Circulating water aquaculture precision regulation method and system based on graph network water resource management

Abstract

The application provides a circulating water aquaculture precision regulation method and system based on a graph network water resource management, relates to the technical field of fishery regulation methods, and comprises sequentially performing high-frequency synchronous collection processing and Kalman filter noise fusion suppression processing on collected multidimensional data to obtain a spatiotemporally aligned and standardized multimodal water quality time series dataset; defining a core water quality parameter type as a node of a dynamic heterogeneous graph to obtain a dynamic heterogeneous graph structure with edge weight containing time delay mutual information entropy and nodes containing multidimensional features; processing the dynamic heterogeneous graph structure by using a gated spatiotemporal convolutional graph neural network to obtain a multi-dimensional water quality state prediction matrix including potential anomaly probability, water quality risk level, water cycle efficiency and biological fitness; and processing by using a multi-objective constraint reinforcement learning algorithm to realize precise fishery circulating water regulation. The application significantly reduces water resource consumption cost and provides key technical support for water saving and emission reduction of land-based industrialized aquaculture.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fishery regulation, in particular, to a circulating water aquaculture precise regulation method and system based on graph network water resource management. BACKGROUND

[0002] At present, land-based factory farming, as a core mode to break away from the dependence on natural water areas and realize the intensification of aquaculture, its efficient use and fine management of water resources have become a key constraint factor for the sustainable development of the industry. Under the land-based farming scenario, water resources cannot rely on natural water replenishment and need to realize water reuse through a circulating water system. The water saving effect and water resource utilization efficiency directly determine the cost and environmental benefits of aquaculture. Therefore, water resource management has become one of the core technical needs of land-based factory farming.

[0003] At present, in the land-based factory aquaculture, the aquaculture water quality regulation system is relatively common, but it still has obvious technical shortcomings in water resource management, and it is difficult to achieve efficient and coordinated regulation:

[0004] First, most systems fail to fully link the dynamic coupling relationship of "water quality-fish population-equipment-water resources", and often treat water resource regulation as a relatively independent link from water quality and fish population status. For example, some systems will increase the daily water exchange rate (the daily water exchange rate of some scenes can reach more than 30% of the total water body) to quickly reduce the ammonia nitrogen in the water body, without considering the fish population metabolism and equipment processing capacity, which easily leads to unreasonable consumption of water resources;

[0005] Second, the existing systems generally lack quantitative modeling capabilities related to water resource consumption and aquaculture system load. When the metabolic waste of a certain aquaculture pond increases due to the increase in fish population density, the system cannot predict the chain effect of this change on the water resource allocation of the entire circulating water system, and relies on global water exchange to cope, further exacerbating water resource consumption;

[0006] Third, the water resource management of the system is relatively extensive, and mostly relies on manual regular adjustment of water exchange parameters. Although some real-time water quality data can be collected, it is not possible to dynamically optimize the water resource allocation strategy in combination with the fish population activity state (such as feeding and swimming conditions), resulting in generally low water resource utilization rate of land-based aquaculture systems. SUMMARY

[0007] The purpose of the present application is to provide a circulating water aquaculture precise regulation method and system based on graph network water resource management to improve the above problems. In order to achieve the above purpose, the technical scheme adopted by the present application is as follows:

[0008] In a first aspect, the present application provides a circulating water aquaculture precise regulation method based on graph network water resource management, comprising:

[0009] Deploy a multi-source sensor array covering the water inlet, biochemical reactor and water outlet of the aquaculture pond, collect multi-dimensional data of the fishery recirculating water system through the multi-source sensor array, sequentially perform high-frequency synchronous collection processing and Kalman filter noise fusion suppression processing on the collected multi-dimensional data, and combine the historical optimal breeding cycle growth coefficient to perform Z-score standardization processing, to obtain a spatiotemporally aligned and standardized multi-modal water quality time series dataset;

[0010] Based on the multi-modal water quality time series dataset, a dynamic heterogeneous graph is constructed: the core water quality parameter types are defined as nodes of the dynamic heterogeneous graph, wherein the node attributes are set as a multi-dimensional feature vector composed of parameter monitoring values, parameter change rates, water-borne organism metabolism correlation coefficients and equipment energy consumption matching degrees; the causal correlation strength within the time delay window between parameters is defined as the edge of the dynamic heterogeneous graph, the edge weight is calculated by time delay mutual information entropy, and a sparse adjacency matrix is constructed by L1 regularization to obtain a dynamic heterogeneous graph structure with time delay mutual information entropy and multi-dimensional features of nodes;

[0011] The dynamic heterogeneous graph structure is processed by using a gated spatio-temporal convolutional graph neural network, which includes aggregating neighbor node information within a specified range by an attention gate module to assign high attention weights to key water quality sensitive nodes; capturing long-period dynamic evolution patterns of parameters by a time convolution module; and fusing a gray prediction model to complete the trend of low-frequency parameters to obtain a multi-dimensional water quality state prediction matrix including potential anomaly probability, water quality risk level, water cycle efficiency and biological fitness;

[0012] Based on the multi-dimensional water quality state prediction matrix, a multi-objective constraint reinforcement learning algorithm is used for processing: using the compliance water quality parameter threshold as the water quality safety hard constraint, and using the preset daily water saving rate as the optimization target soft constraint; combining the chaotic disturbance improved particle swarm optimization method and the fuzzy decision screening mechanism to solve the optimal parameter combination covering valve opening, pump speed, backwashing frequency and oxygenation timing to obtain the precise control instruction set;

[0013] The precise control instruction set is subjected to distributed collaborative control processing of the actuator cluster: the industrial standard communication protocol conversion instruction is used to realize multi-type actuator linkage control through edge computing nodes; a redundant actuator dynamic switching mechanism is introduced to drive the actuator action to complete the periodic state update of the recirculating water system, and precise fishery recirculating water control is realized.

[0014] Preferably, the deployment covers the multi-source sensor array of the aquaculture pond inlet, biochemical reactor and outlet, through the multi-source sensor array to collect multi-dimensional data of the fishery circulating water system, and sequentially perform high-frequency synchronous collection processing and Kalman filter noise fusion suppression processing on the collected multi-dimensional data, and combine the historical optimal growth coefficient of the breeding period to perform Z-score standardization processing, to obtain a spatio-temporally aligned and standardized multi-modal water quality time series dataset, which includes:

[0015] An optical dissolved oxygen sensor and a water temperature sensor are deployed at the inlet of the aquaculture pond, a pH glass electrode and an ORP sensor are deployed inside the biochemical reactor, and a turbidity-ammonia nitrogen integrated sensor is deployed at the outlet, to form a multi-source sensor array covering key monitoring points; through the multi-source sensor array, multi-dimensional raw data of the fishery circulating water system is collected, including water quality parameter data, equipment operation data and fish population monitoring data, to obtain a multi-dimensional raw data set;

[0016] The multi-dimensional raw data is synchronously collected, Kalman filtering is used to suppress noise fusion of the water quality parameter data, and a dynamic threshold based on the fish population metabolism model is introduced for screening, combined with the water quality tolerance range of different growth stages of the fish population, to exclude outliers of the equipment operation data and the fish population monitoring data, to obtain a pre-processed data set;

[0017] The pre-processed data set is subjected to Z-score standardization in combination with the historical optimal growth coefficient of the complete breeding period;

[0018] The Pearson correlation coefficient of any two parameters is calculated to add parameter correlation information to the standardized data, and finally a multi-modal water quality time series dataset that is spatio-temporally aligned and contains parameter correlation information is obtained, wherein the multi-modal water quality time series dataset forms a structured matrix with time steps as rows and monitoring parameters as columns.

[0019] Preferably, a dynamic heterogeneous graph is constructed based on the multi-modal water quality time series dataset: the core water quality parameter type is defined as the node of the dynamic heterogeneous graph, wherein the node attribute is set as a multi-dimensional feature vector composed of parameter monitoring value, parameter change rate, aquatic organism metabolism correlation coefficient and equipment energy consumption matching degree; the causal correlation strength within the time delay window between parameters is defined as the edge of the dynamic heterogeneous graph, the edge weight is calculated by time delay mutual information entropy, and a sparse adjacency matrix is constructed by L1 regularization, to obtain a dynamic heterogeneous graph structure with edge weight containing time delay mutual information entropy and node containing multi-dimensional features, which includes:

[0020] Based on the multi-modal water quality time series data set, water temperature, pH, dissolved oxygen, ammonia nitrogen and turbidity are defined as the core water quality parameter nodes of the dynamic heterogeneous graph; the importance weight of each node is set according to the growth stage of the fish population, and a multi-dimensional attribute vector is constructed for each node, including the real-time monitoring value of the parameter, the 24-hour sliding window change rate, the water-biological metabolism correlation coefficient and the equipment energy consumption matching degree, to obtain the node set with dynamic importance weight and the node attribute vector;

[0021] The causal correlation strength within the parameter time delay window is defined as the edge of the graph, and the time delay window is set according to the response period of the water quality parameters on the growth of the fish population; the correlation strength between any two nodes is calculated by time delay mutual information entropy to obtain the initial edge weight; the edge weight is iteratively corrected every hour according to the latest water quality data and fish behavior data, the initial edge weight matrix is sparsified by L1 regularization, weakly correlated edges are removed, and a sparse adjacency matrix is constructed;

[0022] Integrate the node set, edge set and sparse adjacency matrix to form a dynamic heterogeneous graph, and mark the edges with changing correlation strength to obtain a dynamic heterogeneous graph that can dynamically reflect the changes of the correlation between water quality parameters over time.

[0023] Preferably, the dynamic heterogeneous graph structure is processed by using a gated spatio-temporal convolutional graph neural network, which includes aggregating neighbor node information within a specified range by an attention gate module to assign high attention weights to key water quality sensitive nodes; capturing the long-period dynamic evolution pattern of the parameter by a time convolution module; and fusing a gray prediction model to complete the trend of low-frequency parameters to obtain a multi-dimensional water quality state prediction matrix including potential anomaly probability, water quality risk level, water cycle efficiency and biological fitness, which includes:

[0024] Based on the dynamic heterogeneous graph, a gated spatio-temporal convolutional graph neural network is used for processing; the neighbor node features within a specified range around each node are aggregated by an attention gate module to assign attention weights to key water quality sensitive nodes that affect fish metabolism, such as dissolved oxygen and ammonia nitrogen, and the node features extracted by different layers of the graph neural network are interactively fused to obtain a node feature set that fuses cross-layer information;

[0025] The node feature set is processed in time series dimension by a time convolution module to capture the dynamic evolution pattern of the water quality parameter in a long period and identify the periodicity of the parameter change; the water quality parameter change period in different seasons and different breeding stages is automatically identified, and for water quality parameters with low sampling frequency, a gray prediction model is fused to complete the trend of the parameter change to obtain a node feature set with complete time series characteristics;

[0026] The extracted time sequence feature complete node feature set is classified and regressed by a neural network prediction layer, multi-modal joint prediction is performed by fusing fish behavior features and water quality features, and a multi-dimensional water quality state prediction matrix including potential anomaly probability, water quality risk level, water cycle efficiency and biological fitness is output.

