Water pollutant concentration detection system and method based on big data analysis

Abstract

The invention relates to the technical field of water pollutant concentration detection, and discloses a water pollutant concentration detection system and method based on big data analysis, and the water pollutant concentration detection method based on big data analysis comprises the following steps: constructing a multi-level sensing network, the multi-layer sensing network comprises a fixed monitoring station layer, a water surface unmanned ship layer and an underwater robot layer; carrying out multi-modal data fusion by utilizing a knowledge constraint data assimilation algorithm to generate consistent water quality parameter estimation; a double-layer cognitive structure combining a knowledge graph and a causal model is constructed, and water pollutant propagation understanding and tracing are achieved; continuous optimization of system configuration is realized through a multi-objective reinforcement learning framework; through collaborative optimization of a multi-layer sensing network, the system realizes three-dimensional monitoring of a water body, the monitoring coverage range is expanded, and meanwhile, the high-precision detection capability is kept.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of water pollutant concentration detection, more specifically, it relates to a water pollutant concentration detection system and method based on big data analysis. BACKGROUND

[0002] With the acceleration of industrialization process and the improvement of environmental protection requirements, water pollutant concentration detection technology has become an important research direction in the field of environmental monitoring. At present, water pollutant concentration detection mainly adopts traditional methods based on fixed monitoring sites, mobile monitoring platform technology and data-driven intelligent analysis methods.

[0003] The existing fixed monitoring station technology can provide relatively accurate water quality parameter measurement by deploying high-precision sensing devices at key locations. This method has the advantages of high monitoring accuracy and good data stability, and has been widely used in many water quality monitoring projects. However, the coverage of fixed monitoring stations is limited, and it is difficult to achieve comprehensive monitoring for large water areas relying solely on fixed stations, resulting in monitoring blind spots. Mobile monitoring platform technologies, such as unmanned ships and underwater robots, can expand the monitoring coverage and achieve flexible spatial sampling. These technologies can adjust the monitoring position as needed, to some extent, to compensate for the spatial limitations of fixed stations. However, mobile monitoring platforms are often subject to energy constraints, communication limitations, and other factors, which affect the detection accuracy and real-time performance compared to fixed stations. Data-driven intelligent analysis methods use machine learning, artificial intelligence, and other technologies to analyze and process water quality monitoring data, which can discover potential patterns and abnormal patterns in the data. These methods have advantages in handling large amounts of monitoring data, and can improve the efficiency and accuracy of data analysis. However, purely data-driven methods often ignore the professional knowledge and physical and chemical laws in the field of water quality monitoring, resulting in analysis results that may lack physical meaning or violate basic chemical principles. Traditional expert knowledge-based methods can ensure that the analysis results conform to the field knowledge, but they are inefficient in handling large-scale, multi-dimensional monitoring data, and are difficult to fully utilize the advantages of modern sensing technology and computing technology. Existing water quality monitoring systems face technical challenges in multi-source heterogeneous data fusion. Water quality monitoring involves optical, electrochemical, biological, and other types of sensors, and these sensors have different data formats, sampling frequencies, measurement accuracies, and reliabilities. How to effectively integrate these heterogeneous data and achieve consistent water quality parameter estimation is an important problem faced by current technology. Most existing monitoring systems use static configuration mode, and the topology structure and sampling strategy of the monitoring network remain basically unchanged after deployment. This mode is difficult to adapt to the dynamic changes of water environment and the adjustment of monitoring needs, resulting in unreasonable allocation of monitoring resources and insufficient overall efficiency and adaptability of the system. In terms of pollution source tracking, traditional methods mainly focus on the detection and concentration measurement of pollutants, but have limited ability to identify pollution sources and track diffusion paths. This limitation makes it difficult for existing technology to provide comprehensive technical support for water environment management and pollution prevention and control decision-making.

[0004] Therefore, there is an urgent need for a water pollutant concentration detection technology that can balance monitoring coverage and detection accuracy, organically integrate data-driven and knowledge-driven methods, effectively handle multi-source heterogeneous data, have adaptive adjustment capability, and support pollution source tracing. SUMMARY

[0005] The application provides a water body pollutant concentration detection system and method based on big data analysis, and solves the technical problems of limited monitoring coverage, difficult fusion of multi-source heterogeneous data, insufficient pollution source tracing capability, and poor system adaptability in the related art.

[0006] The application provides a water body pollutant concentration detection method based on big data analysis, comprising:

[0007] A multi-level perception network is constructed, and the multi-level perception network comprises a fixed monitoring station layer, a water surface unmanned ship layer and an underwater robot layer.

[0008] Based on the multi-level perception network, a multi-modal data fusion is performed by using a knowledge-constrained data assimilation algorithm to generate consistent water quality parameter estimation.

