Intelligent monitoring and early warning system and method for agricultural non-point source pollution

Abstract

The invention discloses an intelligent monitoring and early warning system and method for agricultural non-point source pollution, and relates to the technical field of environmental monitoring, accurate prediction of water pollutant concentration is realized through multi-source heterogeneous data acquisition and fusion, a dynamic attention mechanism and a PINN-Transformer coupling model, the system extracts data spatial and temporal characteristics by using an optimized Transformer model, and the method is applied to the intelligent monitoring and early warning of agricultural non-point source pollution. A water pollution diffusion physical constraint is embedded, it is ensured that a prediction result conforms to an actual hydrodynamic law, and based on high-precision spatial-temporal distribution data, an intelligent algorithm is adopted to track a pollution diffusion path and rapidly lock a pollution source; meanwhile, the cellular automaton model simulates pollution risk dynamic diffusion and assists regional risk assessment, the system also combines a block chain technology to carry out credible evidence storage on key monitoring data, and real-time data processing and early warning pushing are realized through a cloud edge collaborative architecture. And an efficient and reliable technical solution is provided for agricultural water environment management and pollution prevention and control.

Description

Technical Field

[0001] The present invention relates to the technical field of environmental monitoring, and in particular to an intelligent monitoring and early warning system and method for agricultural non-point source pollution. Background Technique

[0002] Currently, the monitoring of agricultural non-point source pollution mainly uses methods such as chemical analysis, isotope tracing, and remote sensing monitoring. Chinese invention patent CN115130903B discloses a method for determining the hydrological path and source contribution of agricultural non-point source pollution in a basin. This method identifies the hydraulic connectivity between groundwater and river runoff by determining water quality indicators (such as nitrate nitrogen) and isotope indicators (hydrogen and oxygen isotopes, nitrogen and oxygen isotopes of nitrate nitrogen), and then determines the contribution of each agricultural pollution source to the nitrate nitrogen load of river runoff based on the Bayesian mixing model. Although this method can provide a basis for the prevention and control of agricultural non-point source pollution in the basin, it has deficiencies such as relying on a large amount of offline experimental data, complex analysis processes, and poor real-time performance.

[0003] Another disclosed technology CN116106265A proposes an intelligent monitoring and early warning method and system for agricultural non-point source pollution in a small watershed based on hyperspectral remote sensing. Through monitoring area layout, remote sensing background database construction, etc., it realizes the monitoring and early warning of the spatio-temporal changes of agricultural non-point source pollution, and solves the problem of real-time monitoring in small watersheds. However, this method is limited by remote sensing resolution, meteorological influence, and data update frequency, and it is difficult to accurately capture sudden pollution changes, and its application scope is relatively limited. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies of the prior art and propose an intelligent monitoring and early warning system and method for agricultural non-point source pollution to solve the above-mentioned existing problems.

[0005] The purpose of the present invention is achieved through the following technical solutions: An intelligent monitoring and early warning system for agricultural non-point source pollution includes a multi-source heterogeneous data perception and fusion module, a dynamic attention mechanism module, a pollution source tracing module, a PINN-Transformer coupling model module, a cellular automaton risk propagation model module, a pollution data credible tracing module, and a cloud-edge collaborative computing module; The dynamic attention mechanism module adopts an optimized structure Transformer deep learning model, uses the multi-head self-attention mechanism to model the temporal and spatial features of the multi-source water pollution data processed by the multi-source heterogeneous data perception and fusion module, and improves the prediction accuracy of the water pollution situation by dynamically adjusting the attention weights; The PINN-Transformer coupling model module is used to receive the spatio-temporal feature data processed by the dynamic attention mechanism module, and combines the hydrodynamic and pollutant diffusion physical equations as physical constraints to generate pollutant spatio-temporal distribution and flow field feature data; The pollution source tracing module, based on the pollutant spatio-temporal distribution and flow field characteristic data output by the PINN-Transformer coupling model module, is used to determine the pollutant source location by using an intelligent algorithm that integrates physical information and data-driven methods. The intelligent algorithm includes the following steps: Utilize the flow field characteristic data and obtain the precise flow field dynamics by solving the Navier-Stokes equations; adopt the multi-head attention mechanism to perform differential weight assignment according to the importance of parameters such as flow velocity, flow direction, water depth, and turbulent kinetic energy in the flow field characteristic data; construct an error feedback network based on the pollutant concentration gradient and introduce pre-trained model weights for calibration by combining transfer learning; adopt the graph neural network (GNN) combined with the spatial topology analysis method to identify the pollutant propagation path and dynamically update the topology structure; adopt the deep reinforcement learning method to optimize the source tracing path iteration process according to the dynamic diffusion rules to lock the pollutant source location. The cellular automaton risk propagation model module divides the monitored water area into several grid cells and sets dynamic diffusion rules based on the water pollution diffusion mechanism for each grid cell.

[0006] The multi-source heterogeneous data perception and fusion module includes various water quality sensors and data acquisition devices deployed in the monitored water area, which are used to obtain pollution-related data of the water area, and perform cleaning, calibration, and fusion processing on the obtained multi-source data to output multi-source water pollution data for the dynamic attention mechanism module.

[0007] The time and space characteristics of the multi-source water pollution data modeled by the dynamic attention mechanism module include the changes in water quality parameter concentrations, hydrological parameters, and meteorological parameters over time and space.

[0008] The intelligent algorithm adopted by the pollution source tracing module also uses pollutant concentration gradient inversion to assist in identifying the water pollution diffusion path.

[0009] The dynamic diffusion rules set by the cellular automaton risk propagation model module are that when the pollutant concentration of any grid cell changes, the pollution risk status of its adjacent grid cells is updated in real time according to the dynamic diffusion rules to simulate the diffusion process of pollutants in the entire water area; among them, the dynamic diffusion rules can be adaptively adjusted according to environmental factors such as water flow velocity, rainfall, and terrain.

[0010] The pollution data credible source tracing module includes a blockchain network composed of multiple nodes, which is used to perform hash encryption on the key water pollution data generated during the monitoring process and record it in the distributed ledger in chronological order to ensure that the data is tamper-proof and traceable.

