Environmental monitoring methods, systems, equipment and media for marine aquaculture

Abstract

This invention relates to the field of marine aquaculture technology, specifically providing a method, system, equipment, and medium for environmental monitoring in marine aquaculture. The method includes: acquiring multi-source sensor data; constructing a scene graph based on the multi-source sensor data; the scene graph using sensor monitoring indicators as feature nodes and physical associations, statistical correlations, or time-series dependencies between nodes as edges; constructing a strategy graph; the strategy graph using preset processing strategies as strategy nodes and logical relationships between strategies as edges, and attaching condition labels to each strategy node; calculating the matching degree between the scene graph and the strategy graph, and outputting environmental monitoring results or triggering corresponding processing strategies based on the matching degree. This invention significantly improves the accuracy and automation level of marine aquaculture environmental monitoring through a dual-graph structure design of scene graphs and strategy graphs and a cross-graph matching mechanism.

Description

Technical Field

[0001] This invention belongs to the field of marine aquaculture technology, specifically relating to an environmental monitoring method, system, equipment, and medium for marine aquaculture. Background Technology

[0002] Marine aquaculture is an important economic industry in my country's coastal areas. With the continuous expansion of aquaculture scale and the increase in intensification, risk management of the aquaculture environment has become crucial to ensuring the sustainable development of the industry. Compared with freshwater aquaculture, the marine aquaculture environment is significantly more complex and dynamic: it is affected by multiple factors such as tidal cycles, ocean currents, temperature and salinity stratification, and extreme weather events such as typhoons and rainstorms. As a result, the physicochemical indicators in the water, such as dissolved oxygen, salinity, and pH, exhibit strongly coupled, nonlinear, and spatiotemporally heterogeneous variations.

[0003] Currently, environmental monitoring in marine aquaculture mainly relies on single-point sensor data acquisition and threshold alarm methods. Specifically, existing methods involve deploying sensors for dissolved oxygen, water temperature, salinity, etc., in the aquaculture area to acquire various environmental indicators in real time. When any indicator exceeds a preset safety threshold, the system triggers an alarm and prompts aquaculture personnel to take appropriate measures, such as starting an aerator or opening a water exchange valve.

[0004] However, the above monitoring methods have the following technical drawbacks: First, relying on individual indicators for independent judgment leads to a high false alarm rate. Various indicators in the marine environment are interconnected and influence each other. For example, rising water temperature leads to a decrease in dissolved oxygen saturation, and slowing currents exacerbate the accumulation of pollutants at the bottom, further depleting dissolved oxygen. Existing methods process these indicators in isolation, relying solely on the threshold of a single indicator for judgment. This makes it difficult to distinguish between genuine environmental degradation and instantaneous sensor fluctuations, easily generating false alarms and increasing unnecessary intervention by aquaculture personnel.

[0005] Secondly, it lacks the ability to identify complex environmental scenarios. The same indicator value may represent drastically different risk states under different environmental backgrounds. For example, a dissolved oxygen level of 3.5 mg / L is a severe precursor to hypoxia in calm, hot weather, but may only be a temporary fluctuation in a strong current, cold environment. Existing methods cannot integrate multi-source data to identify environmental scenarios, making it difficult to make accurate risk assessments and decision support.

[0006] Third, there is a disconnect between treatment strategies and monitoring data. In existing technologies, monitoring systems are only responsible for data collection and alarms, while specific treatment strategies rely on the experience and judgment of aquaculture personnel. There are complex logical relationships between different treatment strategies. For example, oxygenation and water exchange are mutually exclusive in some cases, and extreme weather warnings should take precedence over routine adjustments. Existing systems cannot achieve intelligent matching and automated linkage of strategies. Summary of the Invention

[0007] In view of the above-mentioned shortcomings of the prior art, the present invention provides an environmental monitoring method, system, equipment and medium for marine aquaculture to solve the above-mentioned technical problems.

[0008] In a first aspect, the present invention provides an environmental monitoring method for marine aquaculture, comprising: Acquire multi-source sensor data, which includes water quality physicochemical indicators, meteorological and hydrological indicators, and biological sediment indicators; A scene graph is constructed based on the multi-source sensor data. The scene graph uses sensor monitoring indicators as feature nodes and physical associations, statistical correlations or time series dependencies between nodes as edges. Calculate the matching degree between the scene graph and the pre-built strategy graph, and output the environment monitoring results or trigger the corresponding processing strategy based on the matching degree; The strategy graph uses preset processing strategies as strategy nodes, logical relationships between processing strategies as edges, and adds condition labels to each strategy node.

[0009] Secondly, the present invention provides an environmental monitoring system for marine aquaculture, comprising: The data acquisition module is used to acquire multi-source sensor data, which includes water quality physicochemical indicators, meteorological and hydrological indicators, and biological sediment indicators. The data processing module is used to construct a scene graph based on the multi-source sensor data. The scene graph uses sensor monitoring indicators as feature nodes and physical associations, statistical correlations or time series dependencies between nodes as edges. The strategy matching module is used to calculate the matching degree between the scene graph and the pre-built strategy graph, and output the environment monitoring results or trigger the corresponding processing strategy based on the matching degree. The strategy graph uses preset processing strategies as strategy nodes, logical relationships between processing strategies as edges, and adds condition labels to each strategy node.

[0010] Thirdly, a device is provided, comprising: Memory, used to store environmental monitoring programs for marine aquaculture; A processor is used to implement the steps of the marine aquaculture environmental monitoring method as provided in the first aspect when executing the environmental monitoring program for marine aquaculture.

[0011] Fourthly, a computer-readable medium is provided, on which an environmental monitoring program for marine aquaculture is stored, wherein when the environmental monitoring program for marine aquaculture is executed by a processor, the program implements the steps of the environmental monitoring method for marine aquaculture provided in the first aspect.

[0012] The environmental monitoring method, system, equipment, and medium for marine aquaculture provided by this invention have the following beneficial effects: First, it improves the accuracy and reliability of environmental monitoring. This invention constructs a scene graph, fusing multi-source sensor data into a graph representation with a topological structure. This not only preserves the numerical information of each monitoring indicator but also characterizes the intrinsic relationships between indicators through physical correlation edges, statistical correlation edges, and time-series dependency edges. When a sensor experiences instantaneous fluctuations, the scene graph can be verified using information from neighboring nodes, effectively distinguishing between real environmental degradation and equipment interference, significantly reducing the false alarm rate. Simultaneously, the introduction of spatiotemporal auxiliary nodes enables the model to perceive periodic patterns such as tidal phases and seasonal changes, enhancing its adaptability to the complex dynamics of the marine environment.

[0013] Second, it achieves intelligent matching between environmental conditions and treatment strategies. This invention constructs a strategy graph, organizing dispersed treatment strategies into a logically related graph structure, and attaching condition and priority labels to each strategy node. The cross-graph matching step calculates the matching degree between the scenario graph and the strategy graph, automatically identifying the specific scenario of the current environment—whether it's high-temperature, still water with low oxygen, a sudden drop in salinity due to heavy rain, or a precursor to a red tide—and matching the most suitable treatment strategy. This achieves a closed loop from "data acquisition" to "decision support," overcoming the limitations of traditional methods that rely on human experience.

[0014] Third, it enhances the foresight and scientific rigor of risk response. This invention quantifies the gap between the current environmental state and the triggering conditions of various strategies by constructing a constraint field and calculating constraint response coefficients. It can not only determine whether a strategy has been triggered, but also assess "how far away from triggering." When the matching score falls below a preset threshold, the system can issue an early warning, allowing farmers time to respond and transforming passive response into proactive prevention.

[0015] In summary, this invention, through its dual-graph structure design of scene graph and strategy graph and its core mechanism of cross-graph matching, effectively overcomes the technical shortcomings of existing marine aquaculture environmental monitoring methods, such as one-sided judgment of single indicators, weak scene recognition ability, unintelligent strategy matching, and unpredictable intervention effects. It significantly improves the accuracy, automation, and intelligence of aquaculture risk management. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic flowchart of a method according to an embodiment of the present invention.