[0009] Based on the consistent water quality parameter estimation result, a double-layer cognitive structure combining a knowledge graph and a causal model is constructed to realize water body pollutant propagation understanding and tracing.

[0010] Based on the water body pollutant propagation understanding and tracing analysis output, a multi-objective reinforcement learning framework is used to realize continuous optimization of system configuration.

[0011] Further, the step of constructing the multi-level perception network comprises:

[0012] Based on a hydrological geographical feature clustering algorithm, a water area space is divided, and an initial monitoring point distribution is output.

[0013] A multi-level perception strategy is generated by using an interlayer collaborative optimization algorithm.

[0014] An self-organizing learning algorithm is used to dynamically update the perception network structure according to real-time monitoring data and environmental change information.

[0015] Further, the step of multi-modal data fusion comprises:

[0016] The original data of various sensors are standardized and noise filtered.

[0017] A sensor state evaluation algorithm is used to calculate a data credibility score.

[0018] A knowledge-constrained data assimilation algorithm is used to fuse the standardized multi-modal data set, the data credibility score and the domain knowledge base to generate consistent water quality parameter estimation.

[0019] Further, the step of constructing the double-layer cognitive structure combining the knowledge graph and the causal model comprises:

[0020] A water environment knowledge graph is constructed and updated.

[0021] Combine the domain constraints in the knowledge graph and the observation data to construct a causal relationship model of water body pollutant propagation and mutual influence;

[0022] Based on the established causal network, a reverse reasoning mechanism is designed to trace back the pollution source from the observed pollution phenomenon;

[0023] According to the identified pollution source position, pollutant propagation path and key area, the monitoring strategy of the perception layer is dynamically adjusted.

[0024] Further, the causal relationship model adopts a knowledge-constrained causal discovery algorithm, which further includes:

[0025] Conduct confidence evaluation on the preliminary identified variable dependency relationship;

[0026] Combine time delay analysis to determine the timing characteristics of the causal relationship;

[0027] Use cross-validation method to verify the stability and generalization ability of the causal graph structure.

[0028] Further, the steps of the multi-objective reinforcement learning framework include:

[0029] Define the key performance indicators of the system and construct a comprehensive performance evaluation function;

[0030] Construct a mapping relationship model between environmental characteristics and system performance;

[0031] Model the system configuration optimization problem as a multi-objective Markov decision process, and learn the optimal parameter adjustment strategy through multi-objective deep reinforcement learning algorithm;

[0032] Use knowledge distillation technology to transfer the knowledge of complex reinforcement learning model to simple lightweight model, realize the localization of system adaptive adjustment on edge device.

[0033] Further, the multi-objective deep reinforcement learning algorithm adopts a multi-objective Q learning architecture based on decomposition, including:

[0034] Build a multi-head neural network, each output head corresponding to an optimization target;

[0035] Learn the state-action value function through the deep Q network architecture, while maintaining an experience replay buffer to store transition samples;

[0036] Use multi-objective Pareto dominance principle to evaluate and select different system configuration strategies;

[0037] Periodically update the target network parameters through a soft update mechanism to improve learning stability.

[0038] Further, the knowledge distillation technology includes:

[0039] The trained complex deep reinforcement learning model is used as a teacher network, and a lightweight neural network with reduced parameters is constructed as a student network;

[0040] A large number of state samples are generated, and the corresponding Q values and policy distribution are generated by using the teacher network;

[0041] A comprehensive loss function is designed, including the distillation loss between the student network prediction and the teacher network output, and the task loss of the student network interacting with the environment;

[0042] The student network parameters are optimized by gradient descent, so that the student network can be lightweight while retaining the key decision-making ability of the teacher network.

[0043] Further, the sensor state evaluation algorithm comprises:

[0044] The reliability scores of the data of each sensor are dynamically calculated by comprehensively considering the service life of the sensor, the historical accuracy and the environmental adaptability factor;

[0045] Abnormal data points are identified through multiple test mechanisms of physical constraint test, time continuity test and space correlation test;

[0046] According to the type of anomaly, corresponding processing is performed, including interpolation, marking and rejection operations;

[0047] An abnormality detection algorithm based on statistical learning is adopted to automatically identify and handle various types of sensor failure modes.

[0048] The application provides a water pollutant concentration detection system based on big data analysis, which is used to execute the water pollutant concentration detection method based on big data analysis.

[0049] The multi-level perception network module is used to construct a three-dimensional monitoring network including a fixed monitoring station layer, a water surface unmanned ship layer and an underwater robot layer;

[0050] The multi-modal data fusion module is used to fuse and process multi-source heterogeneous sensing data based on a knowledge-constrained data assimilation algorithm, and output unified water quality parameter evaluation results;

[0051] The cognitive reasoning module is used to construct a double-layer cognitive architecture of knowledge graph and causal model, and realize pollution propagation mechanism analysis and pollution source positioning;

[0052] The adaptive optimization module is used to dynamically adjust and continuously optimize the system operation parameters based on a multi-objective reinforcement learning framework.