[0011] The cloud-edge collaborative computing module includes at least one edge computing device deployed on-site in the monitored water area and a cloud server. The edge computing device is responsible for executing some or all of the functions of the multi-source heterogeneous data perception and fusion module and local preliminary analysis, while the cloud server is responsible for executing complex calculations of the dynamic attention mechanism module, PINN-Transformer coupling model module, and pollution source tracing module and global water pollution situation analysis. The edge computing device and the cloud server are connected through a wired or wireless network to achieve real-time transmission and synchronization of monitoring data and analysis results.

[0012] An intelligent monitoring and early warning method for agricultural non-point source pollution includes the following steps: S1. Multi-source heterogeneous data collection and fusion: Using the multi-source heterogeneous data perception and fusion module, collect water pollution-related data from multiple sensors and data sources, and perform cleaning, calibration, and fusion preprocessing on the data to form a unified pollution data input set. S2. Pollution monitoring model analysis: Using the dynamic attention mechanism module, input the fused data into a Transformer deep learning model with a dynamic attention mechanism to extract water pollution features and perform preliminary pollution condition prediction and analysis. S3. Physical constraint model coupling and field data generation: Using the PINN-Transformer coupling model module, introduce a physics-informed neural network (PINN) during the model analysis process and couple it with the Transformer model, integrate the physical model constraints of water pollutant diffusion, and generate spatio-temporal distribution of pollutant concentration and flow field characteristic data. S4. Pollution propagation path tracking and source tracing: Using the pollution source tracing module, according to the spatio-temporal distribution of water pollutants and flow field characteristic data output by the PINN-Transformer coupling model module, perform the following steps to analyze the pollutant propagation path and identify the location of the pollution source: S4.1. Use the flow field characteristic data and determine the exact dynamic characteristics of the flow field by solving the Navier-Stokes equation to establish a water flow motion model. S4.2. Convert the flow field parameters into numerical vector inputs, and perform differential weight allocation according to the importance of parameters such as flow velocity, flow direction, water depth, and turbulent kinetic energy through the multi-head attention mechanism. S4.3. Construct an error feedback network based on the pollutant concentration gradient, and introduce pre-trained model weights using transfer learning technology. S4.4. Adopt a graph neural network (GNN) combined with a spatial topology analysis method to identify the pollutant propagation path and dynamically update the topology structure. S4.5. Optimize the source tracing path iteration process through a deep reinforcement learning method, adaptively adjust the path prediction strategy according to the dynamic diffusion rule, and lock the location of the pollution source. S5. Risk Diffusion Simulation: Using the risk propagation model module of cellular automata, divide the monitored water area into grid cells based on cellular automata, and simulate the diffusion process of pollutants between grid cells according to the preset dynamic diffusion rules to evaluate the pollution risk level of each grid area; S6. Trusted Data Archiving: Using the trusted pollution data traceability module, store and solidify the key water pollution data during the monitoring process through blockchain to ensure that the data is tamper-proof and traceable for query; S7. Early Warning Information Push: When the monitoring indicators exceed the preset threshold or high pollution risks are predicted, automatically generate pollution early warning information and intelligently push it to relevant users through terminal devices.

[0013] In step S5, the dynamic diffusion rules are adaptively adjusted according to environmental factors such as water flow velocity, rainfall, and terrain to update the pollution risk status of adjacent grid cells in real time.

[0014] The beneficial effects of the present invention are as follows: 1. Through the multi-source heterogeneous data perception and fusion module, information from different sensing devices and data sources is integrated, providing a comprehensive and rich data foundation for the model. The dynamic attention mechanism module can deeply explore the complex spatio-temporal dependence relationships among multi-dimensional data such as water quality, hydrology, and meteorology, and dynamically focus on key influencing factors. More importantly, the PINN-Transformer coupling model module closely combines data-driven deep learning with physical laws such as hydrodynamics and pollutant diffusion, ensuring that the generated pollutant concentration field and flow field data not only fit the observations but also conform to physical reality, greatly improving the accuracy and reliability of pollution status prediction and future trend judgment, and laying a solid foundation for accurate early warning.

[0015] 2. Aiming at the characteristics of agricultural non-point source pollution, such as dispersed, concealed, and highly random sources, the pollution traceability module of the present invention adopts an advanced intelligent algorithm that integrates physical information and data-driven methods. This algorithm utilizes the high-precision flow field data generated by the module and can lock the source location faster and more accurately than traditional methods through precise calculation steps such as solving the Navier-Stokes equation, attention weighting, error feedback and transfer learning, GNN topological analysis, and DRL optimization. This multi-technology fusion strategy, especially the combination of physical information and advanced machine learning models, significantly improves the robustness and efficiency of the traceability process under complex hydrological environments and data uncertainties.

[0016] 3. The present invention uses a cellular automaton risk propagation model module to perform spatial simulation of pollution diffusion. Its key advantage lies in the adoption of dynamic diffusion rules that can be adaptively adjusted according to real-time environmental factors (such as water flow velocity, rainfall, terrain, etc.). This enables the risk simulation to no longer be a deduction based on static parameters, but rather to dynamically reflect the impact of real environmental changes on the behavior of pollutant propagation, thereby more accurately assessing future potential pollution risk areas and levels, providing more forward-looking and practically guiding decision-making support for management departments, and realizing the transformation from passive response to active risk prevention and control.

[0017] 4. By introducing a pollution data trusted traceability module, especially using a blockchain network composed of multiple nodes to perform hash encryption and distributed accounting on key water pollution data, the present invention provides strong trust endorsement for the whole process data of monitoring and early warning (including original observations, model outputs, traceability results, early warning records, etc.). The application of blockchain technology ensures the immutability and full traceability of data, effectively solving the problems of data forgery, loss or trust that may exist in traditional monitoring systems. Combining privacy protection verification technologies such as PHEFL-DFV further enhances the quality credibility of data before it is uploaded to the chain, providing a highly trusted data basis for environmental supervision, law enforcement evidence collection, information disclosure and multi-party cooperation.

[0018] 5. The present invention integrates multiple automated and intelligent links such as data collection, fusion, intelligent analysis, physical-information coupling prediction, intelligent traceability, risk simulation, trusted evidence storage and automatic early warning push, greatly reducing manual intervention and improving work efficiency. At the same time, adopting a distributed architecture of a cloud-edge collaborative computing module, tasks such as data preprocessing and local analysis are deployed at the edge, complex model calculations and global analysis are placed in the cloud, and real-time synchronization of data and results is achieved through a high-speed network connection, effectively balancing the computing load, reducing system latency, improving the overall operation efficiency and the rapid response ability to sudden pollution events, and being easy to expand.