[0027] Preferably, the multi-dimensional water quality state prediction matrix is processed by a multi-objective constraint reinforcement learning algorithm: using the compliance water quality parameter threshold as the water quality safety hard constraint, and using the preset daily water saving rate as the optimization target soft constraint; combining the chaotic disturbance improved particle swarm optimization method and the fuzzy decision screening mechanism, the optimal parameter combination covering the valve opening, pump speed, backwashing frequency and oxygenation timing is solved, and the accurate regulation and control instruction set is obtained, which includes:

[0028] According to the multi-dimensional water quality state prediction matrix, the constraint conditions of multi-objective optimization are set; the compliance threshold of pH and dissolved oxygen parameters is set as the water quality safety hard constraint, the daily water saving rate target is adjusted based on the water quality risk level, and the daily water saving rate is set as the optimization target soft constraint, to obtain a multi-objective optimization constraint condition set for initializing the regulation and control parameter population;

[0029] A chaotic disturbance improved particle swarm optimization algorithm is used to initialize the regulation and control parameter population according to the multi-objective optimization constraint conditions, wherein the parameter population includes parameter combinations of valve opening, pump speed, backwashing frequency and oxygenation timing; a parameter coordination constraint factor is set to determine the coordination of valve opening and pump speed, and backwashing frequency and oxygenation timing; the fitness value of each population individual is calculated by iterative optimization, wherein the fitness value is determined by water quality compliance rate, water saving amount and equipment energy consumption, and then a plurality of candidate regulation and control parameters are screened out, and a candidate regulation and control parameter set is obtained;

[0030] A fuzzy decision mechanism is introduced to perform secondary screening on the candidate regulation and control parameter set, and the optimal parameter combination is determined according to the actual operation scene of the aquaculture system; based on the optimal parameter combination, the historical regulation and control effect under the same scene is used as a weight to correct the candidate parameters, and finally the accurate regulation and control instruction set is obtained.

[0031] Preferably, the accurate regulation and control instruction set is subjected to distributed collaborative control processing of an executor cluster: using an industrial standard communication protocol conversion instruction to realize multi-type executor linkage control through an edge computing node; introducing a redundant executor dynamic switching mechanism to drive the executor action to complete the periodic state update of the recirculating water system and realize accurate regulation and control of the fishery recirculating water, which includes:

[0032] Based on the precise regulation instruction set, the instruction is converted into a digital signal recognizable by the actuator by adopting the OPC-UA industrial standard communication protocol; the instructions related to water quality safety are marked with priority, and the digital signal is transmitted to the edge computing node through the industrial Ethernet, wherein the edge computing node checks the signal to obtain the checked actuator control signal;

[0033] The checked actuator control signal is called to transmit control instructions to the electric valve, the variable frequency water pump and the oxygenator, wherein the control instructions include control of electric valve opening adjustment, variable frequency water pump speed adjustment and oxygenator working time sequence switching; the action amplitude is adjusted according to the real-time load of each actuator, the working state data of the actuator is collected in real time, the equipment running state is monitored, and the actuator working state data and the linkage control result are obtained;

[0034] A 1+1 redundant actuator dynamic switching mechanism is introduced, when the main actuator fails, the edge computing node triggers the standby actuator to start based on the actuator working state data; the regulation parameters are adjusted according to the performance difference of the standby actuator, the actuator is driven to act according to the checked actuator control signal, the periodic state update of the circulating water system is completed, and the precise collaborative regulation and control of water quality and water saving are realized.

[0035] In the second aspect, the application also provides a circulating water aquaculture precise regulation and control system based on a graph network water resource management, comprising:

[0036] The acquisition module is used for deploying a multi-source sensor array covering the inlet of the breeding pond, the biochemical reactor and the outlet, collecting multi-dimensional data of the fishery circulating water system through the multi-source sensor array, sequentially performing high-frequency synchronous acquisition processing and Kalman filter noise fusion suppression processing on the collected multi-dimensional data, and combining the historical optimal breeding cycle growth coefficient to perform Z-score standardization processing, to obtain a spatiotemporally aligned and standardized multi-modal water quality time series data set;

[0037] The construction module is used for constructing a dynamic heterogeneous graph based on the multi-modal water quality time series data set: defining the core water quality parameter type as a node of the dynamic heterogeneous graph, wherein the node attribute is set as a multi-dimensional feature vector composed of parameter monitoring values, parameter change rates, water organism metabolism correlation coefficients and device energy consumption matching degrees; defining the causal correlation strength within the time delay window between parameters as an edge of the dynamic heterogeneous graph, calculating the edge weight through time delay mutual information entropy, constructing a sparse adjacency matrix by adopting L1 regularization, and obtaining the dynamic heterogeneous graph structure with edge weight containing time delay mutual information entropy and node containing multi-dimensional features;

[0038] The processing module is used for processing the dynamic heterogeneous graph structure by using a gated spatio-temporal convolution graph neural network, wherein the neighbor node information in a specified range is aggregated by an attention gate module, and high attention weights are assigned to key water quality sensitive nodes; a long-period dynamic evolution mode of parameters is captured by a time convolution module; a gray prediction model is fused to complete the variation trend of low-frequency parameters, and a multi-dimensional water quality state prediction matrix including potential anomaly probability, water quality risk level, water cycle efficiency and biological fitness is obtained;

[0039] The solving module is used for processing based on the multi-dimensional water quality state prediction matrix by using a multi-objective constraint reinforcement learning algorithm: taking the compliant water quality parameter threshold as a water quality safety hard constraint, and taking the preset daily water saving rate as an optimization target soft constraint; a chaotic disturbance improved particle swarm optimization method and a fuzzy decision screening mechanism are combined to solve an optimal parameter combination covering valve opening, pump speed, backwashing frequency and oxygenation timing, and a precise regulation and control instruction set is obtained;

[0040] The driving module is used for performing distributed collaborative control processing of the precise regulation and control instruction set on the executor cluster: using an industrial standard communication protocol conversion instruction, realizing multi-type executor linkage control through an edge computing node; introducing a redundant executor dynamic switching mechanism, driving the executor action to complete the periodic state update of the circulating water system, and realizing precise aquaculture circulating water regulation and control.

[0041] In a third aspect, the present application further provides a circulating water aquaculture precise regulation and control device based on graph network water resource management, comprising:

[0042] A memory is used for storing a computer program;

[0043] A processor is used for executing the computer program to realize the steps of the circulating water aquaculture precise regulation and control method based on graph network water resource management.

[0044] In a fourth aspect, the present application further provides a readable storage medium, wherein the readable storage medium stores a computer program, and the computer program is executed by a processor to realize the steps of the circulating water aquaculture precise regulation and control method based on graph network water resource management.

[0045] The present application has the following beneficial effects:

[0046] The present application takes "efficient use of water resources" as the core goal, and creatively combines graph network technology with the multi-node association characteristics of the circulating water system: by constructing a dynamic heterogeneous graph integrating water quality, fish population, equipment and water resource consumption, the influence weight of each link on water resource utilization is quantified; the gated spatio-temporal convolutional graph neural network is used to predict the water quality change trend, and the passive situation of "increasing water to maintain water quality" is avoided in advance; combined with the multi-objective constraint reinforcement learning algorithm, "maximizing daily water saving rate" and "improving water resource reuse rate" are taken as the core optimization objectives within the water quality safety boundary, and the optimal control strategy is generated; finally, through the collaborative control of the actuator cluster, the precise matching of water resource allocation and equipment operation is realized. The method breaks through the limitation of the prior art "paying attention to water quality and ignoring water saving", realizes the transformation from "extensive water use" to "precise water saving" and from "passive water supplement" to "active water resource management", and can improve the water resource reuse rate of the land-based factory farming system to more than 85%, significantly reducing the water resource consumption cost, and providing key technical support for water saving and emission reduction and sustainable development of land-based factory farming.

[0047] The present application collects data through a multi-source sensor array covering key nodes of the culture pond, combines noise suppression and standardized processing of biological growth adaptation information, adds parameter correlation information, provides a high-quality basis for subsequent modeling, avoids analysis deviation caused by poor data quality, defines core water quality parameters as nodes with multi-dimensional features including "physics-biology-equipment", defines parameter time delay causal correlation as dynamic weight edges, constructs a sparse adjacency matrix, identifies parameter correlation mutations in advance, improves the representation accuracy of the graph model for system dynamic characteristics, provides a structured and dynamic input carrier for spatio-temporal feature extraction, and avoids misjudgment of correlation caused by static modeling; the method can focus on key water quality nodes to improve feature contribution, accurately capture long-period water quality change rules, reduce low-frequency parameter completion errors, generate multi-dimensional water quality state prediction results including abnormal probability, risk level and other information, realize advanced prediction of water quality fluctuations, and reduce water quality exceeding situations.

[0048] The present application takes water quality safety hard constraint and dynamic water saving soft constraint as the guide, combines improved optimization algorithm and fuzzy decision mechanism, ensures that the theoretical optimal parameters adapt to the actual scene, improves the water resource utilization efficiency under the premise of meeting the water quality safety, and reduces the equipment energy consumption; and through industrial standard protocol conversion, edge computing linkage control and redundancy switching mechanism, the compatible control of multi-vendor actuators is realized, the continuous operation of the system is ensured, and finally the breakthrough from "experience-driven extensive regulation and control" to "data-driven precise regulation and control" of the land-based factory farming circulating water is realized, providing key technical support for water saving and emission reduction and sustainable development of the industry.

[0049] Other features and advantages of the present application will be set forth in the following description, and in part will be apparent from the description, or can be learned by practice of the application. The objects and other advantages of the present application will be realized and attained by the structure particularly pointed out in the written description and claims hereof as well as the appended drawings. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some of the embodiments of the present application, and therefore should not be considered as limiting the scope. For those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0051] Figure 1 Flowchart of the precise regulation and control method for recirculating aquaculture based on graph network water resource management described in the embodiments of the present application;

[0052] Figure 2 Structure diagram of the precise regulation and control system for recirculating aquaculture based on graph network water resource management described in the embodiments of the present application;

[0053] Figure 3 Structure diagram of the precise regulation and control device for recirculating aquaculture based on graph network water resource management described in the embodiments of the present application.

[0054] In the figure: 701, acquisition module; 702, construction module; 703, processing module; 704, solving module; 705, driving module; 800, precise regulation and control device for recirculating aquaculture based on graph network water resource management; 801, processor; 802, memory; 803, multimedia assembly; 804, I / O interface; 805, communication assembly. DETAILED DESCRIPTION

[0055] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the following will combine the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are some of the embodiments of the present application, but not all the embodiments. The components of the embodiments of the present application described and shown in the drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0056] It should be noted that similar reference numerals and letters refer to like items in the accompanying drawings, and once an item is defined in one drawing, it need not be further defined and explained in subsequent drawings. Also, in the description of the present application, the terms "first", "second", and the like are used only to distinguish descriptions, and cannot be understood as indicating or implying relative importance.