[0053] The beneficial effects of the present application are that through the cooperative optimization of the multi-level perception network, the system realizes the stereoscopic monitoring of the water body, expands the monitoring coverage range, and at the same time maintains the high-precision detection capability;

[0054] Through the knowledge-constrained data assimilation framework, the system effectively integrates the heterogeneous data of multiple sensors, solves the problem of data inconsistency, and improves the accuracy and reliability of water quality parameter estimation;

[0055] Through the double-layer cognitive structure of the knowledge graph and the causal model, the system can accurately track the location of the pollution source and the propagation path of the pollutant, and the accuracy of the pollution source positioning is improved under the condition of incomplete monitoring data, which provides strong support for water quality management;

[0056] Through the multi-objective reinforcement learning framework, the system can automatically adjust the monitoring strategy and analysis parameters according to the environmental changes, maintain long-term stable high performance, and reduce the performance decline of the system in the seasonal change environment compared with the traditional system;

[0057] The method provided by the present application not only utilizes knowledge to guide data analysis, but also learns new knowledge from data, constantly enriches and updates the knowledge base, and forms a continuously increasing technical asset;

[0058] Through the knowledge distillation and lightweight deployment technology, the system can efficiently run on resource-limited edge devices, reduce the dependence on communication bandwidth and central processing capability, and reduce the overall cost of the system. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 is a flowchart of a water pollutant concentration detection method based on big data analysis in the present application;

[0060] Figure 2 is a line graph comparing the performance of the multi-level perception network and the traditional fixed monitoring system in terms of detection accuracy and coverage range;

[0061] Figure 3 is a column chart comparing the performance of the knowledge-constrained data assimilation framework and the traditional single sensor method in terms of parameter estimation accuracy;

[0062] Figure 4 is a column chart comparing the performance of the knowledge base of the system before and after running for one year in terms of size and accuracy;

[0063] Figure 5 is a column chart comparing the performance of the edge device deployment and the central server deployment of the system in terms of response time and communication load;

[0064] Figure 6is a radar chart showing the performance comparison between the system of the present application and the traditional monitoring system in five key dimensions. DETAILED DESCRIPTION

[0065] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that the discussion of these implementations is merely meant to provide a better understanding of the subject matter described herein and can be changed in ways not specifically discussed which will fall within the scope of the disclosure. Various examples can omit, substitute, or add various procedures or components as appropriate. Also, it should be understood that instead of being discussed in the ordering apparently from the foregoing, discussions can occur in any order.

[0066] A water body pollutant concentration detection method based on big data analysis is disclosed in at least one embodiment of the present application, as shown in Figure 1 The method comprises the following steps:

[0067] Step 1, a multi-level perception network is constructed, which includes a fixed monitoring station layer, a water surface unmanned ship layer and an underwater robot layer;

[0068] According to one embodiment of the present application, this step uses a self-organizing multi-level perception network algorithm to process water body monitoring spatial distribution data to generate an optimized multi-level monitoring network structure. Specifically, the following sub-steps are included:

[0069] Step 1.1, water space division and monitoring point initialization;

[0070] This sub-step inputs water area geographic information data, historical water quality data and expert knowledge, divides the water area space through a hydrographic feature clustering algorithm, and outputs the initial monitoring point distribution. The system first performs grid segmentation on the water area, combines hydrographic features (such as water depth, flow rate, confluence points, etc.) and pollutant concentration change characteristics in historical monitoring data, and uses hierarchical clustering method to divide the water area into multiple sub-regions. Optionally, the system can also use density clustering or fuzzy clustering method to divide the water area to adapt to different types of water body environmental characteristics. In addition, the system preliminarily determines the location of the fixed monitoring station in each sub-region based on the principle of maximum information entropy, and plans the initial deployment area for the mobile monitoring platform. It should be understood that the layout of the initial monitoring points can be adjusted as appropriate according to the specific water area environmental characteristics and monitoring targets.

[0071] Step 1.2, interlayer collaborative perception strategy generation;

[0072] This sub-step inputs multi-level monitoring resource information and monitoring targets, and generates a multi-level perception strategy through an inter-layer collaborative optimization algorithm. It should be noted that the system divides the monitoring network into three levels: a fixed monitoring site layer (providing long-term baseline data), a surface unmanned vessel layer (providing surface coverage monitoring), and an underwater robot layer (providing depth profile monitoring). For each level, the system constructs a multi-objective optimization model based on task importance, energy constraints, and monitoring coverage requirements, and calculates the optimal sampling frequency, spatial position, and movement path of each monitoring unit. In some embodiments, inter-layer collaboration is achieved through a gradient information sharing mechanism, and water quality anomalies or gradient change areas discovered by the upper-level monitoring unit will automatically trigger the detailed observation tasks of the lower-level monitoring unit.

