A deep learning-based watershed non-point source pollution accurate tracing method

Abstract

The application relates to the technical field of watershed water environment pollution tracing, and discloses a watershed non-point source pollution accurate tracing method based on deep learning, which comprises the following steps: a generation and verification double-network collaborative framework comprising a pollution source hypothesis generator and a hypothesis verifier is constructed, and a collaborative optimization mechanism of the pollution source hypothesis generator and the hypothesis verifier is established; generation and probability evaluation of pollution source hypotheses are realized; an evidence chain link tracker is constructed to analyze the relationship between pollution source candidate hypotheses and supporting evidence; the relationship between the pollution source candidate hypotheses and the supporting evidence is further analyzed; analysis results of the relationship between the pollution source candidate hypotheses and the supporting evidence are integrated; and the calculation efficiency and the system performance are optimized; the application provides probability evaluation for each pollution source hypothesis through a Bayesian posterior inference framework, quantizes the uncertainty of the tracing results, and provides reliable reliability indexes for decision making.

Description

Technical Field

[0001] The present invention relates to the technical field of tracing the source of water environment pollution in a river basin, and more specifically, to a method for accurately tracing the source of non-point source pollution in a river basin based on deep learning. Background Art

[0002] With rapid economic development and accelerated urbanization, non-point source pollution (NSP) is becoming increasingly serious in river basins and has become a significant factor affecting water quality. Non-point source pollution, characterized by dispersion, intermittency, and uncertainty, is more difficult to identify and control than point source pollution, posing significant challenges to river basin water environment management.

[0003] Currently, basin non-point source pollution tracing technologies primarily include hydrological model simulation methods, statistical analysis methods, and machine learning methods. Hydrological model simulation methods, such as the SWAT model and the HSPF model, predict the generation and migration of pollutants by constructing physical process models. However, these methods require extensive parameter calibration and have limited model accuracy in complex basin environments. Statistical analysis methods, such as principal component analysis and cluster analysis, can identify the statistical characteristics of pollution sources but have difficulty handling nonlinear relationships and spatiotemporal variations. In recent years, machine learning methods, such as support vector machines and random forests, have been applied to pollution tracing, improving tracing accuracy to a certain extent.

[0004] However, existing technologies still have the following deficiencies in the application of non-point source pollution tracing in watersheds:

[0005] Traditional traceability methods usually adopt deterministic models and can only provide a single traceability result, which cannot fully reflect the complexity and uncertainty of the watershed environment. In the actual watershed environment, there are often multiple potential pollution sources, and the contribution of each pollution source has temporal and spatial variability. A single deterministic result is difficult to fully describe the actual pollution situation; existing methods lack a quantitative evaluation mechanism for the reliability of traceability results. The uncertainty information of traceability results is crucial for formulating effective pollution control strategies, but most traditional methods cannot provide credibility indicators of the results, making it difficult for decision makers to evaluate the reliability of traceability conclusions; there is a lack of transparent correlation mechanism between traceability results and supporting evidence. Traditional methods often regard the traceability process as a "black box" with no evidence to support the results. The method cannot clearly demonstrate the evidentiary basis of the traceability conclusion, which not only affects the interpretability of the results, but also reduces the acceptance of the traceability results in practical applications; the existing methods have limited expressive power when dealing with multi-source and nonlinear pollution problems in complex watershed environments. Non-point source pollution in watersheds often involves the combined effects of multiple pollution source types (such as agricultural non-point sources, urban runoff, livestock and poultry breeding, etc.). Traditional methods find it difficult to simultaneously model the interactive effects of multiple pollution sources; existing methods are more sensitive to noise and outliers in environmental monitoring data. Environmental monitoring data inevitably have data quality problems caused by measurement errors, equipment failures, etc. Traditional methods lack effective anti-noise mechanisms and are easily affected by data quality problems, resulting in unstable traceability results.

[0006] Therefore, there is an urgent need to develop a new watershed non-point source pollution tracing technology that can simultaneously handle issues such as multi-hypothesis generation, credibility assessment, evidence transparency, and noise resistance. Summary of the Invention

[0007] The present invention provides a method for accurately tracing the source of non-point source pollution in a watershed based on deep learning, which solves the technical problems in related technologies that traditional tracing methods only provide a single result, lack credibility quantification and evidence transparency, and have weak noise resistance.

[0008] The present invention provides a method for accurately tracing the source of non-point source pollution in a watershed based on deep learning, comprising:

[0009] Construct a dual-network collaborative framework for pollution source hypothesis generators and hypothesis verifiers, and establish a collaborative optimization mechanism for the pollution source hypothesis generators and hypothesis verifiers;

[0010] Based on a generation-verification dual-network collaborative framework, we use a conditional variational autoencoder and a Bayesian posterior inference framework to generate and probabilistically evaluate pollution source hypotheses. We also construct an evidence link tracker to analyze the relationship between candidate pollution source hypotheses and supporting evidence.

[0011] Based on the generation and probability assessment results of pollution source hypotheses, the relationship between candidate pollution source hypotheses and supporting evidence is further analyzed through the evidence-weighted reasoning module, multi-hypothesis collaborative verification module, and adaptive uncertainty representation technology.

[0012] Integrate the analysis results of the relationship between candidate pollution source hypotheses and supporting evidence to optimize computing efficiency and system performance.