[0019] 6. The present invention specifically solves many problems existing in traditional technologies in aspects such as monitoring data acquisition, complex process simulation, rapid and accurate traceability, data trust guarantee, and system response speed, providing a new, systematic, intelligent and reliable technical solution for the effective monitoring and treatment of the complex environmental problem of agricultural non-point source pollution. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is the system architecture diagram of the present invention; Figure 2 is the working step diagram of the present invention; Figure 3 is the system interaction diagram of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0021] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of the present invention.

[0022] It should be noted that in the following solutions, the orientation concepts of "left", "right", "up", "down", "front", "back", "inside", and "outside" are all relative directions, and will not be listed one by one here.

[0023] Embodiment 1 As Figures 1 to 3 shown, this embodiment provides an intelligent monitoring and early warning system for agricultural non-point source pollution. The intelligent monitoring and early warning system for agricultural non-point source pollution in this embodiment includes: a multi-source heterogeneous data perception and fusion module 10, a dynamic attention mechanism module 20, a pollution source tracing module 30, a PINN-Transformer coupling model module 40, a cellular automaton risk propagation model module 50, a pollution data credible tracing module 60, and a cloud-edge collaborative computing module 70.

[0024] The multi-source heterogeneous data perception and fusion module 10 is used to collect data related to water quality, hydrology, meteorology, and pollution sources from multiple different sources (such as fixed water quality monitoring stations, mobile monitoring devices, remote sensing data, meteorological data sources, etc.) within the agricultural area, and perform necessary cleaning, format unification, calibration in time and space, and fusion processing on these raw data to form a high-quality and consistent multi-dimensional spatio-temporal data input set.

[0025] The dynamic attention mechanism module 20 receives the data processed by the multi-source heterogeneous data perception and fusion module 10. This module adopts an optimized Transformer deep learning model architecture. Its core is to use the multi-head self-attention mechanism to effectively capture the complex long-term time-dependent relationships and spatial correlation features in the input data. By dynamically adjusting the attention weights, this module can automatically focus on the feature information that is more critical for judging the current water pollution situation and predicting future trends, thereby improving the accuracy and robustness of the prediction model.

[0026] The PINN-Transformer coupling model module 40 receives the spatio-temporal feature data output by the dynamic attention mechanism module 20. This module innovatively couples the Physics-Informed Neural Network (PINN) with the Transformer model. It not only uses the Transformer to process the spatio-temporal patterns in the data, but also embeds the physical laws of the known diffusion, migration, and transformation of pollutants in water bodies (such as hydrodynamic equations, advection-diffusion equations, etc.) as strong constraints into the training process of the neural network. In this way, the PINN-Transformer coupling model module 40 can generate the spatio-temporal concentration distribution field of pollutants and the corresponding flow field feature data (such as flow velocity, flow direction, water depth, turbulent kinetic energy, etc.) that conform to both the data patterns and physical laws.

[0027] The pollution source tracing module 30 receives the spatio-temporal distribution of pollutants and the flow field feature data generated by the PINN-Transformer coupling model module 40. The core function of this module is to use an advanced intelligent algorithm that integrates physical information and data-driven methods to accurately trace the source of pollutants. This intelligent algorithm comprehensively uses a variety of technical means, based on the high-precision flow field information and pollutant concentration gradient output by the physical model, and combines machine learning models for path analysis and source location.

[0028] The cellular automaton risk propagation model module 50 discretizes the target monitoring water area (such as rivers, lakes, irrigation areas) in space into a series of interconnected grid cells. Each cell is regarded as a cellular automaton. This module sets dynamic diffusion rules for each cell based on the physical mechanisms of water pollution diffusion (such as advection, diffusion, degradation, etc.) and the influence of neighborhood states, and is used to simulate the propagation process and risk evolution of pollutants in the entire water area grid.

[0029] The pollution data trustworthy tracing module 60 is used to securely and trustworthily store and manage information such as key monitoring data, model outputs, tracing results, and warning records generated during the operation of the system, ensuring the integrity, immutability, and traceability of the data.

[0030] The cloud-edge collaborative computing module 70 provides computing resource support and task scheduling for the entire system, realizing the collaborative processing of tasks such as data processing, model training, and inference calculation between the cloud server and edge computing devices, and optimizing the system response speed and resource utilization efficiency.

[0031] The basic data flow between modules is as follows: After the multi-source heterogeneous data perception and fusion module 10 processes the raw data, it transmits the data to the dynamic attention mechanism module 20 for spatio-temporal feature extraction and modeling. The output of the dynamic attention mechanism module 20 is fed into the PINN-Transformer coupling model module 40 to generate concentration field and flow field data in combination with physical constraints. These field data are then used by the pollution source tracing module 30 for pollution source location. At the same time, the monitoring data and model results can be used by the cellular automaton risk propagation model module 50 for risk simulation and stored evidence by the pollution data credible tracing module 60. The entire process is computationally supported and scheduled by the cloud-edge collaborative computing module 70.

[0032] Specifically, the following formula is adopted: Adaptive physical constraint deep fusion network During the data fitting process, the weights of physical constraints are dynamically adjusted so that the model can not only approximate the observed data but also follow known physical laws.

[0033] Main network f_θ(X): Used to predict the system state (such as the concentration field), where X is the input data and θ is the parameter.

[0034] Physical constraint residual R_i(f_θX): Physical constraints are constructed for the residuals of different physical equations (such as the convection-diffusion equation).

[0035] Weight evaluator network g_ϕ(X, f_θX): Based on the current input and predicted state, dynamically generate the weights λ_i(t) (usually greater than 0) of each physical constraint in the loss function, so that the physical constraints have different influences in different situations.

[0036] Loss function L: By jointly optimizing θ and ϕ, the model can automatically adjust the importance of each physical process during training to ensure that the model not only fits the data but also satisfies the physical laws.

