning. III. Summary of the invention

[0009] The present invention is made in view of the problems existing in the prior art. The present invention has made systematic improvements to the three-dimensional fluorescence spectrometer, significantly improving the detection accuracy and the detection limit of trace pollutants. First, a high-sensitivity spectrometer is selected to replace the high quantum efficiency photomultiplier tube (PMT), which significantly reduces the response time and improves the sensitivity to dark current; an aspheric lens group is introduced into the optical path to reduce the influence of aberration and improve the imaging quality; a built-in temperature control system is used to prevent temperature fluctuations from interfering with the measurement accuracy.

[0010] In addition, the present invention introduces an adaptive wavelength selection algorithm (AWSSA) and intelligent data analysis technology to achieve the analytical goal of "fuzzifying data and clarifying conclusions". AWSSA guides the wavelength selection of new samples by analyzing the characteristic wavelength distribution in historical data sets, significantly improving detection efficiency. A multi-scale convolutional neural network (MS-CNN) is constructed, and the residual connection structure and batch normalization (BN) technology are introduced to enhance model stability. Combined with transfer learning technology, the trained ACNN-MS model has an identification accuracy of 96.5% for specific types of pollutants in a simulated environment, and the generalization ability for unknown pollutants is also improved by 12.3%. In summary, the present invention solves the common problems existing in traditional water pollution tracing technology by integrating efficient sampling schemes, improved three-dimensional fluorescence spectrometer design, and advanced intelligent algorithms.

[0011] In order to achieve the above object, the technical solution of the present invention is a water pollution source analysis method based on adaptive EEM and ACNN-MS, characterized in that the method specifically comprises the following steps:

[0012] Step 1: Collection and pretreatment of water samples. Multiple sampling points were set up in suspected contaminated areas to ensure that the samples were representative. Water samples were collected using sterile polyethylene bottles to avoid cross-contamination between samples. Water samples were filtered through a nylon filter membrane with a pore size of 0.45 μm to remove suspended particles and diluted with ultrapure water to a suitable concentration to meet the requirements of three-dimensional fluorescence (EEM) detection. To eliminate the inner filter effect, the samples were pretreated using the ultraviolet absorption correction method to ensure data accuracy.

[0013] Step 2: Select a dedicated high-sensitivity three-dimensional fluorescence spectrometer that supports an excitation wavelength range of 200 to 400 nm, an emission wavelength range of 250 to 600 nm, and a resolution better than 5 nm.

[0014] Step 3: Design of high-sensitivity fluorescence spectrometer and strategy for optical path improvement; the spectrometer uses a high-quantum-efficiency photomultiplier tube (PMT) to replace the original detector, with a response time requirement of ≤1ns and a dark current of ≤1pA; the light source is a pulsed xenon lamp with an energy density of ≥1J / cm 2 , ensuring sufficient excitation intensity; introducing a new aspheric lens group in the optical path to reduce the influence of aberration and improve the imaging quality; in addition, adding a built-in temperature control system, requiring the working temperature to be stable within the range of ±0.1℃ to prevent temperature fluctuations from interfering with the measurement accuracy; the above improvement measures are used to improve the detection accuracy and the detection limit of trace pollutants and their detection signals;

[0015] Step 4, adaptive wavelength selection algorithm (AWSSA) optimization step; AWSSA analyzes the characteristic wavelength distribution in the historical data set, establishes a prediction model, and guides the wavelength selection of new samples; the algorithm uses a genetic algorithm (GA) for parameter optimization, with the initial population size set to 100, the number of iterations to 50, the crossover probability to 0.8, and the mutation probability to 0.1; the optimal solution is updated after each iteration, gradually approaching the ideal wavelength combination, significantly improving the detection efficiency;

[0016] Step 5, using principal component analysis (PCA) and wavelet transform (WT) for denoising and smoothing; applying PCA to the original three-dimensional fluorescence data to extract the main eigenvectors, retaining more than 95% of the variance information, and effectively compressing the data dimension; then, using WT to perform multi-scale decomposition of each principal component, selecting the db4 wavelet basis function, the number of decomposition layers is 3, and the threshold is set to the soft thresholding method to achieve denoising and smoothing processing and maintain signal integrity;

[0017] Step 6: Construct a multi-scale convolutional neural network (MS-CNN) architecture containing multiple convolutional layers. Each layer uses convolution kernels of different sizes (such as 3×3, 5×5, and 7×7) to extract multi-scale spatial features. The residual connection structure is introduced to avoid the gradient vanishing problem and speed up the training. At the same time, the batch normalization (BN) technology is applied to enhance the model stability. Finally, the final result is output through the Softmax classifier to ensure the prediction accuracy.

