AI fish tank ecological management method and system

Abstract

The invention discloses an AI fish tank ecological management method and system, and the method comprises the steps: collecting multi-source ecological data, carrying out the multi-modal feature extraction, and generating an ecological fingerprint vector; constructing a space-time dynamic causal map based on the ecological fingerprint vector in combination with day and night and operation states; a control action generation space is constructed, deformation constraint evolution is combined, and an anti-fact future path is generated; constructing a strategy form space based on a future path, executing expansion, contraction, splitting and merging, and updating the form; generating a control strategy template based on the updated morphological space, and issuing the control strategy template to equipment for linkage execution; and collecting ecological feedback data after execution, and updating a fingerprint spectrum path form to form a self-adaptive cycle. According to the method, ecological fingerprints are generated according to multi-source ecological data of the fish tank, and a space-time dynamic causal map, an anti-factual future path and a strategy form space are constructed, so that predictable generation and self-adaptive evolution management of a fish tank ecological control strategy are realized.

Description

Technical Field

[0001] This invention relates to the field of ecological management technology at the intersection of artificial intelligence and the Internet of Things, and in particular to an AI-powered fish tank ecological management method and system. Background Technology

[0002] With the increasing popularity of home aquariums and small-scale ornamental aquariums, smart aquarium products have gradually evolved from traditional manual maintenance to automated management integrating sensors, controllers, and mobile devices. Existing solutions typically collect water quality parameters such as water temperature, dissolved oxygen, pH, and turbidity, and combine them with actuators for aeration, filtration, lighting, feeding, and water changes to achieve coordinated control. Some products also incorporate cameras to observe the activity of the fish, thereby enabling remote monitoring, timed feeding, or threshold-triggered control, improving ease of use and reducing the burden of daily maintenance.

[0003] However, aquarium ecosystems are typical dynamically coupled systems, with causal relationships and lag effects that change over time among factors such as water quality parameters, fish behavior, equipment operation, and diurnal lighting. Existing smart aquariums mostly rely on fixed thresholds, fixed rules, or static models for control, lacking the ability to model the dynamic adjustments of ecological causal relationships with changes in diurnal cycles and operational behavior. This leads to difficulties in maintaining consistent stability and adaptability of control strategies under different ecological stages, different load conditions, or changes in external lighting, easily resulting in problems such as over-intervention, delayed response, or false triggering.

[0004] Existing technologies typically rely on single-path predictions or empirical judgments to determine operations, making it difficult to perform counterfactual extrapolation for multiple control actions and to predict future ecological trends. They also fail to provide a comprehensive assessment and comparison of the risks that different strategies may generate within future time windows. Due to the lack of mechanisms to automatically generate and evolve control strategies based on future risk changes, strategies are often fixed as pre-set plans or a limited number of selectable modes. This results in insufficient adaptability to complex situations such as rapid water quality fluctuations, algal community changes, or abnormal behavior, hindering the achievement of stable, sustainable, and evolvable aquarium ecological management.

[0005] Therefore, how to provide an AI-based aquarium ecosystem management method and system is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an AI-powered aquarium ecosystem management method and system. This invention constructs an ecological fingerprint vector by extracting multimodal features from multi-source ecological data, establishes a spatiotemporal dynamic causal graph by combining day / night states and equipment operating states, and generates counterfactual ecological future paths for various control actions based on this graph. Furthermore, it achieves adaptive generation and evolution of control strategies through the construction of a strategy morphology space and elastic deformation processing. This invention fully utilizes artificial intelligence data processing, causal modeling, and intelligent decision-making technologies to achieve accurate characterization of the dynamic laws of the aquarium ecosystem and predictable simulation of the future impact of control behaviors. It offers advantages such as more accurate ecological control, more comprehensive risk assessment, strong strategy adaptability, and high operational stability.

[0007] An AI-based aquarium ecosystem management method according to an embodiment of the present invention includes: Collect multi-source ecological data related to aquarium ecology, extract multi-modal features based on multi-source ecological data, and generate ecological fingerprint vectors; Based on ecological fingerprint vectors and day-night and operational status information from control equipment operating parameter data and environmental parameter data, a spatiotemporal dynamic causal graph of aquarium ecology is constructed. Based on spatiotemporal dynamic causal graphs and ecological fingerprint vectors, a control action generation space is constructed. In the control action generation space, multiple sets of candidate control actions are generated through combination, deformation and constraint evolution. Based on each candidate control action, multiple corresponding counterfactual ecological future paths are generated. Based on multiple counterfactual ecological future paths, a strategy form space is constructed. The strategy form space is subjected to elastic deformation processing based on the future state distribution. The strategy form corresponding to the safe future state is expanded, the strategy form corresponding to the risky future state is contracted, the strategy form exhibiting a multi-peak future state distribution is split, and the strategy form with similar future state distribution characteristics is merged to generate an updated strategy form space. Aquarium ecological control strategy templates are generated based on the updated strategy morphology space and distributed to oxygenation equipment, filtration equipment, lighting equipment, feeding equipment, and water change equipment for execution. The system collects ecological feedback data after the equipment has performed its tasks. Based on this data, it updates the ecological fingerprint vector, spatiotemporal dynamic causal graph, counterfactual ecological future path, and strategy morphology space, forming an adaptive update cycle for the ecological management process.

[0008] Optionally, the multi-source ecological data includes water quality parameter data, fish behavior data, equipment operation data, and environmental status data.

[0009] Optionally, generating the ecological fingerprint vector includes: Time series feature extraction is performed on the water quality parameter data in the multi-source ecological data to obtain a water quality feature vector; Behavioral features are extracted from the fish behavior data in the multi-source ecological data to obtain a fish behavior feature vector. State features are extracted from the equipment operation data and environmental status data in the multi-source ecological data to obtain the equipment environmental feature vector; The water quality feature vector, fish behavior feature vector, and equipment environment feature vector are subjected to feature fusion processing to generate an ecological fingerprint vector that characterizes the overall ecological state of the aquarium.