[0037] Working process The working process of the intelligent monitoring and early warning method for agricultural non-point source pollution in this embodiment includes the following steps: S1. Multi-source heterogeneous data collection and fusion: Start the multi-source heterogeneous data perception and fusion module 10 to collect real-time multi-source heterogeneous data such as water quality (such as COD, ammonia nitrogen, total phosphorus, total nitrogen concentration), hydrology (flow velocity, flow rate, water level), and meteorology (rainfall, wind speed) from various sensors and data platforms deployed in the monitoring area. The multi-source heterogeneous data perception and fusion module 10 performs preprocessing on the collected data, such as denoising, filling missing values, unifying units, time alignment, and spatial interpolation, and fuses them into a unified format spatio-temporal data set.

[0038] S2. Pollution monitoring model analysis: Input the fused data obtained in S1 into the dynamic attention mechanism module 20. The Transformer model within the dynamic attention mechanism module 20 analyzes the spatio-temporal dependence of the data using its multi-head self-attention mechanism, captures the complex patterns of pollution changes, and dynamically adjusts the attention weights to focus on the key influencing factors, outputs a representation containing rich spatio-temporal features, and can conduct a preliminary assessment of the pollution status or short-term prediction.

[0039] S3. Physical constraint model coupling and field data generation: Send the spatio-temporal feature data output by the dynamic attention mechanism module 20 into the PINN-Transformer coupling model module 40. Based on the analysis of the Transformer, the PINN-Transformer coupling model module 40 forcibly introduces the hydrodynamic and pollutant transport equations as the physical constraints of the PINN. By jointly optimizing the data fitting loss and the physical equation residual loss, the PINN-Transformer coupling model module 40 generates a high-precision spatio-temporal distribution map of pollutant concentrations that conforms to physical laws and the corresponding hydrodynamic flow field feature dataset (including parameters such as flow velocity, flow direction, water depth, and turbulent kinetic energy).

[0040] S4. Pollution propagation path tracking and source tracing: The pollution source tracing module 30 is activated and receives the pollutant concentration field and flow field feature data from the PINN-Transformer coupling model module 40. Execute the following intelligent algorithm steps: S4.1: Use the flow field feature data generated by the PINN-Transformer coupling model module 40, and further obtain or verify the accurate flow field dynamics characteristics by solving the Navier-Stokes equation (or its simplified form, such as the Saint-Venant equation), and establish a refined water flow motion model.

[0041] S4.2: Convert the flow field parameters (flow velocity, flow direction, water depth, turbulent kinetic energy) into numerical vectors and input them into a multi-head attention network, which learns the importance of different flow field parameters for pollutant source tracing at different positions and times and assigns differential attention weights.

[0042] S4.3: Construct an error feedback network based on the difference between the observed pollutant concentration gradient and the model-predicted concentration gradient, and use transfer learning technology to load the pre-trained model weights under similar hydro-geomorphic conditions to quickly calibrate and optimize the source tracing model.

[0043] S4.4: Apply the graph neural network (GNN) technology to combine the water network topology (such as river network) and pollutant concentration distribution information, identify the main propagation paths of pollutants, and dynamically update the graph topology and node / edge attributes according to the changes in water flow and concentration.

[0044] S4.5: Adopt the deep reinforcement learning (DRL) framework and regard the source tracing process as a sequential decision-making problem. The agent, based on the current state (concentration distribution, flow field information) and the preset dynamic diffusion rules (from the cellular automaton risk propagation model module 50), learns the optimal path search strategy (such as reverse tracing step size and direction). Through iterative optimization, gradually narrow down the range of the pollution source and finally lock the exact or high-probability location of the pollution source.

[0045] S5. Risk Diffusion Simulation The cellular automaton risk propagation model module 50 is activated and uses the current pollutant concentration distribution data (real-time data from the multi-source heterogeneous data perception and fusion module 10 or the prediction results of the PINN-Transformer coupling model module 40) as the initial state. According to the preset dynamic diffusion rules based on physical processes (considering convection, diffusion, degradation, etc.), simulate the migration and diffusion process of pollutants between adjacent units on the divided water area grid, and predict the pollution risk levels of each grid area in the future for a period of time.

[0046] S6. Data Trusted Depositing During the whole monitoring and early warning process, key data points (such as original monitoring values, model prediction results, source tracing conclusions, early warning event records, etc.) are processed by the pollution data trusted source tracing module 60. For example, generate data fingerprints (hash values) through blockchain technology and record them in the distributed ledger to ensure the anti-tampering and traceability of the data.

[0047] S7. Early Warning Information Push The system continuously monitors the real-time monitoring indicators (from the multi-source heterogeneous data perception and fusion module 10) and the predicted risk levels (from the cellular automaton risk propagation model module 50). Once it is found that an indicator exceeds the preset safety threshold, or it is predicted that a high-pollution risk area may appear in the future, the system will automatically generate structured early warning information (including time, location, pollutant type, concentration / risk level, possible pollution source, etc.), and push the early warning information to relevant management personnel, emergency response teams or users in the affected area through the preset channels (such as text messages, APP push, emails, linkage platform interfaces, etc.).

[0048] Through the deep mining and dynamic weighting of the spatio-temporal features of multi-source data by the dynamic attention mechanism module 20, and the PINN-Transformer coupling model module 40 integrating physical law constraints, the accuracy and physical interpretability of water pollution condition prediction and pollutant concentration field generation are significantly improved, making the early warning more reliable.

[0049] The pollution source tracing module 30 adopts an intelligent algorithm that integrates physical information and data-driven, and comprehensively utilizes a variety of advanced technologies such as high-precision flow field simulation (S4.1), attention mechanism (S4.2), error feedback and transfer learning (S4.3), GNN topology analysis (S4.4), and DRL optimization (S4.5). It can more quickly and accurately locate the source of agricultural non-point source pollution events, providing strong support for pollution control and liability investigation.

[0050] The cellular automaton risk propagation model module 50 can simulate the diffusion process of pollutants, evaluate the potential pollution impact range and risk level (S5), and combine real-time monitoring and prediction (S7) to achieve the transformation from passive response to active prevention, improving the foresight of environmental management.

[0051] Using the pollution data credible traceability module 60 (such as through blockchain technology) to deposit key data (S6), effectively preventing data tampering, ensuring the transparency and credibility of the monitoring and early warning process, and providing a credible basis for environmental supervision and decision-making.

