An intelligent monitoring and analyzing system and method for behavior of aquatic animals based on edge-cloud cooperation

Abstract

The application discloses an intelligent monitoring and analyzing system and method for aquatic animal behavior based on end-edge-cloud cooperation, which comprises: a terminal sensing layer, configured with POE single-wire transmission power supply and collection equipment, used for collecting original video streams of aquatic animals and executing physical control instructions; an edge analysis layer, deployed with a deep learning inference engine, used for local frame-by-frame analysis of the video streams, judgment of the behavior state of the aquatic animals, and uploading of light-weight analysis results to a cloud platform layer only when a state change is detected or at preset periodic time points; and the cloud platform layer, integrated with a large language model and a domain knowledge base, used for converting structured text into a natural language report containing data trend interpretation, environmental abnormality attribution and breeding intervention suggestions. The application greatly reduces transmission bandwidth consumption through light-weight processing of edge-side data, realizes professional semantic mining of monitoring data by using a cloud-side large model, and constructs an automatic closed loop from video collection to intelligent decision-making.

Description

Technical Field

[0001] This invention relates to the field of aquatic animal monitoring and analysis technology, specifically to an intelligent monitoring and analysis system and method for aquatic animal behavior based on edge-cloud collaboration. Background Technology

[0002] Aquatic animal behavior studies are a crucial tool for assessing ecosystem health, aquaculture efficiency, and the impact of environmental pollution. Aquatic animals (such as fish and crayfish) are extremely sensitive to changes in the aquatic environment; their feeding, swimming, and stress behaviors directly reflect water quality conditions such as dissolved oxygen, heavy metal pollution, or sudden temperature changes. In precision aquaculture and aquatic ecosystem protection, real-time and accurate monitoring and interpretation of these behavioral characteristics are of great significance for preventing disease outbreaks, reducing aquaculture risks, and making ecological restoration decisions.

[0003] Currently, aquatic animal monitoring technology is struggling to meet the demands for refined and real-time monitoring, primarily due to four major bottlenecks: 1. Network and cost limitations: Traditional methods rely on uploading massive amounts of high-definition video to the cloud, but outdoor waters often suffer from poor network coverage and high bandwidth costs, resulting in significant transmission delays and hindering real-time response. 2. Insufficient data analysis depth: Existing edge devices typically only support simple motion detection and cannot run complex attitude estimation models, making it difficult to extract detailed kinematic parameters. 3. Low system efficiency: Existing solutions often involve simple data forwarding, lacking intelligent task scheduling and resource allocation mechanisms between devices, and failing to offload tasks to more appropriate levels based on real-time needs. 4. High barrier to result interpretation: Existing systems mostly output raw coordinates or charts, lacking in-depth semantic mining of the data. Non-professionals find it difficult to directly translate velocity curves into intuitive decision-making evidence such as water quality and oxygen deficiency, and manual report writing is required, which is not only inefficient but also prone to data interpretation bias.

[0004] Therefore, there is an urgent need for a monitoring and analysis system and method based on an edge-cloud collaborative architecture that can systematically solve the above problems, achieve efficient, intelligent, and low-cost aquatic animal monitoring, and provide practical scientific tools for aquatic animal breeding management and ecological research. Summary of the Invention

[0005] In view of this, the present invention provides an intelligent monitoring and analysis system and method for aquatic animal behavior based on edge-cloud collaboration, in order to solve the technical problems of large outdoor data transmission delay, insufficient edge analysis depth, and lack of intuitive and professional decision-making suggestions in traditional aquatic animal monitoring technologies.