[0010] Optionally, the construction of the spatiotemporal dynamic causal graph of the aquarium ecosystem includes: An ecological node set is established, and a node data channel is established for each ecological node. The ecological nodes include water quality parameter nodes, fish behavior nodes, equipment operation status nodes, and environmental status nodes. The ecological fingerprint vector is bound to the node data channel of the ecological node to obtain the node input sequence. The node input sequence is divided into two phase segments according to day and night state information, and within each phase segment, it is divided into an operation sub-segment according to operation state information; Within each phase segment and each operation sub-segment, a set of candidate causal edges is generated based on the node input sequence. The generation of the set of candidate causal edges includes: for any pair of ecological nodes, determining the directionality of the candidate causal edges based on the consistency of the order and direction of change of the node data channels within the same time window. For each phase segment and each operation sub-segment, a reachability constraint screening is performed on the set of candidate causal edges. The reachability constraint screening includes: retaining only candidate causal edges that satisfy the causal direction constraints from the device operation status node to the water quality parameter node, from the environmental status node to the water quality parameter node, and from the water quality parameter node to the fish behavior node, and deleting candidate causal edges that contradict the causal direction constraints. Based on the filtered candidate causal edge set, causal subgraphs corresponding to phase segments and operation sub-segments are generated. The causal subgraphs corresponding to different phase segments and different operation sub-segments are then spliced ​​together in chronological order to form a spatiotemporal dynamic causal graph.

[0011] Optionally, the generation of multiple counterfactual ecological future paths based on each candidate control action includes: Based on the spatiotemporal dynamic causal graph and ecological fingerprint vector, the controllable adjustment dimension corresponding to each ecological node is determined, and the value range and change rate limit of the control parameters are established for each controllable adjustment dimension. A control action generation space is constructed. The control action generation space uses the range of control parameters of each controllable and adjustable dimension as the intensity axis, the duration and start time of the control action as the time axis, and the control parameter change rate limit as the deformation boundary. The controllable and adjustable dimensions corresponding to the device operation-driven causal connections marked in the spatiotemporal dynamic causal graph are set as priority expansion dimensions, and the controllable and adjustable dimensions corresponding to the day and night-driven causal connections marked in the spatiotemporal dynamic causal graph are set as phase synchronization dimensions. In the control action generation space, control action primitives are generated. The control action primitives include single-dimensional action primitives and multi-dimensional linkage action primitives. The multi-dimensional linkage action primitives are generated by binding two controllable and adjustable dimensions according to the causal connection direction in the spatiotemporal dynamic causal graph. In the control action generation space, the control action primitives are combined, intensity deformed, time-shifted and segmented spliced ​​to generate multiple sets of candidate control actions. Boundary verification and phase consistency verification are performed on each set of candidate control actions. Candidate control actions that do not meet the control parameter value range, change rate limit, linkage ratio constraint, linkage sequence constraint and phase synchronization constraint are eliminated, and the set of candidate control actions that meet the constraint conditions is retained. For each set of retained candidate control actions, the ecological fingerprint vector is used as the initial ecological state representation, the spatiotemporal dynamic causal graph is used as the input of the ecological causal structure, and the candidate control actions are used as control inputs to perform counterfactual inference to generate counterfactual ecological future paths.

[0012] Optionally, generating the updated strategy morphology space includes: From each set of candidate control actions and their corresponding counterfactual ecological future paths, we extract water quality parameter change features, ecological fingerprint change features, and causal topological structure change features over multiple future time segments to form future state feature data. Based on candidate control actions and future state feature data, a basic morphological layer is constructed for the strategy morphological space. A corresponding basic strategy morphological representation is generated for each group of candidate control actions. The basic morphological layer stores the relationship between each basic strategy morphological representation and the corresponding counterfactual ecological future path. A future state evaluation layer is constructed based on future state feature data. The counterfactual ecological future path corresponding to each basic strategy form representation is statistically analyzed to obtain future state evaluation features. Each basic strategy form representation is assigned to a safe future state category, a risky future state category, or a multi-peak future state category. The future state evaluation features are associated with the corresponding basic strategy form representations and stored in the future state evaluation layer. Based on the future state evaluation features, a deformation control layer is constructed for the strategy morphology space. In the deformation control layer, morphology expansion conditions, morphology contraction conditions, morphology splitting conditions, and morphology merging conditions are set for each basic strategy morphology representation. The morphology expansion condition is associated with the safe future state category, the morphology contraction condition is associated with the risky future state category, the morphology splitting condition is associated with the multi-peak future state category, and the morphology merging condition is associated with multiple basic strategy morphology representations whose similarity to the future state evaluation features exceeds the similarity threshold. Under the control of the deformation control layer, elastic deformation processing is performed on the strategy shape space: For basic strategy forms that meet the conditions for form expansion, the occupied area is increased in the strategy form space. For basic strategy morphology that meets the morphological contraction condition, the occupied area in the strategy morphology space is reduced. The basic strategy form representation that satisfies the form splitting condition is split into multiple new basic strategy form representations in the strategy form space. Multiple basic strategy form representations that meet the form merging conditions are merged into a single basic strategy form representation in the strategy form space. The corresponding records in the basic morphology layer and the future state evaluation layer are updated synchronously to obtain the updated strategy morphology space after elastic deformation processing.

[0013] Optionally, the step of generating aquarium ecological control strategy templates based on the updated strategy morphology space includes: Based on the updated strategy morphology space, basic strategy morphology representations that are associated with the safe future state category and whose stability evaluation characteristics meet the stability conditions are selected in the future state evaluation layer to form a candidate strategy morphology set. For each basic strategy form representation in the candidate strategy form set, read the control dimension combination information, control timing structure information and control intensity structure information stored in the basic form layer of the basic strategy form representation. Determine the control dimension combination information as the set of device control objects, determine the control timing structure information as the control start time, control duration and segmentation order of each device control object, and determine the control intensity structure information as the control parameter values ​​of each device control object in each segment. Based on the set of controlled devices, control start time, control duration, segmentation order, and control parameter values, an aquarium ecological control strategy template is generated. The aquarium ecological control strategy template includes a multi-device joint control configuration for oxygenation devices, filtration devices, lighting devices, feeding devices, and water changing devices, and includes the execution order and parallel execution relationship of each device. The aquarium ecological control strategy template is converted into a set of device control instructions that the devices can recognize. The device control instructions are time-programmed according to the execution order and parallel execution relationship determined in the aquarium ecological control strategy template. The time-programmed device control instructions are then sent to the corresponding oxygenation devices, filtration devices, lighting devices, feeding devices, and water changing devices for execution.