[0013] Furthermore, the steps of constructing the pollution source hypothesis generator include:

[0014] Construct an encoder network that receives watershed observation data and conditional variables and maps the input data to the mean vector and variance vector of the latent space;

[0015] Construct a decoder network that receives latent variables and conditional variables and generates pollution source distribution;

[0016] The encoder network and decoder network are trained using a variational lower bound objective function, which includes reconstruction error and KL divergence.

[0017] Furthermore, the steps of constructing the hypothesis verifier include:

[0018] Construct a pollution source feature extraction sub-network and use a spatial attention convolutional network to extract the spatial distribution characteristics of pollution source hypotheses;

[0019] Construct an observation data feature extraction subnetwork and use a multi-scale temporal convolutional network to extract the spatiotemporal features of the observation data;

[0020] A feature fusion network is constructed, and the correlation between pollution source features and observation data features is calculated through a bidirectional attention mechanism, and the consistency score is output.

[0021] Furthermore, the application steps of the Bayesian posterior inference framework include:

[0022] Calculate the consistency score between each candidate hypothesis of pollution source and the observed data;

[0023] The sum of the consistency scores of all candidate hypotheses of pollution sources is calculated as the normalization factor;

[0024] The consistency score of each pollution source candidate hypothesis is divided by the normalization factor to obtain the posterior probability of the hypothesis.

[0025] Furthermore, the steps of constructing the evidence link tracker include:

[0026] A gradient-based sensitivity analysis method is used to calculate the gradient of the hypothesis validator output with respect to each input feature;

[0027] Quantify the contribution of different observational data to the hypothesis verification results;

[0028] By visualizing the contribution, an evidence support diagram is formed to show the evidentiary basis for the tracing conclusion.

[0029] Furthermore, the steps of constructing the evidence weighted reasoning module include:

[0030] Construct an attention-based evidence quality assessment network to learn to assign attention weights to different observations;

[0031] Build a data feature extractor to extract the feature representation of each observed data point;

[0032] Construct a quality assessment subnetwork to evaluate the quality and consistency of each observation point by calculating the relationship between each observation point and all other observation points;

[0033] Construct a weight normalization layer and convert the initial weights into an attention distribution with a sum of 1 through the softmax function.

[0034] Furthermore, the application steps of the adaptive uncertainty representation technology include:

[0035] Construct a hierarchical uncertainty estimation network to simultaneously output the predicted value and uncertainty estimate of pollution source distribution;

[0036] Modeling parameter distributions through Bayesian neural network design;

[0037] Combined with Monte Carlo sampling, estimate the statistical properties of the prediction distribution;

[0038] Distinguish between aleatory uncertainty and epistemic uncertainty, and provide comprehensive uncertainty information.

[0039] Furthermore, the step of optimizing the computational efficiency includes:

[0040] Apply hypothesis clustering and representative sampling techniques to select a representative subset of the generated pollution source hypotheses for in-depth verification;

[0041] Evaluate the importance of evidence by analyzing the contribution of different observation data to the traceability results and identifying key evidence and redundant evidence;

[0042] Construct a multi-stage reasoning strategy to decompose the complex tracing problem into simplified stages and gradually refine the reasoning process.

[0043] Furthermore, the steps for implementing the evidence importance assessment include:

[0044] Using mutual information theory, calculate the mutual information between each observation data and the pollution source hypothesis;

[0045] Quantify the information value of data;

[0046] The final monitoring points and parameters are determined based on the mutual information value, and an efficient data collection plan is formulated.

[0047] The present invention provides a computer storage medium comprising a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the above-mentioned method for precise tracing of non-point source pollution in a watershed based on deep learning.

[0048] The beneficial effects of the present invention are as follows: compared with the traditional single-result tracing method, the present invention can simultaneously consider multiple possible pollution source hypotheses and provide more comprehensive tracing results, which is particularly suitable for tracing non-point source pollution in complex watershed environments;

[0049] Through the Bayesian posterior inference framework, the present invention can provide a probability assessment for each pollution source hypothesis, quantify the uncertainty of the traceability results, and provide a reliable confidence indicator for decision-making;

[0050] The evidence link tracker makes the relationship between the traceability conclusion and supporting evidence transparent and traceable, enhancing the interpretability and credibility of the results;

[0051] The evidence-weighted reasoning module can automatically identify and highlight key evidence, reduce the impact of noisy data, and improve the stability of the system in complex environments;

[0052] Through hypothesis clustering, evidence importance evaluation and multi-stage reasoning strategy, the present invention achieves high computational efficiency and can meet the pollution source tracing needs of large-scale watersheds;

[0053] Through multi-hypothesis probability distribution and uncertainty assessment, comprehensive information support is provided for river basin pollution control and governance decisions, effectively reducing decision-making risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a flow chart of a method for accurately tracing the source of non-point source pollution in a watershed based on deep learning in the present invention;

[0055] Figure 2 is a line graph showing the change in validator consistency scores over time for different pollution source hypotheses;

[0056] Figure 3 is a scatter plot of the relationship between the epistemic uncertainty estimated by the adaptive uncertainty representation module and the traceability accuracy;

[0057] Figure 4 It is a bar chart showing the relative performance of different computational efficiency optimization techniques and traditional methods in terms of computational time and data requirements;

[0058] Figure 5It is to generate a grouped bar chart to verify the performance differences between the dual-network collaborative framework and the traditional pollution source tracing method on five key indicators;

[0059] Figure 6 It is a radar chart that generates and verifies the comparison between the dual-network collaborative framework and the traditional pollution source tracing method in five key performance indicators. DETAILED DESCRIPTION

[0060] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0061] At least one embodiment of the present invention discloses a method for accurately tracing the source of non-point source pollution in a watershed based on deep learning, such as Figure 1 Shown, including:

[0062] Step 1: Construct a dual-network collaborative framework for pollution source hypothesis generator and hypothesis verifier, and establish a collaborative optimization mechanism for the pollution source hypothesis generator and hypothesis verifier.