[0073] Step 1.3, network self-organization and dynamic adjustment;

[0074] This sub-step inputs real-time monitoring data and environmental change information, and updates the perception network structure through a self-organizing learning algorithm. According to an embodiment of the present application, the system dynamically adjusts the topological structure and sampling strategy of the monitoring network based on the spatiotemporal variation characteristics of water quality parameters. The method provided in the present application introduces a self-organizing control model based on multiple agents, so that each monitoring unit can autonomously adjust its position and working mode according to local observations and information from neighboring units. In addition, the system combines the prediction results of the hydrological model to predict the possible direction of pollutant diffusion, deploy monitoring resources in advance, and realize forward-looking adjustment of the monitoring network. Optionally, in complex water environments, the system can also adopt a topological structure optimization algorithm based on a graph neural network to further improve the adaptability and robustness of the monitoring network.

[0075] like Figure 2 As shown, the multi-layered perception network provided by this application achieves improvements in both detection accuracy and coverage compared to traditional fixed monitoring systems. This line chart visually demonstrates how this system, through a collaboratively optimized monitoring network architecture, expands monitoring coverage while maintaining high-precision detection capabilities, validating the effectiveness of the multi-layered perception network construction method.

[0076] Step 2: Based on the multi-level perception network, the knowledge-constrained data assimilation algorithm is used to fuse multimodal data and generate consistent water quality parameter estimates;

[0077] According to another embodiment of the present application, this step utilizes a knowledge-constrained data assimilation algorithm to process multi-source heterogeneous sensor data and generate consistent water quality parameter estimates. Specifically, it includes the following sub-steps:

[0078] Step 2.1, multimodal sensor data preprocessing;

[0079] The sub-step inputs various sensor raw data, and outputs standardized multi-modal data sets through sensor data standardization and noise filtering algorithms. It should be noted that the system first performs time alignment and spatial registration on data of different types of sensors (such as electrochemical sensors, optical sensors, biological sensors, etc.), solving the problem of inconsistent sampling frequencies and uneven spatial distribution. Therefore, the system uses an adaptive filtering algorithm to remove data noise according to the characteristics of various sensors, and performs dimensionless normalization processing to convert the measurement results of different physical quantities into a unified data representation form. In some embodiments, the system can also apply a multi-scale denoising method based on wavelet transform, which is particularly suitable for processing sensor data containing various frequency noises.

[0080] Step 2.2, sensor reliability evaluation and data quality control;

[0081] The sub-step inputs historical sensor performance data and environmental condition information, and outputs data reliability scores through a sensor state evaluation algorithm. It should be understood that the system builds a sensor reliability evaluation model, considering factors such as sensor service life, historical accuracy, environmental adaptability, etc., to dynamically calculate the reliability scores of each sensor data. In addition, the system identifies abnormal data points through multiple test mechanisms such as physical constraint test, time continuity test and spatial correlation test, and processes them accordingly (such as interpolation, marking or rejection) according to the type of anomaly. In some embodiments, the system can also use an anomaly detection algorithm based on statistical learning to automatically identify and handle various types of sensor failure modes, further improving the accuracy of data quality control.

[0082] Step 2.3, knowledge-constrained multi-source data fusion;

[0083] The sub-step inputs standardized multi-modal data sets, data reliability scores and domain knowledge base, and outputs consistent water quality parameter estimates through a knowledge-constrained data assimilation algorithm. According to the method provided in the present application, the system first extracts relevant physical and chemical laws (such as ion balance relationship, pollutant degradation kinetics model, etc.) from the knowledge base as constraint conditions. As can be seen, the system builds a fusion model based on the variational data assimilation framework, integrates multi-source sensor data, reliability scores and knowledge constraints into a unified optimization problem, and obtains the optimal water quality parameter estimates that satisfy the physical and chemical laws by minimizing the constrained cost function. The fusion process automatically balances the two sources of data-driven and knowledge-driven information, giving priority to data when data is sufficient, and increasing the weight of knowledge constraints when data is sparse or uncertain. Alternatively, the system can also use the ensemble Kalman filter method as an alternative implementation of data assimilation, which is particularly suitable for handling high-dimensional state space water quality parameter estimation problems.

