River ecological restoration data dynamic monitoring and scheme optimization method and system

Abstract

The invention relates to the technical field of ecological environment monitoring and intelligent decision making, and discloses a river ecological restoration data dynamic monitoring and scheme optimization method and system, and the method comprises the steps: obtaining the five-dimensional data of hydrology, water quality, bottom mud, biology and bank zones through a multi-modal sensing terminal; constructing a hydrodynamic-water quality-ecology coupled dynamic graph neural network model, and deducing an ecological health index of each section in real time; based on a reinforcement learning strategy generator, combining a degradation section identification result and a restoration knowledge base, outputting an optimization restoration measure combination and issuing the optimization restoration measure combination to an execution unit; and updating the strategy network through effect feedback to realize continuous optimization. The system comprises a sensing terminal cluster, a data preprocessing unit, an evaluation engine, an identification module, a knowledge base, a strategy generator, an instruction issuing unit and a feedback module. According to the method and the device, the transformation of river channel restoration from passive response to active intelligent regulation is realized, and the restoration accuracy and timeliness are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of ecological environment monitoring and intelligent decision-making technology, and in particular to a method for dynamic monitoring of river ecological restoration data and optimization of solutions. Background Technology

[0002] With the deepening of ecological civilization construction, river ecological restoration has become an important part of water environment management. Traditional river restoration relies heavily on manual inspections and static assessments, making it difficult to achieve real-time perception and dynamic response to complex hydrological, water quality, and biological elements. This results in delayed restoration plans, inefficient resource allocation, and difficulty in quantifying restoration effectiveness. In recent years, ecological restoration assessment and optimization technologies that integrate multi-source monitoring data and intelligent algorithms have gradually emerged. However, in the context of rivers—a typical linear, flowing, and variable water system—there is still a lack of comprehensive solutions that combine high spatiotemporal resolution, multi-element coupled modeling, closed-loop feedback mechanisms, and adaptive optimization capabilities.

[0003] Among these, dynamic monitoring and optimization methods for ecological restoration driven by multi-source data have become a research hotspot. These methods aim to construct a technical framework covering the entire chain of "perception-assessment-decision-execution" by integrating remote sensing, IoT sensors, hydrological models, and artificial intelligence algorithms, to support the precise formulation and dynamic adjustment of ecological restoration measures. However, existing technologies still face multiple challenges in practical applications: on the one hand, river systems are characterized by strong fluidity, hydrological periodicity, and ecological connectivity, with significant spatiotemporal heterogeneity in processes such as water quality evolution, pollutant migration, and biological community succession; on the other hand, most current monitoring systems focus only on single-dimensional indicators (such as dissolved oxygen, COD, or vegetation cover), lacking the ability to model the coupling relationships of multiple elements such as hydrodynamics, water quality, sediment, and organisms, making it difficult to accurately depict the overall state of the ecosystem.

[0004] The aforementioned limitations make it difficult for existing solutions to support the transformation and upgrading of river ecological restoration from "experience-based judgment" to "data-driven" and from "local intervention" to "system regulation." There is an urgent need for an intelligent decision-making system for river ecological restoration that can integrate multi-dimensional heterogeneous data and has dynamic state perception and adaptive optimization capabilities. Summary of the Invention

[0005] This invention aims to address the core problems in existing technologies, such as delayed restoration, inefficient resource allocation, and unquantifiable results, caused by the lack of high spatiotemporal resolution perception of river water bodies, coupled modeling capabilities for hydrodynamic, water quality, and ecological factors, and the absence of adaptive closed-loop optimization mechanisms for restoration schemes. To this end, this invention proposes a method and system for dynamic monitoring of river ecological restoration data and optimization of restoration schemes.

[0006] According to one aspect of the present invention, a method for dynamic monitoring of river ecological restoration data and optimization of schemes is provided, comprising:

[0007] Multi-dimensional raw monitoring data of the river ecosystem is obtained by deploying multi-modal sensing terminals throughout the river. The multi-dimensional raw monitoring data includes hydrological flow data, water quality physicochemical index data, sediment pollutant concentration data, aquatic biological community structure data, and riparian vegetation cover status data.

[0008] The multi-dimensional raw monitoring data are subjected to spatiotemporal alignment and outlier removal to generate a standardized time-series dataset of river ecological status.