[0063] As the foundation of this method, this step constructs the Generate-Verify Dual-Network Collaborative Framework (GVD-TPS), laying the foundation for subsequent multi-hypothesis generation, probabilistic assessment, and uncertainty quantification. This framework comprises two core components: a pollution source hypothesis generator and a hypothesis verifier, which implement the multi-hypothesis generation and credibility assessment process. It should be understood that this collaborative framework provides the system with the foundation to simultaneously consider the possibility of multiple pollution sources. Specifically, it includes the following sub-steps:

[0064] Step 1.1, construct the pollution source hypothesis generator network;

[0065] According to one embodiment of the present application, the pollution source hypothesis generator uses a conditional variational autoencoder structure to generate multiple possible pollution source distributions based on observed data. The specific structure is as follows:

[0066] The encoder part receives the watershed observation data (including water quality parameters, meteorological conditions, terrain information, etc.) and conditional variables (including time, hydrological conditions, etc.), through multiple layers of convolution and fully connected layers, the input data is mapped to the mean vector of the latent space and variance vector Latent variables From the normal distribution via the reparameterization technique The decoder receives the latent variables and condition variables , through transposed convolution and fully connected layers, generate possible pollution source distribution .

[0067] The specific structure of the conditional variational autoencoder includes: the input layer receives the dimension The observation data ( is the number of monitoring parameters, and denotes the height and width dimensions of the spatial grid respectively) and the dimensions are The conditional vector of ;

[0068] The encoder consists of 4 residual convolution blocks, each of which consists of two layers It consists of a convolutional layer, a batch normalization layer, and a ReLU activation function, with the number of channels being 32, 64, 128, and 256, respectively;

[0069] The fully connected layer converts the feature map into a mean vector of two latent spaces of length 128. and the logarithm of the latent space variance The decoder contains 4 transposed convolution blocks, each of which doubles the size of the feature map. The number of channels is 128, 64, 32, .

[0070] In specific applications, this conditional variational autoencoder can generate a distribution of possible pollution sources based on different hydrological conditions (such as wet / dry seasons) as conditional variables. For example, in a case study on nitrate pollution source tracing during the wet season of a river basin, the model, by inputting water quality data from 24 monitoring points and the conditional encoding of "wet season," successfully generated 10 possible nitrate pollution source distributions, covering different combinations of the three main pollution sources: agricultural, urban, and industrial areas.

[0071] Alternatively, the pollution source hypothesis generator can employ other generative model structures, such as generative adversarial networks (GANs) or diffusion models. In some implementations, the most appropriate generator structure can be selected based on the characteristics of a specific watershed. For example, for small watersheds with limited data, a simplified variational autoencoder with fewer parameters can be used; whereas for large watersheds with complex topography, a more robust diffusion model can be used as the generator.

[0072] The network training adopts a variational lower bound objective function, which includes reconstruction error and KL divergence. It encourages the model to make the latent space distribution close to the standard normal distribution while ensuring the reconstruction quality, thereby supporting diversified sampling.

[0073] Specifically, the calculation process of the variational lower bound objective function (ELBO, Evidence Lower BOund) is as follows:

[0074] First, the reconstruction error is calculated, that is, the difference between the predicted pollution source distribution and the true pollution source distribution, usually using mean square error or cross entropy loss;

[0075] Then the KL divergence between the latent variable distribution and the standard normal distribution is calculated to quantify the degree of difference between the two probability distributions;

[0076] Finally, the two parts are weighted and combined to obtain the final loss function. The objective function can be expressed as:

[0077] ;

[0078] in, represents the variational lower bound objective function; Indicates that the distribution Lower pair expected value; Represents a given latent variable Time pollution source distribution The conditional probability of represents the KL divergence operator; represents the posterior distribution learned by the encoder; represents the standard normal prior distribution; The separator representing the KL divergence.

[0079] Step 1.2, build the hypothesis verifier network;

[0080] According to one embodiment of the present application, the hypothesis verifier network is used to evaluate the consistency of each pollution source hypothesis with the observed data, and adopts a multimodal feature fusion structure. and observational data As input, the two are feature extracted and fused to calculate the consistency score .

[0081] The multimodal feature fusion architecture consists of three main components: a pollution source feature extraction subnetwork, an observation data feature extraction subnetwork, and a feature fusion network. The pollution source feature extraction subnetwork uses a spatial attention convolutional network, which contains three spatial attention convolutional blocks. Each block consists of a spatial self-attention layer and two standard convolutional layers. The spatial self-attention layer calculates the correlation strength between different locations on the feature map, enabling the network to capture long-range spatial dependencies.

[0082] The observation data feature extraction sub-network adopts a multi-scale temporal convolutional network, which contains three parallel kernels with different kernel sizes ( 、 、 ) to capture spatiotemporal patterns at different scales.