[0018] Figure 2 This is a schematic flowchart illustrating the construction scenario of a method according to an embodiment of the present invention.

[0019] Figure 3 This is a visual schematic diagram of the trigger management strategy of a method according to an embodiment of the present invention.

[0020] Figure 4 This is a schematic block diagram of a system according to an embodiment of the present invention.

[0021] Figure 5 This is a schematic diagram of the structure of a device provided in an embodiment of the present invention. Detailed Implementation

[0022] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0023] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0024] The environmental monitoring method for marine aquaculture provided in this embodiment of the invention is executed by computer equipment, and correspondingly, the environmental monitoring system for marine aquaculture runs on the computer equipment.

[0025] Figure 1 This is a schematic flowchart illustrating a method according to an embodiment of the present invention. Wherein, Figure 1 The implementing entity can be an environmental monitoring system for marine aquaculture. Depending on different needs, the order of the steps in this flowchart can be changed, and some steps can be omitted.

[0026] like Figure 1 As shown, the method includes: S1. Acquire multi-source sensor data, which includes water quality physicochemical indicators, meteorological and hydrological indicators, and biological sediment indicators; S2. Construct a scene graph based on the multi-source sensor data. The scene graph uses sensor monitoring indicators as feature nodes and physical associations, statistical correlations, or time series dependencies between nodes as edges. S3. Calculate the matching degree between the scene graph and the pre-built strategy graph, and output the environment monitoring results or trigger the corresponding processing strategy based on the matching degree; The strategy graph uses preset processing strategies as strategy nodes, logical relationships between processing strategies as edges, and adds condition labels to each strategy node.

[0027] In one embodiment of the present invention, based on step S1, the following will provide a possible embodiment and describe its specific implementation in a non-limiting manner.

[0028] S101. Acquisition and Deployment of Multi-Source Sensor Data Given the unique characteristics of marine aquaculture environments, the deployment of sensor networks must consider three principles: three-dimensional layering, spatial representativeness, and resistance to harsh environments.

[0029] Water quality physicochemical indicators are the core of environmental monitoring, mainly including dissolved oxygen, water temperature, salinity, pH, ammonia nitrogen, nitrite, chlorophyll a, and chemical oxygen demand. Due to the significant temperature and salinity stratification in marine waters, this invention employs a three-dimensional stratified deployment strategy: Surface sensor: Deployed 0.5 meters below the water surface, primarily monitoring surface water that is significantly affected by atmospheric reoxygenation and sunlight. This sensor mainly collects dissolved oxygen, water temperature, chlorophyll a, and light intensity.

[0030] Mid-layer sensors: Deployed in the middle layer of the water body (e.g., at 1 / 2 of the water depth) as a transition layer to monitor changes in the thermocline. These sensors primarily collect data on water temperature, salinity, and dissolved oxygen.

[0031] Bottom-layer sensors: These are installed 0.5 meters above the bottom sediment. This area is where uneaten food and feces accumulate, making it highly susceptible to oxygen deficiency and sediment deterioration. These sensors primarily collect data on dissolved oxygen, water temperature, salinity, pH, ammonia nitrogen, nitrite, and chemical oxygen demand (COD).

[0032] All of the above sensors are fixed using multi-parameter water quality monitoring buoys or underwater probe arrays. Each sensor node is equipped with a unique ID and collects data at a preset frequency (such as once every 10 minutes), which is then transmitted back to the data center in real time via a wireless network.

[0033] Meteorological and hydrological indicators are used to characterize the impact of the external environment on aquaculture water bodies, and mainly include wind speed, wind direction, light intensity, tide level, and current velocity: Wind speed and direction: Install automatic weather stations on the shore or buoys of the aquaculture area. The anemometer and wind vane are installed 10 meters above the sea level and the data is collected every 5 minutes.

[0034] Light intensity: A photosynthetically active radiation sensor was deployed on the weather station platform to calculate the potential oxygen production capacity of algae photosynthesis.

[0035] Tidal level and flow velocity: Pressure level gauges and acoustic Doppler current profilers are deployed in the main channel of the aquaculture area or in the cage area. Tidal level data is used to identify the phases of high tide, low tide, and ebb tide; flow velocity data is used to assess the water exchange capacity and pollutant diffusion rate. Flow velocity sensors should focus on bottom flow velocity, as it directly relates to the resuspension of bottom deposits and dissolved oxygen replenishment.

[0036] Biological sediment indicators are used for early warning of red tide outbreaks and sediment aging: Harmful algal cell density: obtained using in-situ algae monitoring instruments or periodic manual sampling. Specifically, algal fluorescence probes are deployed in the aquaculture area to monitor changes in chlorophyll a concentration in real time; when chlorophyll a exceeds a preset threshold (e.g., 10 μg / L), an automatic water sampling device is triggered, and the cell density of harmful algae (unit: cells / L) is analyzed using microscopy or flow cytometry.

[0037] Sediment sulfide content and redox potential: Sediment samples are collected periodically using a sediment sampler or by deploying in-situ sediment monitoring probes. Sulfide content is measured using a sulfide ion-selective electrode, and redox potential is measured using a platinum electrode. These two indicators are used to determine the degree of sediment decay. When the sulfide content exceeds 0.3 mg / g or the redox potential is below -100 mV, it indicates that the sediment has severely deteriorated and a sediment cleanup strategy needs to be initiated.

[0038] S102. Data Preprocessing and Standardization The raw sensor data collected needs to be preprocessed to remove outliers and fill in missing values ​​to ensure the accuracy of subsequent map construction: Outlier detection and removal: The Isolation Forest algorithm is used to detect outliers in the data. For example, when the dissolved oxygen value remains unchanged for two consecutive hours, or when there is a sudden change far exceeding the normal range, it is determined to be a sensor malfunction or biological adhesion interference, and is removed, and the data is filled with the historical average value of the previous period.

[0039] Missing value handling: For short-term missing values ​​(less than 2 hours), linear interpolation is used to fill them in; for long-term missing values ​​(more than 2 hours), data from nearby sensors within the same aquaculture grid or historical data from the same period (considering tidal phase matching) are used to fill them in.

[0040] Data standardization: Normalize all sensor data to the [0,1] interval to eliminate the influence of different units on model calculation.

[0041] In one embodiment of the present invention, based on step S2, please refer to... Figure 2 The following provides a possible embodiment and its specific implementation is described in a non-limiting manner.

[0042] S201. The multi-source sensor data is numerically processed to construct an initial feature vector for graph computation; wherein, each sensor monitoring index or spatiotemporal attribute is defined as a node, and a multi-dimensional feature vector containing the current real-time value of the index, preset historical statistical values, and dynamic change rate is generated for each node.

[0043] In this embodiment, the acquired multi-source sensor data is first quantified, with each monitoring indicator or spatiotemporal attribute defined as an independent graph node, and a multi-dimensional feature vector generated for each node. The node types and feature vector structures are as follows: Water quality physicochemical node feature vectors: Taking the dissolved oxygen node as an example, its 32-dimensional feature vector consists of the following components: Current real-time values ​​(4 dimensions): including instantaneous values ​​of dissolved oxygen in the surface layer, middle layer, and bottom layer (unit: mg / L), as well as the vertical gradient difference between the three layers (bottom layer minus surface layer).

[0044] Preset historical statistical values ​​(8 dimensions): including the mean, maximum, minimum, and standard deviation of the past 24 hours; the mean of the same tidal phase over the past 7 days; and the percentile value relative to the same historical period (e.g., the current value is at the 90th percentile of historical data).

[0045] Dynamic change rate (4 dimensions): including the change rate over the past 1 hour, the change rate over the past 3 hours, the diurnal fluctuation range (daytime peak minus nighttime trough), and the difference from the ideal saturated dissolved oxygen.