[0018] Step 7: The attention mechanism dynamically adjusts the weights of each layer. The attention mechanism is introduced to dynamically adjust the weights of each layer according to the results of each iteration, highlighting important feature areas and suppressing noise interference. The attention score is calculated using a weighted summation method, and the weight is controlled by an adaptive learning rate to ensure that the model can flexibly respond to different types of pollution source characteristics.

[0019] Step 8: Improve generalization capabilities. Combined with transfer learning technology, use pre-trained models to initialize network parameters and quickly adapt to new types of pollution. In addition, regularly update the training data set to cover more types of pollutants, further improve the generalization and robustness of the model, and ensure long-term stable performance.

[0020] It is further defined that the water sample pretreatment described in step 1 also includes adding an appropriate amount of fluorescence quencher, such as ascorbic acid, at a concentration of 0.1 to 1 mM to eliminate background fluorescence interference and improve the signal-to-noise ratio.

[0021] It is further defined that the improved adaptive three-dimensional fluorescence spectrum design described in step 3 also involves the introduction of an automatic focusing function, using a laser rangefinder to monitor the liquid level in the sample pool in real time to ensure that the focus position is consistent during each measurement, thereby improving data consistency.

[0022] Further defined, the PCA and WT noise reduction processing described in step 5 also includes applying independent component analysis (ICA) to separate the various independent sources in the mixed signal, further improving the data quality and ensuring the accuracy of subsequent analysis.

[0023] Further defining, the attention mechanism described in step 7 also includes introducing a local sensitive hashing (LSH) algorithm, using To speed up similarity search It can improve the model training efficiency and is particularly suitable for rapid response requirements under large-scale data sets.

[0024] The present invention is beneficial in that:

[0025] 1. High-precision detection: Optimized sampling and pre-processing, combined with an improved three-dimensional fluorescence spectrometer, significantly improve the accuracy of identifying trace pollutants.

[0026] 2. Intelligent data analysis: AWSSA and ACNN-MS models are introduced to achieve adaptive wavelength selection and feature extraction, improve detection efficiency and traceability accuracy, and enhance model generalization capabilities.

[0027] 3. Real-time response and warning: Through efficient sampling schemes and rapid analysis technology, the response cycle can be shortened, timely prevention and control of water pollution and accurate early warning can be achieved, and water resource security can be guaranteed. IV. Description of the drawings

[0028] In order to more clearly illustrate the specific embodiments of the present invention, the drawings used in the description of the specific embodiments are briefly described below.

[0029] Figure 1 This is a flow chart of a water pollution source analysis method based on adaptive EEM and ACNN-MS.

[0030] Figure 2 Comparison diagram of the three-dimensional fluorescence signal of water samples scanned before (left) and after (right) the improvement of the three-dimensional fluorescence spectrometer.

[0031] Figure 3 A comparison chart of the accuracy of correlation identification of water samples before (left) and after (right) using the adaptive wavelength selection algorithm (AWSSA). V. Specific implementation methods

[0032] The present invention is described in detail below with reference to the accompanying drawings and embodiments:

[0033] Embodiment 1:

[0034] A method for water pollution source analysis based on adaptive EEM and ACNN-MS, characterized in that the method specifically comprises the following steps:

[0035] Step 1, water sample collection and pretreatment, set up multiple sampling points in suspected contaminated areas to ensure that the samples are representative; use sterile polyethylene bottles to collect water samples to avoid cross-contamination between samples; filter the water samples through a nylon filter membrane with a pore size of 0.45μm to remove suspended particles, and dilute with ultrapure water to a suitable concentration to meet the requirements of three-dimensional fluorescence (EEM) detection; water sample pretreatment also includes adding an appropriate amount of fluorescence quencher, such as ascorbic acid, at a concentration of 0.1 to 1mM to eliminate background fluorescence interference and improve the signal-to-noise ratio; in order to eliminate the inner filter effect, the ultraviolet absorption correction method is used to pretreat the samples to ensure data accuracy;

[0036] Step 2: Select a dedicated high-sensitivity three-dimensional fluorescence spectrometer that supports an excitation wavelength range of 200 to 400 nm, an emission wavelength range of 250 to 600 nm, and a resolution better than 5 nm.