[0083] The feature fusion network first calculates the attention matrix between the two sets of features through a bidirectional attention mechanism, then undergoes cross-fusion and adaptive gating, and finally outputs the consistency score through a fully connected layer.

[0084] In some implementations, the hypothesis validator network can be augmented with a time series processing module, such as a Long Short-Term Memory (LSTM) or Gated Recurrent Unit (GRU) to process time series data, thereby better capturing the temporal characteristics of pollutant migration and diffusion. Furthermore, based on the characteristics of different watersheds, the validator network can also integrate Geographic Information System (GIS) data, processing geographic information such as topography and land use types through additional feature extraction branches to further improve verification accuracy.

[0085] In practical applications, the validator network can effectively distinguish between reasonable and unreasonable pollution source hypotheses. For example, in a case study of phosphorus pollution in an agricultural watershed, when the input was a hypothesis of upstream farmland pollution source consistent with the observed data, the validator gave a high consistency score of 0.89. However, when the input was a hypothesis of downstream industrial pollution source inconsistent with the observed data, the validator gave a low consistency score of 0.21, demonstrating its effectiveness.

[0086] like Figure 2 As shown in the figure, a multi-series line chart of the time-varying consistency scores for different pollution source hypotheses visually illustrates the hypothesis verifier's evaluation results for three different pollution source hypotheses (upstream farmland, downstream industrial zone, and urban area). This chart verifies that the hypothesis verifier can effectively distinguish between reasonable and unreasonable pollution source hypotheses over time, providing a reliable assessment basis for pollution source tracing.

[0087] The feature extraction part uses sub-networks specialized for different types of data: for the pollution source hypothesis , use the spatial attention convolutional network to extract its spatial distribution features; for the observation data , uses a multi-scale temporal convolutional network to extract spatiotemporal features. The fusion layer uses a bidirectional attention mechanism to calculate the correlation between pollution source features and observation data features, capture the degree of consistency between the two, and finally output a consistency score through a fully connected layer.

[0088] Step 1.3: Establish a collaborative optimization mechanism between the pollution source hypothesis generator and the hypothesis verifier;

[0089] In order to achieve the coordinated optimization of the pollution source hypothesis generator and the hypothesis verifier, this application constructs an adversarial learning mechanism. The goal is to generate a hypothesis that can be verified Evaluate high-quality pollution source hypotheses, while hypothesis validators The goal is to accurately distinguish between real pollution sources and generated pollution source hypotheses.

[0090] The loss function of collaborative optimization consists of two parts:

[0091] Pollution source hypothesis generator loss The pollution source hypothesis generator is encouraged to produce hypotheses that can achieve high verification scores;

[0092] Assuming validator loss The hypothesis validator is encouraged to give high scores to real pollution sources and low scores to generated fake pollution sources.

[0093] By alternately optimizing the two networks, the generation and verification capabilities can be improved simultaneously.

[0094] Optionally, in some implementations, the collaborative optimization mechanism can employ different training strategies. For example, a cyclic consistency loss can be introduced to require that the pollution source distribution reconstructed by the pollution source hypothesis generator remain consistent with the original distribution. Alternatively, a progressive training strategy can be employed, first training a low-resolution model and then gradually increasing the resolution to improve model performance and stability. Furthermore, knowledge distillation techniques can be employed to use pre-trained hydrodynamic model knowledge to guide the training of the pollution source hypothesis generator and hypothesis verifier.

[0095] Step 2: Based on the generation-verification dual-network collaborative framework, we use the conditional variational autoencoder and Bayesian posterior inference framework to generate and probabilistically evaluate pollution source hypotheses, and construct an evidence link tracker to analyze the relationship between candidate pollution source hypotheses and supporting evidence.

[0096] Based on the generation-verification dual-network collaborative framework constructed in step 1, this step utilizes the trained framework to generate and probabilistically evaluate multiple possible pollution source hypotheses. By fully leveraging the generation capabilities of the pollution source hypothesis generator and the evaluation capabilities of the hypothesis verifier in step 1, this step can generate a variety of pollution source candidate hypotheses and conduct reliable probabilistic evaluations on them. It includes the following sub-steps:

[0097] Step 2.1, use the conditional variational autoencoder to generate multiple pollution source hypotheses;

[0098] Based on the trained conditional variational autoencoder, a set of candidate hypotheses of pollution sources is generated by sampling multiple times in the latent space:

[0099] ;

[0100] in, represents the set of candidate hypothesis of pollution sources; Indicates the A generated pollution source hypothesis; Represents the number of pollution source hypotheses generated.

[0101] By adjusting the sampling strategy, the generated hypothesis set is ensured to be diverse and representative, and to cover the possible pollution source solution space.

[0102] In practice, from the prior distribution mid-sampling Condition variables:

[0103] ;

[0104] in, represents the prior distribution of the conditional variable; Indicates the Condition variables; represents the total number of conditional variables sampled;

[0105] Combined with observational data , through the generator generate Pollution source assumptions:

[0106] ;

[0107] in Indicates that the parameter is The generator function of Represents the parameters of the generator; Indicates the A generated pollution source hypothesis; represents observation data; Indicates the Condition variables; Represents the number of pollution source hypotheses generated.

[0108] At the same time, clustering techniques such as kernel density estimation are applied to perform redundant filtering and representative sampling on the generated hypotheses to reduce the computational cost of the verification stage.