[0046] Similarly, the water temperature node feature vector includes the stratified temperature values ​​of the surface, middle, and bottom layers, vertical temperature difference, daily variation rate, and deviation from historical temperatures for the same period. The salinity node feature vector includes the current value, the trend over the past 6 hours (used to identify sudden drops caused by heavy rainfall), and the deviation from the expected value matched with the tidal phase.

[0047] Meteorological and hydrological node feature vectors: The wind speed node feature vector includes: current wind speed (m / s), average wind speed over the past 1 hour, maximum gust wind speed, and prevailing wind direction angle (0-360° encoded as sine and cosine components).

[0048] The light intensity node feature vector includes: current photosynthetically active radiation (μmol·m⁻¹) -2 ·s -1 The cumulative illumination over the past 24 hours, the average illumination over the past hour, and the product of the day / night indicator (1 for daytime and 0 for nighttime) and the illumination value.

[0049] The feature vectors of tidal level and current velocity nodes include: current tidal level (m), tidal phase encoding (high tide = 0, low tide = 1, low tide = 2, encoded as a 3-dimensional vector by one-hot encoding), current current velocity (m / s), bottom current velocity, and current velocity change trend over the past 3 hours.

[0050] Biological substrate node feature vectors: The feature vector of the harmful algal cell density node includes: current algal cell density (cells / L), density change rate over the past 24 hours, chlorophyll a concentration, and algal species identifier (such as the proportion of harmful algal species).

[0051] The feature vector of the sediment sulfide content node includes: the current sulfide concentration (mg / g), the redox potential (mV), and the sediment health index derived from both (between 0 and 1, the lower the value, the more decayed the sediment).

[0052] Spatiotemporal auxiliary node feature vectors: Tidal phase nodes: The 24 hours of a day are divided into 12 tidal phases (each phase lasts 2 hours), and sine and cosine codes are used to form a 2D vector to capture the periodicity of the tides.

[0053] Seasonal nodes: The four seasons are encoded as 4-dimensional one-hot vectors to represent the long-term climate background.

[0054] Aquaculture area grid coordinate nodes: The aquaculture sea area is divided into 10×10 grids, each grid corresponds to a spatial node, and its feature vector includes the normalized coordinates of the grid center latitude and longitude.

[0055] S202. Construct a scene graph with a topological structure using the nodes as vertices and the relationships between nodes as edges; wherein the edges include at least one type based on physical laws, statistical correlation, or time dependence, and assign a weight value representing the strength of the relationship to each type of edge.

[0056] After completing node feature vectorization, this embodiment constructs edges based on the relationships between nodes, forming a scene graph with a topological structure. Edge construction includes the following three types: Physical laws are constructed along the way: Based on prior knowledge of oceanography and aquaculture, construct physical law-related edges: Water temperature-dissolved oxygen: Within each spatial grid, water temperature nodes are connected to dissolved oxygen nodes, with an edge weight of 0.9, representing the physical negative correlation between increased water temperature and decreased dissolved oxygen saturation concentration.

[0057] Flow velocity-pollutant diffusion: Connect the flow velocity node with the ammonia nitrogen and nitrite nodes, and set the edge weight to 0.8 to characterize the positive correlation between increased flow velocity and pollutant diffusion.

[0058] Light-Chlorophyll a: Connect the light node to the chlorophyll a node with an edge weight of 0.85 to represent the positive correlation between enhanced light and algal photosynthesis.

[0059] Tidal Phase-Salinity: Connect the tidal phase node to all water quality nodes with an edge weight of 0.7 to characterize the periodic impact of tidal fluctuations on salinity and other water quality indicators.

[0060] Interlayer connection edge: Within the same vertical water column, the surface water temperature node is connected to the bottom water temperature node, and the edge weight is set to 0.75 to represent the heat exchange between the upper and lower water layers; the bottom dissolved oxygen is connected to the surface dissolved oxygen, with a weight of 0.6, to represent the replenishment of dissolved oxygen by vertical mixing.

[0061] Construction of statistical correlation edges: Based on historical monitoring data (such as data from the past year), calculate the Pearson correlation coefficient between each pair of nodes: Select all nodes within the same spatial grid and extract their historical time series data.

[0062] Calculate the correlation coefficient r for each pair of nodes, with a value in the range [-1, 1].

[0063] reserve Pairs of nodes with a value greater than 0.3 are considered as candidate edges, and... As edge weight.

[0064] For example, if historical data shows that chlorophyll a and pH value are consistently strongly positively correlated (r=0.85), then a statistical correlation edge with a weight of 0.85 is added between the two.

[0065] Time series dependency edge construction: Based on mutual information analysis of time series data, edges reflecting causal delay relationships are constructed: The time-delay mutual information method is used to calculate the impact of the historical value of node X on the future value of node Y.

[0066] For each pair of nodes, calculate the mutual information value MI(τ) under different time delays τ (e.g., τ = 1 hour, 3 hours, 6 hours).

[0067] Retain the node pairs with the maximum MI(τ) > 0.1, and use the normalized MI(τ) as the edge weight, while recording the optimal delay time τ as the edge attribute.

[0068] For example, if the analysis finds that the dissolved oxygen in the bottom layer decreases significantly after the bottom layer velocity decreases by 3 hours (MI=0.25), then a time-dependent edge with a weight of 0.25 is added between the velocity node and the dissolved oxygen node, and a delay attribute τ=3h is attached.

[0069] Multi-edge fusion: When there are multiple edges of different types between the same pair of nodes, this embodiment adopts the following fusion strategy: Maximum value retention: Keep the edge with the largest weight and discard the other edges to ensure a simple graph structure.

[0070] Alternatively, weighted fusion: the weights of multiple edges are summed by weighting according to preset coefficients (0.5 for physical law edge, 0.3 for statistical edge, and 0.2 for time edge) to obtain composite edge weights.

[0071] Ultimately, all nodes and the constructed edges together form a complete scene graph G=(V,E), where V is the set of nodes, E is the set of edges, and each edge e ij With weight w ij .

[0072] S203. Input the scene graph into a graph neural network, and use the edge weights to aggregate and update the feature vectors of neighboring nodes through graph convolutional layers, so that each node integrates the environmental information of its neighboring nodes.

[0073] In the input layer, the initial feature vector of each node is... (Dimension 32) serves as the input to the graph neural network. Node feature matrix. , where N is the total number of nodes.

[0074] This embodiment sets up two layers of graph convolution, and the calculation process for each layer is as follows: First layer: For each node v, aggregate the features of all its neighboring nodes u∈N(v) using a mean aggregator: Among them, w uv W represents the edge weight. (1) σ is a trainable weight matrix, and σ is the ReLU activation function. The output feature dimension is increased to 64 dimensions.

[0075] The second layer introduces an attention mechanism, assigning different attention coefficients to different neighbors:

[0076] Here, ∥ represents vector concatenation, and a is the attention parameter vector. The output feature dimension is increased to 128 dimensions.

[0077] After two layers of graph convolution, the feature vector of each node has been fused with information from its neighboring nodes. For example, the final feature of the bottom dissolved oxygen node not only includes its own monitoring value, but also information from related nodes such as bottom water temperature, bottom flow velocity, upper dissolved oxygen, and tidal phase, thus achieving a comprehensive perception of the environmental state.

[0078] S204. Perform global pooling on all node features after graph convolution aggregation to generate a fixed-dimensional scene graph embedding vector that represents the overall state of the current aquaculture environment.

[0079] After completing graph convolution aggregation, this embodiment uses global pooling to generate fixed-dimensional scene graph embedding vectors to represent the overall state of the current aquaculture environment.

[0080] Global average pooling averages the final feature vectors of all nodes in one dimension:

[0081] We obtain a 128-dimensional scene graph embedding vector E. scene .

[0082] Global max pooling, in order to preserve the most significant environmental features, can be performed simultaneously to extract the maximum value in each dimension, and then concatenated with the average pooling result to obtain a 256-dimensional enhanced embedding vector.