[0014] Optionally, the adaptive update cycle that forms the ecological management process includes: Collect ecological feedback data after the equipment is executed. The ecological feedback data includes water quality parameter feedback data, fish behavior feedback data, actual equipment execution status data, and actual equipment execution time sequence data. Align the ecological feedback data with the aquarium ecological control strategy template in time to obtain the aligned feedback dataset. The aligned feedback dataset is re-input into the multimodal feature extraction process used to generate the ecological fingerprint vector, generating an updated ecological fingerprint vector, which is then used as the input for the next round of constructing the spatiotemporal dynamic causal graph. The spatiotemporal dynamic causal graph is updated based on the aligned feedback dataset. The causal connections, causal type labels, and time delay labels between ecological nodes in each time segment are adjusted to obtain the updated spatiotemporal dynamic causal graph. The updated spatiotemporal dynamic causal graph is then used as the causal structure input for the next round of generating counterfactual ecological future paths. Based on the updated ecological fingerprint vector and the updated spatiotemporal dynamic causal graph, the counterfactual ecological future path is corrected, and the updated counterfactual ecological future path set is obtained. The updated counterfactual ecological future path set is written into the future state evaluation layer to update the strategy morphology space, forming an adaptive update cycle of the ecological management process.

[0015] An AI aquarium ecosystem management system according to an embodiment of the present invention includes the following modules: The ecological fingerprint generation module is used to collect multi-source ecological data and extract multi-modal features to generate ecological fingerprint vectors. The causal graph construction module is used to construct a spatiotemporal dynamic causal graph based on ecological fingerprint vectors, day and night state information, and operational state information. The action generation and counterfactual deduction module is used to construct a control action generation space based on spatiotemporal dynamic causal graphs and ecological fingerprint vectors, generate candidate control actions, and generate corresponding counterfactual ecological future paths. The strategy form construction and deformation module is used to obtain the updated strategy form space based on the counterfactual ecological future path; The strategy template generation and execution module is used to generate control strategy templates based on the updated strategy morphology space and distribute them to the oxygenation, filtration, lighting, feeding and water exchange equipment for execution. The feedback adaptive update module is used to collect ecological feedback data and update the ecological fingerprint vector, spatiotemporal dynamic causal graph, counterfactual ecological future path and strategy morphology space, forming an adaptive update cycle.

[0016] The beneficial effects of this invention are: This invention extracts multimodal features from multi-source ecological data, including water quality parameters, fish behavior, equipment operating status, and environmental day-night information, to generate ecological fingerprint vectors, thus achieving a unified representation of the aquarium's ecological state. Furthermore, it constructs a spatiotemporal dynamic causal graph, enabling the causal connections between ecological nodes to dynamically adjust with time segments, day-night phases, and operational phases. This overcomes the shortcomings of existing technologies that rely on static thresholds or fixed rules and struggle to adapt to dynamic changes in ecological causal relationships, thereby improving the accuracy of ecological state characterization and the reliability of control basis.

[0017] This invention further constructs a control action generation space based on spatiotemporal dynamic causal graphs and ecological fingerprint vectors. Within this space, candidate control actions are generated through combination, deformation, and constraint evolution. For each set of candidate control actions, a corresponding counterfactual ecological future path is generated. This enables the system to perform multi-path deduction on the future water quality parameter sequences, ecological fingerprint sequences, and causal topological structure sequences of different control actions before execution, thereby obtaining the future ecological situation distribution. This significantly enhances the ability to assess the global risk of control strategies and reduces the risk of ecological fluctuations caused by response lag, false triggering, or excessive intervention.

[0018] This invention constructs a strategy morphology space based on multiple counterfactual ecological future paths, and performs elastic deformation processing on the strategy morphology space through morphological expansion, morphological contraction, morphological splitting, and morphological merging to generate an updated strategy morphology space. From this, an aquarium ecological control strategy template is generated and distributed. Combined with the ecological feedback data after execution, the ecological fingerprint vector, spatiotemporal dynamic causal map, counterfactual ecological future paths, and strategy morphology space are adaptively updated to form a sustainable iterative ecological management closed loop. This enables the control strategy to have the ability to self-organize and self-evolve with changes in the ecological future state, improving the stability and adaptability of long-term operation, and achieving the beneficial effects of more accurate, stable, and evolvable ecological control. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an AI aquarium ecological management method proposed in this invention; Figure 2 This is a schematic diagram of the structure of an AI aquarium ecological management system proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figure 1 An AI-powered aquarium ecosystem management method includes: Collect multi-source ecological data related to aquarium ecology, extract multi-modal features based on multi-source ecological data, and generate ecological fingerprint vectors; Based on ecological fingerprint vectors and day-night and operational status information from control equipment operating parameter data and environmental parameter data, a spatiotemporal dynamic causal graph of aquarium ecology is constructed. Based on spatiotemporal dynamic causal graphs and ecological fingerprint vectors, a control action generation space is constructed. In the control action generation space, multiple sets of candidate control actions are generated through combination, deformation and constraint evolution. Based on each candidate control action, multiple corresponding counterfactual ecological future paths are generated. Based on multiple counterfactual ecological future paths, a strategy form space is constructed. The strategy form space is subjected to elastic deformation processing based on the future state distribution. The strategy form corresponding to the safe future state is expanded, the strategy form corresponding to the risky future state is contracted, the strategy form exhibiting a multi-peak future state distribution is split, and the strategy form with similar future state distribution characteristics is merged to generate an updated strategy form space. Aquarium ecological control strategy templates are generated based on the updated strategy morphology space and distributed to oxygenation equipment, filtration equipment, lighting equipment, feeding equipment, and water change equipment for execution. The system collects ecological feedback data after the equipment has performed its tasks. Based on this data, it updates the ecological fingerprint vector, spatiotemporal dynamic causal graph, counterfactual ecological future path, and strategy morphology space, forming an adaptive update cycle for the ecological management process.

[0022] In this embodiment, the multi-source ecological data includes water quality parameter data, fish behavior data, equipment operation data, and environmental status data.