[0109] Generator Function Indicates that the parameter is The mapping function of the conditional variational self-encoder generator is as follows: First, the observation data The mean is obtained by mapping the encoder to the latent space and variance ; Then use the reparameterization technique to get the distribution Sampling to obtain latent variables ; Finally, the latent variable and condition variables Common input decoder to generate the corresponding pollution source distribution This process can be expressed as:

[0110] ;

[0111] ;

[0112] in Indicates that the parameter is The generator function of Represents the generator parameters; represents observation data; Indicates the Condition variables; represents a normal distribution; represents the latent variable; Indicates the A generated pollution source hypothesis; Indicates that the parameter is decoder; A function representing the mean of the encoder output; A function representing the variance of the encoder output; represents sampling from a distribution.

[0113] The entire generation process is to transform the observation data and condition variables The process of mapping to the pollution source distribution space.

[0114] Step 2.2, apply the Bayesian posterior inference framework to calculate the hypothesis probability distribution;

[0115] The generated pollution source hypothesis set is input into the hypothesis verifier network, and the consistency score of each hypothesis with the observed data is calculated:

[0116] ;

[0117] in, Indicates that the parameter is The hypothesis validator function of Represents the parameters of the hypothesis validator; Indicates the A generated pollution source hypothesis; Represents observation data.

[0118] Based on these consistency scores, we apply the Bayesian posterior inference framework to calculate the probability distribution of each hypothesis:

[0119] The Bayesian posterior inference framework is applied to determine the probability of each pollution source hypothesis. First, the consistency score of each hypothesis with the observed data is calculated, and then the consistency score is divided by the sum of the consistency scores of all hypotheses to obtain the normalized probability value of the hypothesis. Specifically, for the first The posterior probability calculation process of a pollution source hypothesis is as follows: first, the consistency score between the hypothesis and the observed data is calculated through the hypothesis verification network, which reflects the degree of match between the hypothesis and the actual observed data; then, all The sum of the consistency scores of the hypotheses is used as a normalization factor; finally, the The consistency score of each hypothesis is divided by the normalization factor to obtain the posterior probability of the hypothesis. This calculation method ensures that the sum of all hypothesis probabilities is 1, forming a complete probability distribution, so that the hypotheses can be quantitatively compared.

[0120] Hypothesis Validator Function Indicates that the parameter is The evaluation function of the hypothesis verifier network is used to measure the pollution source hypothesis With observational data The specific implementation of this function includes: first extracting the pollution source hypothesis through the spatial attention convolutional network Feature representation ; At the same time, the observation data is extracted through the multi-scale temporal convolutional network Feature representation ; Then use the bidirectional attention mechanism to calculate and The attention weight matrix between ; Then based on the attention weight matrix Perform feature fusion to obtain fusion features ; Finally, through the three-layer fully connected network Mapped to a single consistency score, ranging from The larger the value, the more consistent the hypothesis is with the observed data. The entire assessment process is to quantify the degree of match between the pollution source hypothesis and the actual observed data.

[0121] This probability distribution reflects the relative likelihood of different pollution source hypotheses, providing a probabilistic basis for subsequent decision-making. For hypotheses with higher probabilities, their spatial distribution and temporal evolution characteristics can be further refined and analyzed.

[0122] Step 2.3: Construct an evidence link tracker to analyze the evidence support relationship;

[0123] The evidence link tracker is used to analyze the relationship between each pollution source hypothesis and supporting evidence to improve the interpretability of the traceability results. , analyze the strength of its correlation with various observational data, and identify the key evidence that contributes most to the hypothesis.

[0124] In the implementation, a gradient-based sensitivity analysis method is used to calculate the gradient of the hypothesis verifier output with respect to each input feature, quantifying the contribution of different observations to the hypothesis verification results. By visualizing these contributions, an evidence support diagram is formed, which intuitively displays the evidentiary basis for the traceability conclusion.

[0125] Step 3: Based on the generation and probability assessment results of pollution source hypotheses, the relationship between candidate pollution source hypotheses and supporting evidence is further analyzed through the evidence weighted reasoning module, multi-hypothesis collaborative verification module, and adaptive uncertainty representation technology;

[0126] Based on the multiple hypothesis generation and probability assessment in step 2, this step further quantifies the credibility and uncertainty of the traceability results. By conducting in-depth analysis of the multiple pollution source hypotheses and their probability distributions generated in step 2, this step can provide more comprehensive credibility information and uncertainty representation, providing a reliable basis for decision-making. It includes the following sub-steps:

[0127] Step 3.1: Construct an evidence-weighted reasoning module to identify key evidence;

[0128] According to one embodiment of the present application, the evidence-weighted reasoning module analyzes the quality and reliability of different evidence in the observation data and assigns different weights to them, thereby improving the system's sensitivity to high-quality data and reducing the impact of noisy data.

[0129] In terms of implementation, we build an evidence quality assessment network based on the attention mechanism, which learns to assign attention weights to different observation data. These weights are used in the verifier network to weightedly fuse different evidences. Prior information on data quality, such as sensor reliability and data collection conditions, is incorporated into the weight learning process to further improve the rationality of weight distribution.