[0083] Dynamic graph updates: This embodiment uses a sliding window mechanism to update the scene graph every hour. Node feature update: Replace with the latest collected sensor data and recalculate derived features such as the rate of change.

[0084] Edge weight fine-tuning: Recalculate the weights of statistically relevant edges and time-dependent edges based on data from the most recent 24 hours, so that the graph structure can dynamically adapt to changes in the environment.

[0085] A snapshot of the graph from the past 24 hours is retained for subsequent time series analysis and counterfactual inference.

[0086] In one example, a monitoring network including sensors for dissolved oxygen, water temperature, salinity, current velocity, and chlorophyll a was deployed in a marine cage aquaculture area. At a certain moment, the system collected the following data: bottom dissolved oxygen 4.2 mg / L, bottom water temperature 28.5℃, bottom current velocity 0.03 m / s, surface chlorophyll a 15 μg / L, and the tidal phase was mid-ebb tide.

[0087] According to the method of this embodiment: Feature vectorization: Each node generates a 32-dimensional feature vector containing the current value, historical statistics, and rate of change.

[0088] Graph topology construction: Based on physical laws, water temperature nodes are connected to dissolved oxygen nodes (weight 0.9), and flow velocity nodes are connected to dissolved oxygen nodes (weight 0.8); based on statistical correlation, chlorophyll a nodes are connected to pH nodes (weight 0.7); based on time dependence, flow velocity nodes and dissolved oxygen nodes are connected with an edge delayed by 3 hours (weight 0.6).

[0089] Graph convolutional aggregation: The bottom dissolved oxygen node aggregates information such as water temperature (high temperature reduces dissolved oxygen), flow velocity (low flow velocity limits reoxygenation), and tide (low tide weakens exchange), generating enhanced node features.

[0090] Graph embedding generation: Global pooling yields a 128-dimensional scene graph embedding vector, which comprehensively represents the environmental conditions of "high temperature, low flow, low tide, and low dissolved oxygen".

[0091] In one embodiment of the present invention, based on step S3, the following will provide a possible embodiment and describe its specific implementation in a non-limiting manner.

[0092] This implementation details how to construct a predefined processing strategy into a computable graph structure and encode it into a policy graph embedding vector that represents the semantics of policy combination. The core function of the policy graph is to transform the scattered, text-described processing strategies into a structured graph representation that can be used for machine reasoning.

[0093] 1. Strategy Node Definition and Attribute Settings In this embodiment, the preset processing strategy is first defined as a node in the strategy graph, and a condition label and a priority label are attached to each node.

[0094] (1) Classification of strategy node types Based on the function and urgency of the processing strategy, the strategy nodes are divided into the following four categories: Emergency Response Nodes: These are used to address emergencies that may endanger the survival of farmed organisms and must be executed immediately. Specific nodes include: Aerator Linkage Node: Activating aeration equipment when dissolved oxygen falls below the safe threshold. Water Change Valve Control Node: Introducing fresh seawater when salinity drops sharply or ammonia nitrogen levels exceed standards. Red Tide Emergency Node: Initiating physical isolation or chemical control procedures when harmful algae density is detected to exceed standards.

[0095] Optimization and adjustment points: These are used to improve the aquaculture environment and enhance growth efficiency, and can be implemented periodically or as needed. Specifically, they include: Bottom sediment cleaning point: When sulfide accumulation in the bottom sediment exceeds the standard, bottom aeration or dredging operations are initiated. Net cage attachment cleaning point: When excessive attachment to the net cages obstructs water flow, the nets are cleaned or replaced.

[0096] Early warning nodes: These are used to send alerts to aquaculture personnel without directly triggering equipment operation. Specifically, they include: Extreme weather early warning nodes: issuing warnings when extreme weather events such as typhoons or heavy rains are predicted. Oxygen deficiency risk early warning nodes: issuing early warnings when dissolved oxygen levels approach safe thresholds and environmental conditions are unfavorable.

[0097] Special environmental nodes: used for counterfactual simulations to model the impact of extreme environments on aquaculture systems. These include: Typhoon simulation node: simulating the environmental impacts of a typhoon, such as wind speed, wave height, and bottom upwelling. Heavy rain simulation node: simulating the impacts of heavy rain, such as a sudden drop in salinity and the input of land-based pollutants.

[0098] (2) Conditional label configuration Attach a condition label to each policy node to define the threshold or range for policy triggering: Aerator linkage node: The condition label is "Dissolved oxygen < 3.0 mg / L" or "Dissolved oxygen diurnal fluctuation > 2.0 mg / L". The former is a hard trigger condition, and the latter is an auxiliary judgment condition.

[0099] Water exchange valve control node: The condition label is "Salinity < 20‰" (freshwater intrusion) or "Ammonia Nitrogen > 0.5 mg / L". For the salinity condition, the additional change rate sub-condition "decline exceeding 10‰ within 24 hours" is used to identify sudden drops caused by heavy rainfall.

[0100] Red tide emergency node: The conditions are "chlorophyll a > 20 μg / L" and "harmful algal cell density > 500 cells / L". Both conditions must be met simultaneously to trigger the event.

[0101] Sediment cleaning node: The condition label is "Sediment sulfide > 0.3 mg / g" or "Oxidation-reduction potential < -100 mV".

[0102] Extreme weather warning nodes: The condition label is "wind speed > 15 m / s" (typhoon warning) or "24-hour rainfall > 50 mm" (rainstorm warning).

[0103] Oxygen deficiency risk warning node: The condition label is "dissolved oxygen < 4.0 mg / L" and "water temperature > 28℃". This condition is more lenient than the emergency trigger threshold and is used for early warning.

[0104] (3) Priority label settings Set a priority label for each strategy node for weight allocation during subsequent matching: High Priority: Emergency Response Node, with a priority value set to 1.0. This type of strategy involves life safety or significant property loss and must be prioritized.

[0105] Medium priority: Optimize adjustment nodes and extreme weather warning nodes, with a priority value set to 0.6. This type of strategy affects aquaculture efficiency and should be considered only after meeting urgent needs.

[0106] Low priority: Regular early warning nodes, with a priority value set to 0.3. This type of policy is only for informational purposes and is not enforced.

[0107] 2. Strategy Edge Construction and Logical Relationship Modeling After completing the node definition, this implementation method constructs directed edges based on the logical relationships between policy nodes to form the topological structure of the policy graph.

[0108] (1) Triggering relationship edge Triggering relationship edges indicate that when the conditions of one policy node are met, the execution of another policy node will be triggered: Hypoxia Risk Warning - Oxygenator Linkage: When the hypoxia risk warning node is activated (dissolved oxygen is close to the threshold), if the conditions further deteriorate, the oxygenator linkage node will be triggered. The edge direction is from the warning node to the response node, and the weight is set to 1.0, indicating a strong triggering relationship.

[0109] Extreme Weather Warning - Water Exchange Valve Control: If a typhoon or rainstorm warning is triggered and accompanied by a sudden drop in salinity, the water exchange valve control node will be activated. However, since water exchange under extreme weather conditions may introduce polluted water bodies, a cautionary execution flag is added to this triggering relationship, and the edge weight is set to 0.7.

[0110] (2) Dependency edge Dependency edges represent the order constraints or resource constraints of strategy execution: Sequential dependency edge: The sediment cleaning node depends on the aerator linkage node. Specifically, the aerator must be started before cleaning the sediment to prevent acute poisoning caused by sulfides released during the cleaning process. The edge direction is from aerator linkage to sediment cleaning, with a weight of 0.9, indicating a strong dependency.

[0111] Resource Mutual Exclusion Edge: A mutual exclusion relationship exists between the aerator linkage node and the water exchange valve control node. Due to the limited power resources in the farm, simultaneously activating a high-power aerator and a high-flow water exchange valve may overload the power grid. In this embodiment, a negative-weighted edge is constructed between the two, with a weight of -0.8, indicating mutual exclusion. During strategy matching, if both nodes are activated simultaneously, the matching score will be penalized.