[0023] In this embodiment, generating the ecological fingerprint vector includes: Time series feature extraction is performed on the water quality parameter data in the multi-source ecological data to obtain a water quality feature vector; Behavioral features are extracted from the fish behavior data in the multi-source ecological data to obtain a fish behavior feature vector. State features are extracted from the equipment operation data and environmental status data in the multi-source ecological data to obtain the equipment environmental feature vector; The water quality feature vector, fish behavior feature vector, and equipment environment feature vector are subjected to feature fusion processing to generate an ecological fingerprint vector that characterizes the overall ecological state of the aquarium.

[0024] In this embodiment, the construction of the spatiotemporal dynamic causal graph of the aquarium ecosystem includes: An ecological node set is established, and a node data channel is established for each ecological node. The ecological nodes include water quality parameter nodes, fish behavior nodes, equipment operation status nodes, and environmental status nodes. The ecological fingerprint vector is bound to the node data channel of the ecological node to obtain the node input sequence. The node input sequence is divided into two phase segments based on day / night state information. Within each phase segment, it is further divided into an operation sub-segment based on operation state information. Specifically, the division of the node input sequence into two phase segments based on day / night state information is as follows: A day-night state determination rule is established. The ambient light characteristics and timestamp characteristics corresponding to each moment in the node input sequence are input into the day-night state determination rule to obtain the day-night state label corresponding to each moment in the node input sequence. The day-night state determination rule is as follows: Using ambient light intensity as the criterion, a light intensity threshold for distinguishing between day and night states and a minimum duration for avoiding the influence of instantaneous fluctuations on the determination results are preset. When the ambient light intensity continuously reaches or exceeds the light intensity threshold and the duration is not less than the minimum duration, the corresponding time period is determined as day. When the ambient light intensity continuously falls below the light intensity threshold and the duration is not less than the minimum duration, the corresponding time period is determined as night. When the ambient light intensity switches from meeting the determination conditions for day to meeting the determination conditions for night, or from meeting the determination conditions for night to meeting the determination conditions for day, the moment when the determination conditions first switch occurs is determined as the day-night phase switching boundary. The node input sequence is divided into a first phase segment and a second phase segment according to the day and night state markers. The first phase segment is the day phase segment, which contains node input sequence segments corresponding to consecutive times when the day and night state markers are day. The second phase segment is the night phase segment, which contains node input sequence segments corresponding to consecutive times when the day and night state markers are night. Boundary alignment processing is performed on the first phase segment and the second phase segment respectively. The moment when the day and night state markers change at adjacent times is taken as the phase switching boundary. The node input sequence at the phase switching boundary is truncated or spliced ​​according to a preset window to obtain the day phase segment and night phase segment that are continuous in time and do not overlap. Within each phase segment and each operation sub-segment, a set of candidate causal edges is generated based on the node input sequence. The generation of the set of candidate causal edges includes: for any pair of ecological nodes, determining the directionality of the candidate causal edges based on the consistency of the order and direction of change of the node data channels within the same time window. For each phase segment and each operation sub-segment, a reachability constraint screening is performed on the set of candidate causal edges. The reachability constraint screening includes: retaining only candidate causal edges that satisfy the causal direction constraints from the device operation status node to the water quality parameter node, from the environmental status node to the water quality parameter node, and from the water quality parameter node to the fish behavior node, and deleting candidate causal edges that contradict the causal direction constraints. Based on the filtered candidate causal edge set, causal subgraphs corresponding to phase segments and operation segments are generated. The causal subgraphs corresponding to different phase segments and different operation segments are spliced ​​together in time order to form a spatiotemporal dynamic causal graph. When the day-night state label or operation state label changes, the causal subgraph with the corresponding label is switched, so that the topology of the spatiotemporal dynamic causal graph switches dynamically with the day-night state and operation state.

[0025] In this embodiment, the generation of multiple counterfactual ecological future paths based on each candidate control action includes: Based on the spatiotemporal dynamic causal graph and ecological fingerprint vector, the controllable adjustment dimension corresponding to each ecological node is determined, and the value range and change rate limit of the control parameters are established for each controllable adjustment dimension. A control action generation space is constructed. The control action generation space uses the range of control parameters of each controllable and adjustable dimension as the intensity axis, the duration and start time of the control action as the time axis, and the control parameter change rate limit as the deformation boundary. The controllable and adjustable dimensions corresponding to the device operation-driven causal connections marked in the spatiotemporal dynamic causal graph are set as priority expansion dimensions, and the controllable and adjustable dimensions corresponding to the day and night-driven causal connections marked in the spatiotemporal dynamic causal graph are set as phase synchronization dimensions. In the control action generation space, control action primitives are generated. The control action primitives include single-dimensional action primitives and multi-dimensional linkage action primitives. The multi-dimensional linkage action primitives are generated by binding two controllable adjustable dimensions according to the causal connection direction in the spatiotemporal dynamic causal graph. The single-dimensional action primitives are generated by discretizing any controllable adjustable dimension within its adjustable range according to the adjustment granularity to form a single-dimensional action value set, and combining any value in the single-dimensional action value set with its corresponding execution duration, execution start time, and execution change rate limit. In the control action generation space, the control action primitives are combined, intensity deformed, time-shifted and segmented spliced ​​to generate multiple sets of candidate control actions. Boundary verification and phase consistency verification are performed on each set of candidate control actions. Candidate control actions that do not meet the control parameter value range, change rate limit, linkage ratio constraint, linkage sequence constraint and phase synchronization constraint are eliminated, and the set of candidate control actions that meet the constraint conditions is retained. For each retained set of candidate control actions, using the ecological fingerprint vector as the initial ecological state representation and the spatiotemporal dynamic causal graph as the input to the ecological causal structure, the candidate control actions are used as control inputs to perform counterfactual inference, generating counterfactual ecological future paths. Specifically, the counterfactual inference is performed by using candidate control actions as control inputs as follows: Determine the prediction time range and the prediction time step of the counterfactual inference, and divide the prediction time range into multiple prediction moments according to the prediction time step; At each simulation moment, read the control values, execution duration, execution start time and execution change rate limit of each controllable adjustment dimension in the candidate control action corresponding to that simulation moment, and map the control values ​​to the control input of the simulation moment; At each simulation moment, the node features in the spatiotemporal dynamic causal graph that are causally connected to the control input are updated based on the control input, and the update effect is propagated in time according to the causal connection direction and time delay label of the spatiotemporal dynamic causal graph to obtain the predicted node state at the simulation moment. Based on the predicted node state at the time of the simulation, a predicted ecological fingerprint vector for the time of the simulation is generated, and the predicted ecological fingerprint vector is used as the initial ecological state representation for the next time of the simulation. This process is repeated until all simulation times within the predicted time range are completed. By combining the predicted node state sequence, the predicted ecological fingerprint vector sequence, and the causal topological structure sequence derived from the predicted node state at each simulation time, a counterfactual ecological future path corresponding to the candidate control action is formed.