[0057] Embodiment 1:

[0058] The embodiment provides a precise regulation method for recirculating aquaculture based on a graph network water resource management.

[0059] Referring to Figure 1 , the method includes steps S100, S200, S300, S400 and S500.

[0060] S100, deploy a multi-source sensor array covering the water inlet of the culture pond, the biochemical reactor and the water outlet, collect multi-dimensional data of the fishery recirculating water system through the multi-source sensor array, sequentially perform high-frequency synchronous collection processing and Kalman filter noise fusion suppression processing on the collected multi-dimensional data, and perform Z-score standardization processing in combination with a historical optimal culture cycle growth coefficient, to obtain a spatiotemporally aligned and standardized multi-modal water quality time series data set.

[0061] It can be understood that the present step S100 includes S101, S102, S103 and S104.

[0062] S101, deploy an optical dissolved oxygen sensor and a water temperature sensor at the water inlet of the culture pond, deploy a pH glass electrode and an ORP sensor inside the biochemical reactor, and deploy a turbidity-ammonia nitrogen integrated sensor at the water outlet, to form a multi-source sensor array covering key monitoring points; collect multi-dimensional original data of the fishery recirculating water system through the multi-source sensor array, wherein the multi-dimensional original data includes water quality parameter data, equipment operation data and fish school monitoring data, to obtain a multi-dimensional original data set.

[0063] It should be noted that the water quality parameter data includes water temperature, pH, dissolved oxygen, ammonia nitrogen and turbidity, the equipment operation data includes water pump frequency, aerator power and filter pressure difference, and the fish school monitoring data includes cluster density and swimming speed.

[0064] S102, synchronously collect the multi-dimensional original data, perform noise fusion suppression on the water quality parameter data by using Kalman filtering, introduce a dynamic threshold screening based on a fish school metabolism model, perform outlier rejection on the equipment operation data and the fish school monitoring data in combination with a water quality tolerance range of different growth stages of the fish school, to obtain a preprocessed data set.

[0065] S103, in combination with the historical optimal growth coefficient of the complete cultivation cycle, the pre-processed data set is subjected to Z-score standardization, and the calculation formula is as follows:

[0066]

[0067] In the formula, is the normalized multi-modal water quality time series data, is the original collection value of the water quality parameter, is the mean value of the corresponding parameter in the historical period, is the historical optimal growth coefficient, is the standard deviation of the corresponding parameter in the historical period;

[0068] S104, by calculating the Pearson correlation coefficient of any two parameters, the standardized data is added with parameter correlation degree information, and finally the multi-modal water quality time series data set with space-time alignment and parameter correlation degree information is obtained, wherein the multi-modal water quality time series data set is formed into a structured matrix with time steps as rows and monitoring parameters as columns.

[0069] It should be noted that, with the core target of obtaining high-quality, space-time consistent multi-dimensional data, the first step is to deploy a multi-source sensor array and collect data. At the key nodes of the land-based factory farming system, sensors are deployed according to functional requirements: optical dissolved oxygen sensors, water temperature sensors and turbidity sensors are deployed at the inlet of the breeding pond to monitor the initial water quality of the inlet water and avoid external pollution impact; pH glass electrodes, ORP sensors and ammonia nitrogen ion selective electrodes are deployed inside the biochemical reactor to track the regulation effect of microbial metabolism on water quality in real time; turbidity-COD integrated sensors and total phosphorus sensors are deployed at the outlet to control the water quality of discharge or recycling. At the same time, equipment operation data (such as water pump flow, backwash valve state) and biological data (such as fish swimming speed, feeding frequency) are collected synchronously to build a "water quality-equipment-biology" multi-dimensional raw data set, laying a foundation for subsequent processing.

[0070] Specifically, combined with the characteristics of water quality parameters, the sampling interval is set, and for second-level fluctuation parameters such as dissolved oxygen, a 5-10 second sampling interval is adopted, and time stamp calibration is used to ensure the space-time alignment of data at each node, avoiding misjudgment of correlation analysis caused by time deviation. For sensor drift, electromagnetic interference and other problems, Kalman filter algorithm is used for processing: a dynamic model of water quality parameters (such as dissolved oxygen decay curve) is established, the parameter value at the next moment is predicted and weighted with the actual collection value, and abnormal fluctuations are filtered out; at the same time, according to the parameter change characteristics, the filter window is adjusted, and for slowly changing parameters such as water temperature, a 30-minute window is used, and for fast-changing parameters such as ORP, a 5-minute window is used, to ensure that the noise suppression effect is adapted to different parameter characteristics.

[0071] The core of the Z-score standardization and historical data fusion link is to eliminate dimensional differences and integrate biological growth adaptation information. By analyzing the adaptation relationship between the survival rate and weight gain rate of the fish population in the past complete cultivation cycle and the water quality parameters, such as the dissolved oxygen adaptation value corresponding to the historical optimal generation coefficient of 1.2 in the juvenile period and the historical optimal generation coefficient of 0.9 in the adult fish period, the standardized data not only eliminates the dimension, but also meets the biological growth demand. The standardized data is stored in a time sequence matrix (rows represent time steps and columns represent parameter types), and parameter correlation labels (such as the Pearson correlation coefficient of dissolved oxygen and water temperature) are added, providing initial correlation basis for subsequent dynamic heterogeneous graph construction, realizing the transformation of data from “physical dimension unification” to “analysis usability”.

[0072] S200, constructing a dynamic heterogeneous graph based on a multi-modal water quality time series dataset: defining core water quality parameter types as nodes of the dynamic heterogeneous graph, wherein the node attributes are set as a multi-dimensional feature vector composed of parameter monitoring values, parameter change rates, water-borne biological metabolic correlation coefficients, and equipment energy consumption matching degrees; defining the causal correlation strength within the time delay window between parameters as the edge of the dynamic heterogeneous graph, calculating the edge weight by time delay mutual information entropy, constructing a sparse adjacency matrix by L1 regularization, and obtaining the dynamic heterogeneous graph structure with time delay mutual information entropy in the edge weight and multi-dimensional features in the node.

[0073] It can be understood that the present step S200 includes S201, S202 and S203.

[0074] S201, based on the multi-modal water quality time series dataset, defining water temperature, pH, dissolved oxygen, ammonia nitrogen and turbidity as core water quality parameter nodes of the dynamic heterogeneous graph; setting importance weights adjusted according to the growth stage of the fish population for each node, and constructing a multi-dimensional attribute vector for each node, including parameter real-time monitoring value, 24-hour sliding window change rate, water-borne biological metabolic correlation coefficient and equipment energy consumption matching degree, obtaining a node set with dynamic importance weights and node attribute vectors;

[0075] S202, defining the causal correlation strength within the time delay window between parameters as the edge of the graph, and setting the time delay window according to the response period of the water quality parameters on the growth of the fish population; calculating the correlation strength between any two nodes by time delay mutual information entropy to obtain the initial edge weight; iteratively correcting the edge weight according to the latest water quality data and fish population behavior data every hour, sparsifying the initial edge weight matrix by L1 regularization, removing weakly correlated edges, and constructing a sparse adjacency matrix, whose weight calculation formula is as follows:

[0076]

[0077] In the formula, is the edge weight between node i and node j, The time parameter at t And The time parameter The time delay mutual information entropy, The time delay window, The L1 regularization processing of the weight matrix W is performed;

[0078] S203, integrate the node set, the edge set and the sparse adjacency matrix to form a dynamic heterogeneous graph, and mark the edges with changing correlation strength to obtain a dynamic heterogeneous graph that can dynamically reflect the change of the correlation between the water quality parameters over time.

[0079] It should be noted that G=(V, E, M); wherein G is a dynamic heterogeneous graph, V is a node set (including all core water quality parameter nodes), E is an edge set (composed of effective causal correlation edges between parameters), and M is a sparse adjacency matrix (elements are edge weights processed by sparsification).

[0080] In this step, the node definition needs to focus on the core water quality parameters of the land-based factory farming system, and combined with the farming needs and data availability, five types of core parameters, water temperature, pH, dissolved oxygen, ammonia nitrogen and turbidity, are selected as the nodes of the dynamic heterogeneous graph - these five types of parameters directly determine the fish survival environment and the operating efficiency of the recirculating water system: water temperature affects the metabolic rate of fish and microbial activity, pH is related to water chemical balance and fish gill respiratory function, dissolved oxygen is a key factor for fish survival, ammonia nitrogen is the main pollutant of fish metabolism, and turbidity reflects the suspended solids content of water (affects the filtration system load). The node attribute is not a single parameter value, but a four-dimensional feature vector containing "physical characteristics-biological correlation-equipment correlation":

[0081] 1. Parameter monitoring value: that is, the real-time parameter value after Z-score standardization (such as the standardized dissolved oxygen value of 1.2), which reflects the basic physical quantity of the current water quality state;

[0082] 2. Parameter change rate: the parameter change slope calculated by a 24-hour sliding window (such as an increase of 0.1 mg / L per hour of ammonia nitrogen), which captures the dynamic evolution trend of the parameter to avoid static judgment relying only on instantaneous value;

[0083] 3. Aquatic organism metabolism correlation coefficient: obtained by fitting the Pearson correlation coefficient of fish feeding intensity, swimming speed and water quality parameters (such as the correlation coefficient of dissolved oxygen and fish feeding intensity is 0.8), which quantifies the influence of water quality on organisms;

[0084] 4. Equipment energy consumption matching degree: calculate the linear regression coefficient of water quality parameters and key equipment (water pump, aerator) energy consumption (such as an increase of 5% in water pump energy consumption per 1 NTU increase in turbidity), and establish the correlation between water quality and equipment operating cost.

[0085] The property design breaks through the limitation of traditional graph model "node only contains physical parameters", integrates biological and equipment dimension information into the node, and makes the graph structure more suitable for the actual needs of land-based aquaculture "water quality-biology-equipment" coordinated regulation.

[0086] In step S202, L1 regularization (penalty coefficient λ is usually set to 0.05-0.1) is applied to the initial weight matrix, and weakly associated edges with weights less than the threshold value are removed (for example, if the correlation weight between turbidity and pH is less than 0.1, the edge is removed), and a sparse adjacency matrix is constructed. This can reduce the complexity of the graph model and avoid interference from weakly associated edges in subsequent feature extraction. In land-based aquaculture systems, not all parameters have significant causal relationships (for example, the direct correlation between water temperature and total phosphorus is very weak), and sparse processing can focus on core associations (for example, ammonia-nitrogen-dissolved oxygen, turbidity-pump energy consumption).