[0037] Step 3: Design of high-sensitivity fluorescence spectrometer and strategy for optical path improvement; the spectrometer uses a high-quantum-efficiency photomultiplier tube (PMT) to replace the original detector, with a response time requirement of ≤1ns and a dark current of ≤1pA; the light source is a pulsed xenon lamp with an energy density of ≥1J / cm 2, ensuring sufficient excitation intensity; introducing a new aspheric lens group in the optical path to reduce the influence of aberration and improve the imaging quality; in addition, adding a built-in temperature control system, requiring the working temperature to be stable within the range of ±0.1℃, to prevent temperature fluctuations from interfering with the measurement accuracy; the improved adaptive three-dimensional fluorescence spectrum design also involves the introduction of an automatic focusing function, using a laser rangefinder to monitor the liquid level in the sample pool in real time, ensuring that the focus position is consistent during each measurement, and improving data consistency; the above-mentioned improvement measures are used to improve the detection accuracy and the detection limit of trace pollutants and their detection signals;

[0038] Step 4, adaptive wavelength selection algorithm (AWSSA) optimization step; AWSSA analyzes the characteristic wavelength distribution in the historical data set, establishes a prediction model, and guides the wavelength selection of new samples; the algorithm uses a genetic algorithm (GA) for parameter optimization, with the initial population size set to 100, the number of iterations to 50, the crossover probability to 0.8, and the mutation probability to 0.1; the optimal solution is updated after each iteration, gradually approaching the ideal wavelength combination, significantly improving the detection efficiency;

[0039] Step 5, using principal component analysis (PCA) and wavelet transform (WT) for denoising and smoothing; applying PCA to the original three-dimensional fluorescence data to extract the main eigenvectors, retaining more than 95% of the variance information, and effectively compressing the data dimension; then, using WT to perform multi-scale decomposition of each principal component, selecting the db4 wavelet basis function, the number of decomposition layers is 3, and the threshold is set to the soft thresholding method to achieve denoising and smoothing processing and maintain signal integrity; wherein, PCA and WT denoising processing also includes the application of independent component analysis (ICA) to separate the independent sources in the mixed signal, further improve the data quality, and ensure the accuracy of subsequent analysis.

[0040] Step 6: Construct a multi-scale convolutional neural network (MS-CNN) architecture containing multiple convolutional layers. Each layer uses convolution kernels of different sizes (such as 3×3, 5×5, and 7×7) to extract multi-scale spatial features. The residual connection structure is introduced to avoid the gradient vanishing problem and speed up the training. At the same time, the batch normalization (BN) technology is applied to enhance the model stability. Finally, the final result is output through the Softmax classifier to ensure the prediction accuracy.

[0041] Step 7: The attention mechanism dynamically adjusts the weights of each layer. The attention mechanism is introduced to dynamically adjust the weights of each layer according to the results of each iteration, highlighting important feature areas and suppressing noise interference. The attention score is calculated using a weighted summation method, and the weight is controlled by an adaptive learning rate to ensure that the model can flexibly respond to different types of pollution source characteristics. The attention mechanism also includes the introduction of the Local Sensitive Hashing (LSH) algorithm to accelerate the similarity search process and improve the model training efficiency, which is particularly suitable for rapid response requirements under large-scale data sets.

[0042] Step 8: Improve generalization capabilities. Combined with transfer learning technology, use pre-trained models to initialize network parameters and quickly adapt to new types of pollution. In addition, regularly update the training data set to cover more types of pollutants, further improve the generalization and robustness of the model, and ensure long-term stable performance.

[0043] The method for water pollution source analysis based on adaptive EEM and ACNN-MS described in Example 1 is shown in the attached Figure 1 .

[0044] Embodiment 2:

[0045] The method of water pollution source analysis based on adaptive EEM and ACNN-MS was implemented in an industrial park in Hebei Province. Ten sampling points were set up in the park, and precise three-dimensional fluorescence spectra of water samples were collected regularly for analysis.

[0046] The three-dimensional fluorescence spectrometer has been redesigned and the optical path has been improved. The response value and sensitivity have been significantly improved. The detection results of the effluent of a certain enterprise are shown in the attached figure. Figure 2 As shown: before the improvement, the spectrum was fuzzy, with more noise, weak signal peak and difficult to identify; after the improvement, the signal peak is very obvious, the background is relatively clean, and the noise interference is shielded.