[0026] In this embodiment, generating the updated strategy morphology space includes: From each set of candidate control actions and their corresponding counterfactual ecological future paths, we extract water quality parameter change features, ecological fingerprint change features, and causal topological structure change features over multiple future time segments to form future state feature data. A basic morphological layer is constructed based on candidate control actions and future state feature data to build a strategy morphological space. A corresponding basic strategy morphological representation is generated for each group of candidate control actions. The basic morphological layer stores the association between each basic strategy morphological representation and its corresponding counterfactual ecological future path. Specifically, the generation of the corresponding basic strategy morphological representation for each group of candidate control actions is as follows: For each group of candidate control actions, extract its action structure features, which include the types of controllable adjustment dimensions involved, the adjustment range of each adjustment dimension, the execution duration, the execution start time, and the linkage relationship between each adjustment dimension. Future state feature data is extracted from the counterfactual ecological future path corresponding to the candidate control action. The future state feature data includes the future water quality parameter change trend, the future ecological fingerprint vector change trajectory, the number of changes in the future causal topology, and the time interval features between each change. The action structure features and the future state feature data are concatenated and normalized to form a strategy feature description vector with a unified dimension. The strategy feature description vector is subjected to feature mapping processing and mapped to a preset strategy morphology space coordinate system to obtain a basic strategy morphology representation used to characterize the overall regulation characteristics of candidate control actions; A future state evaluation layer is constructed based on future state feature data. The counterfactual ecological future path corresponding to each basic strategy form representation is statistically analyzed to obtain future state evaluation features. Each basic strategy form representation is assigned to a safe future state category, a risky future state category, or a multi-peak future state category. The future state evaluation features are associated with the corresponding basic strategy form representations and stored in the future state evaluation layer. Based on the future state evaluation features, a deformation control layer is constructed for the strategy morphology space. In the deformation control layer, morphology expansion conditions, morphology contraction conditions, morphology splitting conditions, and morphology merging conditions are set for each basic strategy morphology representation. The morphology expansion condition is associated with the safe future state category, the morphology contraction condition is associated with the risky future state category, the morphology splitting condition is associated with the multi-peak future state category, and the morphology merging condition is associated with multiple basic strategy morphology representations whose similarity to the future state evaluation features exceeds the similarity threshold. Under the control of the deformation control layer, elastic deformation processing is performed on the strategy shape space: For basic strategy forms that meet the conditions for form expansion, the occupied area is increased in the strategy form space. For basic strategy morphology that meets the morphological contraction condition, the occupied area in the strategy morphology space is reduced. The basic strategy form representation that satisfies the form splitting condition is split into multiple new basic strategy form representations in the strategy form space. Multiple basic strategy form representations that meet the form merging conditions are merged into a single basic strategy form representation in the strategy form space. The corresponding records in the basic morphology layer and the future state evaluation layer are updated synchronously to obtain the updated strategy morphology space after elastic deformation processing.

[0027] In this embodiment, generating the aquarium ecological control strategy template based on the updated strategy morphology space includes: Based on the updated strategy morphology space, basic strategy morphology representations that are associated with the safe future state category and whose stability evaluation characteristics meet the stability conditions are selected in the future state evaluation layer to form a candidate strategy morphology set. For each basic strategy form representation in the candidate strategy form set, read the control dimension combination information, control timing structure information and control intensity structure information stored in the basic form layer of the basic strategy form representation. Determine the control dimension combination information as the set of device control objects, determine the control timing structure information as the control start time, control duration and segmentation order of each device control object, and determine the control intensity structure information as the control parameter values ​​of each device control object in each segment. Based on the set of controlled devices, control start time, control duration, segmentation order, and control parameter values, an aquarium ecological control strategy template is generated. The aquarium ecological control strategy template includes a multi-device joint control configuration for oxygenation devices, filtration devices, lighting devices, feeding devices, and water changing devices, and includes the execution order and parallel execution relationship of each device. The aquarium ecological control strategy template is converted into a set of device control instructions that the devices can recognize. The device control instructions are time-programmed according to the execution order and parallel execution relationship determined in the aquarium ecological control strategy template. The time-programmed device control instructions are then sent to the corresponding oxygenation devices, filtration devices, lighting devices, feeding devices, and water changing devices for execution.

[0028] In this embodiment, the formation of an adaptive update cycle for the ecological management process includes: Collect ecological feedback data after the equipment is executed. The ecological feedback data includes water quality parameter feedback data, fish behavior feedback data, actual equipment execution status data, and actual equipment execution time sequence data. Align the ecological feedback data with the aquarium ecological control strategy template in time to obtain the aligned feedback dataset. The aligned feedback dataset is re-input into the multimodal feature extraction process used to generate the ecological fingerprint vector, generating an updated ecological fingerprint vector, which is then used as the input for the next round of constructing the spatiotemporal dynamic causal graph. The spatiotemporal dynamic causal graph is updated based on the aligned feedback dataset. The causal connections, causal type labels, and time delay labels between ecological nodes in each time segment are adjusted to obtain the updated spatiotemporal dynamic causal graph. The updated spatiotemporal dynamic causal graph is then used as the causal structure input for the next round of generating counterfactual ecological future paths. Based on the updated ecological fingerprint vector and the updated spatiotemporal dynamic causal graph, the counterfactual ecological future path is corrected, and the updated counterfactual ecological future path set is obtained. The updated counterfactual ecological future path set is written into the future state evaluation layer to update the strategy morphology space, forming an adaptive update cycle of the ecological management process.