[0087] Specifically, in this step S203, the node set (containing four-dimensional feature vectors), the edge set (containing time-delay causal relationships), and the sparse adjacency matrix (containing regularized weights) are integrated to form a dynamic heterogeneous graph, where "dynamic" refers to two aspects:

[0088] 1. Dynamic update of weights: Based on the latest multi-modal time series data collected every hour, the time-delay mutual information entropy and edge weight are recalculated to capture the changes in parameter correlation over time (for example, when fish metabolism is active during the day, the correlation weight between ammonia-nitrogen and dissolved oxygen increases; at night, when metabolism slows down, the weight decreases);

[0089] 2. Dynamic adjustment of topological structure: When the correlation strength of a certain parameter suddenly increases / decreases (for example, after backwashing, the correlation weight between turbidity and filter pressure difference increases sharply), the existence state of the edge is automatically adjusted to ensure that the graph structure always reflects the real-time correlation of the system.

[0090] The final dynamic heterogeneous graph not only quantifies the causal relationships and spatiotemporal characteristics between parameters, but also integrates biological and equipment dimension information, laying a foundation for subsequent gated spatiotemporal convolutional graph neural networks to accurately extract spatiotemporal features.

[0091] S300, a gated spatiotemporal convolutional graph neural network is used to process the dynamic heterogeneous graph structure, which includes aggregating neighbor node information within a specified range through an attention gate module to assign high attention weights to key water quality sensitive nodes; capturing long-period dynamic evolution patterns of parameters through a time convolution module; and integrating a gray prediction model to complete the trend of low-frequency parameters, obtaining a multi-dimensional water quality state prediction matrix including potential anomaly probability, water quality risk level, water cycle efficiency, and biological fitness.

[0092] It can be understood that step S300 includes S301, S302, and S303.

[0093] S301, based on a dynamic heterogeneous graph, a gated spatio-temporal convolutional graph neural network is used for processing; the attention gate module is used to aggregate the features of the neighbor nodes within a specified range around each node, and the attention weight is assigned to the key water quality sensitive nodes that affect the metabolism of fish groups due to dissolved oxygen and ammonia nitrogen, and the node features extracted by different layers of the graph neural network are interactively fused to obtain a set of node features with fused cross-layer information;

[0094] S302, the time convolution module is used to process the node feature set in the time sequence dimension, capture the dynamic evolution mode of the water quality parameters in a long period, and identify the periodicity of parameter changes; the water quality parameter change period in different seasons and different breeding stages is automatically identified, and for the water quality parameters with low sampling frequency, the gray prediction model is used to complete the trend, and a node feature set with complete time sequence characteristics is obtained;

[0095] In this step, the applicable scenario constraint is that the gray prediction model is only enabled when the low-frequency parameters (such as turbidity and ammonia nitrogen) meet any of the following conditions: ① sampling interval ≥ 8 hours (single-day sampling ≤ 3 times); ② sample size < 30 groups (such as the system just starts or the recovery stage after the sensor fails); if the above conditions are not met, the LSTM time series prediction model (input features include the same period data of the previous 7 days and the change trend of adjacent parameters) is used to complete the trend. The theoretical basis of the gray prediction model (such as GM(1,1) model) is a "small sample, poor information, and uncertain system", and its prediction accuracy depends on the premise of "weak randomness and obvious trend":

[0096] When the sample size is ≥ 30 groups, the data has statistical significance, and the "information completion" advantage of gray prediction disappears. At this time, LSTM, ARIMA, and other models can capture complex time series rules through multi-feature learning, and the prediction error is 15%-30% lower than that of gray prediction (according to the comparison experiment data of multiple models in "Gray System Theory and Its Application").

[0097] When the sampling interval is < 8 hours (such as 1-4 hours / time), the single-day sampling amount is ≥ 6 times, and the data fluctuates frequently in the short term (such as the sudden rise of ammonia nitrogen after feeding and the sudden change of dissolved oxygen after oxygenation), the "accumulation generation" processing of the gray prediction will smooth the key fluctuation information, resulting in a trend prediction deviation of more than 10%; when the sampling interval is ≥ 8 hours, the data fluctuation period is more matched with the breeding physiological period (such as fish group metabolism and microbial decomposition), and the trend is stronger, and the gray prediction error can be controlled within 5%.

[0098] S303, the neural network prediction layer is used to classify and regress the extracted time sequence feature complete node feature set, fuse fish behavior features and water quality features for multi-modal joint prediction, output a multi-dimensional water quality state prediction matrix including potential anomaly probability, water quality risk level, water cycle efficiency and biological fitness, and the calculation formula of the generation process is as follows:

[0099]

[0100] wherein, is a multi-dimensional water quality state prediction matrix, is a processing result of the gated spatio-temporal convolutional graph neural network on the dynamic heterogeneous graph G, is a feature fusion operator, is a prediction result of the grey prediction model on the low-frequency parameter node subset in the heterogeneous graph.

[0101] It should be noted that the neighbor node aggregation radius (usually 3) is set according to the physical meaning of the parameter correlation, that is, only the 3 or fewer neighbor nodes having a direct causal relationship with the current node (such as the neighbors of the dissolved oxygen node being the ammonia nitrogen, water temperature, and aerator power nodes) are aggregated to avoid meaningless long-distance node features interference; the attention coefficient is calculated by a Sigmoid activation function, and the key water quality sensitive nodes (dissolved oxygen, ammonia nitrogen) are assigned a basic weight that is 3 times that of ordinary nodes (such as the attention coefficient of the dissolved oxygen node being initialized to 0.3, and the water temperature node being initialized to 0.1), and then dynamically adjusted in combination with the edge weight (time delay mutual information entropy value) between nodes - the higher the edge weight (such as the edge weight between ammonia nitrogen and dissolved oxygen being 0.8), the greater the feature contribution weight of the neighbor node to the current node; the neighbor features are aggregated in a "weighted sum + residual connection" manner to ensure the effectiveness of feature aggregation. This module breaks through the limitation of traditional graph convolution "treating all neighbor nodes equally", so that the model pays more attention to key parameters that have a significant impact on the system, and improves the relevance of feature extraction.

[0102] In step S302, a 1D time convolution kernel (size set to 5, step set to 2) is used, and the convolution kernel size corresponds to "5 time steps" (such as 1 hour per time step, and the convolution kernel covers 5 hours of parameter changes), which can effectively capture the fluctuation pattern in a short period; the step is set to 2, which can reduce the data dimension while retaining the key time sequence information, and improve the calculation efficiency; and for the parameter evolution pattern in a long period (such as 48 hours), a dilation convolution (dilation coefficient set to 2) is introduced, which can expand the time sequence perception range (such as when dilation = 2, a 5-size convolution kernel can cover 10 time step information) by inserting "holes" in the convolution kernel, without increasing the number of parameters to capture long-period trends; at the same time, a ReLU activation function is used to enhance the nonlinear fitting ability of the model, and to process the nonlinear changes of water quality parameters (such as the nonlinear degradation of ammonia nitrogen in the nitrification reaction); subsequently, the global average pooling is used to compress the time dimension features into a fixed length vector, which is aligned with the node features in the spatial dimension, to provide a uniform dimension input for the subsequent prediction layer. Therefore, this module realizes the dual capture of short-period fluctuations and long-period trends through the combination of "regular convolution + dilation convolution", solving the problem of weak long-period perception ability of traditional time sequence models.

[0103] In this step, low-frequency parameters (such as ammonia nitrogen and total phosphorus) are screened out by a sampling frequency threshold (e.g., sampling frequency < 0.1 Hz, i.e., less than 1 sample per 10 seconds), and only these parameters are completed to avoid redundant processing of high-frequency parameters (such as dissolved oxygen and pH, with a sampling frequency > 1 Hz); the historical time series data of low-frequency parameters (such as the past 7 days of ammonia nitrogen sampling data) are accumulated and processed to generate a differential equation for the GM(1,1) model, the model parameters are solved by the least squares method, and the future time steps (such as the ammonia nitrogen value every hour) are predicted to obtain high-frequency completed data; then the completed data predicted by GM(1,1) is spliced with the high-frequency parameter features output by GTCN, and the data dimension alignment and information fusion are realized through the feature fusion layer (using a multi-layer perception), to ensure the consistency of the completed data and the original data (such as the completed ammonia nitrogen value needs to comply with the time sequence rule of "postprandial peak"). This completion mechanism solves the pain point of "discontinuous time series of low-frequency parameters", provides complete time series input for GTCN, and ensures the accuracy of the multi-dimensional water quality state prediction matrix. The "space-time-completion" fusion features output by the above module are input into the fully connected prediction layer, and through softmax classification (predicting potential abnormal probability, water quality risk level) and linear regression (predicting water cycle efficiency, biological fitness), a multi-dimensional prediction matrix of 4 columns of features (rows correspond to future time steps, such as 24 hours in the future, 1 row per hour; columns correspond to 4 types of prediction indicators) is output. The final generated prediction matrix provides a "water quality-water saving-biology" multi-dimensional decision basis for the subsequent multi-objective constraint reinforcement learning algorithm, realizing the transformation from "data features" to "control targets".

[0104] In this step, the fish behavior data in "iteratively correcting the edge weight every hour based on the latest water quality data and fish behavior data" includes fish swimming frequency (unit: times / minute, monitoring range [0, 50]), feeding rate (unit: %, monitoring range [0, 100]), and cluster coefficient (value range [0, 1], the closer to 1, the more dense the cluster).

[0105] Iteration correction trigger threshold: when the change in any core water quality parameter (water temperature, pH, dissolved oxygen, etc.) within 1 hour exceeds ±5%, or the fish swimming frequency fluctuates by ±10 times / minute, or the feeding rate fluctuates by ±15%, the edge weight iteration correction is triggered; if the threshold is not reached, only the data is updated but the edge weight is not corrected, balancing the calculation efficiency and the timeliness of the correlation.

[0106] S400, based on the multi-dimensional water quality state prediction matrix, using multi-objective constraint reinforcement learning algorithm for processing: with the compliance water quality parameter threshold as the water quality safety hard constraint, with the preset daily water saving rate as the optimization target soft constraint; combined with the chaos disturbance improved particle swarm optimization method and fuzzy decision screening mechanism, the optimal parameter combination covering valve opening, pump speed, backwashing frequency and oxygenation timing is solved, and the accurate regulation instruction set is obtained.

[0107] It can be understood that the step S400 includes S401, S402 and S403.