[0047] First, the adaptive wavelength selection algorithm (AWSSA) combined with the genetic algorithm (GA) is used to optimize the wavelength combination and improve the detection efficiency. Then, the principal component analysis (PCA) and wavelet transform (WT) are used to reduce noise and smooth the data, retaining more than 95% of the variance information, and combined with independent component analysis (ICA) to separate mixed signals. Subsequently, a multi-scale convolutional neural network (MS-CNN) is constructed, and convolution kernels of different sizes are used to extract multi-scale features. Residual connection and batch normalization (BN) technology are introduced to enhance the stability of the model. In order to dynamically adjust the feature weights, the attention mechanism is introduced, and the local sensitive hashing (LSH) algorithm is combined to accelerate the similarity search. The correlation between specific pollutant-discharging enterprises and downstream sampling points (including enterprise outlets, sewage treatment plant integrated pools and total outlets) is verified.

[0048]

[0049] As can be seen from the table above, the B8 sewage discharge enterprise has a very high correlation with the total discharge water sample of the sewage treatment plant, indicating that the wastewater of the plant has high toxicity and difficulty in degradation, and remains stable during the sewage treatment process. Therefore, B8 is the main source of pollution in the park. In addition, A3, B5, and B7 also have certain contributions. On the contrary, B12, C1, and C3 contribute less to pollution.

[0050] The above results had poor recognition ability before the use of the adaptive wavelength selection algorithm (AWSSA), and could not distinguish the contribution difference between B8 and A3, B5, and B7. After the use of AWSSA, the traceability accuracy was significantly improved, and the correlation was significantly improved (see Appendix). Figure 3 ).

[0051] The source analysis work guided the corresponding control measures, including limiting the wastewater discharge of the coking plant and strengthening the operation supervision of the wastewater treatment facilities. After the implementation of the control measures, the electrochemical parameters of the wastewater were regularly monitored to evaluate the control effect. The results showed that the disorderly discharge of wastewater in the park was effectively controlled and the water quality was significantly improved.

[0052] Embodiment 3:

[0053] Using the operation flow in Example 1, the source analysis of agricultural pollution in a small watershed in Beijing was carried out. Among the 12 sampling sections in the small watershed, the water sample spectra of the upstream sampling points 1 to 8 have a high inheritance; the correlation of samples 9 to 12 is strong; however, due to the mixed interference of multiple species in the water, the pollutants in samples 9 to 12 are difficult to distinguish from the inheritance in samples 1 to 8. By introducing the improved adaptive EEM and ACNN-MS water pollution source analysis method, the specific traceability comparison results are shown in the following table:

[0054] Sample No. 1 2 3 4 5 6 7 8 Standard Deviation 14.90 13.09 15.19 17.98 20.31 17.37 15.23 18.14 Sample No. 9 10 11 12 13 14 15 16 Standard Deviation 30.15 26.28 30.22 33.08 42.08 68.75 24.08 16.79

[0055] Furthermore, through similarity analysis, it can be inferred that samples 15 and 16, and samples 11 and 12 have more prominent inheritance, and this inheritance relationship comes from samples 2 and 7, proving that there is a high correlation between samples 2→11→15; 7→12→16, that is, the pollution at the downstream points of No. 11 and No. 12 may come from the pollution sources at points 2 and 7.

[0056] The specific implementation methods described above are only used to specifically illustrate the spirit of the present invention, and the protection scope of the present invention is not limited thereto. For those skilled in the art, other implementation methods can certainly be easily made by changing, replacing or modifying the technical contents disclosed in this specification, and these other implementation methods should all be covered within the protection scope of the present invention.