[0112] (3) Deducing the relationship Inference relationship edges are used for counterfactual analysis, connecting special environment nodes with regular strategy nodes: Typhoon Simulation Node - All Water Quality Nodes: The typhoon simulation node connects water quality nodes such as dissolved oxygen, flow velocity, and salinity through simulation. The edge weight is set to 0.7, indicating that a typhoon may cause a decrease in dissolved oxygen at the bottom layer (causing the bottom layer to rise and cause the oxygen-deficient water mass to surge), an increase in flow velocity, and fluctuations in salinity.

[0113] Heavy Rainfall Simulation Node - Salinity Node: The heavy rainfall simulation node is connected to the salinity node through a derivation edge, with the edge weight set to 0.9, representing the strong impact of heavy rainfall on salinity.

[0114] (4) Edge attribute record In addition to weight, each edge has the following attributes: Relationship type: Triggered, Dependent, Mutually Exclusive, Inferential. Delay time: For triggered relationships, a delay attribute can be attached, such as executing the response strategy 30 minutes after the warning is triggered. Condition expression: For complex triggered relationships, a logical expression can be attached, such as "(Condition A AND Condition B) - Strategy C".

[0115] The strategy node attributes are transformed into a 32-dimensional numerical feature vector, structured as follows: Type Encoding (8 dimensions): 4-dimensional one-heat encoding represents the major category (emergency response, optimization regulation, early warning, special environment), and 4-dimensional encoding represents the sub-type (such as aerator, water exchange valve, etc.). Condition Threshold Normalization (8 dimensions): Trigger thresholds are normalized to [0,1]. For example, the dissolved oxygen threshold of 3.0 mg / L (range 0-10) is normalized to 0.3. Condition Hardness Identifier (4 dimensions): Hard constraints [1,0,0,0], soft constraints [0,1,0,0], combined constraints [0,0,1,0], early warning category [0,0,0,1]. Priority Encoding (4 dimensions): Priority values ​​(1.0, 0.6, 0.3) are mapped to embedding vectors. Execution Cost (4 dimensions): Normalized resource consumption value (0-1). Historical Trigger Frequency (4 dimensions): Normalized representation of the number of triggers in past periods.

[0116] Taking the aerator node as an example: Type [0,0,0,1,0,0,0,0] (Emergency Response + Aerator Subclass), Threshold 0.3, Hardness [1,0,0,0], Priority Mapping [0.9,0.1,0.0,0.0], Execution Cost 0.4, Historical Trigger [0,0,0,0].

[0117] A graph attention network encoding strategy is employed. The input node feature matrix is... Set up two layers of graph convolution: First layer: Aggregate features of incoming edge neighbors, considering edge weights w qp The output is 64-dimensional.

[0118] Second layer: Introduce an attention mechanism and add a directional coefficient d. pq (Incoming edge 1.0, outgoing edge 0.5), negative edges are processed separately and their contribution is negative, outputting 128 dimensions.

[0119] After two convolutional layers, the node features are fused with logical association information (e.g., the aerator node incorporates information from the hypoxia warning and water exchange valve nodes). Finally, global max pooling is used to extract the most salient features, resulting in a 128-dimensional policy graph embedding vector E_policy; alternatively, the average pooling result can be concatenated to obtain a 256-dimensional augmentation vector.

[0120] The strategy graph is not static; this implementation supports dynamic updates: Seasonal adjustment: Adjusting the thresholds of strategy nodes based on the aquaculture needs of different seasons. Experience feedback update: Adjusting the weights of strategy edges based on historical execution results. Strategy addition and deletion: Supporting aquaculture personnel to add custom strategy nodes through the human-machine interface; the system automatically calculates the statistical correlation between the new node and existing nodes and constructs initial edges.

[0121] S301. Calculate the matching degree between the scene graph and the pre-constructed strategy graph, including: S301.1 Construction of Conditional Constraint Fields The core function of the condition constraint field is to transform the condition labels of each policy node in the policy graph into a quantifiable constraint space, providing weight benchmarks and response coefficients for subsequent matching calculations.

[0122] (1) Constraint level classification and weight configuration Based on the condition labels of the policy nodes, constraints are divided into two main categories: hard constraints and soft constraints, and a basic weight is assigned to each type of constraint. Hard constraints: These are constraints that, if violated, would lead to the death of farmed organisms or significant property damage, with a weight range of 0.8-1.0. Specifically, they include: Dissolved oxygen thresholds: safety threshold >5.0 mg / L, emergency threshold <3.0 mg / L. The emergency threshold has a weight of 1.0; the safety threshold has a weight of 0.9. Harmful algal cell density threshold: >500 cells / L is the red tide warning threshold, with a weight of 0.95. Salinity change threshold: a change of >10‰ within 24 hours, with a weight of 0.9. Typhoon wind speed threshold: >25 m / s, with a weight of 1.0.

[0123] Soft constraints: These refer to constraints that affect aquaculture efficiency but pose no fatal risk, with a weighting range of 0.4-0.7. Specifically, these include: Flow velocity range: Recommended 0.1-0.5 m / s; below 0.1 m / s, it easily leads to pollutant accumulation, weight 0.6; above 0.5 m / s, it easily leads to excessive energy consumption by aquaculture organisms, weight 0.5. pH range: Recommended 7.8-8.3; exceeding this range affects growth but will not cause immediate death, weight 0.5. Light intensity: Affects algal reproduction, with no direct risk, weight 0.4.

[0124] (2) Calculation of constraint tightness For each policy node, this embodiment further calculates the tightness of its constraint, which is the distance between the current value of the scene graph node and the threshold of the policy node. The smaller the distance, the closer the current state is to the trigger condition, and the tighter the constraint.

[0125] For continuous value constraints, the tightness coefficient is calculated as follows:

[0126] Where V is the current value of the scene graph node, T is the threshold of the strategy node, and R is the reasonable range of variation of the indicator (e.g., dissolved oxygen range of 0-10 mg / L).

[0127] For enumerated value constraints, the tightness coefficient is either 0 or 1: 1 if the condition is met, and 0 if the condition is not met.

[0128] (3) Generation of constraint response coefficients By combining the constraint weights and their tightness, constraint response coefficients are generated to characterize the degree to which the current environmental state satisfies the policy conditions.

[0129] in, The basic weights for the constraints.

[0130] For nodes with multiple conditions (such as red tides where both chlorophyll a and cell density must be satisfied simultaneously), the constraint response coefficient is taken as the minimum or weighted average of the response coefficients for each condition. This implementation adopts the minimum value strategy to reflect the "weakest link effect": .

[0131] S301.2 Generation of Cross-Graph Alignment Relationships The goal of the cross-graph alignment step is to establish the correspondence between scene graph nodes and policy graph nodes, and to generate cross-graph fusion features that integrate the features of both graphs.

[0132] (1) Node correlation calculation For each pair of scene graph nodes S i And strategy graph node P j This implementation method calculates its comprehensive relevance score: Semantic matching degree The semantic matching degree is determined based on node type matching. Nodes with the same index type have a semantic matching degree of 1.0 (e.g., the dissolved oxygen node in the scenario diagram and the aerator linkage node in the strategy diagram, both of which involve dissolved oxygen); the semantic matching degree of nodes with different types is determined according to the preset association matrix, with a value range of 0-0.5 (e.g., the flow velocity node and the sediment cleaning node, because the flow velocity affects sediment accumulation, the matching degree is set to 0.4); the matching degree of completely unrelated nodes is 0.

[0133] Strategy Node Weight This refers to the priority label value of the policy node (high priority 1.0, medium priority 0.6, low priority 0.3).

[0134] Constraint response coefficient: C obtained from the previous step.