[0009] Based on the aforementioned river ecological state time series dataset, a dynamic graph neural network model of the hydrodynamic-water quality-ecological coupling process is constructed. This model uses river cross-sections as nodes and water flow direction and ecological connectivity as edges, and uses gated recurrent units and attention mechanisms to transmit spatiotemporal messages, and performs real-time inference of the ecological health index of each cross-section.

[0010] Based on the ecological health index output by the dynamic graph neural network model, and combined with the preset ecological restoration target threshold system, the key degraded sections and dominant limiting factors in the current river ecosystem are identified.

[0011] The restoration strategy knowledge base is invoked. The restoration strategy knowledge base contains a set of structured restoration measures, including vegetation configuration schemes, artificial wetland layout parameters, aeration and oxygenation intensity, sediment dredging range, ecological floating island density and hydrodynamic control instructions.

[0012] The location information, dominant limiting factor type, current hydrological situation stage, and ecological restoration target of the key degraded section are encoded into a state vector, which is then input into a restoration strategy generator based on a deep deterministic strategy gradient algorithm. The output is an optimized combination of restoration measures and its implementation parameters for the current working condition.

[0013] The optimized remediation measures and their implementation parameters are sent to the corresponding execution units, which include an automatic dosing device, an adjustable aeration device, an ecological floating island deployment platform, and a shoreline irrigation control system.

[0014] Within a set time window after the remediation measures are implemented, the multi-dimensional original monitoring data is collected again, the change in the ecological health index before and after the remediation is calculated, and the change is fed back to the remediation strategy generator as a reward signal to update its strategy network parameters, thereby achieving continuous adaptive optimization of the remediation strategy.

[0015] As one embodiment of the present invention, the multimodal sensing terminal includes a fixed water quality buoy array deployed longitudinally along the river channel, a river patrol drone equipped with a multispectral camera and lidar, an embedded shore-based hydrological station, and an underwater bioacoustic monitoring probe.

[0016] As one embodiment of the present invention, the dynamic graph neural network model of the hydrodynamic-water quality-ecological coupling process adopts a three-layer architecture: the first layer is a node feature encoding layer, which maps the multi-dimensional monitoring data of each river section into a 128-dimensional node embedding vector; the second layer is a spatiotemporal message passing layer, which uses a gated recurrent unit to perform time-series modeling of material transport and ecological impact between upstream and downstream sections, and uses an attention mechanism to weight the information contribution of different adjacent sections; the third layer is an ecological health index decoding layer, which outputs the comprehensive ecological health index of each section, with a value ranging from 0 to 1, and a higher value indicating a better ecological state; the comprehensive ecological health index EHI is synthesized through linear weighting, EHI = ×Water quality compliance rate+ ×Biological Integrity Index+ ×hydrological stability coefficient+ × Coastal zone structural integrity, where weighting coefficients , , , The weights are obtained from a preset weight table based on the river functional zoning, and the weight coefficients are dynamically adjusted according to the river functional zoning.

[0017] As one embodiment of the present invention, the ecological restoration target threshold system is set according to the national "Surface Water Environmental Quality Standard" and the local river ecological function positioning, and is divided into four functional zones: drinking water source protection zone, ecological conservation zone, landscape and recreation zone, and agricultural water use zone. Each functional zone corresponds to different water quality targets, biodiversity requirements, and hydrological fluctuation tolerance. The identification of key degraded sections adopts the sliding window detection method. When the ecological health index of three consecutive cross sections is lower than the threshold of the corresponding functional zone and shows a downward trend, it is determined to be a degraded section. The dominant limiting factor is determined by ranking the importance of features. Specifically, the integrated gradient method is used to calculate the contribution of each monitoring variable to the ecological health index prediction result. The variable with the highest contribution is the dominant limiting factor.

[0018] In one embodiment of the present invention, the restoration strategy knowledge base is stored in a distributed database. Each restoration measure record includes a measure type identifier, an applicable ecological problem label, hydrological constraints, implementation cost, expected restoration cycle, and historical effectiveness verification data. The restoration strategy generator includes an actor network and a commentator network. The actor network receives the current state vector and outputs a continuous action vector, which, after discretization and decoding, corresponds to a specific combination of restoration measures. The commentator network evaluates the long-term cumulative reward of the action in the current state. The long-term cumulative reward is linearly weighted by the improvement of the ecological health index, the inverse of the measure implementation cost, and the resource consumption efficiency.