[0108] S401, according to the multi-dimensional water quality state prediction matrix, the constraint conditions of multi-objective optimization are set; the compliance threshold of pH and dissolved oxygen parameters is set as the water quality safety hard constraint, the daily water saving rate target is adjusted based on the water quality risk level, and the daily water saving rate is set as the optimization target soft constraint, to obtain the multi-objective optimization constraint condition set for regulating parameter population initialization;

[0109] S402, using the chaos disturbance improved particle swarm optimization algorithm, initializing the regulation parameter population according to the multi-objective optimization constraint conditions, wherein the parameter population includes parameter combinations of valve opening, pump speed, backwashing frequency and oxygenation timing; setting parameter coordination constraint factor, determining the coordination of valve opening and pump speed, backwashing frequency and oxygenation timing; the fitness value of each population individual is calculated by iterative optimization, wherein the fitness value is determined by water quality compliance rate, water saving amount and equipment energy consumption, and then a plurality of candidate regulation parameters are selected, and a candidate regulation parameter set is obtained;

[0110] S403, introducing fuzzy decision mechanism to perform secondary screening on the candidate regulation parameter set, and determining the optimal parameter combination combined with the actual operation scene of the breeding system; taking the optimal parameter combination as the benchmark, the historical regulation effect under the same scene is taken as the weight to correct the candidate parameters, and finally the accurate regulation instruction set is obtained, wherein the calculation formula of the optimal parameter solution is as follows:

[0111]

[0112] In the formula, is the regulation instruction set, is the deviation loss function of water quality state and prediction matrix P under the action of regulation parameter C, and C is the regulation parameter, is the regulation parameter constraint interval.

[0113] It should be noted that the hard constraint is centered on "water quality parameter compliance", based on the "potential abnormal probability" and "water quality risk level" in the multi-dimensional water quality state prediction matrix, the safety threshold of key water quality parameters is converted into an algorithm-unsolvable constraint condition. The specific constraint index needs to be consistent with the characteristics of the breeding species and industry standards, for example: pH constraint: the set range is 6.8-8.5 (exceeding this range will cause gill damage and metabolic disorder of fish), the constraint form is "predicted pH value ∈ [6.8, 8.5]"; dissolved oxygen constraint: the lower limit is set to 5 mg / L (lower than this value will cause fish hypoxia stress), the constraint form is "predicted dissolved oxygen value ≥ 5 mg / L"; ammonia nitrogen constraint: the upper limit is set to 0.5 mg / L (higher than this value will produce toxicity), the constraint form is "predicted ammonia nitrogen value ≤ 0.5 mg / L". The hard constraint is embedded in the reward mechanism of reinforcement learning through the "penalty function" - if the predicted water quality corresponding to a group of control parameters breaks the hard constraint, it is directly assigned a negative infinite reward value, ensuring that the algorithm prioritizes avoiding water quality risks during the optimization process.

[0114] Among them, the soft constraint takes "preset daily water saving rate" as the core optimization target, combined with the "water cycle efficiency" in the multi-dimensional water quality state prediction matrix, the water saving demand is converted into a quantifiable optimization index. Unlike the "black and white" of hard constraint, soft constraint has dynamic adjustment: basic target setting: according to the water resource circulation capacity of the breeding system (such as membrane filtration efficiency, sedimentation tank capacity), the daily water saving rate is set to 80% (i.e. daily water replenishment ≤ 20% of total water body); dynamic adjustment logic: if the predicted water quality risk level is 1-2 (low risk), the daily water saving rate target is adjusted to 85%, prioritizing water saving; if the predicted risk level is 4-5 (high risk), the daily water saving rate target is adjusted to 75%, prioritizing water quality safety. The soft constraint is integrated into reinforcement learning through "reward weight" - the closer the actual water saving rate is to the target value, the higher the reward value, guiding the algorithm to optimize in the direction of optimal water saving under the premise of meeting the hard constraint.

[0115] For the optimization of control parameter combination, an improved particle swarm optimization algorithm is used to achieve efficient solution: particle coding and population initialization: valve opening, pump speed, backwashing frequency, oxygenation timing and other control parameters are coded as particle position vectors, and the population is initialized within the parameter constraint range; among them, in the particle iteration process, a disturbance factor is generated through chaotic mapping to fine-tune the position of particles trapped in local optimum and break out of local optimum trap; then design the fitness function by combining water quality compliance rate and water saving rate, dynamically adjust the weight of the two to balance the target, update the individual optimum and global optimum through iteration, and obtain the theoretical optimal parameter combination until the algorithm converges.

[0116] Specifically, the theoretical optimal parameters are secondarily optimized in combination with the actual breeding scene to generate instructions: from the equipment load adaptability, fish group stress avoidance, and energy consumption economy three dimensions, an index is established to convert qualitative demand into quantitative membership; then the membership of each index and the comprehensive evaluation value are calculated, and the final optimal parameter combination that meets the actual scene demand is screened; finally, the final optimal parameters are converted into a standardized signal format recognizable by the actuator to form a structured and precise control instruction set for subsequent actuator cluster control.

[0117] S500, distributed collaborative control processing of the actuator cluster for the precise control instruction set: converting the instructions by using the industrial standard communication protocol, realizing the linkage control of multiple types of actuators through the edge computing node; introducing a redundant actuator dynamic switching mechanism to drive the actuator action to complete the periodic state update of the recirculating water system and realize the precise aquaculture recirculating water control.

[0118] It can be understood that the present step S500 includes S501, S502 and S503.

[0119] S501, based on the precise control instruction set, the instructions are converted into digital signals recognizable by the actuators by using the OPC-UA industrial standard communication protocol; the instructions related to water quality safety are marked with priority, and the digital signals are transmitted to the edge computing node through the industrial Ethernet, wherein the edge computing node verifies the signals to obtain the verified actuator control signals;

[0120] S502, the verified actuator control signals are called to transmit control instructions to electric valves, variable frequency water pumps and oxygenators, wherein the control instructions include control of electric valve opening adjustment, variable frequency water pump speed adjustment and oxygenator working time sequence switching; the action amplitude is adjusted according to the real-time load of each actuator, the working state data of the actuator is collected in real time, the device running state is monitored, and the actuator working state data and linkage control results are obtained;

[0121] S503, a 1+1 redundant actuator dynamic switching mechanism is introduced, when the main actuator fails, the edge computing node triggers the standby actuator to start based on the actuator working state data; the control parameters are adjusted according to the performance difference of the standby actuator, the actuator is driven to act according to the verified actuator control signals, the periodic state update of the recirculating water system is completed, and the precise collaborative control of water quality and water saving is realized.

[0122] It should be noted that the OPCUA industrial Ethernet protocol (supporting multi-vendor device interoperability) is adopted, the logical parameters such as "valve opening percentage" and "pump speed set value" in the instruction set are coded into binary control signals that can be parsed by actuators (such as 4-20mA analog signals for electric valves and Modbus-RTU protocol register write instructions for frequency conversion water pumps), and "instruction frame check code" (such as CRC16 check) is embedded in the conversion process to avoid byte loss or tampering during signal transmission. Based on the water quality safety level, the instruction priority is divided (referring to the water quality risk level predicted by S300), and a "static priority + dynamic adjustment" mechanism is adopted - the instructions related to fish survival such as oxygenation timing adjustment and emergency water valve control are set as the highest priority (priority 1), and non-urgent instructions such as backwash frequency adjustment are set as low priority (priority 3). When high-priority instructions and low-priority instructions are issued at the same time, the "instruction preemption mechanism" of the edge computing node is triggered, the execution of low-priority instructions is suspended, high-priority instructions are responded to preferentially, and the safety bottom line of the system is guaranteed.

[0123] In this embodiment, after the edge computing node receives the protocol-converted signals, it first performs "parameter range verification" (such as determining whether the valve opening is within the 0-100% effective range and whether the pump speed exceeds the rated range of the device), and eliminates invalid instructions. Then, through the "device address mapping table" (pre-stored correspondence between actuator IP address and control port), the instructions are distributed to the target actuators to avoid misoperation caused by address mismatch. For actuators with coupling relationship (such as "water inlet valve - frequency conversion water pump - oxygenator"), the "timing coordination module" of the edge computing node sets the action delay threshold - for example, after the water inlet valve opening is increased by 30%, the frequency conversion water pump is started after a delay of 20 seconds to avoid excessive water flow impact on the sensor data. At the same time, the "action feedback signals" of each actuator (such as valve opening feedback value, water pump current feedback) are collected, and the "instruction set value" and "actual feedback value" are compared in real time. If the deviation exceeds the allowed range (such as the actual valve opening deviation from the set value > 5%), the "secondary correction instruction" is automatically triggered to ensure the control accuracy. Then, the "device load monitoring algorithm" is introduced, the running parameters of the actuators (such as water pump working current, valve motor temperature) are collected in real time, and the current load rate of the device (actual load / rated load) is calculated. When the load rate of a certain actuator exceeds 80% (high load threshold) for 10 consecutive minutes, the edge computing node automatically adjusts the action of the associated actuators (such as distributing part of the water circulation task to the standby water pump in the same area) to avoid the life attenuation caused by long-term overload of a single device.

[0124] In this step S503, the same type of backup actuator is configured for the key actuator (such as the main oxygenation machine and the core water circulation pump). The edge computing node monitors the running state of the main actuator through "heartbeat signal detection" (a state query instruction is sent every 5 seconds); when the main actuator does not feedback the heartbeat signal for 3 times in succession, or feedbacks the "fault code" (such as motor overcurrent fault, valve jamming fault), the "non-disturbance switching logic" is triggered - first preheat the standby actuator to standby state, then cut off the power supply of the main actuator, at the same time, the control instruction is seamlessly switched to the standby actuator, the switching process time is <1 second, to avoid significant fluctuations in water quality parameters; finally, according to the preset regulation and control period (such as once every hour), after the actuator completes a round of instruction action, the edge computing node automatically collects "water quality data after regulation and control" (interfaces with S100 sensor array), "actuator running data", generates "regulation and control effect log" (including instruction content, execution result, water quality change trend), and uploads to the cloud database; at the same time, based on the log data, update the "actuator performance attenuation model" (such as record the change trend of valve action response time with the use time), provide data support for subsequent actuator maintenance or parameter correction, finally realize the closed-loop management of "regulation and control-feedback-optimization".

[0125] Embodiment 2:

[0126] As shown in Figure 2 The present embodiment provides a precise regulation and control system for circular water aquaculture based on graph network water resource management, please refer to Figure 2 The system comprises:

[0127] The acquisition module 701 is used for deploying a multi-source sensor array covering the water inlet of the breeding pond, the biochemical reactor and the water outlet, acquiring multi-dimensional data of the fishery circular water system through the multi-source sensor array, sequentially performing high-frequency synchronous acquisition processing and Kalman filter noise fusion suppression processing on the acquired multi-dimensional data, and combining the historical optimal breeding period growth coefficient to perform Z-score standardization processing, to obtain a spatiotemporally aligned and standardized multi-modal water quality time series data set;

[0128] The construction module 702 is used for constructing a dynamic heterogeneous graph based on the multi-modal water quality time series data set: defining the core water quality parameter types as nodes of the dynamic heterogeneous graph, wherein the node attributes are set as a multi-dimensional feature vector composed of parameter monitoring values, parameter change rates, water-borne organism metabolism correlation coefficients and device energy consumption matching degrees; defining the causal correlation strength within the time delay window between parameters as the edges of the dynamic heterogeneous graph, calculating the edge weights through time delay mutual information entropy, constructing a sparse adjacency matrix by L1 regularization, and obtaining the dynamic heterogeneous graph structure with edge weights containing time delay mutual information entropy and nodes containing multi-dimensional features;

[0129] The processing module 703 is configured to process the dynamic heterogeneous graph structure by using the gated spatio-temporal convolution graph neural network, including aggregating neighbor node information in a specified range by an attention gate module, assigning a high attention weight to a key water quality sensitive node, capturing a long-period dynamic evolution mode of parameters by a time convolution module, and fusing a gray prediction model to complete a variation trend of a low-frequency parameter to obtain a multi-dimensional water quality state prediction matrix including a potential anomaly probability, a water quality risk level, a water cycle efficiency, and a biological fitness.