[0084] As Figure 3As shown, the knowledge-constrained data assimilation framework provided by the present application realizes improvement in parameter estimation accuracy compared to traditional single sensor methods. The histogram comparison shows that under complex water quality conditions, through the knowledge-constrained multi-source data fusion technology, the system can effectively integrate heterogeneous sensor data, improve the accuracy and reliability of water quality parameter estimation, and verify the technical advantages of the data fusion method of the present application.

[0085] Step 3, based on the consistent water quality parameter estimation result, a double-layer cognitive structure combining knowledge graph and causal model is constructed to realize water pollutant propagation understanding and tracing;

[0086] According to an embodiment of the present application, this step uses a double-layer cognitive structure algorithm of knowledge graph and causal model to process spatio-temporal water quality data and knowledge base, and generates water pollutant propagation understanding and tracing results. Specifically, the following sub-steps are included:

[0087] Step 3.1, water environment knowledge graph construction and update;

[0088] This sub-step inputs water environment field knowledge and historical monitoring data, and outputs a structured water environment knowledge base through a knowledge graph construction and update algorithm. The method provided by the present application first integrates multi-source knowledge (such as hydrological geographical knowledge, pollutant characteristics, industrial discharge characteristics, and ecological system characteristics), establishes the ontology structure of the water environment field knowledge graph, defines entity types, relationship types, and attributes. In addition, structured knowledge is extracted from literature, reports, and historical data through knowledge extraction technology to fill the knowledge graph. It should be noted that the system designs an incremental learning mechanism to continuously update and improve the content of the knowledge graph based on newly acquired monitoring data and observation results, realizing the dynamic evolution of the knowledge base. Optionally, the system can also apply a knowledge reasoning mechanism to deduce implicit relationships and rules based on existing knowledge, further enriching the semantic content of the knowledge graph.

[0089] As shown, Figure 4 The improvement in the size and accuracy of the knowledge base before and after the system runs for one year verifies the effectiveness of the knowledge graph self-evolution mechanism of the present application. The histogram intuitively shows that through incremental learning and knowledge reasoning technology, the system can continuously learn and extract knowledge from new monitoring data, constantly enrich and optimize the knowledge base content, form a positive cycle of data and knowledge, and realize the continuous appreciation of knowledge assets.

[0090] Step 3.2, knowledge-enhanced causal relationship reasoning;

[0091] The sub-step inputs spatiotemporal water quality data and water environment knowledge graph, and outputs water body pollutant causal network through a knowledge-constrained causal discovery algorithm. It should be noted that the system combines the domain constraints in the knowledge graph and the observation data to construct a causal relationship model of the propagation and mutual influence of water body pollutants. Specifically, the system first identifies potential causal associations using conditional independence testing in spatiotemporal data, and then verifies and corrects them in combination with the physical and chemical laws in the knowledge graph to eliminate false correlations. By integrating data-driven statistical associations and knowledge-driven causal constraints, the system generates a water body pollutant causal network with physical meaning, revealing the propagation mechanism and influence path among pollutants. In some embodiments, the system can also use a causal relationship verification method based on counterfactual reasoning to further improve the reliability of causal discovery.

[0092] According to embodiments of the present application, the knowledge-constrained causal discovery algorithm is implemented as follows: First, the algorithm uses a time series-based conditional independence test method (such as a kernel-based conditional independence test) to preliminarily identify the dependency relationships between variables; then, it extracts relevant physical and chemical laws (such as degradation conditions for certain pollutants, reaction mechanisms, etc.) from the knowledge graph as prior constraints; next, the algorithm constructs a scoring function containing domain constraints to evaluate candidate causal structures, which considers both data fitting and knowledge consistency; finally, it uses optimization methods such as greedy search or MCMC sampling to find the optimal causal graph structure. It should be understood that in specific application scenarios, for example, when detecting an abnormal increase in ammonia nitrogen concentration in a river section, the algorithm can combine upstream discharge records, hydrological conditions, and biochemical reaction knowledge to infer the possible pollution sources (such as industrial discharge or agricultural non-point source pollution) and their propagation paths, providing accurate information for pollution control.

[0093] Step 3.3, pollution source tracking and early warning;

[0094] This sub-step inputs water body pollution causal network and real-time monitoring data, and outputs pollution source location and propagation path through causal reasoning and counterfactual analysis algorithm. According to the method provided by the present application, the system designs a reverse reasoning mechanism based on the established causal network to trace back possible pollution sources from observed pollution phenomena. Specifically, the system models the pollution source tracking problem as an inference problem on a Bayesian network, combines with a hydrodynamic model for constraint, and calculates the probability distribution of each potential source point through a back propagation algorithm. In addition, the system uses counterfactual analysis to evaluate the reasonableness of different pollution source hypotheses, simulates the pollutant diffusion process under different source release scenarios, and matches with actual observations to further improve the accuracy of source tracking. Optionally, the system can also use a particle filtering-based dynamic source tracking method, which is particularly suitable for handling nonlinear diffusion processes of pollutants under complex hydrological conditions.