[0019] In one embodiment of the present invention, the execution unit connects to the central control platform via an industrial Internet of Things (IoT) protocol and receives structured instructions containing the type of measure, location of action, intensity of action, and duration of action; the automatic dosing device accurately dispenses microbial agents or flocculants according to the instructions; the adjustable aeration equipment dynamically adjusts the number of aeration heads opened and the air supply pressure according to the water depth and dissolved oxygen demand; the ecological floating island deployment platform is equipped with prefabricated floating island units, which automatically deploy and anchor on the water surface of the degraded section according to the instructions; the shoreline irrigation control system links soil moisture sensors and meteorological forecast data to start drip irrigation or sprinkler irrigation according to the water demand pattern of vegetation.

[0020] As one embodiment of the present invention, the calculation formula of the reward signal is: Reward signal = α × ΔEHI + β / C + γ × η, where ΔEHI is the difference in ecological health index before and after restoration, C is the total implementation cost of this restoration measure, η is the water resource and energy utilization efficiency, and α, β, and γ are preset weight coefficients, with values ​​of 0.6, 0.3, and 0.1, respectively; the update of the policy network parameters adopts an experience replay buffer mechanism, in which a batch of historical interaction samples are randomly selected from the buffer for gradient backpropagation in each iteration to ensure the stability and generalization ability of policy learning.

[0021] According to another aspect of the present invention, a dynamic monitoring and scheme optimization system for river ecological restoration data is provided, comprising:

[0022] Multimodal sensing terminals are used to acquire multi-dimensional raw monitoring data of river ecosystems;

[0023] The data preprocessing unit is used to perform spatiotemporal alignment and outlier removal on the raw monitoring data to generate a standardized time-series dataset of river ecological status.

[0024] A dynamic graph neural network evaluation engine is used to build dynamic graph neural network models and deduce the ecological health index of each section.

[0025] The degradation zone identification module is used to identify key degradation zones and dominant limiting factors based on the ecological health index and a preset ecological restoration target threshold system.

[0026] A remediation strategy knowledge base for storing a structured set of remediation measures;

[0027] The restoration strategy generator takes the key degraded section information, dominant limiting factors, current hydrological situation and ecological restoration goals as input, and outputs an optimized combination of restoration measures and their implementation parameters.

[0028] The execution instruction issuing unit issues the optimized repair measures combination and its implementation parameters to the corresponding execution unit;

[0029] The effect verification and feedback module collects new monitoring data after the remediation measures are implemented, calculates the change in the ecological health index, and feeds back the change as a reward signal to the remediation strategy generator to update its strategy network parameters.

[0030] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0031] Compared to traditional single-point water quality monitoring, this invention achieves unified spatiotemporal scale monitoring of five major elements of river hydrology, water quality, bottom sediment, organisms, and shoreline through multimodal terminal clusters and standardized spatiotemporal grids, thus solving the problem of fragmented sensing data.

[0032] Compared to static or single-element models, the dynamic graph neural network model proposed in this invention achieves accurate prediction of the spatiotemporal evolution of ecological health status by simulating the transport and transformation process of materials along the river network.

[0033] Compared to open-loop repair with a fixed solution, this invention uses reinforcement learning to create a closed loop, enabling the repair strategy to continuously and adaptively optimize based on actual results. This ensures repair effectiveness while achieving optimized control of cost and resource consumption.

[0034] Attached Figure Description Figure 1 This is a flowchart of the overall technical solution of the present invention. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this invention clearer and more complete, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention in any way.

[0036] A method for dynamic monitoring of river ecological restoration data and optimization of schemes includes the following steps:

[0037] S1: Obtain multi-dimensional raw monitoring data of the river ecosystem through multimodal sensing terminals deployed throughout the river basin;

[0038] S2: Perform spatiotemporal alignment and outlier removal on the multi-dimensional raw monitoring data to generate a standardized time-series dataset of river ecological status;

[0039] S3: Based on the aforementioned river ecological state time series dataset, a dynamic graph neural network model (DGNN) for the hydrodynamic-water quality-ecological coupling process is constructed. This model uses river cross-sections as nodes. The construction of edges includes two parts: directed edges based on the direction of water flow, pointing from the upstream cross-section to the downstream cross-section; and undirected edges based on ecological connectivity. When the correlation coefficient of sediment pollutant concentration between two cross-sections is greater than 0.7, or the similarity of aquatic vegetation types is greater than 80%, an undirected edge is established between them. The model also uses gated recurrent units and attention mechanisms to transmit spatiotemporal messages and perform real-time inference of the ecological health index of each cross-section.