[0130] The solving module 704 is configured to process the multi-dimensional water quality state prediction matrix by using a multi-objective constraint reinforcement learning algorithm, taking a compliant water quality parameter threshold as a water quality safety hard constraint, taking a preset daily water saving rate as an optimization target soft constraint, combining a chaotic disturbance improved particle swarm optimization method and a fuzzy decision screening mechanism, solving an optimal parameter combination including a valve opening degree, a pump speed, a backwashing frequency, and an oxygenation timing, and obtaining a precise regulation and control instruction set.

[0131] The driving module 705 is configured to perform distributed collaborative control processing of the precise regulation and control instruction set on an executor cluster, convert instructions by using an industrial standard communication protocol, realize multi-type executor linkage control through an edge computing node, introduce a redundant executor dynamic switching mechanism, drive executor actions to complete periodic state updates of the recirculating aquaculture system, and realize precise regulation and control of the recirculating aquaculture system.

[0132] Specifically, the collection module 701 includes the following.

[0133] The constituent unit is configured to deploy an optical dissolved oxygen sensor and a water temperature sensor at an inlet of a culture pond, deploy a pH glass electrode and an ORP sensor inside a biochemical reactor, deploy a turbidity-ammonia nitrogen integrated sensor at an outlet, and constitute a multi-source sensor array covering key monitoring points; collect multi-dimensional original data of the recirculating aquaculture system through the multi-source sensor array, including water quality parameter data, equipment operation data, and fish population monitoring data, and obtain a multi-dimensional original data set.

[0134] The first processing unit is configured to synchronously collect the multi-dimensional original data, use Kalman filtering to perform noise fusion and suppression on the water quality parameter data, introduce a dynamic threshold screening based on a fish population metabolism model, and combine a water quality tolerance range of different growth stages of the fish population to perform outlier rejection on the equipment operation data and the fish population monitoring data, and obtain a preprocessed data set.

[0135] The second processing unit is configured to combine a historical optimal growth coefficient of a complete culture period to implement Z-score standardization on the preprocessed data set, and a calculation formula is as follows.

[0136]

[0137] In the formula, is the standardized multi-modal water quality time series data, is the original water quality parameter collection value, is the mean value of the corresponding parameter in the historical period, is the historical optimal growth coefficient, is the standard deviation of the corresponding parameter in the historical period;

[0138] The obtaining unit is used to increase the parameter correlation degree information of the standardized data by calculating the Pearson correlation coefficient of any two parameters, and finally obtain a multi-modal water quality time series data set with time-space alignment and parameter correlation degree information, wherein the multi-modal water quality time series data set forms a structured matrix with time steps as rows and monitoring parameters as columns.

[0139] Specifically, the construction module 702 comprises:

[0140] The first construction unit is used to define water temperature, pH, dissolved oxygen, ammonia nitrogen and turbidity as core water quality parameter nodes of a dynamic heterogeneous graph based on the multi-modal water quality time series data set, set importance weights adjusted according to the growth stage of the fish population for each node, and construct a multi-dimensional attribute vector for each node, including the parameter real-time monitoring value, the 24-hour sliding window change rate, the water organism metabolism correlation coefficient and the equipment energy consumption matching degree, to obtain a node set with dynamic importance weights and a node attribute vector.

[0141] The second construction unit is used to define the causal correlation strength in the parameter time delay window as the edge of the graph, and the time delay window is set according to the response period of the water quality parameter on the fish population growth; the correlation strength between any two nodes is calculated by time delay mutual information entropy to obtain an initial edge weight; the edge weight is iteratively corrected according to the latest water quality data and fish population behavior data every hour, the initial edge weight matrix is sparsified by L1 regularization, and weakly correlated edges are removed to construct a sparse adjacency matrix, and the weight calculation formula is as follows:

[0142]

[0143] In the formula, is the edge weight between node i and node j, is the time delay mutual information entropy of the parameter and at time t, is the time delay mutual information entropy of the parameter at time t, is the L1 regularization processing of the weight matrix W;

[0144] The integration unit is used to integrate the node set, the edge set and the sparse adjacency matrix to form a dynamic heterogeneous graph, and mark the edges with changing correlation strength to obtain a dynamic heterogeneous graph that can dynamically reflect the change of the correlation relationship between water quality parameters with time.

[0145] Specifically, the processing module 703 comprises:

[0146] The fusion unit is configured to process based on the dynamic heterogeneous graph by using a gated spatio-temporal convolution graph neural network, aggregate neighbor node features within a specified range around each node by using an attention gating module, assign attention weights to key water quality sensitive nodes that affect fish metabolism of dissolved oxygen and ammonia nitrogen, and interactively fuse node features extracted by different layers of the graph neural network to obtain a node feature set with fused cross-layer information.

[0147] The recognition unit is configured to perform time series dimension processing on the node feature set by using a time convolution module, capture dynamic evolution patterns of water quality parameters in a long period, and identify periodic rules of parameter changes.

[0148] The prediction unit is configured to perform classification and regression on the extracted node feature set with time series characteristics by using a neural network prediction layer, perform multi-modal joint prediction by fusing fish behavior features and water quality features, and output a multi-dimensional water quality state prediction matrix including potential anomaly probability, water quality risk level, water cycle efficiency, and biological fitness.

[0149]

[0150] In the formula, the multi-dimensional water quality state prediction matrix is is a processing result of the gated spatio-temporal convolution graph neural network on the dynamic heterogeneous graph G, is a feature fusion operator, is a prediction result of the gray prediction model on a low-frequency parameter node subset in the heterogeneous graph.

[0151] Specifically, the solving module 704 comprises:

[0152] The setting optimization unit is configured to set constraint conditions of multi-objective optimization according to the multi-dimensional water quality state prediction matrix, set compliance thresholds of pH and dissolved oxygen parameters as water quality safety hard constraints, adjust a daily water saving rate target based on a water quality risk level, set the daily water saving rate as a soft constraint of an optimization target, and obtain a multi-objective optimization constraint condition set for initializing the regulation parameter population.

[0153] ​The setting determination unit is configured to initialize a population of control parameters according to multi-objective optimization constraints by using a chaos disturbance improved particle swarm optimization algorithm, wherein the population of parameters includes a parameter combination of valve opening, pump speed, backwashing frequency and oxygenation timing; a parameter coordination constraint factor is set to determine the coordination of valve opening and pump speed, backwashing frequency and oxygenation timing; the fitness value of each population individual is calculated by iterative optimization, wherein the fitness value is determined by the water quality compliance rate, water saving amount and equipment energy consumption, and then a plurality of candidate control parameter groups are screened out, and a candidate control parameter set is obtained;

[0154] The screening and correction unit is configured to introduce a fuzzy decision mechanism to perform secondary screening on the candidate control parameter set, and determine an optimal parameter combination according to the actual operation scene of the aquaculture system; the optimal parameter combination is taken as a reference, and the historical control effect under the same scene is taken as a weight to correct the candidate parameters, and finally a precise control instruction set is obtained, wherein the calculation formula of the optimal parameter solution is as follows:

[0155]

[0156] In the formula, is the control instruction set, is the deviation loss function of the water quality state and the prediction matrix P under the action of the control parameter C, and C is the control parameter, is the control parameter constraint interval.

[0157] Specifically, the driving module 705 includes:

[0158] The transmission unit is configured to convert the instructions into digital signals recognizable by the actuators based on the precise control instruction set by using the OPC-UA industrial standard communication protocol; mark the instructions related to water quality safety with priority, and transmit the digital signals to the edge computing node through the industrial Ethernet, wherein the edge computing node verifies the signals to obtain verified actuator control signals;

[0159] The calling unit is configured to call the verified actuator control signals, and transmit control instructions to the electric valve, variable frequency water pump and oxygenator, wherein the control instructions include control of electric valve opening adjustment, variable frequency water pump speed adjustment and oxygenator working timing switching; adjust the action amplitude according to the real-time load of each actuator, and collect the working state data of the actuator in real time to monitor the equipment operation state, and obtain the actuator working state data and the linkage control result;

[0160] The adjusting unit is used for introducing a 1+1 redundant executor dynamic switching mechanism. When a main executor fails, an edge computing node triggers a standby executor to start based on executor working state data. According to performance differences of the standby executor, adjusting parameters are adjusted to drive the executor to act according to the checked executor control signal, complete periodic state updating of the circulating water system, and realize precise collaborative regulation and control of water quality and water saving.

[0161] It should be noted that, as for the system in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment related to the method, and will not be described in detail here.

[0162] Embodiment 3:

[0163] Corresponding to the above method embodiment, the present embodiment also provides a circulating water aquaculture precise regulation and control device based on graph network water resource management. The circulating water aquaculture precise regulation and control device based on graph network water resource management described below can be correspondingly referred to the circulating water aquaculture precise regulation and control method based on graph network water resource management described above.

[0164] Figure 3 is a block diagram of a circulating water aquaculture precise regulation and control device 800 based on graph network water resource management according to an exemplary embodiment. As Figure 3 shown, the circulating water aquaculture precise regulation and control device 800 based on graph network water resource management includes a processor 801 and a memory 802. The circulating water aquaculture precise regulation and control device 800 based on graph network water resource management also includes one or more of a multimedia component 803, an I / O interface 804, and a communication component 805.

[0165] The processor 801 is configured to control overall operation of the equipment 800 for graph network-based water resource management of recirculating aquaculture precise regulation, to complete all or part of the steps in the above-mentioned method for graph network-based water resource management of recirculating aquaculture precise regulation. The memory 802 is configured to store various types of data to support the operation of the equipment 800 for graph network-based water resource management of recirculating aquaculture precise regulation, which can include, for example, instructions for any application or method operating on the equipment 800 for graph network-based water resource management of recirculating aquaculture precise regulation, and application-related data, such as contact data, sent and received messages, pictures, audio, video, and the like. The memory 802 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk. The multimedia component 803 can include a screen and an audio component. The screen can be a touch screen, for example, and the audio component is configured to output and / or input audio signals. For example, the audio component can include a microphone configured to receive external audio signals. The received audio signals can be further stored in the memory 802 or transmitted through the communication component 805. The audio component also includes at least one speaker for outputting audio signals. The I / O interface 804 provides an interface between the processor 801 and other interface modules, which can be a keyboard, mouse, or button, etc. These buttons can be virtual buttons or physical buttons. The communication component 805 is configured to enable wired or wireless communication between the equipment 800 for graph network-based water resource management of recirculating aquaculture precise regulation and other devices. Wireless communication, such as Wi-Fi, Bluetooth, near field communication (NFC), 2G, 3G or 4G, or a combination of one or more of them, so the corresponding communication component 805 can include a Wi-Fi module, a Bluetooth module or an NFC module.