[0130] The evidence quality assessment network employs a self-attention architecture and consists of three main components: a data feature extractor, a quality assessment subnetwork, and a weight normalization layer. The data feature extractor utilizes a three-layer fully connected network (256-128-64 nodes) to extract feature representations for each observation. The quality assessment subnetwork employs a self-attention mechanism to assess the quality and consistency of each observation by calculating its relationship with all other observations, outputting initial weights. The weight normalization layer converts the initial weights into an attention distribution summing to 1 using a softmax function. Furthermore, the network incorporates sensor metadata features (such as device type, calibration date, and data collection environment). A parallel feature branch extracts metadata features and fuses them with the observed data features for a more accurate data quality assessment.

[0131] In the above process, weight normalization uses the exponential normalization method to convert the initial weight values ​​into a probability distribution form with a sum of 1. The specific calculation process is as follows: First, a natural exponential transformation is performed on the initial weight value of each observation point, mapping the weight value from the real number domain to the positive real number domain, ensuring that all converted weights are positive; then, the sum of the weights of all observation points after the exponential transformation is calculated as the normalization factor; finally, the exponential transformation weight of each observation point is divided by the normalization factor to obtain the final attention weight of the observation point. This normalization method has two important characteristics: first, the relative size relationship between weights is maintained, and observation points with larger weights still have larger weight values ​​after normalization; second, the sum of the generated weights is 1, which is convenient for direct use in the subsequent weighted fusion process without the need for additional normalization.

[0132] In some implementations, the evidence-weighted reasoning module can also incorporate Bayesian networks to establish a causal relationship model between evidence, further improving the ability to identify key evidence. Alternatively, the module can also employ reinforcement learning methods, treating the allocation of evidence weights as a strategy to be optimized. By continuously optimizing the weight allocation strategy through interaction with the environment, the system's adaptability can be enhanced.

[0133] In practical applications, the evidence quality assessment network can effectively identify and reduce the impact of outliers. For example, in a study of copper pollution source tracing in a certain watershed, the system automatically assigned low weights (0.03-0.05) to monitoring points affected by heavy rain, while assigning high weights (0.15-0.20) to stable and reliable monitoring points. Ultimately, the tracing accuracy increased by 26%, demonstrating the effectiveness of the module.

[0134] Step 3.2: Implement multi-hypothesis collaborative verification to improve overall traceability reliability;

[0135] The multi-hypothesis collaborative verification module further improves the reliability of traceability results by analyzing the relationships between multiple hypotheses. Based on the concept of ensemble learning, this module treats multiple hypotheses as "experts" and improves the reliability of conclusions through their collective judgment.

[0136] In the implementation, we construct a hypothesis relationship diagram and analyze the similarities and complementarities between hypotheses. For highly similar hypothesis groups, we analyze their common characteristics as strong evidence for tracing the source. For significantly different hypotheses, we analyze the differences in their applicable conditions to gain a deeper understanding of the impact of environmental changes on tracing results.

[0137] Step 3.3, apply adaptive uncertainty representation techniques to quantify the confidence of inference;

[0138] The Adaptive Uncertainty Representation module dynamically adjusts uncertainty estimates based on the strength of evidence, providing an uncertainty representation that matches the actual credibility. This module distinguishes between aleatoric uncertainty (caused by randomness) and epistemic uncertainty (caused by incomplete knowledge), providing comprehensive uncertainty information for decision-making.

[0139] In practice, a hierarchical uncertainty estimation network is constructed to simultaneously output predicted values ​​and uncertainty estimates for pollution source distributions. This network uses a Bayesian neural network design to model parameter distributions and, combined with Monte Carlo sampling, estimates the statistical properties of the predicted distributions, including aleatory uncertainty and epistemic uncertainty.

[0140] The core of the hierarchical uncertainty estimation network is the Bayesian neural network structure. The weights in the network are no longer deterministic values, but probabilistic distributions. The specific implementation uses three layers of Bayesian fully connected layers (512-256-128 nodes), and the weight parameters of each layer are parameterized by mean and variance. The variational inference method is used to approximate the posterior distribution during training, and Monte Carlo sampling (usually The network output consists of three components: the predicted mean (distribution of the most likely pollution sources), aleatoric uncertainty (uncertainty caused by the inherent randomness of the data) and epistemic uncertainty (Uncertainty caused by lack of model knowledge).

[0141] During the prediction process, the system uses multiple Monte Carlo sampling cycles to obtain stable and reliable estimates. First, the system randomly generates multiple sets of network parameters, each corresponding to a possible configuration of the Bayesian neural network. Then, for the input observation data, the system uses each set of network parameters to make predictions, generating multiple predictions. The system then calculates the average of all these predictions as the final predicted mean, which represents the most likely distribution of pollution sources.

[0142] To calculate aleatory uncertainty, the system first obtains the uncertainty estimates directly output from each prediction, which reflect the inherent randomness of the data itself; then, the system calculates the weighted average of these uncertainty estimates under all sampling times to obtain the final aleatory uncertainty index, which represents the degree of prediction variability caused by the randomness of the data itself even if the model is completely correct.

[0143] To calculate epistemic uncertainty, the system evaluates the dispersion of predictions under different parameter configurations. Specifically, the system first calculates the difference between each prediction and the overall prediction mean; then squares these differences to amplify the degree of deviation; finally, it calculates the average of these squared differences across all sampling times to obtain the final epistemic uncertainty index, which reflects the prediction uncertainty caused by the model's insufficient understanding of the current input data. When the input data is similar to the data seen during model training, the prediction results under different parameter configurations will be more consistent, resulting in lower epistemic uncertainty; however, when the input data deviates from the training distribution, the prediction results under different parameter configurations will vary significantly, resulting in higher epistemic uncertainty, reminding decision makers to view the prediction results with caution.