[0095] Step 3.4, cognitive result feedback and perception optimization;

[0096] This sub-step inputs pollution source tracing results and propagation prediction, and outputs updated monitoring strategies through a perception resource optimization algorithm. As can be seen, the system dynamically adjusts the monitoring strategies of the perception layer based on the analysis results of the cognitive layer, forming a closed-loop feedback mechanism. Specifically, the system re-allocates monitoring resources based on the identified pollution source locations, pollutant propagation paths, and key areas, and strengthens the monitoring intensity of high-risk areas and key propagation paths. In addition, the system predicts the diffusion trend of pollutants based on the causal model and deploys monitoring resources to the areas that may be affected in advance to achieve predictive monitoring. It should be noted that this cognitive-guided perception mechanism enables the monitoring network to more efficiently allocate resources and improve the response speed and coverage efficiency of pollution incidents. In some embodiments, the system can also use an attention mechanism-based resource allocation strategy to automatically identify and focus on the water area that needs monitoring the most.

[0097] Step 4, based on the understanding of water pollutant propagation and the output of the tracing analysis, the system realizes the continuous optimization of the system configuration through a multi-objective reinforcement learning framework;

[0098] According to another embodiment of the present application, this step uses a multi-objective reinforcement learning framework algorithm to process historical performance data and environmental change information to generate a continuous optimization strategy for system configuration. Specifically, the following sub-steps are included:

[0099] Step 4.1, system performance evaluation and target function construction;

[0100] This sub-step inputs system historical performance indicators and user requirements, and outputs a balanced target function through a multi-objective weight optimization algorithm. It should be noted that the system defines a set of key performance indicators, including detection accuracy, coverage range, energy efficiency, time delay, and cost, to comprehensively evaluate the running status of the water quality monitoring system. Therefore, according to the demand characteristics and user preferences of different application scenarios, the system dynamically adjusts the weight coefficients of each indicator to construct a comprehensive performance evaluation function. This evaluation function reflects the overall performance level of the current system configuration, providing a target for subsequent optimization. In some embodiments, the system can also use a multi-index fusion method based on fuzzy comprehensive evaluation to better handle the fuzzy relationships and uncertainties between performance indicators.

[0101] Step 4.2, environmental adaptability learning;

[0102] The sub-step inputs environmental change data and system response history, and outputs an environmental feature-system performance mapping model through an environmental adaptability learning algorithm. According to the method provided in the present application, the system collects and analyzes the correlation data of environmental factors (such as seasonal changes, weather conditions, hydrological feature changes, etc.) affecting water quality monitoring and system performance, and constructs a mapping relationship model between environmental features and system performance. It should be understood that the model can predict the expected performance of different system configurations under specific environmental conditions, providing a decision basis for system adaptability adjustment. In addition, the system continuously updates and improves this mapping relationship through a continuous learning mechanism, improving the adaptability to new environmental conditions. Optionally, the system can use transfer learning technology to utilize the knowledge accumulated in known water environments to accelerate the adaptation process to new water environments.

[0103] Step 4.3, multi-objective reinforcement learning optimization;

[0104] The sub-step inputs the system target function and the environmental feature-system performance mapping model, and outputs the optimized system configuration strategy through a multi-objective reinforcement learning algorithm. According to the embodiments of the present application, the system models the system configuration optimization problem as a multi-objective Markov decision process, where the state space includes environmental conditions and the current system configuration, the action space includes adjustable system parameters, and the reward function is based on the target function defined in step 4.1. Through a multi-objective deep reinforcement learning algorithm, the system learns the optimal parameter adjustment strategy to automatically select the appropriate system configuration under different environmental conditions, achieving continuous optimization of performance. It can be seen that this learning process uses techniques such as experience replay and target network to improve learning efficiency and stability.

[0105] The specific implementation of the multi-objective deep reinforcement learning algorithm is as follows: the algorithm uses a decomposition-based multi-objective Q-learning architecture to decompose the multi-objective optimization problem into multiple single-objective sub-problems. First, the system constructs a multi-head neural network, with each output head corresponding to an optimization target (such as detection accuracy, energy consumption, etc.); then, a deep Q-network (DQN, DeepQ-Network) architecture is used to learn the state-action value function, while an experience replay buffer is used to store transition samples; next, the multi-objective Pareto dominance principle is used to evaluate and select different system configuration strategies; finally, a soft update mechanism is used to periodically update the target network parameters to improve learning stability. In specific application scenarios, such as in environments with large seasonal changes in reservoirs, the algorithm can automatically adjust parameters such as monitoring frequency, sensor type, and sampling location according to changes in environmental features such as water temperature and dissolved oxygen, optimizing system energy consumption and maintenance costs while ensuring detection accuracy.