[0135] The formula for calculating the overall relevance score is: (2) Alignment Relationship Filtering This implementation method sets up a dual threshold to filter effective alignment relationships: Correlation threshold: Retain Score ij The alignment relationship is >0.5. The above example of aerator linkage (0.98) meets the condition, and the example of bottom cleaning (0.084) is filtered out.

[0136] Constraint response threshold: Retain alignment relationships with C>0.3 to avoid relationships that do not meet the conditions from being included in subsequent calculations.

[0137] The selected alignment relationships form a set A, where each alignment relationship a∈A contains a scene graph node index i, a policy graph node index j, and a relevance score. ij And the corresponding constraint response coefficient C.

[0138] (3) Cross-graph fusion feature generation Based on the selected alignment relationships, this embodiment generates a cross-graph fusion feature vector to characterize the interaction information between the scene graph and the policy graph: Node feature extraction: For each alignment relation 'a', extract the corresponding scene graph node embedding vector. (128-dimensional) and policy graph node embedding vectors (128 dimensions).

[0139] Feature splicing and weighting: and The vectors are concatenated into a 256-dimensional vector and scored based on their relevance. ij Weighted fusion is performed using these weights:

[0140] Here, ⊕ represents the vector concatenation operation.

[0141] Normalization processing: for L2 normalization is performed to obtain the final cross-graph fusion feature vector (256 dimensions).

[0142] S301.3 Spatiotemporal Graph Neural Network Matching This implementation uses a spatiotemporal graph neural network as the matching model, and calculates a comprehensive matching score by integrating scene graph embedding, policy graph embedding and cross-graph fusion features.

[0143] The input to the matching model consists of three parts: Scene graph sequence: Take scene graph snapshots from the past T=24 hours, with each time step t corresponding to one scene graph. After encoding by a graph neural network, a scene graph embedding vector sequence is obtained. This sequence contains information on the spatiotemporal evolution of environmental conditions, such as diurnal variations in dissolved oxygen and the influence of tidal cycles.

[0144] Policy graph embedding: Static policy graph embedding vectors It remains unchanged throughout the matching process (unless the strategy library is updated).

[0145] Cross-graph fusion features: Cross-graph fusion features corresponding to each time step t It is generated by the alignment relationship between the scene graph nodes and the strategy graph nodes at that moment.

[0146] The spatiotemporal graph neural network designed in this embodiment contains three core layers: Spatial graph convolutional layers are used to capture the spatial dependencies between sensor nodes within the same time step: Input: Scene graph at each time step t The node features are the original sensor features (32-dimensional).

[0147] Operation: A graph convolutional network is used to aggregate the neighbor node information of each node. Similar to the graph convolution in Example 1, but here the focus is on feature extraction in the spatial dimension.

[0148] Output: Spatially augmented node features, which are then globally pooled to obtain a spatially aware scene graph embedding. .

[0149] Temporal convolutional layers are used to capture the changing patterns of various environmental indicators over time. Input: Historical time series data for each node, with a length of 24 hours.

[0150] Operation: Temporal convolutional networks or gated recurrent units are used to extract temporal features. Temporal convolutional networks expand the receptive field through dilated convolutions, enabling them to capture dependencies at different time scales, such as rapid changes at the hourly level and periodic changes at the daily level.

[0151] Output: Temporally encoded node features, which are then globally pooled to obtain a temporally aware scene graph embedding. .

[0152] The spatiotemporal interaction layer is used to jointly model spatiotemporal dynamics and integrate policy graph information and cross-graph features: enter: , , , .

[0153] Operation: First, and By splicing, spatiotemporal fusion features are obtained. .Will , , By concatenating the three elements, a 384-dimensional comprehensive feature vector is obtained.

[0154] A multi-head attention mechanism is introduced to calculate the contribution weights of different feature components to the matching score. The attention weights are dynamically adjusted based on the current environmental state; for example, when dissolved oxygen is low, feature dimensions related to dissolved oxygen receive higher weights.

[0155] The integrated features are mapped to matching scores through a fully connected layer.

[0156] Multi-channel scoring output: To distinguish between constraints of different natures, this embodiment sets up a multi-channel scoring mechanism: Emergency Response Channel: Focuses on hard constraint indicators (dissolved oxygen, harmful algae density, typhoon wind speed). This channel uses the Sigmoid activation function and outputs the emergency probability P. emergency ∈ [0,1], representing the probability of triggering an emergency response in the current environment. For example, when dissolved oxygen is 2.8 mg / L, P emergency Approximately 1.0; when dissolved oxygen is 5.5 mg / L, P emergency Close to 0.0.

[0157] Optimization suggestion channel: Focuses on soft constraint indicators (flow rate, pH, light intensity). This channel uses the Tanh activation function and outputs the optimization magnitude A. optimization ∈[-1,1]. Positive values ​​indicate that the current indicator is too low and needs to be increased, negative values ​​indicate that it is too high and needs to be decreased, and absolute values ​​indicate the degree of deviation. For example, when the flow velocity is 0.03 m / s (0.1 m / s lower than the recommended lower limit), A optimization Output +0.7, it is recommended to increase the flow rate.

[0158] Overall Matching Score: The final overall matching score S is obtained by weighted summing of the outputs from the two channels. match : Where H is the historical matching stability index (calculated based on the variance of matching scores over the past 24 hours), and α, β, and γ are preset weights (e.g., α=0.6, β=0.3, γ=0.1). match The value range is [0,1]. A higher value indicates a better match between the current environment and the strategy library, that is, a more ideal environment or a lower risk.

[0159] The spatiotemporal graph neural network is trained in an end-to-end manner: Training data: Environmental change events are extracted from historical monitoring data, and aquaculture experts annotate the "ideal match score" (based on post-event evaluation) for each moment. Simultaneously, data augmentation techniques are used to apply random interventions to historical data, generating counterfactual samples to enhance the model's generalization ability to unseen scenarios.

[0160] Loss function: The mean squared error is used as the main loss function to calculate the difference between the predicted matching score and the expert-annotated score. Simultaneously, an auxiliary loss function is introduced, requiring the model to accurately predict dissolved oxygen changes over the next hour, thereby enhancing the model's ability to understand temporal dynamics.

[0161] Real-time inference: In actual operation, a matching score calculation is performed every 10 minutes. A sliding window mechanism is used, taking the data from the past 24 hours as input and outputting the comprehensive matching score S at the current moment. match .

[0162] S302. Output environmental monitoring results or trigger corresponding processing strategies based on the matching degree. Please refer to [link / reference]. Figure 3 .

[0163] S302.1 Comprehensive Matching Score Grading Output (1) Setting of hierarchical thresholds Based on the actual needs of aquaculture management, the following four grading ranges are set: Excellent range (S) match ≥0.9): The environmental conditions are perfectly suited, all monitoring indicators are within the ideal range, there is no risk, and no adjustment measures are required. Good range (0.7≤S) match <0.9: The overall environmental condition is good, but some indicators show slight deviations (e.g., flow rate slightly lower than the recommended value), requiring attention but no immediate intervention is needed. Risk range (0.5≤S) match <0.7: The environmental condition presents a significant risk, with some key indicators approaching or slightly exceeding safety thresholds (e.g., dissolved oxygen 4.2 mg / L), requiring short-term adjustments. Hazardous Zone (S) match <0.5: The environmental condition has deteriorated severely, and the hard constraints have been violated (such as dissolved oxygen below 3.0 mg / L or red tide outbreak), and emergency intervention must be taken immediately.

[0164] (2) Visualization of grading results On the aquaculture management platform interface, this embodiment displays the comprehensive matching score and grading status in the form of a dashboard: a circular progress bar shows the current score of 0.68, in yellow (corresponding to the risk range). The grading label reads: "Current Environmental Status: Risk (Requires Attention)". The trend arrow shows the changing trend of the matching score over the past 24 hours (rising, falling, or remaining stable), for example, "decreased by 15% in the past 3 hours".