[0040] S4: Based on the ecological health index output by the dynamic graph neural network model, and combined with the preset ecological restoration target threshold system, identify the key degraded sections and dominant limiting factors in the current river ecosystem.

[0041] S5: Call the restoration strategy knowledge base, which contains a set of structured restoration measures including vegetation configuration schemes, artificial wetland layout parameters, aeration and oxygenation intensity, sediment dredging range, ecological floating island density and hydrodynamic control instructions.

[0042] S6: Encode the location information, dominant limiting factor type, current hydrological situation stage, and ecological restoration target of the key degraded section into a state vector, input it into the restoration strategy generator based on the Deep Deterministic Strategy Gradient Algorithm (DDPG), and output the optimized restoration measure combination and its implementation parameters for the current working condition; the current hydrological situation stage is divided into high-water period, normal-water period, and low-water period according to the historical flow data of the river channel, and its encoding information can be obtained from the hydrological database or real-time monitoring data;

[0043] S7: The optimized remediation measures combination and its implementation parameters are sent to the corresponding execution unit, the execution unit including an automatic dosing device, an adjustable aeration device, an ecological floating island deployment platform and a shoreline irrigation control system;

[0044] S8: Within the time window set after the remediation measures are implemented, the multi-dimensional original monitoring data is collected again, the change in the ecological health index before and after the remediation is calculated, and the change is fed back to the remediation strategy generator as a reward signal to update its strategy network parameters and realize the continuous adaptive optimization of the remediation strategy.

[0045] In step S1, the multimodal sensing terminal includes a fixed water quality buoy array deployed longitudinally along the river channel, a river patrol drone equipped with a multispectral camera and lidar, an embedded shore-based hydrological station, and an underwater bioacoustic monitoring probe. The fixed water quality buoy array is deployed at 500-meter intervals, and each buoy integrates a dissolved oxygen sensor, ammonia nitrogen sensor, total phosphorus sensor, turbidity meter, flow velocity and direction meter, and water temperature probe, with a sampling frequency of once per minute. The river patrol drone performs two full-coverage flight missions daily along a preset route to acquire images of river surface reflectance, vegetation normalization index, distribution of floating debris, and riverbank erosion traces. The underwater bioacoustic monitoring probe uses passive acoustic monitoring technology to identify fish species, numbers, and activity rhythms, generating a biodiversity index. All sensing terminals are equipped with a Global Navigation Satellite System (GNSS) positioning module and a time synchronization module to ensure that the collected data has a unified spatiotemporal reference.

[0046] In step S2, the specific operations for spatiotemporal alignment and outlier removal of the multi-dimensional raw monitoring data include: First, based on the timestamps and geographical location information of each sensing terminal, heterogeneous data are mapped to a unified river cross-section grid system. This grid system uses the river centerline as a reference and divides the data into cross-section units every 100 meters along the water flow direction. Each cross-section unit is laterally divided into three sub-regions: left bank, main channel, and right bank. Second, time interpolation alignment is performed on data from different sensors within the same cross-section unit. Cubic spline interpolation is used to unify low-frequency data, such as UAV imagery, and high-frequency data, such as buoy readings, to one time point per minute. Third, anomaly detection is performed on each variable at each time point. An unsupervised algorithm based on local outlier factors is used to identify observations that deviate from the normal fluctuation range and replace them with the weighted average of neighboring time points and neighboring cross-sections. Finally, the processed data is organized into a three-dimensional tensor structure according to cross-section units, time points, and variable types to form a standardized river ecological state time series dataset.

[0047] In step S3, the dynamic graph neural network model of the hydrodynamic-water quality-ecological coupling process adopts a three-layer architecture. The first layer is the node feature encoding layer, which maps the multidimensional monitoring data of each river section into a 128-dimensional node embedding vector. This mapping is implemented through a fully connected feedforward network, with the input being the values ​​of all variables within the section unit at the current time step, and the output being a fixed-dimensional embedding vector. The second layer is the spatiotemporal message passing layer, which uses gated recurrent units to perform time-series modeling of material transport and ecological impact between upstream and downstream sections, and uses an attention mechanism to weight the information contribution of different adjacent sections. The third layer is the ecological health index decoding layer, which outputs the comprehensive ecological health index of each section. The attention mechanism adopts a variant based on graph attention networks, which is applied to the set of upstream adjacent sections. and downstream adjacent cross sections Calculate the attention coefficient separately. For upstream messages, the attention coefficient is calculated as follows: in, Represents a node The set of all upstream adjacent cross sections; They are nodes and In the hidden state of the previous time step ( (to hide the state dimension). The weight matrix is ​​a learnable weight matrix; This is a learnable attention vector; This represents a vector concatenation operation. Downstream attention coefficients. The calculation method is similar, but it uses independent learnable parameters. and The final state of node i at time t is:

[0048] .