[0106] Step 4.4, knowledge distillation and lightweight deployment;

[0107] The sub-step inputs a complex optimization model and edge device constraints, and outputs a lightweight decision model through a knowledge distillation algorithm. According to an embodiment of the present application, in order to realize system optimization decision on an edge device with limited computing resources, the system adopts a knowledge distillation technology to transfer the knowledge of the complex reinforcement learning model trained in step 4.3 to a lightweight model with simple structure. Specifically, through a teacher-student network framework, the lightweight student network imitates the decision-making behavior of the complex teacher network, reducing the computational complexity and memory requirement while maintaining the key decision-making ability. It should be understood that such a lightweight model can be directly deployed on an edge device to realize localized system adaptive adjustment, reduce dependence on the central server, and improve system response speed and robustness. Optionally, the system can also use model quantization and structure pruning technologies to further reduce the hardware resource requirement of model deployment.

[0108] According to the method provided by the present application, the specific implementation of the knowledge distillation algorithm is as follows: first, the complex deep reinforcement learning model trained in step 4.3 is used as a teacher network, and a lightweight neural network with reduced parameter quantity is constructed as a student network; then, a large number of state samples are generated, and the teacher network is used to generate corresponding Q values or policy distribution; then, a comprehensive loss function is designed, including the distillation loss (such as KL divergence) between the student network prediction and the teacher network output, and the task loss of the student network interacting with the environment; finally, the student network parameters are optimized through gradient descent, so that the key decision-making ability of the teacher network is retained while being lightweight. It should be noted that in specific application scenarios, for example, on energy-limited edge devices such as water quality monitoring buoys, the lightweight model can autonomously adjust the sampling frequency, sensor working mode and data transmission strategy according to the local observed water quality parameter changes, realizing intelligent resource configuration optimization without relying on cloud computing.

[0109] As shown in Figure 5 Through knowledge distillation and lightweight deployment technology, the deployment of the system on the edge device realizes optimization in response time and communication load compared with the central server deployment. The column chart intuitively shows the technical advantages of the edge computing architecture, verifying the effectiveness of the lightweight deployment method of the present application. By sinking the optimization decision-making ability to the edge device, the system not only improves the response speed, but also reduces the dependence on network bandwidth, enhancing the robustness and reliability of the system.

[0110] As shown in Figure 6As shown, the present application is superior to the traditional monitoring system in five key dimensions: detection accuracy, coverage, pollution source tracing, environmental adaptability, and energy efficiency. The radar chart comprehensively demonstrates the comprehensive technical advantages of the system, verifying the breakthrough in water quality monitoring technology achieved through the knowledge-physical dual fusion architecture. The balanced improvement of various performance indicators fully proves the advancement and practicality of the technical solution of the present application.

[0111] A water body pollutant concentration detection system based on big data analysis is used to perform the water body pollutant concentration detection method based on big data analysis described above, comprising:

[0112] A multi-level perception network module is used to construct a stereoscopic monitoring network including a fixed monitoring station layer, a water surface unmanned ship layer, and an underwater robot layer.

[0113] A multi-modal data fusion module is used to fuse and process multi-source heterogeneous sensing data based on a knowledge-constrained data assimilation algorithm, and output unified water quality parameter evaluation results.

[0114] A cognitive reasoning module is used to construct a double-layer cognitive architecture of a knowledge graph and a causal model, and realize pollution propagation mechanism analysis and pollution source positioning.

[0115] An adaptive optimization module is used to dynamically adjust and continuously optimize system operation parameters based on a multi-objective reinforcement learning framework.

[0116] The embodiments of the present application are described above, but the embodiments are not limited to the specific implementation described above, which is only illustrative and not limiting. Those skilled in the art can make more forms of equivalent embodiments under the inspiration of the embodiments, which are all within the protection scope of the embodiments.

Claims

1. A method for detecting concentration of pollutants in a water body based on big data analysis, characterized in that, The method comprises the following steps: a multi-level perception network is constructed, the multi-level perception network comprising a fixed monitoring station layer, a water surface unmanned ship layer and an underwater robot layer; based on the multi-level perception network, multi-modal data fusion is performed by using a knowledge-constrained data assimilation algorithm to generate consistent water quality parameter estimation; based on the consistent water quality parameter estimation result, a double-layer cognitive structure combining a knowledge graph and a causal model is constructed to realize water pollutant propagation understanding and tracing; based on the water pollutant propagation understanding and tracing output, through a multi-objective reinforcement learning framework, continuous optimization of system configuration is realized.