[0165] S302.2 Key Feature Information Extraction To enhance the interpretability of the monitoring results, this embodiment extracts key features that affect the matching score based on the attention mechanism in the spatiotemporal graph neural network and converts them into natural language descriptions.

[0166] (1) Attention weight extraction In the spatiotemporal interaction layer of the spatiotemporal graph neural network, the contribution weight of each input feature dimension to the final matching score is recorded. Specifically: for each node in the scene graph, its attention coefficient after spatial graph convolution and temporal convolution is calculated. For each alignment relationship in the cross-graph fusion features, its relevance score is recorded as a contribution. After normalizing all contributions, the top 3 key features are selected.

[0167] (2) Natural Language Generation Based on the node type and value of key features, a natural language description is generated using a preset template. For example, if the top-ranked key feature is "bottom dissolved oxygen node", with a current value of 4.2 mg / L and a contribution of 0.35, the description generated is: "Main risk factor: Bottom dissolved oxygen is low (4.2 mg / L), below the safety line of 5.0 mg / L, approaching the risk of hypoxia." (3) Generation of comprehensive monitoring report The grading results, key feature descriptions, and matching score trends are integrated to generate a complete monitoring report, which is then pushed to the farmer's mobile app or management platform interface.

[0168] S302.3 Strategy Linkage Trigger When the matching score is lower than the preset threshold, or when the hard constraints are met, this embodiment automatically triggers the corresponding processing strategy to achieve the linkage control of the equipment.

[0169] (1) Configuration of linkage rules This embodiment pre-defines triggering rules in the strategy graph, and each rule includes conditions, actions, and security protection mechanisms: When the overall matching score is below 0.6, or the dissolved oxygen concentration drops below 3.0 mg / L, the system automatically triggers the aerator linkage strategy, starting aerators 1 and 2, with a preset running time of 120 minutes. Simultaneously, to ensure stable equipment operation, the system stipulates a minimum interval of 10 minutes between two consecutive triggers. Furthermore, there is a mutual exclusion relationship between the aerators and the water exchange valves; if the water exchange valve is detected to be open, the aerator start command will be temporarily suspended.

[0170] When the overall matching score falls below 0.5, or the salinity drops sharply to below 20‰, the system triggers a water exchange valve control strategy, automatically opening the inlet valve to 30% and continuing for 60 minutes. Before executing this operation, the system performs a safety check based on weather forecast data. If heavy rain is predicted within the next 2 hours, the operation is delayed to avoid introducing low-salinity water and exacerbating stress.

[0171] When the dissolved oxygen concentration is below 4.0 mg / L and the water temperature is above 28℃, the system triggers an oxygen deficiency risk warning strategy, automatically sending a warning SMS to the mobile phone of the linked aquaculture personnel. To avoid information overload, the system is set to receive a maximum of one warning message per hour for the same number.

[0172] When the sulfide content in the bottom sediment exceeds 0.3 mg / g, the system triggers a bottom sediment cleaning strategy. However, this strategy does not directly control the equipment; instead, it generates a cleaning task order on the aquaculture management platform, prompting aquaculture personnel to perform manual operations. Given that bottom sediment cleaning is a high-risk operation, the system requires manual confirmation on the platform before subsequent cleaning steps can be executed.

[0173] (2) Implementation of equipment control interface This embodiment establishes a connection with field devices through an IoT platform and sends control commands using a standard API interface: Aerator control: POST / api / aerator?action=start&id=1,2&duration=120, returns the device startup status and estimated startup time.

[0174] Warning notification: Send a template message via SMS gateway or push service, such as "[Aquaculture Warning] Dissolved oxygen in cage No. 1 has dropped to 2.8 mg / L, and the aerator has been automatically activated. Please pay attention." (3) Safety protection mechanism To prevent frequent start-ups and shutdowns or misoperations of the equipment, this implementation method adopts the following safety protection measures: Time interval limit: The minimum time interval between two consecutive triggers from the same device is 10 minutes. If the trigger condition is met again within 10 minutes, the system will only log the information but will not take any action.

[0175] Equipment mutual exclusion check: Before executing any action, check for any mutual exclusion relationships. For example, before starting the aerator, check the status of the water exchange valve. If the valve is already open and the opening degree is greater than 20%, then postpone the start of the aerator and send a prompt: "The water exchange valve is open. To avoid power overload, the aerator start is postponed. Manual confirmation is recommended." Manual confirmation mechanism: For high-risk operations (such as bottom cleaning or complete water change), the system only generates a task order and pushes it to the management platform. The aquaculture personnel must click to confirm before the operation can be carried out.

[0176] Abnormal rollback: If the scene graph node values ​​do not improve or even worsen within a preset time (e.g., 30 minutes) after the device is executed, the system will automatically stop the current operation and issue an alarm.

[0177] S302.4 Output combined with counterfactual reasoning Below the monitoring report, resilience scores for different intervention strategies for the current environment are displayed: Preliminary assessment of intervention effects: Option A (starting the bottom aerator): 78% resilience score; dissolved oxygen is expected to rise to 5.2 mg / L after 1 hour. Option B (opening the water exchange valve): 45% resilience score; however, it should be noted that water exchange may cause short-term stress, and the current tide is low, so the effect of water exchange is limited. Option C (cleaning up the net cage attachments): 32% resilience score; the effect is slow and not suitable for emergencies.

[0178] Recommended option: Option A (start the bottom aerator).

[0179] The formula for calculating elasticity score is defined as follows:

[0180] Among them, S base S is the current match score. counter To determine the counterfactual matching score after intervention, S max The ideal match score (usually 1.0) is used. A higher resilience score indicates a more significant improvement effect of the intervention strategy.

[0181] This embodiment performs counterfactual analysis on various possible intervention strategies, calculates their respective resilience scores, and sorts them from highest to lowest. Simultaneously, it considers the execution costs of the strategies (such as energy consumption and time) to generate a comprehensive cost-effectiveness index. For example: Strategy Cost-Effectiveness Assessment: Starting the bottom aerator: 78% elasticity score, energy cost 50 yuan / hour, cost-effectiveness 1.56% / yuan. Starting the surface aerator: 52% elasticity score, energy cost 30 yuan / hour, cost-effectiveness 1.73% / yuan. Opening the water change valve: 45% elasticity score, energy cost 20 yuan / hour, cost-effectiveness 2.25% / yuan, but note that water changes may introduce contaminants. Overall Recommendation: The surface aerator offers the highest cost-effectiveness and lowest risk, and is recommended as the first choice.

[0182] S302.5 Closed-Loop Feedback and Dynamic Update This embodiment achieves complete closed-loop control: after the device is activated, the node values ​​in the scene graph are updated in real time, and the system automatically recalculates the matching score and elasticity score to verify the intervention effect. After the device is started, the sensors continuously collect data, and the scene graph is updated every 10 minutes.

[0183] In some embodiments, the marine aquaculture environmental monitoring system may include multiple functional modules composed of computer program segments. The computer programs for each program segment in the marine aquaculture environmental monitoring system may be stored in the memory of a computer device and executed by at least one processor to perform (see details). Figure 1 (Description) The function of environmental monitoring in marine aquaculture.

[0184] In this embodiment, the marine aquaculture environmental monitoring system can be divided into multiple functional modules according to its functions, such as... Figure 4 As shown. The module referred to in this invention is a series of computer program segments that can be executed by at least one processor and perform a fixed function, and is stored in memory. In this embodiment, the functions of each module will be described in detail in subsequent embodiments.

[0185] The data acquisition module is used to acquire multi-source sensor data, which includes water quality physicochemical indicators, meteorological and hydrological indicators, and biological sediment indicators. The data processing module is used to construct a scene graph based on the multi-source sensor data. The scene graph uses sensor monitoring indicators as feature nodes and physical associations, statistical correlations or time series dependencies between nodes as edges. The strategy matching module is used to calculate the matching degree between the scene graph and the pre-built strategy graph, and output the environment monitoring results or trigger the corresponding processing strategy based on the matching degree. The strategy graph uses preset processing strategies as strategy nodes, logical relationships between processing strategies as edges, and adds condition labels to each strategy node.