[0049] In step S4, the ecological restoration target threshold system is set according to the national "Surface Water Environmental Quality Standard" and the local river ecological function positioning, and is divided into four functional zones: drinking water source protection zone, ecological conservation zone, landscape and recreation zone, and agricultural water use zone. Each functional zone corresponds to different water quality targets, biodiversity requirements, and hydrological fluctuation tolerance. The identification of key degraded sections uses a sliding window detection method; when the ecological health index of three consecutive cross-sections is lower than the corresponding functional zone threshold and shows a downward trend, it is determined to be a degraded section. The dominant limiting factor is determined by ranking the importance of features, specifically using an integrated gradient method to calculate the contribution of each monitoring variable to the ecological health index prediction result. The formula for calculating the integrated gradient is: in, This is the input feature vector of the current cross-sectional element. The baseline eigenvector is the moving average of all variables over the past 30 days during the same hydrological condition phase at this cross section. This is the ecological health index prediction function of a dynamic graph neural network. Integral variable. It varies between 0 and 1. For each monitored variable i, the variable corresponding to the largest absolute value of its IG value is the dominant limiting factor.

[0050] In step S5, the restoration strategy knowledge base is stored in a distributed database, and its construction begins with the structured extraction of historical successful restoration cases, expert experience, and literature data. Each restoration measure record includes a measure type identifier, applicable ecological problem tags, hydrological constraints, implementation costs, expected restoration cycle, and historical effectiveness verification data. The restoration strategy knowledge base is updated through a feedback loop: the actual effect data after each restoration action forms a new case record, which, after evaluation, is used to correct the "expected effectiveness" and "cost model" parameters of the original record, thus achieving self-evolution of the knowledge base.

[0051] In step S6, the restoration strategy generator includes an actor network and a commentator network. The actor network receives a state vector composed of the degraded section location code, the dominant limiting factor type code, the current hydrological situation stage code, the ecological restoration target code, and the ecological health index sequence of the most recent seven days. It then outputs a continuous action vector. The dimension of the action vector equals the total number of restoration measures, and each dimension ranges from 0 to 1, representing the intensity of the measure. This action vector, after discretization and decoding, corresponds to a specific combination of restoration measures. The decoding rule is as follows:

[0052] For the k-th dimension action value ,like If ≤ 0.5, the measure will not be activated; if If the value is greater than 0.5, the measure will be activated, and its implementation parameters will be as follows: Calculated according to the preset linear mapping function: in, and The predefined physical feasibility lower and upper bounds for the k-th parameter of this remediation measure are stored in the remediation strategy knowledge base.

[0053] The critic network evaluates the long-term cumulative reward Q value under a given state and action. The Q value target is calculated by weighting three parts: the improvement of the ecological health index, the inverse of the implementation cost, and the resource utilization efficiency.

[0054] In step S7, the execution unit connects to the central control platform via the Industrial Internet of Things (IIoT) protocol and receives structured instructions containing the type of measure, location of action, intensity of action, and duration of action. The automatic dosing device accurately dispenses microbial agents or flocculants according to the instructions, with a dosing accuracy error not exceeding 5%. The adjustable aeration equipment dynamically adjusts the number of aeration heads and the air supply pressure based on water depth and dissolved oxygen requirements, with the air supply pressure adjustment range from 0.1 to 0.5 MPa. The ecological floating island deployment platform, equipped with prefabricated floating island units, automatically deploys and anchors on the surface of the degraded section of water according to instructions, with the anchoring depth automatically adjusted according to the water depth. The shoreline irrigation control system, linked to soil moisture sensors and weather forecast data, initiates drip irrigation or sprinkler irrigation according to the vegetation's water requirements, with an irrigation volume error not exceeding 10%.