2. The method for detecting concentration of pollutants in water bodies based on big data analysis according to claim 1, characterized in that, The step of constructing the multi-level perception network comprises: water space is divided based on a hydrological geographical feature clustering algorithm to output an initial monitoring point distribution; a multi-level perception strategy is generated through an interlayer cooperative optimization algorithm; the perception network structure is dynamically updated according to real-time monitoring data and environmental change information by using a self-organizing learning algorithm.

3. The method of claim 1, wherein, The step of multi-modal data fusion comprises: standardization and noise filtering processing are performed on raw data of multiple sensors; data credibility scores are calculated based on a sensor state evaluation algorithm; the standardized multi-modal data set, the data credibility scores and a domain knowledge base are fused by using the knowledge-constrained data assimilation algorithm to generate the consistent water quality parameter estimation.

4. The method for detecting concentration of pollutants in water bodies based on big data analysis according to claim 1, characterized in that, The step of constructing the double-layer cognitive structure combining the knowledge graph and the causal model comprises: a water environment knowledge graph is constructed and updated; a causal relationship model of water pollutant propagation and mutual influence is constructed in combination with domain constraints in the knowledge graph and observation data; based on the established causal network, a reverse reasoning mechanism is designed to trace the pollution source from the observed pollution phenomenon; the monitoring strategy of the perception layer is dynamically adjusted according to the identified pollution source position, pollutant propagation path and key area.

5. The method of claim 4, wherein the method is based on big data analysis. The causal relationship model adopts a knowledge-constrained causal discovery algorithm, and the knowledge-constrained causal discovery algorithm further comprises: confidence evaluation is performed on the preliminary identified variable dependency relationship; time delay analysis is combined to determine the timing characteristics of the causal relationship; a cross-validation method is used to verify the stability and generalization ability of the causal graph structure.

6. The method for detecting concentration of pollutants in water bodies based on big data analysis according to claim 1, characterized in that, The step of the multi-objective reinforcement learning framework comprises: system key performance indicators are defined, and a comprehensive performance evaluation function is constructed; a mapping relationship model between environmental characteristics and system performance is constructed; the system configuration optimization problem is modeled as a multi-objective Markov decision process, and an optimal parameter adjustment strategy is learned by using a multi-objective deep reinforcement learning algorithm; knowledge distillation technology is used to transfer the knowledge of a complex reinforcement learning model to a lightweight model with simple structure, realizing localized system adaptive adjustment on an edge device.

7. The method of claim 6, wherein the method is based on big data analysis. The multi-objective deep reinforcement learning algorithm adopts a decomposition-based multi-objective Q learning architecture, comprising: a multi-head neural network is constructed, each output head corresponding to an optimization target; a state action value function is learned through a deep Q network architecture, and an experience replay buffer is used to store transition samples; a multi-objective Pareto dominance principle is used to evaluate and select different system configuration strategies; a soft update mechanism is used to periodically update target network parameters to improve learning stability.

8. The method of claim 6, wherein the method is characterized by, The knowledge distillation technology comprises: The trained complex deep reinforcement learning model is used as a teacher network, and a lightweight neural network with reduced parameters is used as a student network; A large number of state samples are generated, and the teacher network is used to generate corresponding Q values and policy distribution; A comprehensive loss function is designed, including the distillation loss between the student network prediction and the teacher network output, and the task loss of the student network interacting with the environment; The student network parameters are optimized by gradient descent, so that the student network can be lightweight while retaining the key decision-making ability of the teacher network.

9. The method of claim 3, wherein the method is characterized by, The sensor state evaluation algorithm comprises: The reliability score of each sensor data is dynamically calculated by considering the service life, historical accuracy and environmental adaptability of the sensor; An abnormal data point is identified through a multiple test mechanism including physical constraint test, time continuity test and spatial correlation test; According to the type of anomaly, corresponding processing is carried out, including interpolation, marking and rejection operations; An abnormal detection algorithm based on statistical learning is used to automatically identify and handle various types of sensor failure modes. 10.A water body pollutant concentration detection system based on big data analysis, characterized in that, A water pollution concentration detection method based on big data analysis is used to execute any one of claims 1-9, comprising: A multi-level perception network module is used to construct a three-dimensional monitoring network including a fixed monitoring station layer, a water surface unmanned ship layer and an underwater robot layer; A multi-modal data fusion module is used to fuse and process multi-source heterogeneous sensor data based on a knowledge-constrained data assimilation algorithm, and output a unified water quality parameter evaluation result; A cognitive reasoning module is used to construct a double-layer cognitive architecture of knowledge graph and causal model, and realize pollution propagation mechanism analysis and pollution source positioning; An adaptive optimization module is used to dynamically adjust and continuously optimize the system operation parameters based on a multi-objective reinforcement learning framework.

Images