[0052] The multi-source heterogeneous data perception and fusion module 10 realizes the automatic acquisition and integration of data (S1); the entire system operates efficiently with the support of the cloud-edge collaborative computing module 70; the application of intelligent algorithms (S4) and automatic early warning push (S7) greatly reduces the manual burden and improves the overall intelligent level and emergency response efficiency of agricultural non-point source pollution monitoring and early warning work.

[0053] In summary, through the integration of a variety of advanced sensing, modeling, algorithm, and information technologies, this embodiment constructs a powerful and superior intelligent monitoring and early warning system for agricultural non-point source pollution and its supporting methods, which can effectively address the challenges of difficult monitoring, difficult source tracing, and lagging early warning of agricultural non-point source pollution.

[0054] Embodiment 2 As Figures 1 to 3 shown, based on the intelligent monitoring and early warning system and method for agricultural non-point source pollution described in Embodiment 1, this embodiment further optimizes and specifically implements the technical details of some key modules and steps, thereby further improving the performance and intelligent level of the system. The basic system architecture and overall work process of this embodiment are basically the same as those of Embodiment 1.

[0055] In this embodiment, the functions and implementation details of the following modules are deepened: Modeling details of the dynamic attention mechanism module 20: Compared with the description in Embodiment 1, when the dynamic attention mechanism module 20 in this embodiment performs spatio-temporal feature modeling, the following multi-category multi-source water pollution-related data is clearly taken into consideration: The concentration of water quality parameters includes real-time or historical concentration value sequences of key water quality indicators such as, but not limited to, chemical oxygen demand (COD), ammonia nitrogen (NH3-N), total phosphorus (TP), total nitrogen (TN), dissolved oxygen (DO), pH value, conductivity, etc.

[0056] Hydrological parameters include parameters reflecting the hydrodynamic state such as the flow velocity, flow rate, water level, cross-sectional area, water depth, etc. of a river or water body.

[0057] Meteorological parameters include meteorological factors such as rainfall, rainfall intensity, rainfall duration, air temperature, wind speed, wind direction, etc. that may affect the processes of pollutant generation, scouring, and transport. The dynamic attention mechanism module 20 can deeply analyze the patterns of these different types of parameters changing over time and space respectively, as well as the comprehensive impact of their complex interactions on the water pollution situation through its multi-head self-attention mechanism, so as to construct a more comprehensive and refined pollution dynamic model.

[0058] Algorithm enhancement of the pollution source tracing module 30: Based on the intelligent algorithm that fuses physical information and data-driven used by the pollution source tracing module 30 described in Embodiment 1, this embodiment further clarifies that this intelligent algorithm also utilizes the pollutant concentration gradient inversion technology to assist in identifying the diffusion path and potential source area of water pollution. Specifically, when the pollution source tracing module 30 executes its algorithm (especially during the path identification or optimization iteration process), it will analyze the spatio-temporal distribution field of pollutant concentration generated by the PINN-Transformer coupling model module 40. By calculating the spatial gradient of the concentration field, the areas and directions with rapid concentration changes are identified, and these gradient information can inversely indicate the possible source directions of pollutants or the diffusion paths of high-concentration plumes. The gradient inversion result can be used as additional features input into the graph neural network (GNN), or as constraint conditions or heuristic information incorporated into the policy optimization of deep reinforcement learning (DRL), so as to improve the robustness and accuracy of the source tracing algorithm under complex hydrodynamic conditions or data sparse situations.

[0059] Details of the dynamic rules of the cellular automaton risk propagation model module 50: This embodiment elaborates more specifically on the dynamic diffusion rules set for the cellular automaton risk propagation model module 50. This rule not only simulates the basic convection-diffusion process, and its core lies in the adaptive adjustment ability: Example of rule content. The state update rule function for each grid cell (cell) can be expressed as: C(i, t+1) = f(Ci, t, {Cj, t|j∈Ni}, Vi, t, Di,t, Ri, t, Envt). Where C(i, t) is the concentration of cell i at time t, N(i) is the set of neighbors of cell i, V is the water flow velocity vector (affecting the convection term, from the PINN-Transformer coupling model module 40), D is the diffusion coefficient (affecting the diffusion term), R is the degradation rate (affecting the reaction term), and Env(t) represents the time-varying environmental factors.

[0060] Adaptive adjustment mechanism. The parameters in this rule function (such as the weight affecting convection, the diffusion coefficient D, the degradation rate R, etc.) are not fixed, but are dynamically adjusted according to real-time environmental factors. For example: When an increase in rainfall is detected (data from the source heterogeneous data perception and fusion module 10), the rule will correspondingly increase the contribution of surface runoff to pollutant input, or increase the diffusion coefficient in certain areas.

[0061] When the water flow velocity V changes (data from the PINN-Transformer coupling model module 40), the convection term in the rule will be updated in real time to accurately reflect the speed and direction of pollutants carried by the water flow.

[0062] Topographic factors (such as slope, river channel curvature) are also encoded into the rule, affecting local water flow and diffusion patterns. This adaptive adjustment makes the simulation of pollutant diffusion closer to the real physical process and improves the accuracy of risk assessment.

[0063] Example 2 uses a causal intervention attention mechanism Integrate causal effect estimation into the traditional attention mechanism, so that the attention score can reflect the true causal relationship between input features, rather than just correlation.

[0064] Standard attention score: Causal effect estimator h_ψ: Using the input X and the corresponding Key vector K_j, estimate its average causal effect (ACE) on the target variable.

[0065] Weighted fusion: Gating mechanism: By introducing causal information into attention calculation, not only can the interpretability of the model be improved, but also the model's ability to capture causal relationships can be enhanced in certain scenarios, resulting in more robust performance.

[0066] Pollution Source Tracing Guided by the Reverse Diffusion Probability Model Using a reverse process similar to the diffusion model, infer the probability distribution of the pollution source based on the observed pollution concentration field.

[0067] Forward Process: Describe the evolution process from the unknown source field C0 to the observed field C through a series of diffusion processes. T of.

[0068] The reverse process model pθ(Ct−1∣Ct,Contextt) is parameterized as a neural network (such as UNet), and uses physical information (such as FlowField_t) to predict the statistics of the previous moment (such as the mean μ_θ and variance Σ_θ) or directly predict the noise ε_θ at each time step.