[0029] refer to Figure 2 An AI-powered aquarium ecosystem management system includes the following modules: The ecological fingerprint generation module is used to collect multi-source ecological data and extract multi-modal features to generate ecological fingerprint vectors. The causal graph construction module is used to construct a spatiotemporal dynamic causal graph based on ecological fingerprint vectors, day and night state information, and operational state information. The action generation and counterfactual deduction module is used to construct a control action generation space based on spatiotemporal dynamic causal graphs and ecological fingerprint vectors, generate candidate control actions, and generate corresponding counterfactual ecological future paths. The strategy form construction and deformation module is used to obtain the updated strategy form space based on the counterfactual ecological future path; The strategy template generation and execution module is used to generate control strategy templates based on the updated strategy morphology space and distribute them to the oxygenation, filtration, lighting, feeding and water exchange equipment for execution. The feedback adaptive update module is used to collect ecological feedback data and update the ecological fingerprint vector, spatiotemporal dynamic causal graph, counterfactual ecological future path and strategy morphology space, forming an adaptive update cycle. Example 1

[0030] To verify the feasibility of this invention in practice, it was applied to a freshwater ornamental fish tank in a home. The tank, located in the living room of a typical residence, has a volume of 180 liters and employs a combination of top and external filters. It houses 15 tropical ornamental fish, supplemented with a small amount of submerged plants. The tank is equipped with dissolved oxygen, temperature, pH, and turbidity sensors, and a camera to monitor the fish's activity. Control equipment includes aeration, filtration, lighting, automatic feeding, and automatic water changing systems. Indoor lighting is influenced by both natural and artificial light, exhibiting significant diurnal variations.

[0031] Before using the method of this invention, the aquarium employed a common intelligent control method, controlling oxygenation and filtration equipment through fixed thresholds, and controlling lighting and feeding according to a fixed schedule. In actual operation, although water quality parameters were mostly within safe ranges, problems such as rapid drop in dissolved oxygen, short-term accumulation of ammonia nitrogen, and frequent start-ups and shutdowns of equipment easily occurred after feeding and at night. The fish exhibited mild stress behavior at certain times, and the overall ecological stability was limited.

[0032] In this embodiment, an AI-powered aquarium ecosystem management method of the present invention is deployed and run on an aquarium control terminal. The system continuously collects multi-source ecological data related to the aquarium ecosystem, including water quality parameter data, fish behavior video data, equipment operating status data, and environmental day-night status data. Based on the above multi-source ecological data, multimodal feature extraction is performed to generate an ecological fingerprint vector that characterizes the overall ecological state of the aquarium.

[0033] Based on the ecological fingerprint vector, and combined with the day-night and operational status information from the control equipment operating parameter data and environmental parameter data, the system constructs a spatiotemporal dynamic causal graph of the aquarium ecosystem. This allows the causal relationship between water quality changes, fish behavior changes, and equipment operation to be dynamically adjusted according to day-night changes and operational behavior. Subsequently, the system constructs a control action generation space based on the spatiotemporal dynamic causal graph and the ecological fingerprint vector. In this space, multiple sets of candidate control actions are generated through the combination, transformation, and constraint evolution of control intensity, execution sequence, and multi-device linkage relationships. For each set of candidate control actions, a corresponding counterfactual ecological future path is generated to describe the future trends of water quality parameters and ecological state under different control actions.

[0034] The system further constructs a strategy morphology space based on multiple counterfactual ecological future paths and performs elastic deformation processing on the strategy morphology space based on future state distribution. This causes strategy morphologies corresponding to safe future states to expand in the space, while strategy morphologies corresponding to risky future states to contract. Strategy morphologies exhibiting multi-peak future state distributions are split, and strategy morphologies with similar future state distribution characteristics are merged, thereby generating an updated strategy morphology space. Based on this updated strategy morphology space, the system generates aquarium ecological control strategy templates and distributes these templates to oxygenation equipment, filtration equipment, lighting equipment, feeding equipment, and water change equipment for execution.

[0035] During the execution of the control strategy template by the device, the system continuously collects ecological feedback data after the device executes the strategy, and updates the ecological fingerprint vector, spatiotemporal dynamic causal graph, counterfactual ecological future path and strategy morphology space based on the ecological feedback data, forming an adaptive update cycle of the ecological management process, so that the control strategy can evolve with the long-term ecological state of the aquarium.

[0036] Under the same aquarium conditions, the operating effects of traditional threshold control methods and the method of this invention were compared. The methods were run continuously for 21 days and key ecological indicators were recorded.

[0037] Table 1 Comparison of aquarium ecological operation indicators under different control methods Indicator Name Traditional control methods Method of the present invention Average dissolved oxygen (mg / L) 6.2 6.6 Lowest dissolved oxygen at night (mg / L) 4.9 5.4 Water quality recovery time after feeding (hours) 6.5 4.2 Peak ammonia nitrogen (mg / L) 0.22 0.14 Average number of times the oxygenation equipment is started and stopped per day (times) 9.1 5.8 Average number of times the filtration equipment is started and stopped per day (times) 6.4 4.1 Frequency of abnormal fish behavior (times / day) 1.3 0.6 As shown in Table 1, the dissolved oxygen level in the aquarium water was improved overall after adopting the method of this invention. The average dissolved oxygen increased from 6.2 mg / L under the traditional control method to 6.6 mg / L, indicating that the system can more rationally adjust the working intensity of the aeration equipment during daily operation. At the same time, the lowest dissolved oxygen level at night increased from 4.9 mg / L to 5.4 mg / L, indicating that this invention can regulate the dissolved oxygen level in advance before the nighttime decline worsens, effectively reducing the risk of nighttime hypoxia and improving the stability of the aquarium ecosystem during the diurnal cycle.

[0038] In terms of water quality recovery capability, this invention also demonstrates significant advantages. The water quality recovery time after feeding was shortened from 6.5 hours under traditional control methods to 4.2 hours. This reflects that after feeding, the system can identify water quality change trends in advance based on ecological fingerprint vectors and spatiotemporal dynamic causal maps, and adopt more suitable control strategies, accelerating the return of the water body from a short-term disturbed state to a stable state. At the same time, the peak ammonia nitrogen level decreased from 0.22 mg / L to 0.14 mg / L, indicating that this invention, through the synergistic effect of controlling feeding, filtration, and water exchange, effectively inhibits the phased accumulation of harmful metabolites and reduces the risk of water quality deterioration.