Claims

1. A water pollution source analysis method based on adaptive EEM and ACNN-MS, characterized in that The method specifically comprises the following steps: Step 1: Collection and pretreatment of water samples. Multiple sampling points were set up in suspected contaminated areas to ensure that the samples were representative. Water samples were collected using sterile polyethylene bottles to avoid cross-contamination between samples. Water samples were filtered through a nylon filter membrane with a pore size of 0.45 μm to remove suspended particles and diluted with ultrapure water to a suitable concentration to meet the requirements of three-dimensional fluorescence (EEM) detection. To eliminate the inner filter effect, the samples were pretreated using the ultraviolet absorption correction method to ensure data accuracy. Step 2: Select a dedicated high-sensitivity three-dimensional fluorescence spectrometer that supports an excitation wavelength range of 200 to 400 nm, an emission wavelength range of 250 to 600 nm, and a resolution better than 5 nm. Step 3: Design of high-sensitivity fluorescence spectrometer and strategy for optical path improvement; replace the original detector with a high-quantum-efficiency photomultiplier tube (PMT) for the spectrometer, with a response time requirement of ≤1ns and a dark current of ≤1pA; use a pulsed xenon lamp as the light source with an energy density of ≥1J / cm 2 , ensuring sufficient excitation intensity; introducing a new aspheric lens group in the optical path to reduce the influence of aberration and improve the imaging quality; in addition, adding a built-in temperature control system, requiring the working temperature to be stable within the range of ±0.1℃ to prevent temperature fluctuations from interfering with the measurement accuracy; the above improvement measures are used to improve the detection accuracy and the detection limit of trace pollutants and their detection signals; Step 4, adaptive wavelength selection algorithm (AWSSA) optimization step; AWSSA analyzes the characteristic wavelength distribution in the historical data set, establishes a prediction model, and guides the wavelength selection of new samples; the algorithm uses a genetic algorithm (GA) for parameter optimization, with the initial population size set to 100, the number of iterations to 50, the crossover probability to 0.8, and the mutation probability to 0.1; the optimal solution is updated after each iteration, gradually approaching the ideal wavelength combination, significantly improving the detection efficiency; Step 5, using principal component analysis (PCA) and wavelet transform (WT) for denoising and smoothing; applying PCA to the original three-dimensional fluorescence data to extract the main eigenvectors, retaining more than 95% of the variance information, and effectively compressing the data dimension; then, using WT to perform multi-scale decomposition of each principal component, selecting the db4 wavelet basis function, the number of decomposition layers is 3, and the threshold is set to the soft thresholding method to achieve denoising and smoothing processing and maintain signal integrity; Step 6: Construct a multi-scale convolutional neural network (MS-CNN) architecture containing multiple convolutional layers. Each layer uses convolution kernels of different sizes (such as 3×3, 5×5, and 7×7) to extract multi-scale spatial features. The residual connection structure is introduced to avoid the gradient vanishing problem and speed up the training. At the same time, the batch normalization (BN) technology is applied to enhance the model stability. Finally, the final result is output through the Softmax classifier to ensure the prediction accuracy. Step 7: The attention mechanism dynamically adjusts the weights of each layer. The attention mechanism is introduced to dynamically adjust the weights of each layer according to the results of each iteration, highlighting important feature areas and suppressing noise interference. The attention score is calculated using a weighted summation method, and the weight is controlled by an adaptive learning rate to ensure that the model can flexibly respond to different types of pollution source characteristics. Step 8: Improve generalization capabilities. Combined with transfer learning technology, use pre-trained models to initialize network parameters and quickly adapt to new types of pollution. In addition, regularly update the training data set to cover more types of pollutants, further improve the generalization and robustness of the model, and ensure long-term stable performance.

2. A method for water pollution source analysis based on adaptive EEM and ACNN-MS according to claim 1, characterized in that: The water sample pretreatment described in step 1 also includes adding an appropriate amount of fluorescence quencher, such as ascorbic acid, at a concentration of 0.1 to 1 mM to eliminate background fluorescence interference and improve the signal-to-noise ratio.

3. A method for water pollution source analysis based on adaptive EEM and ACNN-MS according to claim 1, characterized in that: The improved adaptive three-dimensional fluorescence spectroscopy design described in step 3 also involves the introduction of an automatic focusing function, using a laser rangefinder to monitor the liquid level in the sample pool in real time to ensure that the focus position is consistent during each measurement, thereby improving data consistency.

4. A method for water pollution source analysis based on adaptive EEM and ACNN-MS according to claim 1, characterized in that: The PCA and WT noise reduction processing described in step 5 also includes the application of independent component analysis (ICA) to separate the individual independent sources in the mixed signal, further improving the data quality and ensuring the accuracy of subsequent analysis.

5. The method for water pollution source analysis based on adaptive EEM and ACNN-MS according to claim 1, characterized in that: The attention mechanism described in step 7 also includes the introduction of the Locality Sensitive Hashing (LSH) algorithm to accelerate the similarity search process. Improve model training efficiency, especially suitable for rapid response requirements under large-scale data sets.

Images