[0166] In an example embodiment, the recirculating aquaculture precision regulation device 800 based on graph network water resource management can be implemented by one or more Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), controller, microcontroller, microprocessor or other electronic elements for executing the above-mentioned recirculating aquaculture precision regulation method based on graph network water resource management.

[0167] In another example embodiment, a computer readable storage medium including program instructions is also provided, which, when executed by a processor, implements the steps of the above-mentioned recirculating aquaculture precision regulation method based on graph network water resource management. For example, the computer readable storage medium can be the above-mentioned memory 802 including program instructions, and the above-mentioned program instructions can be executed by the processor 801 of the recirculating aquaculture precision regulation device 800 based on graph network water resource management to complete the above-mentioned recirculating aquaculture precision regulation method based on graph network water resource management.

[0168] Embodiment 4:

[0169] Corresponding to the above method embodiments, the present embodiment also provides a readable storage medium, and the readable storage medium described below can be referred to in conjunction with the above-mentioned recirculating aquaculture precision regulation method based on graph network water resource management.

[0170] The computer program stored on the readable storage medium, when executed by the processor, implements the steps of the above-mentioned recirculating aquaculture precision regulation method based on graph network water resource management of the method embodiments.

[0171] The readable storage medium can be specifically a U disk, a mobile hard disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk or an optical disk, and various readable storage media that can store program codes.

[0172] In summary, this invention systematically solves the core problems of existing land-based factory aquaculture recirculating water regulation, namely "data fragmentation, static modeling, delayed prediction, extensive regulation, and unstable execution," through the synergistic effect of multi-source data acquisition and standardization, dynamic heterogeneous graph construction, spatiotemporal prediction, multi-objective optimization, and actuator cluster collaborative control. This provides key technical support for the industry's water conservation, emission reduction, and sustainable development.

[0173] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0174] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. 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 be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for precise regulation of recirculating aquaculture based on graph network water resource management, characterized in that, The method comprises the following steps: Deploy a multi-source sensor array covering the inlet, biochemical reactor and outlet of the aquaculture pond, collect multi-dimensional data of the fishery circulating water system through the multi-source sensor array, sequentially perform high-frequency synchronous collection processing and Kalman filter noise fusion suppression processing on the collected multi-dimensional data, and combine the historical optimal growth coefficient of the breeding period for Z-score standardization processing to obtain a spatiotemporally aligned and standardized multi-modal water quality time series dataset; A dynamic heterogeneous graph is constructed based on the multi-modal water quality time series dataset: the core water quality parameter types are defined as nodes of the dynamic heterogeneous graph, wherein the node attributes are set as a multi-dimensional feature vector composed of parameter monitoring values, parameter change rates, water organism metabolism correlation coefficients and equipment energy consumption matching degrees; the causal correlation strength within the time delay window between parameters is defined as the edge of the dynamic heterogeneous graph, the edge weight is calculated by time delay mutual information entropy, and a sparse adjacency matrix is constructed by L1 regularization to obtain a dynamic heterogeneous graph structure with node multi-dimensional features and edge time delay mutual information entropy; A gated spatio-temporal convolutional graph neural network is used to process the dynamic heterogeneous graph structure, which includes aggregating neighbor node information within a specified range through an attention gate module to assign high attention weights to key water quality sensitive nodes; capturing long-period dynamic evolution patterns of parameters through a time convolution module; and fusing a gray prediction model to complete the trend of low-frequency parameters to obtain a multi-dimensional water quality state prediction matrix including potential anomaly probability, water quality risk level, water cycle efficiency and biological fitness; Based on the multi-dimensional water quality state prediction matrix, a multi-objective constraint reinforcement learning algorithm is used for processing: the compliance water quality parameter threshold is used as a water quality safety hard constraint, and a preset daily water saving rate is used as an optimization target soft constraint; a chaotic disturbance improved particle swarm optimization method and a fuzzy decision screening mechanism are combined to solve the optimal parameter combination covering valve opening degree, pump speed, backwashing frequency and oxygenation timing to obtain a precise control instruction set; The precise control instruction set is subjected to distributed collaborative control processing of an actuator cluster: the industrial standard communication protocol conversion instruction is used to realize multi-type actuator linkage control through an edge computing node; a redundant actuator dynamic switching mechanism is introduced to drive the actuator to act to complete the periodic state update of the circulating water system and realize precise fishery circulating water regulation and control.

2. The method according to claim 1, wherein, The multi-source sensor array covering the inlet, biochemical reactor and outlet of the aquaculture pond collects multi-dimensional data of the fishery circulating water system, sequentially performs high-frequency synchronous collection processing and Kalman filter noise fusion suppression processing on the collected multi-dimensional data, and combines the historical optimal growth coefficient of the breeding period for Z-score standardization processing to obtain a spatiotemporally aligned and standardized multi-modal water quality time series dataset, which includes: An optical dissolved oxygen sensor and a water temperature sensor are arranged at an inlet of a culture pond, a pH glass electrode and an ORP sensor are arranged inside a biochemical reactor, and a turbidity-ammonia nitrogen integrated sensor is arranged at an outlet, to form a multi-source sensor array covering key monitoring points; multi-dimensional original data of the fishery recirculating water system, including water quality parameter data, equipment operation data and fish population monitoring data, are collected through the multi-source sensor array to obtain a multi-dimensional original data set; The multi-dimensional original data is synchronously collected, the water quality parameter data is subjected to noise fusion suppression by Kalman filtering, and a dynamic threshold screening based on a fish population metabolism model is introduced; in combination with the water quality tolerance range of different growth stages of the fish population, the equipment operation data and the fish population monitoring data are subjected to outlier rejection to obtain a pre-processed data set; The pre-processed data set is subjected to Z-score standardization in combination with the historical optimal growth coefficient of the entire culture cycle, and the calculation formula is as follows: wherein, is the standardized multi-modal water quality time series data, is the raw collected value of the water quality parameter, is the mean value of the corresponding parameter in the historical period, is the historical optimal growth coefficient, is the standard deviation of the corresponding parameter in the historical period; The Pearson correlation coefficient of any two parameters is calculated to add parameter correlation information to the standardized data, and finally a multi-modal water quality time series data set that is spatio-temporally aligned and contains parameter correlation information is obtained, wherein the multi-modal water quality time series data set forms a structured matrix with time steps as rows and monitoring parameters as columns.

3. The method according to claim 1, wherein, The multi-modal water quality time series data set is used to construct a dynamic heterogeneous graph: the core water quality parameter types are defined as nodes of the dynamic heterogeneous graph, wherein the node attributes are set as a multi-dimensional feature vector composed of parameter monitoring values, parameter change rates, aquatic organism metabolism correlation coefficients and equipment energy consumption matching degrees; the causal correlation strength within the time delay window between parameters is defined as the edge of the dynamic heterogeneous graph, the edge weight is calculated by time delay mutual information entropy, and a sparse adjacency matrix is constructed by L1 regularization to obtain a dynamic heterogeneous graph structure with node multi-dimensional feature and edge time delay mutual information entropy, which includes: Based on the multi-modal water quality time series data set, water temperature, pH, dissolved oxygen, ammonia nitrogen and turbidity are defined as core water quality parameter nodes of the dynamic heterogeneous graph; an importance weight adjusted according to the growth stage of the fish population is set for each node, and a multi-dimensional attribute vector is constructed for each node, including real-time monitoring values, 24-hour sliding window change rates, aquatic organism metabolism correlation coefficients and equipment energy consumption matching degrees, to obtain a node set with dynamic importance weight and node attribute vector; The causal correlation strength within the time delay window between parameters is defined as the edge of the graph, and the time delay window is set according to the response period of the water quality parameters on the growth of the fish population; the correlation strength between any two nodes is calculated by time delay mutual information entropy to obtain an initial edge weight; the edge weight is iteratively corrected every hour according to the latest water quality data and fish behavior data, the initial edge weight matrix is sparsified by L1 regularization, weakly correlated edges are removed, and a sparse adjacency matrix is constructed, and the weight calculation formula is as follows: wherein, is the edge weight between node i and node j, is the parameter at time t and is the parameter at time t is the delay mutual information entropy, is the delay window, is the L1 regularization on the weight matrix W; The node set, edge set and sparse adjacency matrix are integrated to form a dynamic heterogeneous graph, and edges with changing correlation strength are marked to obtain a dynamic heterogeneous graph that can dynamically reflect the change of the correlation between water quality parameters over time.

4. The precise regulation method of recirculating aquaculture based on graph network water resources management according to claim 1, characterized in that, The dynamic heterogeneous graph is processed by using the gated spatio-temporal convolution graph neural network, including assigning high attention weight to key water quality sensitive nodes by an attention gate module to aggregate neighbor node information in a specified range; capturing long-period dynamic evolution patterns of parameters by a time convolution module; and obtaining a multi-dimensional water quality state prediction matrix including potential anomaly probability, water quality risk level, water cycle efficiency and biological fitness by fusing a gray prediction model to complete the trend of low-frequency parameters. Based on the dynamic heterogeneous graph, the gated spatio-temporal convolution graph neural network is used for processing; the attention gate module is used to aggregate the neighbor node features in a specified range around each node, and the attention weight is assigned to the key water quality sensitive nodes that affect the fish metabolism of dissolved oxygen and ammonia nitrogen, the node features extracted by different layers of the graph neural network are interactively fused to obtain a node feature set that fuses cross-layer information; The time convolution module is used for time series dimension processing on the node feature set, to capture the dynamic evolution patterns of water quality parameters in a long period, and to identify the periodicity of parameter changes; the water quality parameter change period in different seasons and different breeding stages is automatically identified, and the gray prediction model is fused to complete the trend of the water quality parameters with low sampling frequency, to obtain a node feature set with complete time series characteristics; The neural network prediction layer is used for classification and regression on the extracted node feature set with complete time series characteristics, the fish behavior features and water quality features are fused for multi-modal joint prediction, and a multi-dimensional water quality state prediction matrix including potential anomaly probability, water quality risk level, water cycle efficiency and biological fitness is output, and the calculation formula of the generation process is as follows: wherein, is a multi-dimensional water quality state prediction matrix, is the processing result of the gated spatio-temporal convolutional graph neural network on the dynamic heterogeneous graph G, is a feature fusion operator, is the prediction result of the grey prediction model on the low-frequency parameter node subset in the heterogeneous graph.