[0069] Reverse Sampling: Starting from C T and gradually sampling back to C0 according to the learned reverse steps, which is the inferred pollution source distribution.

[0070] Drawing on the idea of the diffusion model, it can give probabilistic source tracing results when dealing with complex pollution diffusion problems, providing a basis for environmental monitoring and governance.

[0071] Working Process The working process of this embodiment follows the steps S1 to S7 described in Embodiment 1. On this basis, the execution of specific steps has the following enhanced details: S2. Pollution Monitoring Model Analysis: When the dynamic attention mechanism module 20 executes this step, it will particularly focus on and model the spatio-temporal dynamics and their interactions of multi-dimensional parameters such as water quality, hydrology, and meteorology.

[0072] S4. Pollution Propagation Path Tracing and Source Tracing: When the pollution source tracing module 30 executes the intelligent algorithms of S4.1 to S4.5, it will apply the pollutant concentration gradient inversion technology in parallel or serially, and integrate the gradient information into the path recognition (S4.4) or strategy optimization (S4.5) link to improve the source tracing accuracy and efficiency.

[0073] S5. Risk Diffusion Simulation: When the cellular automaton risk propagation model module 50 executes the simulation, its core dynamic diffusion rules will be adaptively adjusted in real time according to the current environmental factors (such as the rainfall information obtained from S1 and the flow field data obtained from S3), ensuring that the simulation process can reflect the impact of real environmental changes on the pollutant diffusion behavior.

[0074] Through the refined modeling of multi-dimensional parameters such as water quality, hydrology, and meteorology by the dynamic attention mechanism module 20, the system can more profoundly understand the complex impacts of various factors on the water pollution process, thereby providing more accurate pollution status assessment and more insightful short-term predictions.

[0075] By supplementing and utilizing the pollutant concentration gradient inversion technology in the intelligent algorithm of the pollution source tracing module 30, it provides an additional information source and verification means for the tracing process. Especially when facing complex and variable hydrodynamic conditions or insufficient monitoring data, it can significantly improve the accuracy of pollution source location and the reliability of the results.

[0076] The cellular automaton risk propagation model module 50 adopts dynamic diffusion rules that can adaptively adjust according to real-time environmental factors, making the risk diffusion simulation in step S5 closer to physical reality, significantly improving the accuracy of risk assessment and the timeliness of future risk prediction, and winning valuable time for taking preventive measures.

[0077] In summary, through the deepening of the key module algorithms and model details in Example 2, the entire intelligent monitoring and early warning system for agricultural non-point source pollution is further enhanced in terms of analysis depth, source tracing accuracy, and authenticity of risk prediction, and the overall performance is more superior.

[0078] Example 3 As Figures 1 to 3 shown, based on the systems and methods described in Example 1 and Example 2, this example further elaborates in detail the physical deployment architecture of the system, the front-end data acquisition method, and the specific technical implementation to ensure data credibility, thereby demonstrating the deployment form, operation efficiency, and data credibility advantages of the present invention in practical applications. This example shares the basic system functions and core algorithm processes of Example 1 and 2.

[0079] In this example, the deployment form of the system and the specific implementation methods of related modules are refined as follows: Physical implementation and deployment of the multi-source heterogeneous data perception and fusion module 10: To achieve comprehensive and accurate data perception, the multi-source heterogeneous data perception and fusion module 10 in this example includes the following components at the physical level and is strategically deployed: Multiple water quality sensors: At key sections of the monitored water area, downstream of sewage outlets, entrances to water source protection areas, confluence points of tributaries, etc., a sensor array is deployed, including multi-parameter water quality online analyzers (measuring COD, ammonia nitrogen, TP, TN, DO, pH, conductivity, turbidity, etc.), specific pollutant sensors (such as heavy metal sensors, pesticide residue sensors), flow meters, water level gauges, etc.

[0080] Data acquisition device: Each sensor or sensor group is equipped with an intelligent data acquisition and transmission unit (such as DTU or RTU). These devices are responsible for periodically collecting sensor readings, performing preliminary data quality verification (such as range check, outlier rejection), caching data, and uploading the data through the built-in communication module (such as GPRS / 4G / 5G / NB-IoT / LoRa). These sensors and acquisition devices constitute the on-site perception network of the multi-source heterogeneous data perception and fusion module 10, ensuring the real-time, continuous, multi-point, and multi-parameter acquisition of data related to water pollution monitoring in the monitored water area. The software part of the multi-source heterogeneous data perception and fusion module 10 (running at the edge or in the cloud) is responsible for further cleaning, time synchronization, spatial interpolation, and multi-source data fusion processing of the uploaded raw data, and finally forming a unified data set for subsequent modules (such as the dynamic attention mechanism module 20) to use.

[0081] Blockchain implementation of the pollution data trustworthy traceability module 60: To ensure the authenticity and immutability of the monitoring data, the pollution data trustworthy traceability module 60 of this embodiment is implemented using blockchain technology. Specifically, a consortium blockchain or private blockchain network is constructed: Blockchain network: It is jointly composed of multiple nodes deployed on the servers of different institutions (such as environmental protection departments, water conservancy departments, agricultural departments, scientific research units, and third-party trustworthy institutions).

[0082] Trustworthy deposit process: During the operation of the system, key water pollution data generated, such as the monitoring data verified and fused by the multi-source heterogeneous data perception and fusion module 10, the important prediction results output by the PINN-Transformer coupling model module 40, the pollution source information determined by the pollution traceability module 30, and the early warning records triggered by the system (generated in step S7), etc., after being generated or confirmed, will be sent to the pollution data trustworthy traceability module 60. This module extracts the core information of these data, calculates its hash value (such as using the SHA-256 algorithm), and packages information such as the hash value, data description, and timestamp into a transaction. These transactions are added to a block of a distributed ledger in chronological order after being consensus-verified by the nodes in the network and linked to the previous block, forming an immutable chain structure. In this way, any data record stored on the blockchain has the characteristics of anti-tampering, verifiability, and traceability, greatly enhancing the credibility of the monitoring data. When it is necessary to verify the data, only need to recalculate the hash value of the original data and compare it with the hash value recorded on the chain.