[0186] Figure 5 The environmental monitoring method for marine aquaculture provided in this application embodiment can be applied to equipment. Those skilled in the art will understand that the equipment structure involved in the embodiments of this invention does not constitute a limitation on the equipment; the equipment may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements.

[0187] The present invention also provides a computer medium, wherein the computer medium may store a program, which, when executed, may include some or all of the steps provided in the embodiments of the present invention. The medium may be a magnetic disk, an optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0188] Although the present invention has been described in detail with reference to the accompanying drawings and preferred embodiments, the present invention is not limited thereto. Various equivalent modifications or substitutions can be made to the embodiments of the present invention by those skilled in the art without departing from the spirit and essence of the invention, and such modifications or substitutions should all be within the scope of the present invention. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should also be covered within the protection scope of the present invention.

Claims

1. An environmental monitoring method for marine aquaculture, characterized in that, include: Acquire multi-source sensor data, which includes water quality physicochemical indicators, meteorological and hydrological indicators, and biological sediment indicators; A scene graph is constructed based on the multi-source sensor data. The scene graph uses sensor monitoring indicators as feature nodes and physical associations, statistical correlations or time series dependencies between nodes as edges. Calculate the matching degree between the scene graph and the pre-built strategy graph, and output the environmental monitoring results or trigger the corresponding processing strategy based on the matching degree; The strategy graph uses preset processing strategies as strategy nodes, logical relationships between processing strategies as edges, and adds condition labels to each strategy node.

2. The method according to claim 1, characterized in that, Acquire multi-source sensor data, including: Collect water quality physicochemical indicators, meteorological and hydrological indicators, and biological and sedimentary indicators; The water quality physicochemical indicators include at least one of dissolved oxygen, water temperature, salinity, pH value, ammonia nitrogen, nitrite, chlorophyll a and chemical oxygen demand, and sensor nodes are deployed in the surface, middle and bottom layers according to a preset three-dimensional layering strategy during data collection. The meteorological and hydrological indicators include at least one of wind speed, wind direction, light intensity, tide level, and current velocity; The biological substrate indicators include at least one of harmful algal cell density, substrate sulfide content, and redox potential.

3. The method according to claim 1, characterized in that, Constructing a scene map based on the multi-source sensor data includes: The multi-source sensor data is numerically processed to construct an initial feature vector for graph computation; wherein each sensor monitoring index or spatiotemporal attribute is defined as a node, and a multi-dimensional feature vector containing the current real-time value of the index, preset historical statistical values, and dynamic change rate is generated for each node. Using the nodes as vertices and the relationships between nodes as edges, a scene graph with a topological structure is constructed; wherein, the edges include at least one type constructed based on physical laws, statistical correlation, or time dependence, and each type of edge is assigned a weight value representing the strength of the relationship; The scene graph is input into a graph neural network, and the feature vectors of neighboring nodes are aggregated and updated using the edge weights through the graph convolutional layer, so that each node integrates the environmental information of its neighboring nodes. Global pooling is performed on all node features after graph convolution aggregation to generate a fixed-dimensional scene graph embedding vector that represents the overall state of the current aquaculture environment.

4. The method according to claim 3, characterized in that, Define each sensor monitoring metric or spatiotemporal attribute as a node, including: Based on the source and nature of multi-source sensor data, the corresponding nodes are divided into four types: water quality physicochemical nodes, meteorological and hydrological nodes, biological substrate nodes, and spatiotemporal auxiliary nodes. Create an independent node entity for each specific sensor monitoring point or each specific spatiotemporal attribute; each node entity has a unique node identifier and is associated with the corresponding sensor time-series data stream or spatiotemporal attribute value; Each node entity is bound to an attribute set, which includes a node type label, a spatial location label, and a time label. The spatial location label includes the grid coordinates of the surface, middle, bottom, or aquaculture area. The time label includes the current timestamp, tidal phase value, and season identifier.

5. The method according to claim 3, characterized in that, Using the nodes as vertices and the relationships between nodes as edges, construct a scene graph with a topological structure, including: Based on preset association types, edge instances are created between node pairs that meet the conditions; wherein, the association types include physical law association, statistical correlation association, and time series dependency association; Iterate through all node pairs. When a node pair has an inherent association based on a physical mechanism, create a physical law edge for the node pair and assign it a first weight according to the association strength. Obtain node time series data within a preset historical period, calculate the Pearson correlation coefficient between each pair of nodes, and when the absolute value of the correlation coefficient is greater than a preset threshold, create a statistical correlation edge and use the absolute value of the correlation coefficient as the second weight. Calculate the mutual information value of each pair of nodes under different time delays, determine the time delay that maximizes the mutual information, and when the maximum mutual information value is greater than a preset threshold, create a time series dependency edge and use the normalized mutual information value as the third weight. When multiple types of edges exist between the same pair of nodes, the edge with the highest weight is retained. All node entities and the created edge instances are assembled to form a graph structure data containing a set of nodes and an adjacency matrix. The adjacency matrix records the connection relationship between each pair of nodes and the corresponding edge weight.

6. The method according to claim 1, characterized in that, The method for constructing the strategy graph includes: The preset processing strategy is defined as a strategy node, and a condition label and a priority label are attached to each strategy node; wherein, the strategy node includes at least one of emergency response node, optimization adjustment node, early warning node and special environment node; Directed edges are constructed based on the logical relationships between policy nodes, and each edge is assigned a weight representing the relationship type; wherein, the logical relationship includes at least one of triggering relationship, dependency relationship and mutual exclusion relationship; The condition label of each policy node is converted into a numerical feature vector, which includes policy type encoding, condition threshold normalization value, and condition hardness identifier. The policy graph is encoded using a graph neural network, and the feature information of neighboring policy nodes is aggregated through graph convolutional layers to generate a policy graph embedding vector that represents the semantics of policy combination.

7. The method according to claim 1, characterized in that, Calculating the matching degree between the scene graph and the pre-constructed policy graph includes: The condition labels of each policy node in the policy graph are transformed into a quantifiable constraint space, weights are assigned to different constraint types, and the constraint response coefficient between the current value of the scene graph node and the threshold of the policy node is calculated. Calculate the correlation score between scene graph nodes and policy graph nodes, filter alignment relationships that meet a preset threshold, and generate cross-graph fusion features that fuse scene graph features and policy graph features based on the alignment relationships; The scene graph embedding vector, the policy graph embedding vector, and the cross-graph fusion features are input into a pre-defined spatiotemporal graph neural network model to calculate the comprehensive matching score between the scene graph and the policy graph.

8. An environmental monitoring system for marine aquaculture, characterized in that, include: The data acquisition module is used to acquire multi-source sensor data, which includes water quality physicochemical indicators, meteorological and hydrological indicators, and biological sediment indicators. The data processing module is used to construct a scene graph based on the multi-source sensor data. The scene graph uses sensor monitoring indicators as feature nodes and physical associations, statistical correlations or time series dependencies between nodes as edges. The strategy matching module is used to calculate the matching degree between the scene graph and the pre-built strategy graph, and output the environment monitoring results or trigger the corresponding processing strategy based on the matching degree. The strategy graph uses preset processing strategies as strategy nodes, logical relationships between processing strategies as edges, and adds condition labels to each strategy node.

9. An environmental monitoring device for marine aquaculture, characterized in that, include: Memory, used to store environmental monitoring programs for marine aquaculture; A processor is used to implement the steps of the marine aquaculture environmental monitoring method as described in any one of claims 1-7 when executing the marine aquaculture environmental monitoring program.

10. A computer-readable medium storing a computer program, characterized in that, The readable medium stores an environmental monitoring program for marine aquaculture, which, when executed by a processor, implements the steps of the environmental monitoring method for marine aquaculture as described in any one of claims 1-7.

Images