[0144] The unique feature of the hierarchical uncertainty estimation network is its adaptability: when the input data distribution is similar to the training set, the network outputs low epistemic uncertainty and more accurate aleatoric uncertainty. However, when the input data distribution deviates from the training set, the network outputs high epistemic uncertainty, reminding decision makers that the results are of limited credibility. For example, in a case study of heavy metal pollution source tracing in a watershed surrounding a city, the system provided reliable tracing results with low epistemic uncertainty (0.05-0.10) for typical areas with sufficient data. However, for newly developed areas with sparse data, the system issued high epistemic uncertainty (0.40-0.60), indicating the need for increased monitoring efforts in that area.

[0145] like Figure 3 The scatter plot of epistemic uncertainty and traceability accuracy clearly demonstrates the relationship between epistemic uncertainty estimated by the Adaptive Uncertainty Representation module and traceability accuracy. The data was divided into two groups: data-rich and data-sparse regions. The plot verifies that the module can provide reasonable uncertainty estimates based on data quality. In data-rich regions, epistemic uncertainty is low and negatively correlated with traceability accuracy. In data-sparse regions, epistemic uncertainty is generally high, reminding decision makers to use traceability results in such areas with caution, and additional monitoring data may be needed.

[0146] Step 4: Integrate the analysis results of the relationship between the candidate hypothesis of pollution sources and supporting evidence to optimize computing efficiency and system performance;

[0147] After completing the basic functions of steps 1 to 3, such as Figure 4 The performance comparison chart for computational efficiency optimization techniques shows the relative performance of different computational efficiency optimization techniques (hypothesis clustering and representative sampling, evidence importance assessment, and multi-stage inference strategies) compared to traditional methods in terms of computational time and data requirements. This chart intuitively demonstrates that the optimization techniques used in this implementation can reduce computational resource requirements, making it possible for large-scale watershed applications.

[0148] This step uses a variety of optimization techniques to improve the system's computational efficiency and overall performance. To address the computational burden and data redundancy issues that may be encountered in the previous steps, this step provides a series of optimization methods to enable the system to efficiently handle complex pollution source tracing problems in large-scale watersheds. It includes the following sub-steps:

[0149] Step 4.1, apply hypothesis clustering and representative sampling techniques to improve computational efficiency;

[0150] Through hypothesis clustering and representative sampling techniques, a representative subset is selected from the numerous generated pollution source hypotheses for in-depth verification, reducing the computational burden.

[0151] In the implementation, the generated hypotheses are first clustered in the latent space to identify the main hypothesis groups. Then, representative samples are selected from each group to ensure that the selected sample set covers the main area of ​​the hypothesis space while remaining small in size.

[0152] Step 4.2: Implement evidence importance assessment to reduce redundant data collection;

[0153] The evidence importance assessment module analyzes the contribution of different observation data to the traceability results, identifies key evidence and redundant evidence, optimizes data collection strategies, and reduces unnecessary monitoring costs.

[0154] This module uses mutual information theory to calculate the mutual information between each observation and the pollution source hypothesis, quantifying the information value of the data. Based on this, the most valuable monitoring points and parameters can be determined, and efficient data collection plans can be developed to reduce redundant data collection while maintaining traceability.

[0155] During the mutual information calculation process, the system quantifies the information value by evaluating the statistical dependence between the observations and the pollution source hypothesis. The specific process involves three main steps: First, the system estimates the joint probability distribution of the observations and the pollution source hypothesis, which describes the probability of their simultaneous occurrence. Simultaneously, the marginal probability distributions of the observations and the pollution source hypothesis are estimated separately, representing the probability of their individual occurrences. Next, the system calculates the ratio of the joint probability distribution to the product of the marginal probability distributions. This ratio reflects the degree of mutual dependence between the observations and the pollution source hypothesis, with higher ratios indicating stronger dependence. Finally, the system takes the logarithm of this ratio and performs a weighted average based on the joint probability distribution to obtain the final mutual information value. Higher mutual information values ​​indicate a greater contribution of the observation to the identification of the pollution source, higher information value, and higher priority for collection. Observations with low mutual information values ​​may provide redundant or irrelevant information, and their sampling priority or frequency can be reduced. By setting an appropriate mutual information threshold, the system can scientifically distinguish between monitoring data with high and low information value, thereby optimizing the overall data collection strategy.

[0156] Step 4.3: Construct a multi-stage reasoning strategy to reduce computational complexity;

[0157] The multi-stage reasoning strategy decomposes the complex tracing problem into multiple simplified stages, gradually refining the reasoning process and reducing the overall computational complexity.

[0158] In the first stage, a simplified model is used to quickly screen low-probability areas, narrowing the search scope. In the second stage, a refined model is applied to high-probability areas to obtain more accurate source tracing results. In the final stage, a detailed analysis of the most likely hypotheses is conducted, including uncertainty assessment and evidence tracking. This multi-stage strategy reduces computational complexity and enables the system to efficiently handle complex pollution source tracing problems in large-scale watersheds.

[0159] like Figure 5 As shown, the performance comparison grouped bar chart of the generation and verification dual network collaborative framework and the traditional method compares the performance differences between the GVD-TPS framework of this embodiment and the traditional pollution source tracing method in five key indicators: multi-hypothesis generation capability, uncertainty quantification, evidence transparency, noise resistance and computational efficiency, clearly demonstrating the comprehensive advantages of this patented technical solution.