[0083] Architecture and task allocation of the cloud-edge collaborative computing module 70: This embodiment adopts a cloud-edge collaboration architecture to optimize the computing performance and response speed of the system. The specific implementation of the cloud-edge collaborative computing module 70 is as follows: Edge computing devices: At least one edge computing device (such as an embedded industrial computer, edge server) is deployed near the monitoring site (such as beside the monitoring station, regional control center). These devices have certain computing and storage capabilities.

[0084] Cloud servers: A high-performance cloud server cluster is deployed in a remote data center or public / private cloud platform, with powerful computing resources (CPU / GPU) and massive storage capabilities.

[0085] Task allocation: Edge side: It is mainly responsible for executing some functions of the multi-source heterogeneous data perception and fusion module 10, especially the acquisition, cleaning, calibration of real-time data, and simple fusion calculations; at the same time, it performs local preliminary analysis, such as data validity checks, basic statistics, simple threshold judgments, or runs lightweight anomaly detection models. This can quickly process on-site data, reducing the amount of data transmission and latency.

[0086] Cloud side: It is mainly responsible for executing computationally intensive tasks and tasks that require global information, including running the dynamic attention mechanism module 20 (complex Transformer model), PINN-Transformer coupling model module 40 (involving physical equation solving and coupled training), pollution source tracing module 30 (including complex machine learning and optimization algorithms), conducting comprehensive analysis of the global water pollution situation, and running the complex cellular automaton risk propagation model module 50.

[0087] Network connection: The edge computing device and the cloud server are connected through a stable and reliable wired (such as optical fiber) or wireless (such as 5G) network, ensuring that monitoring data can be uploaded to the cloud in real time, and the analysis results and control instructions from the cloud can also be sent to the edge side or user terminal in a timely manner. This architecture realizes the reasonable allocation of computing tasks and real-time synchronization of data.

[0088] Working process The working process of this embodiment generally follows the steps S1 to S7 described in Embodiments 1 and 2. In terms of cloud-edge collaboration and data trust guarantee, it is specifically reflected as follows: S1. Data acquisition and fusion: The data collected by on-site sensors is first collected by the acquisition device deployed at the edge (belonging to the multi-source heterogeneous data perception and fusion module 10), and preliminarily processed and analyzed by the edge computing device (belonging to the cloud-edge collaborative computing module 70), and then uploaded to the cloud server in real time.

[0089] S2, S3, S4. Model analysis and traceability: These core computationally intensive steps are mainly executed on the cloud server (belonging to the cloud-edge collaborative computing module 70), leveraging the powerful computing capabilities of the cloud to complete complex model training, inference, and optimization calculations.

[0090] S6. Trusted data storage and certification: At key nodes such as after S1 data verification, after S3 model output, after S4 traceability completion, and after S7 warning generation, the relevant data or its digest information is sent to the contaminated data trusted traceability module 60 for hash encryption and chain storage through the blockchain network.

[0091] Overall process: The entire working process relies on the stable network connection and computing architecture provided by the cloud-edge collaborative computing module 70 to achieve real-time data flow, effective task allocation, and rapid result feedback.

[0092] Improved data collection quality and timeliness: By deploying a diverse sensor array and intelligent data collection devices at the monitoring site (the specific implementation of the multi-source heterogeneous data perception and fusion module 10), more comprehensive, accurate, and timely on-site raw data can be obtained, laying a solid data foundation for subsequent precise analysis and warning.

[0093] Significantly improved system operation efficiency and response speed: The architecture design of the cloud-edge collaborative computing module 70 sinks data preprocessing and preliminary analysis tasks to the edge, reducing the burden on the cloud and network transmission pressure; at the same time, it uses the powerful computing power of the cloud to process complex models, achieving optimized allocation of computing resources. This division of labor and cooperation mode significantly improves the overall operation efficiency of the system, reduces end-to-end latency, and enhances the rapid response ability to pollution events.

[0094] Ensured the authority and immutability of monitoring data: The contaminated data trusted traceability module 60 uses blockchain technology to perform distributed and encrypted storage of key data, technically ensuring the integrity, originality, and immutability of the data. This greatly improves the credibility and authority of monitoring data, providing strong and reliable technical support for environmental supervision and law enforcement, accident liability determination, public information disclosure, etc.

[0095] Enhanced the scalability and robustness of the system: The cloud-edge collaborative architecture makes the system easy to expand, and it can conveniently add monitoring points (add edge devices) or enhance the cloud computing power; the distributed characteristics of the blockchain network (the contaminated data trusted traceability module 60) also improve the robustness of data storage.

[0096] In summary, through the specific designs of the system's physical deployment, data acquisition front-end, computing architecture, and data management method in Embodiment 3, not only high-quality data input and efficient system operation are ensured, but more importantly, a reliable data trust mechanism is established, making the present invention more efficient, scalable, robust, and credible in practical applications.

[0097] The above are only the preferred embodiments of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein, should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications, and environments, and can be changed within the scope of the concept described herein through the above teachings or the technology or knowledge in related fields. As long as the changes and variations made by those skilled in the art do not depart from the spirit and scope of the present invention, they should all be within the protection scope of the appended claims of the present invention.

Claims

1. An intelligent monitoring and early warning system for agricultural non-point source pollution, characterized in that, It includes a multi-source heterogeneous data perception and fusion module (10), a dynamic attention mechanism module (20), a pollution source tracing module (30), a PINN-Transformer coupling model module (40), a cellular automaton risk propagation model module (50), a pollution data trustworthy source tracing module (60), and a cloud-edge collaborative computing module (70); The dynamic attention mechanism module (20) adopts an optimized structure Transformer deep learning model, uses the multi-head self-attention mechanism to model the temporal and spatial features of the multi-source water pollution data processed by the multi-source heterogeneous data perception and fusion module (10), and improves the prediction accuracy of the water pollution situation by dynamically adjusting the attention weights; The PINN-Transformer coupling model module (40) is used to receive the spatio-temporal feature data processed by the dynamic attention mechanism module (20), and combines the hydrodynamic and pollutant diffusion physical equations as physical constraints to generate pollutant spatio-temporal distribution and flow field feature data; The pollution source tracing module (30), based on the pollutant spatio-temporal distribution and flow field feature data output by the PINN-Transformer coupling model module (40), is used to determine the pollutant source location by using an intelligent algorithm that fuses physical information and data-driven; The intelligent algorithm includes the following steps: Utilize the flow field feature data, and obtain the accurate flow field dynamics characteristics by solving the Navier-Stokes equation; adopt the multi-head attention mechanism to perform differential weight allocation according to the importance of the parameters of flow velocity, flow direction, water depth, and turbulent kinetic energy in the flow field feature data; construct an error feedback network based on the pollutant concentration gradient, and introduce pre-trained model weights for calibration by combining transfer learning; adopt a graph neural network (GNN) combined with a spatial topology analysis method to identify the pollutant propagation path and dynamically update the topology structure; adopt a deep reinforcement learning method to optimize the source tracing path iteration process according to the dynamic diffusion rule to lock the pollutant source location; The cellular automaton risk propagation model module (50) divides the monitored water area into several grid cells, and sets dynamic diffusion rules based on the water pollution diffusion mechanism for each grid cell.