[0039] At the level of equipment operation and biological behavior, this invention also demonstrates higher operational efficiency and eco-friendliness. The average daily start-up and shutdown frequency of the aeration equipment decreased from 9.1 times to 5.8 times, and the average daily start-up and shutdown frequency of the filtration equipment decreased from 6.4 times to 4.1 times, indicating that the system can reduce unnecessary frequent start-ups and shutdowns, making equipment operation smoother, helping to extend equipment life and reduce energy consumption. The frequency of abnormal fish behavior decreased from 1.3 times / day to 0.6 times / day, indicating that the fluctuation range of the ecological environment is reduced and the stress response of the fish is alleviated, further verifying the practical effect of this invention in improving the ecological stability and biological comfort of the aquarium.

[0040] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An AI-based aquarium ecosystem management method, characterized in that, include: Collect multi-source ecological data related to aquarium ecology, extract multi-modal features based on multi-source ecological data, and generate ecological fingerprint vectors; Based on ecological fingerprint vectors and day-night and operational status information from control equipment operating parameter data and environmental parameter data, a spatiotemporal dynamic causal graph of aquarium ecology is constructed. Based on spatiotemporal dynamic causal graphs and ecological fingerprint vectors, a control action generation space is constructed. In the control action generation space, multiple sets of candidate control actions are generated through combination, deformation and constraint evolution. Based on each candidate control action, multiple corresponding counterfactual ecological future paths are generated. Based on multiple counterfactual ecological future paths, a strategy form space is constructed. The strategy form space is subjected to elastic deformation processing based on the future state distribution. The strategy form corresponding to the safe future state is expanded, the strategy form corresponding to the risky future state is contracted, the strategy form exhibiting a multi-peak future state distribution is split, and the strategy form with similar future state distribution characteristics is merged to generate an updated strategy form space. Aquarium ecological control strategy templates are generated based on the updated strategy morphology space and distributed to oxygenation equipment, filtration equipment, lighting equipment, feeding equipment, and water change equipment for execution. The system collects ecological feedback data after the equipment has performed its tasks. Based on this data, it updates the ecological fingerprint vector, spatiotemporal dynamic causal graph, counterfactual ecological future path, and strategy morphology space, forming an adaptive update cycle for the ecological management process.

2. The AI ​​aquarium ecological management method according to claim 1, characterized in that, The multi-source ecological data includes water quality parameter data, fish behavior data, equipment operation data, and environmental status data.

3. The AI ​​aquarium ecological management method according to claim 1, characterized in that, The generation of the ecological fingerprint vector includes: Time series feature extraction is performed on the water quality parameter data in the multi-source ecological data to obtain a water quality feature vector; Behavioral features are extracted from the fish behavior data in the multi-source ecological data to obtain a fish behavior feature vector. State features are extracted from the equipment operation data and environmental status data in the multi-source ecological data to obtain the equipment environmental feature vector; The water quality feature vector, fish behavior feature vector, and equipment environment feature vector are subjected to feature fusion processing to generate an ecological fingerprint vector that characterizes the overall ecological state of the aquarium.

4. The AI ​​aquarium ecological management method according to claim 1, characterized in that, The spatiotemporal dynamic causal graph for constructing the aquarium ecosystem includes: An ecological node set is established, and a node data channel is established for each ecological node. The ecological nodes include water quality parameter nodes, fish behavior nodes, equipment operation status nodes, and environmental status nodes. The ecological fingerprint vector is bound to the node data channel of the ecological node to obtain the node input sequence. The node input sequence is divided into two phase segments according to day and night state information, and within each phase segment, it is divided into an operation sub-segment according to operation state information; Within each phase segment and each operation sub-segment, a set of candidate causal edges is generated based on the node input sequence. The generation of the set of candidate causal edges includes: for any pair of ecological nodes, determining the directionality of the candidate causal edges based on the consistency of the order and direction of change of the node data channels within the same time window. For each phase segment and each operation sub-segment, a reachability constraint screening is performed on the set of candidate causal edges. The reachability constraint screening includes: retaining only candidate causal edges that satisfy the causal direction constraints from the device operation status node to the water quality parameter node, from the environmental status node to the water quality parameter node, and from the water quality parameter node to the fish behavior node, and deleting candidate causal edges that contradict the causal direction constraints. Based on the filtered candidate causal edge set, causal subgraphs corresponding to phase segments and operation sub-segments are generated. The causal subgraphs corresponding to different phase segments and different operation sub-segments are then spliced ​​together in chronological order to form a spatiotemporal dynamic causal graph.

5. The AI ​​aquarium ecological management method according to claim 1, characterized in that, The generation of multiple counterfactual ecological future paths based on each candidate control action includes: Based on the spatiotemporal dynamic causal graph and ecological fingerprint vector, the controllable adjustment dimension corresponding to each ecological node is determined, and the value range and change rate limit of the control parameters are established for each controllable adjustment dimension. A control action generation space is constructed. The control action generation space uses the range of control parameters of each controllable and adjustable dimension as the intensity axis, the duration and start time of the control action as the time axis, and the control parameter change rate limit as the deformation boundary. The controllable and adjustable dimensions corresponding to the device operation-driven causal connections marked in the spatiotemporal dynamic causal graph are set as priority expansion dimensions, and the controllable and adjustable dimensions corresponding to the day and night-driven causal connections marked in the spatiotemporal dynamic causal graph are set as phase synchronization dimensions. In the control action generation space, control action primitives are generated. The control action primitives include single-dimensional action primitives and multi-dimensional linkage action primitives. The multi-dimensional linkage action primitives are generated by binding two controllable and adjustable dimensions according to the causal connection direction in the spatiotemporal dynamic causal graph. In the control action generation space, the control action primitives are combined, intensity deformed, time-shifted and segmented spliced ​​to generate multiple sets of candidate control actions. Boundary verification and phase consistency verification are performed on each set of candidate control actions. Candidate control actions that do not meet the control parameter value range, change rate limit, linkage ratio constraint, linkage sequence constraint and phase synchronization constraint are eliminated, and the set of candidate control actions that meet the constraint conditions is retained. For each set of retained candidate control actions, the ecological fingerprint vector is used as the initial ecological state representation, the spatiotemporal dynamic causal graph is used as the input of the ecological causal structure, and the candidate control actions are used as control inputs to perform counterfactual inference to generate counterfactual ecological future paths.