[0160] like Figure 6 As shown in the figure, the performance radar chart of the generation and verification dual-network collaborative framework and the traditional method comprehensively demonstrates the comparison between the GVD-TPS framework and the traditional pollution source tracing method in five key performance indicators in the form of a radar chart, which intuitively reflects the comprehensive technical advantages of this patented technical solution, especially its outstanding performance in multi-hypothesis generation capability and evidence transparency.

[0161] A computer storage medium includes a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the above-mentioned method for accurately tracing the source of non-point source pollution in a watershed based on deep learning.

[0162] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A method for accurately tracing the source of non-point source pollution in a watershed based on deep learning, characterized by: include: Construct a dual-network collaborative framework for pollution source hypothesis generators and hypothesis verifiers, and establish a collaborative optimization mechanism for the pollution source hypothesis generators and hypothesis verifiers; The steps to build the pollution source hypothesis generator include: Construct an encoder network that receives watershed observation data and conditional variables and maps the input data to the mean vector and variance vector of the latent space; Construct a decoder network that receives latent variables and conditional variables and generates pollution source distribution; The encoder network and decoder network are trained using a variational lower bound objective function, which includes reconstruction error and KL divergence; The steps to build a hypothesis validator include: Construct a pollution source feature extraction sub-network and use a spatial attention convolutional network to extract the spatial distribution characteristics of pollution source hypotheses; Construct an observation data feature extraction subnetwork and use a multi-scale temporal convolutional network to extract the spatiotemporal features of the observation data; Construct a feature fusion network, calculate the correlation between pollution source features and observation data features through a bidirectional attention mechanism, and output a consistency score; Based on a generation-verification dual-network collaborative framework, we use a conditional variational autoencoder and a Bayesian posterior inference framework to generate and probabilistically evaluate pollution source hypotheses. We also construct an evidence link tracker to analyze the relationship between candidate pollution source hypotheses and supporting evidence. Based on the generation and probability assessment results of pollution source hypotheses, the relationship between candidate pollution source hypotheses and supporting evidence is further analyzed through the evidence-weighted reasoning module, multi-hypothesis collaborative verification module, and adaptive uncertainty representation technology. Integrate the analysis results of the relationship between candidate pollution source hypotheses and supporting evidence to optimize computing efficiency and system performance.

2. The method for accurately tracing the source of non-point source pollution in a watershed based on deep learning according to claim 1 is characterized in that: The application steps of the Bayesian posterior inference framework include: Calculate the consistency score between each candidate hypothesis of pollution source and the observed data; The sum of the consistency scores of all candidate hypotheses of pollution sources is calculated as the normalization factor; The consistency score of each pollution source candidate hypothesis is divided by the normalization factor to obtain the posterior probability of the hypothesis.

3. The method for accurately tracing the source of non-point source pollution in a watershed based on deep learning according to claim 1 is characterized in that: The steps for constructing the evidence chain tracker include: A gradient-based sensitivity analysis method is used to calculate the gradient of the hypothesis validator output with respect to each input feature; Quantify the contribution of different observational data to the hypothesis verification results; By visualizing the contribution, an evidence support diagram is formed to show the evidentiary basis for the tracing conclusion.

4. The method for accurately tracing the source of non-point source pollution in a watershed based on deep learning according to claim 1 is characterized in that: The steps of constructing the evidence weighted reasoning module include: Construct an attention-based evidence quality assessment network to learn to assign attention weights to different observations; Build a data feature extractor to extract the feature representation of each observed data point; Construct a quality assessment subnetwork to evaluate the quality and consistency of each observation point by calculating the relationship between each observation point and all other observation points; Construct a weight normalization layer and convert the initial weights into an attention distribution with a sum of 1 through the softmax function.

5. The method for accurately tracing the source of non-point source pollution in a watershed based on deep learning according to claim 1 is characterized in that: The application steps of the adaptive uncertainty representation technology include: Construct a hierarchical uncertainty estimation network to simultaneously output the predicted value and uncertainty estimate of pollution source distribution; Modeling parameter distributions through Bayesian neural network design; Combined with Monte Carlo sampling, estimate the statistical properties of the prediction distribution; Distinguish between aleatory uncertainty and epistemic uncertainty, and provide comprehensive uncertainty information.

6. The method for accurately tracing the source of non-point source pollution in a watershed based on deep learning according to claim 1 is characterized in that: The step of optimizing computational efficiency comprises: Apply hypothesis clustering and representative sampling techniques to select a representative subset of the generated pollution source hypotheses for in-depth verification; Evaluate the importance of evidence by analyzing the contribution of different observation data to the traceability results and identifying key evidence and redundant evidence; Construct a multi-stage reasoning strategy to decompose the complex tracing problem into simplified stages and gradually refine the reasoning process.

7. The method for accurately tracing the source of non-point source pollution in a watershed based on deep learning according to claim 6 is characterized in that: The steps for implementing the evidence importance assessment include: Using mutual information theory, calculate the mutual information between each observation data and the pollution source hypothesis; Quantify the information value of data; The final monitoring points and parameters are determined based on the mutual information value, and an efficient data collection plan is formulated.

8. A computer storage medium, characterized in that It includes a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, it is used to implement a method for accurately tracing the source of non-point source pollution in a watershed based on deep learning as described in any one of claims 1 to 7.

Images