[0055] In step S8, the reward signal Q is calculated using the formula: Q = α * ΔEHI + β / C + γ * η, where ΔEHI is the difference in ecological health index before and after restoration, C is the total implementation cost of this restoration measure, η is the water resource and energy utilization efficiency, and α, β, and γ are preset weight coefficients, with values ​​of 0.6, 0.3, and 0.1, respectively. The update of the strategy network parameters adopts an experience replay buffer mechanism with a buffer capacity of 100,000 interaction records. In each iteration, a batch of 256 historical interaction samples is randomly selected from the buffer for gradient backpropagation. Both the actor network and the commentator network adopt a dual-network structure. The target network parameters gradually approach the main network parameters through a soft update method, with the update coefficient set to 0.005.

[0056] The system comprises a multimodal sensing terminal cluster, a data preprocessing unit, a dynamic graph neural network evaluation engine, a degraded section identification module, a restoration strategy knowledge base, a restoration strategy generator, an execution command issuance unit, and an effect verification and feedback module. All modules are deployed on a cloud-edge collaborative computing platform using a microservice architecture. Edge nodes are deployed at shore-based control stations along the river, responsible for raw data preprocessing and emergency command issuance; the cloud platform handles complex model training, strategy generation, and global optimization. Data transmission employs encrypted communication protocols to ensure the security of monitoring data and control commands. The system supports integration with external business systems such as water conservancy, environmental protection, and municipal administration, enabling the sharing and collaborative governance of ecological restoration data.

[0057] Example 1

[0058] This embodiment focuses on the downstream section of a city's receiving river to specifically illustrate the implementation of the present invention.

[0059] Several multi-functional water quality monitoring buoys were deployed along the river, forming a series of monitoring cross-section nodes. Each node integrates sensors to monitor key parameters such as water temperature, dissolved oxygen, and ammonia nitrogen in real time (sampling frequency: 5 minutes / time). Simultaneously, periodic drone inspections were conducted to assess the condition of the riverbank.

[0060] All monitoring data is uploaded to edge computing nodes. The nodes perform spatiotemporal fusion processing: dividing the data into cross-sectional units centered on the monitoring point locations, mapping multi-source data to a unified spatiotemporal grid; using interpolation to unify the data time granularity; and using anomaly detection algorithms to clean the data, ultimately forming a standard time-series dataset of river ecological status.

[0061] Each monitoring section is defined as a graph node. The edges connecting the nodes are constructed based on the actual water flow direction (from upstream to downstream), forming a directed graph. The weights of the edges are initialized to the water flow transmission time from the upstream section to the downstream section (based on historical flow velocity estimates).

[0062] The node feature vectors contain parameters such as water quality and hydrology. The model employs a structure combining graph attention networks and gated recurrent units to aggregate spatial associations and memorize temporal dynamics. Historical data is used to train the model, enabling it to infer the ecological health index of each cross-section over a future period based on current and historical conditions. This index is a quantified value that integrates multiple sub-indicators, including water quality, biology, hydrology, and shoreline characteristics.

[0063] The system continuously runs the aforementioned model to achieve real-time simulations. When the model simulation results show that the ecological health index of a certain continuous river segment is consistently below its functional zone target threshold and shows a deteriorating trend, the system automatically identifies the area as a "critical degradation segment." At the same time, the model outputs the "dominant limiting factors" (e.g., dissolved oxygen deficiency, ammonia nitrogen exceedance, etc.) that cause this degradation through feature importance analysis (such as ensemble gradient method).

[0064] Once the diagnosis is complete (identifying key degraded sections and dominant limiting factors), the system immediately triggers an intelligent decision-making process. First, the diagnostic results (section location, type of limiting factor), current hydrological situation, ecological restoration goals, and recent ecological health index sequences are encapsulated into a structured "state vector." This vector fully represents the "problem scenario" that needs to be restored.

[0065] The encapsulated state vector is input into a pre-trained reinforcement learning policy generator (using a deep deterministic policy gradient framework). The actor network in the policy generator receives this vector and outputs a continuous action vector. Each dimension of this action vector corresponds to the recommended strength of a potential remediation measure (such as aeration, bacterial inoculation, or deployment of ecological floating islands). Subsequently, the system transforms the action vector into executable, specific remediation measure combinations based on predetermined decoding rules (e.g., activating the measure if the strength exceeds a threshold, and calculating specific parameters using a linear mapping).

[0066] The generated remediation instructions (such as "start the aeration equipment at 80% of the rated power between sections A and B, and add compound microbial agent at a rate of 5L / h") are sent to the corresponding automatic execution units (such as adjustable aerators, automatic dosing devices, ecological floating island deployment platforms, etc.) through the Internet of Things protocol, driving them to complete physical intervention.