5. The precise regulation method of recirculating aquaculture based on graph network water resources management according to claim 1, characterized in that, Based on the multi-dimensional water quality state prediction matrix, a multi-objective constraint reinforcement learning algorithm is used for processing: the compliance water quality parameter threshold is used as a water quality safety hard constraint, and a preset daily water saving rate is used as an optimization target soft constraint; a chaotic disturbance improved particle swarm optimization method and a fuzzy decision screening mechanism are combined to solve the optimal parameter combination covering valve opening, pump speed, backwashing frequency and oxygenation timing, to obtain a precise control instruction set, including: According to the multi-dimensional water quality state prediction matrix, the constraint conditions of multi-objective optimization are set; the compliance thresholds of pH and dissolved oxygen parameters are set as water quality safety hard constraints, the daily water saving rate target is adjusted based on the water quality risk level, and the daily water saving rate is set as an optimization target soft constraint, to obtain a multi-objective optimization constraint condition set for initializing the control parameter population; A chaotic disturbance improved particle swarm optimization algorithm is used to initialize the control parameter population according to the multi-objective optimization constraint conditions, wherein the parameter population includes parameter combinations of valve opening, pump speed, backwashing frequency and oxygenation timing; a parameter coordination constraint factor is set to determine the coordination of valve opening and pump speed, and backwashing frequency and oxygenation timing; the fitness value of each population individual is calculated by iterative optimization, wherein the fitness value is determined by the water quality compliance rate, water saving amount and equipment energy consumption, and then a plurality of candidate control parameters are selected, and a candidate control parameter set is obtained. A fuzzy decision mechanism is introduced to perform secondary screening on the candidate regulation parameter set, and the optimal parameter combination is determined in combination with the actual operation scene of the breeding system; taking the optimal parameter combination as the benchmark, the historical regulation effect under the same scene is used as the weight to modify the candidate parameters, and finally a precise regulation instruction set is obtained, and the calculation formula of the optimal parameter solving is as follows: wherein, is a set of control instructions, is a deviation loss function of water quality state and prediction matrix P under the action of control parameter C, C is a control parameter, is a control parameter constraint interval.

6. The precise regulation method of recirculating aquaculture based on graph network water resources management according to claim 1, characterized in that, The distributed collaborative control processing of the executor cluster for the precise regulation instruction set: adopting an industrial standard communication protocol conversion instruction, realizing multi-type executor linkage control through an edge computing node; introducing a redundant executor dynamic switching mechanism, driving the executor action to complete the periodic state update of the recirculating water system, and realizing precise aquaculture recirculating water regulation, which includes: Based on the precise regulation instruction set, the instruction is converted into a digital signal recognizable by the executor by adopting the OPC-UA industrial standard communication protocol; the instructions related to water quality safety are marked with priority, and the digital signal is transmitted to the edge computing node through industrial Ethernet, wherein the edge computing node verifies the signal to obtain the verified executor control signal; The verified executor control signal is called to transmit control instructions to electric valves, variable frequency water pumps and oxygenators, wherein the control instructions include control of electric valve opening adjustment, variable frequency water pump speed adjustment and oxygenator working time sequence switching; according to the real-time load adjustment amplitude of each executor, the working state data of the executor is collected in real time, the equipment running state is monitored, and the executor working state data and linkage control result are obtained; Introducing a 1+1 redundant executor dynamic switching mechanism, when the main executor fails, the edge computing node triggers the standby executor based on the executor working state data; according to the performance difference of the standby executor, the regulation parameter is adjusted, the executor is driven to act according to the verified executor control signal, the periodic state update of the recirculating water system is completed, and the precise collaborative regulation of water quality and water saving is realized.

7. The precise regulation system for recirculating aquaculture based on graph network water resources management according to claim 1, characterized in that, It includes: The acquisition module is used to deploy a multi-source sensor array covering the water inlet of the breeding pond, the biochemical reactor and the water outlet, collect multi-dimensional data of the aquaculture recirculating water system through the multi-source sensor array, perform high-frequency synchronous acquisition processing and Kalman filter noise fusion suppression processing on the collected multi-dimensional data in turn, and perform Z-score standardization processing combined with the historical optimal breeding cycle growth coefficient to obtain a spatiotemporally aligned and standardized multi-modal water quality time series data set; The construction module is used to construct a dynamic heterogeneous graph based on the multi-modal water quality time series data set: the core water quality parameter type is defined as the node of the dynamic heterogeneous graph, wherein the node attribute is set as a multi-dimensional feature vector composed of parameter monitoring value, parameter change rate, water organism metabolism correlation coefficient and equipment energy consumption matching degree; the causal correlation strength within the time delay window between parameters is defined as the edge of the dynamic heterogeneous graph, the edge weight is calculated by time delay mutual information entropy, a sparse adjacency matrix is constructed by L1 regularization, and the dynamic heterogeneous graph structure with edge weight containing time delay mutual information entropy and node containing multi-dimensional features is obtained; The processing module is used for processing the dynamic heterogeneous graph structure by using the gated spatio-temporal convolutional graph neural network, wherein the neighbor node information in a specified range is aggregated by an attention gate module, and high attention weights are assigned to key water quality sensitive nodes; a long-period dynamic evolution mode of parameters is captured by a time convolution module; and a gray prediction model is fused to complete the variation trend of low-frequency parameters, so as to obtain a multi-dimensional water quality state prediction matrix including potential anomaly probability, water quality risk level, water cycle efficiency and biological fitness; The solving module is used for processing based on the multi-dimensional water quality state prediction matrix by using a multi-objective constraint reinforcement learning algorithm: taking the compliant water quality parameter threshold as a water quality safety hard constraint, and taking a preset daily water saving rate as an optimization target soft constraint; combining a chaotic disturbance improved particle swarm optimization method and a fuzzy decision screening mechanism, solving an optimal parameter combination covering valve opening, pump speed, backwashing frequency and oxygenation timing, and obtaining a precise regulation and control instruction set; The driving module is used for performing distributed collaborative control processing of the precise regulation and control instruction set on the executor cluster: converting instructions by using an industrial standard communication protocol, realizing multi-type executor linkage control through an edge computing node; introducing a redundant executor dynamic switching mechanism, driving the executor action to complete the periodic state update of the circulating water system, and realizing the precise fishery circulating water regulation and control.

8. The precise regulation system for recirculating aquaculture based on graph network water resources management according to claim 7, characterized in that, The collection module, wherein the collection module comprises: The constituent unit is used for deploying an optical dissolved oxygen sensor and a water temperature sensor at an inlet of a culture pond, deploying a pH glass electrode and an ORP sensor inside a biochemical reactor, and deploying a turbidity-ammonia nitrogen integrated sensor at an outlet, so as to constitute a multi-source sensor array covering key monitoring points; multi-dimensional original data of the fishery circulating water system are collected through the multi-source sensor array, including water quality parameter data, equipment operation data and fish group monitoring data, so as to obtain a multi-dimensional original data set; The first processing unit is used for synchronously collecting the multi-dimensional original data, using Kalman filtering to perform noise fusion suppression on the water quality parameter data, and introducing a dynamic threshold screening based on a fish group metabolism model, and combining the water quality tolerance range of different growth stages of the fish group, to perform outlier rejection on the equipment operation data and the fish group monitoring data, so as to obtain a preprocessed data set; The second processing unit is used for implementing Z-score standardization on the preprocessed data set in combination with a historical optimal growth coefficient of a complete culture cycle, and the calculation formula is as follows: wherein, is the standardized multi-modal water quality time series data, is the raw collected value of the water quality parameter, is the mean value of the corresponding parameter in the historical period, is the historical optimal growth coefficient, is the standard deviation of the corresponding parameter in the historical period; The obtaining unit is used for adding parameter correlation information to the standardized data by calculating the Pearson correlation coefficient of any two parameters, and finally obtaining a multi-modal water quality time series data set with spatio-temporal alignment and parameter correlation information, wherein the multi-modal water quality time series data set forms a structured matrix with time steps as rows and monitoring parameters as columns.

9. The precise regulation system for recirculating aquaculture based on graph network water resources management according to claim 7, characterized in that, The construction module, wherein the construction module comprises: The first construction unit is configured to define water temperature, pH, dissolved oxygen, ammonia nitrogen and turbidity as core water quality parameter nodes of a dynamic heterogeneous graph based on a multi-modal water quality time series dataset; set an importance weight of each node adjusted according to a growth stage of a fish school; and construct a multi-dimensional attribute vector for each node, including a parameter real-time monitoring value, a 24-hour sliding window change rate, a water organism metabolism correlation coefficient and a device energy consumption matching degree, to obtain a node set with a dynamic importance weight and a node attribute vector; The second construction unit is configured to define a causal correlation strength in a time delay window between parameters as an edge of the graph, and set the time delay window according to a response period of the water quality parameters on the growth of the fish school; calculate the correlation strength between any two nodes through time delay mutual information entropy to obtain an initial edge weight; iteratively correct the edge weight according to the latest water quality data and fish school behavior data every hour; perform sparse processing on the initial edge weight matrix through L1 regularization, remove weak correlation edges, and construct a sparse adjacency matrix, the weight calculation formula of which is as follows: wherein, is the edge weight between node i and node j, is the parameter at time t and is the parameter at time t is the delay mutual information entropy, is the delay window, is the L1 regularization on the weight matrix W; The integration unit is configured to integrate the node set, the edge set and the sparse adjacency matrix to form a dynamic heterogeneous graph, and mark edges with changing correlation strengths to obtain a dynamic heterogeneous graph that can dynamically reflect changes in the correlation between water quality parameters over time.

10. The precise regulation system for recirculating aquaculture based on graph network water resources management according to claim 7, characterized in that, The processing module includes: The fusion unit is configured to process based on the dynamic heterogeneous graph using a gated spatio-temporal convolutional graph neural network; aggregate neighbor node features within a specified range around each node through an attention gate module, assign attention weights to key water quality sensitive nodes that affect fish school metabolism of dissolved oxygen and ammonia nitrogen, and interactively fuse node features extracted by different layers of the graph neural network to obtain a node feature set with fused cross-layer information; The recognition unit is configured to process the node feature set in a time series dimension through a time convolution module, capture dynamic evolution patterns of water quality parameters in a long period, and identify periodic rules of parameter changes; automatically identify water quality parameter change periods in different seasons and different breeding stages, and fuse a grey prediction model to complete the change trend of water quality parameters with low sampling frequency to obtain a node feature set with complete time series characteristics; The prediction unit is configured to classify and regress the extracted node feature set with complete time series characteristics through a neural network prediction layer, perform multi-modal joint prediction by fusing fish school behavior features and water quality features, and output a multi-dimensional water quality state prediction matrix including a potential anomaly probability, a water quality risk level, a water cycle efficiency and a biological fitness, and the calculation formula of the generation process is as follows: wherein, is a multi-dimensional water quality state prediction matrix, is the processing result of the gated spatio-temporal convolutional graph neural network on the dynamic heterogeneous graph G, is a feature fusion operator, is the prediction result of the grey prediction model on the low-frequency parameter node subset in the heterogeneous graph.

Images