2. The intelligent monitoring and early warning system for agricultural non-point source pollution according to claim 1, characterized in that: The multi-source heterogeneous data perception and fusion module (10) includes various water quality sensors and data acquisition devices deployed in the monitored water area, which are used to obtain the pollution-related data of the water area, and perform cleaning, calibration, and fusion processing on the obtained multi-source data to output the multi-source water pollution data for the dynamic attention mechanism module (20) to use.

3. An intelligent monitoring and early warning system for agricultural non-point source pollution according to claim 1, characterized in that: The temporal and spatial features of the multi-source water pollution data modeled by the dynamic attention mechanism module (20) include the changes of water quality parameter concentrations, hydrological parameters, and meteorological parameters over time and space.

4. An intelligent monitoring and early warning system for agricultural non-point source pollution according to claim 1, characterized in that: The intelligent algorithm adopted by the pollution source tracing module (30) also uses pollutant concentration gradient inversion to assist in identifying the water pollution diffusion path.

5. An intelligent monitoring and early warning system for agricultural non-point source pollution according to claim 1, characterized in that: The dynamic diffusion rule set by the cellular automaton risk propagation model module (50) is that when the pollutant concentration of any grid cell changes, the pollution risk status of its adjacent grid cells is updated in real time according to the dynamic diffusion rule to simulate the diffusion process of pollutants in the entire water area; wherein, the dynamic diffusion rule can be adaptively adjusted according to environmental factors such as water flow velocity, rainfall, and terrain.

6. The intelligent monitoring and early warning system for agricultural non-point source pollution according to claim 1, characterized in that: The pollution data credible traceability module (60) includes a blockchain network composed of multiple nodes, which is used to perform hash encryption on the key water pollution data generated during the monitoring process and record it in the distributed ledger in chronological order to ensure that the data is tamper-proof and traceable for query.

7. An intelligent monitoring and early warning system for agricultural non-point source pollution according to claim 1, characterized in that: The cloud-edge collaborative computing module (70) includes at least one edge computing device deployed on-site in the monitored water area and a cloud server. Among them, the edge computing device is responsible for executing some or all of the functions of the multi-source heterogeneous data perception and fusion module (10) and local preliminary analysis, and the cloud server is responsible for executing the complex calculations of the dynamic attention mechanism module (20), the PINN-Transformer coupling model module (40), and the pollution traceability module (30) and the global water pollution situation analysis; the edge computing device and the cloud server are connected through a wired or wireless network to achieve real-time transmission and synchronization of monitoring data and analysis results.

8. An intelligent monitoring and early warning method for agricultural non-point source pollution, characterized in that: It includes the following steps: S1. Multi-source heterogeneous data collection and fusion: Using the multi-source heterogeneous data perception and fusion module (10), collect water pollution-related data from multiple sensors and data sources, and perform cleaning, calibration, and fusion preprocessing on the data to form a unified pollution data input set; S2. Pollution monitoring model analysis: Using the dynamic attention mechanism module (20), input the fused data into a Transformer deep learning model with a dynamic attention mechanism to extract water pollution features and perform preliminary pollution condition prediction analysis; S3. Physical constraint model coupling and field data generation: Using the PINN-Transformer coupling model module (40), introduce a physics-informed neural network (PINN) and couple it with the Transformer model during the model analysis process, incorporate the physical model constraints of water pollutant diffusion, and generate pollutant concentration spatio-temporal distribution and flow field characteristic data; S4. Pollution propagation path tracking and traceability: Using the pollution traceability module (30), according to the water pollutant spatio-temporal distribution and flow field characteristic data output by the PINN-Transformer coupling model module (40), perform the following steps to analyze the pollutant propagation path and identify the source location: S4.

1. Use the flow field characteristic data and determine the exact dynamic characteristics of the flow field by solving the Navier-Stokes equation to establish a water flow motion model; S4.

2. Convert the flow field parameters into numerical vector inputs, and perform differential weight allocation through the multi-head attention mechanism according to the importance of parameters such as flow velocity, flow direction, water depth, and turbulent kinetic energy; S4.

3. Construct an error feedback network based on the pollutant concentration gradient and introduce the pre-trained model weights using transfer learning technology; S4.

4. Adopt a graph neural network (GNN) combined with a spatial topology analysis method to identify the pollutant propagation path and dynamically update the topological structure; S4.

5. Optimize the source tracing path iteration process through a deep reinforcement learning method, adaptively adjust the path prediction strategy according to the dynamic diffusion rule, and lock the location of the pollution source; S5. Risk diffusion simulation: Use the cellular automaton risk propagation model module (50) to divide the monitored water area into grid cells based on cellular automata, and simulate the diffusion process of pollutants between the grid cells according to the preset dynamic diffusion rule to evaluate the pollution risk level of each grid area; S6. Data trusted storage and evidence: Use the pollution data trusted source tracing module (60) to store and solidify the key water pollution data during the monitoring process through the blockchain to ensure that the data is tamper-proof and traceable for query; S7. Early warning information push: When the monitoring index exceeds the preset threshold or a high pollution risk is predicted, automatically generate pollution early warning information and intelligently push it to relevant users through terminal devices.

9. The intelligent monitoring and early warning method for agricultural non-point source pollution according to claim 8, characterized in that: In the step S5, the dynamic diffusion rule is adaptively adjusted according to environmental factors such as water flow velocity, rainfall, and terrain to update the pollution risk status of adjacent grid cells in real time.

Images