[0067] After a preset evaluation period (e.g., 24 hours) has elapsed since the restoration measures were implemented, the system restarts the perception and simulation process: new monitoring data is collected, input into the dynamic relationship network model, and the ecological health index of the area after restoration is calculated.

[0068] The system calculates the change in the ecological health index (ΔEHI) before and after restoration. This change, combined with information such as the cost of implementing the measures, is used to calculate the "reward signal" for this decision. Subsequently, the complete interaction experience (state vector, action vector, reward signal, new state vector) is stored in an experience pool and used to periodically update the network parameters of the policy generator. Through extensive interactive learning, the system's decision-making strategy is continuously optimized, forming a complete adaptive closed loop of "perception-diagnosis-decision-execution-evaluation-learning".

[0069] This embodiment clearly demonstrates a complete technology chain from data to decision-making and then to optimization. The deduction results of the dynamic relationship network model are the sole basis for triggering and driving subsequent intelligent decisions, while the feedback of the decision-making effect, in turn, optimizes the decision-making strategy. The two are tightly coupled through "state vectors" and "reward signals," forming an inseparable synergistic whole, jointly solving the core technical problems of "disconnect between evaluation and decision-making" and "inability to dynamically optimize the solution" in traditional methods, and achieving significant and unexpected technical results.

[0070] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for dynamic monitoring of river ecological restoration data and optimization of schemes, characterized in that, include: S1: Acquire multi-dimensional raw monitoring data of the river ecosystem through multi-modal sensing terminals deployed throughout the river channel; S2: Perform spatiotemporal alignment and outlier removal processing on the multi-dimensional raw monitoring data to generate a standardized time-series dataset of river ecological status; S3: Based on the time-series dataset of river ecological status, construct a dynamic graph neural network model of the hydrodynamic-water quality-ecological coupling process. This model uses river cross-sections as nodes, water flow direction and ecological connectivity as edges, and uses gated recurrent units and attention mechanisms for spatiotemporal message transmission to infer the ecological health index of each cross-section in real time; S4: Based on the derived ecological health index and the preset ecological restoration target threshold system, identify the key degraded sections in the current river ecosystem and use the feature importance ranking method to determine the dominant limiting factors. S5: Call the restoration strategy knowledge base that stores a set of structured restoration measures; S6: Encode the location information, dominant limiting factor type, current hydrological situation stage and ecological restoration target of the key degraded section into a state vector, input it into the restoration strategy generator built based on the deep deterministic strategy gradient algorithm, and output the optimized combination of restoration measures and its implementation parameters; S7: Send the combination of optimized repair measures and its implementation parameters to the corresponding execution unit; S8: Within the time window set after the remediation measures are implemented, the multi-dimensional original monitoring data is collected again and the change in the ecological health index before and after the remediation is calculated. The change is used to form a reward signal, which is fed back to the remediation strategy generator to update its strategy network parameters and realize the continuous adaptive optimization of the remediation strategy.

2. The method for dynamic monitoring and scheme optimization of river ecological restoration data according to claim 1, characterized in that, The preprocessing includes: mapping heterogeneous data to a unified river cross-section grid system based on the timestamps and geographic location information of each sensing terminal. This grid system uses the river centerline as a reference and divides cross-section units along the water flow direction. Each cross-section unit is laterally divided into three sub-regions: left bank, main channel, and right bank. Time interpolation is performed on data from different sensors within the same cross-section unit to align them. Cubic spline interpolation is used to unify low-frequency and high-frequency data to one time point per minute. Anomaly detection is performed on each variable at each time point. An unsupervised algorithm based on local outlier factors is used to identify outliers, and they are replaced with the weighted average of neighboring time points and neighboring cross-sections. The processed data is organized into a three-dimensional tensor structure according to cross-section unit, time point, and variable type to form a standardized time-series dataset of river ecological status.

3. The method for dynamic monitoring and scheme optimization of river ecological restoration data according to claim 2, characterized in that, The construction of the dynamic graph neural network model includes: encoding multidimensional monitoring data of each river section at a certain time point into a 128-node feature vector through a fully connected feedforward network; using the river topology as a graph, aggregating the hidden state information from upstream and downstream adjacent sections at the previous time step using a gated recurrent unit, and distinguishing the importance weights of different neighboring section information through an attention mechanism to update the hidden state of the current section; and decoding and outputting the comprehensive ecological health index of each section based on the hidden state at the final time step. The comprehensive ecological health index is a weighted synthesis of four sub-indicators: water quality compliance rate, biological integrity index, hydrological stability coefficient, and shoreline structure integrity, with the weight coefficients dynamically adjusted according to the river functional zoning.