6. The AI ​​aquarium ecological management method according to claim 1, characterized in that, The generation of the updated strategy morphology space includes: From each set of candidate control actions and their corresponding counterfactual ecological future paths, we extract water quality parameter change features, ecological fingerprint change features, and causal topological structure change features over multiple future time segments to form future state feature data. Based on candidate control actions and future state feature data, a basic morphological layer is constructed for the strategy morphological space. A corresponding basic strategy morphological representation is generated for each group of candidate control actions. The basic morphological layer stores the relationship between each basic strategy morphological representation and the corresponding counterfactual ecological future path. A future state evaluation layer is constructed based on future state feature data. The counterfactual ecological future path corresponding to each basic strategy form representation is statistically analyzed to obtain future state evaluation features. Each basic strategy form representation is assigned to a safe future state category, a risky future state category, or a multi-peak future state category. The future state evaluation features are associated with the corresponding basic strategy form representations and stored in the future state evaluation layer. Based on the future state evaluation features, a deformation control layer is constructed for the strategy morphology space. In the deformation control layer, morphology expansion conditions, morphology contraction conditions, morphology splitting conditions, and morphology merging conditions are set for each basic strategy morphology representation. The morphology expansion condition is associated with the safe future state category, the morphology contraction condition is associated with the risky future state category, the morphology splitting condition is associated with the multi-peak future state category, and the morphology merging condition is associated with multiple basic strategy morphology representations whose similarity to the future state evaluation features exceeds the similarity threshold. Under the control of the deformation control layer, elastic deformation processing is performed on the strategy shape space: For basic strategy forms that meet the conditions for form expansion, the occupied area is increased in the strategy form space. For basic strategy morphology that meets the morphological contraction condition, the occupied area in the strategy morphology space is reduced. The basic strategy form representation that satisfies the form splitting condition is split into multiple new basic strategy form representations in the strategy form space. Multiple basic strategy form representations that meet the form merging conditions are merged into a single basic strategy form representation in the strategy form space. The corresponding records in the basic morphology layer and the future state evaluation layer are updated synchronously to obtain the updated strategy morphology space after elastic deformation processing.

7. The AI ​​aquarium ecological management method according to claim 1, characterized in that, The process of generating aquarium ecological control strategy templates based on the updated strategy morphology space includes: Based on the updated strategy morphology space, basic strategy morphology representations that are associated with the safe future state category and whose stability evaluation characteristics meet the stability conditions are selected in the future state evaluation layer to form a candidate strategy morphology set. For each basic strategy form representation in the candidate strategy form set, read the control dimension combination information, control timing structure information and control intensity structure information stored in the basic form layer of the basic strategy form representation. Determine the control dimension combination information as the set of device control objects, determine the control timing structure information as the control start time, control duration and segmentation order of each device control object, and determine the control intensity structure information as the control parameter values ​​of each device control object in each segment. Based on the set of controlled devices, control start time, control duration, segmentation order, and control parameter values, an aquarium ecological control strategy template is generated. The aquarium ecological control strategy template includes a multi-device joint control configuration for oxygenation devices, filtration devices, lighting devices, feeding devices, and water changing devices, and includes the execution order and parallel execution relationship of each device. The aquarium ecological control strategy template is converted into a set of device control instructions that the devices can recognize. The device control instructions are time-programmed according to the execution order and parallel execution relationship determined in the aquarium ecological control strategy template. The time-programmed device control instructions are then sent to the corresponding oxygenation devices, filtration devices, lighting devices, feeding devices, and water changing devices for execution.

8. The AI ​​aquarium ecological management method according to claim 1, characterized in that, The adaptive update cycle that forms the ecological management process includes: Collect ecological feedback data after the equipment is executed. The ecological feedback data includes water quality parameter feedback data, fish behavior feedback data, actual equipment execution status data, and actual equipment execution time sequence data. Align the ecological feedback data with the aquarium ecological control strategy template in time to obtain the aligned feedback dataset. The aligned feedback dataset is re-input into the multimodal feature extraction process used to generate the ecological fingerprint vector, generating an updated ecological fingerprint vector, which is then used as the input for the next round of constructing the spatiotemporal dynamic causal graph. The spatiotemporal dynamic causal graph is updated based on the aligned feedback dataset. The causal connections, causal type labels, and time delay labels between ecological nodes in each time segment are adjusted to obtain the updated spatiotemporal dynamic causal graph. The updated spatiotemporal dynamic causal graph is then used as the causal structure input for the next round of generating counterfactual ecological future paths. Based on the updated ecological fingerprint vector and the updated spatiotemporal dynamic causal graph, the counterfactual ecological future path is corrected, and the updated counterfactual ecological future path set is obtained. The updated counterfactual ecological future path set is written into the future state evaluation layer to update the strategy morphology space, forming an adaptive update cycle of the ecological management process.

9. An AI aquarium ecosystem management system, comprising the AI ​​aquarium ecosystem management method according to any one of claims 1 to 8, characterized in that, Includes the following modules: The ecological fingerprint generation module is used to collect multi-source ecological data and extract multi-modal features to generate ecological fingerprint vectors. The causal graph construction module is used to construct a spatiotemporal dynamic causal graph based on ecological fingerprint vectors, day and night state information, and operational state information. The action generation and counterfactual deduction module is used to construct a control action generation space based on spatiotemporal dynamic causal graphs and ecological fingerprint vectors, generate candidate control actions, and generate corresponding counterfactual ecological future paths. The strategy form construction and deformation module is used to obtain the updated strategy form space based on the counterfactual ecological future path; The strategy template generation and execution module is used to generate control strategy templates based on the updated strategy morphology space and distribute them to the oxygenation, filtration, lighting, feeding and water exchange equipment for execution. The feedback adaptive update module is used to collect ecological feedback data and update the ecological fingerprint vector, spatiotemporal dynamic causal graph, counterfactual ecological future path and strategy morphology space, forming an adaptive update cycle.

Images