4. The method for dynamic monitoring and scheme optimization of river ecological restoration data according to claim 3, characterized in that, The identification of key degradation zones and dominant limiting factors includes: using a sliding window detection method, when the ecological health index of three consecutive cross sections is lower than the threshold of the corresponding functional area and shows a downward trend, the area is identified as a key degradation zone; using an integrated gradient method to calculate the contribution of each input monitoring variable in the dynamic graph neural network model to the predicted value of the ecological health index, and determining the variable with the largest absolute value of contribution as the dominant limiting factor.

5. The method for dynamic monitoring and scheme optimization of river ecological restoration data according to claim 4, characterized in that, The restoration strategy knowledge base is a distributed database, in which each restoration measure record includes a measure type identifier, applicable ecological problem label, hydrological condition constraints, implementation cost, expected restoration cycle, and historical effectiveness verification data fields.

6. The method for dynamic monitoring and scheme optimization of river ecological restoration data according to claim 5, characterized in that, The operation of the restoration strategy generator includes: the state vector is composed of codes for the location of the degraded section, the type of the dominant limiting factor, the current hydrological situation stage, the ecological restoration target, and the ecological health index sequence of the most recent seven days; the restoration strategy generator includes an actor network and a commentator network, the actor network receives the state vector and outputs a continuous action vector, each dimension of which corresponds to the activation intensity of a restoration measure; the continuous action vector is converted into specific combinations of restoration measures and implementation parameters through a discretization decoding rule; the decoding rule is: For the k-th dimension action value ,like If ≤ 0.5, the measure will not be activated; if If the value is greater than 0.5, the measure is activated, and its parameters are calculated. The calculation formula is as follows: in, and The physical feasibility lower and upper bounds of the k-th parameter of the remediation measure are predefined in the knowledge base; the commentator network evaluates the long-term cumulative reward Q value under a given state and action, and the Q value target is calculated by weighting three parts: the improvement of the ecological health index, the inverse of the implementation cost, and the resource utilization efficiency.

7. The method for dynamic monitoring and scheme optimization of river ecological restoration data according to claim 6, characterized in that, In the discretized decoding rule, the formula for calculating the long-term cumulative reward Q value is: Q = α * ΔEHI + β / C + γ *η, where ΔEHI is the expected improvement value of the ecological health index, C is the total cost of the measures, η is the resource utilization efficiency, and α, β, and γ are preset weight coefficients, with values ​​of 0.6, 0.3, and 0.1, respectively.

8. The method for dynamic monitoring and scheme optimization of river ecological restoration data according to any one of claims 1-7, characterized in that, The execution unit includes, but is not limited to, an automatic dosing device for adding microbial agents, an adjustable aeration device for regulating dissolved oxygen, an ecological floating island platform for deploying aquatic plants, and a shoreline intelligent irrigation control system based on soil moisture sensors and meteorological data linkage.

9. A dynamic monitoring and optimization system for river ecological restoration data, characterized in that, include: Multimodal sensing terminals are used to acquire multi-dimensional raw monitoring data of river ecosystems; The data preprocessing unit is used to perform spatiotemporal alignment and outlier removal on the raw monitoring data to generate a standardized time series dataset of river ecological status; the dynamic graph neural network evaluation engine is used to build a dynamic graph neural network model and deduce the ecological health index of each section. The degradation segment identification module is used to identify key degradation segments and dominant limiting factors; A remediation strategy knowledge base for storing a structured set of remediation measures; The repair strategy generator, based on the deep deterministic policy gradient algorithm, is used to generate optimized repair measures combinations and parameters; the execution instruction issuing unit is used to issue control instructions to the automatic execution unit; the effect verification and feedback module is used to calculate the repair effect reward and feed it back to the strategy generator to update the network parameters.

10. The river ecological restoration data dynamic monitoring and scheme optimization system according to claim 9, characterized in that, The system adopts a cloud-edge collaborative architecture; The data preprocessing unit and the execution instruction issuing unit are deployed on edge computing nodes along the river to achieve rapid data preprocessing and real-time instruction response; the dynamic graph neural network evaluation engine, the repair strategy generator, and the knowledge base are deployed on a cloud server to perform complex model training and global strategy optimization.

Images