[0006] On the one hand, the present invention provides an intelligent monitoring and analysis system for aquatic animal behavior based on edge-cloud collaboration, comprising a terminal sensing layer, an edge analysis layer and a cloud platform layer connected in sequence; The terminal sensing layer, deployed at the monitoring site, is equipped with POE single-wire transmission power supply and data acquisition equipment, used to collect raw video streams of aquatic animals and execute control commands; The edge analysis layer is equipped with a deep learning inference engine to perform localized frame-by-frame analysis of the original video stream, extract the temporal coordinates of key points on the body of aquatic animals and calculate kinematic parameters, determine the behavioral state of aquatic animals based on kinematic parameters, and upload structured text data containing kinematic parameters and behavioral state labels to the cloud platform layer only when a change in behavioral state is detected or at preset periodic time points. The cloud platform layer is equipped with a natural language processing module, which receives the structured text data and uses a pre-built domain knowledge base and prompt word templates to map the structured text data into a natural language report containing environmental problem attribution and decision-making suggestions.

[0007] Furthermore, the terminal perception layer includes: a visual perception unit, a PoE switch, and an instruction execution unit; A visual sensing unit is used to collect behavioral data of aquatic animals at the monitoring site; A PoE switch is used to simultaneously power and transmit data to a visual sensing unit using a single network cable. The instruction execution unit is used to adjust the setting parameters of the visual perception unit in real time according to the control instructions of the edge analysis layer.

[0008] Furthermore, the edge computing device of the edge analysis layer is equipped with a parallel computing acceleration unit to support real-time inference of the deep learning model; the edge analysis layer includes a pose estimation unit and a behavior analysis unit. The attitude estimation unit is used to identify the coordinates of specific anatomical key points of aquatic animals at different time points, and to align the temporal coordinate data of the key points with environmental parameters using a unified timestamp. The behavior analysis module is equipped with a behavior analysis algorithm unit, which is used to calculate kinematic parameters based on the temporal coordinate data of key points, and to determine the behavioral state of aquatic animals based on the kinematic parameters.

[0009] Furthermore, the attitude estimation unit includes an inference model unit and a time-stamping synchronization unit; The inference model unit adopts a pose estimation model based on deep learning; The time-stamp synchronization unit is used to align the extracted key point time-series coordinate data with environmental parameters using a unified timestamp.

[0010] Furthermore, the step of calculating kinematic parameters based on the temporal coordinate data of key points and determining the behavioral state of aquatic animals based on the kinematic parameters includes: Calculate the instantaneous velocity V based on the time-series coordinate data of key points. iThe calculation formula is: , Among them, (Xi, Yi) and ( , ) represents the coordinates of the same key point in adjacent frames, and ΔT is the length of the interval between the two frames; The methods for classifying the behavioral states of aquatic organisms based on preset speed thresholds include: When V0>V i When V1 is greater than V1, it is determined to be a state of stress; when V i When V<V2, it is considered to be in a static state; when V2≤V i A value ≤ V1 is considered normal. Where V2 < V1 < V0, V 0、 V1 and V2 are preset speed thresholds.

[0011] Furthermore, the cloud platform layer includes a macro trend analysis engine, a report generation module, and a model training module; The macro trend analysis engine is used to aggregate, mine, and analyze data across time and space scales based on the lightweight analysis results uploaded by the edge analysis layer, and to identify macro trends in population behavior patterns and environmental relevance. The report generation module is used to construct dynamic prompt words based on the data uploaded by the edge analysis layer through semantic mapping and context embedding technology, and call a large language model to generate a structured natural language report. The model training module is used to retrain the deep learning model using historical monitoring data, remove redundant parameters through model distillation, compress the model volume, and send it to the edge analysis layer to update the deep learning pose estimation module.

[0012] Furthermore, dynamic prompt words are constructed through semantic mapping and context embedding techniques to drive large-scale language models to generate structured natural language reports, including: The structured text data uploaded by the edge analysis layer is transformed into intermediate semantic tags for natural language description through preset mapping rules; Based on the intermediate semantic tags, the associated environmental factors are retrieved from the pre-set behavior-environment attribution knowledge graph and used as expert knowledge constraints. Define system roles, combine system role definitions with intermediate semantic tags, expert knowledge constraints and preset output templates to assemble dynamic prompt words, and input them into a large language model; A large-scale language model is used to generate structured text containing data interpretation, anomaly attribution, and intervention suggestions based on the dynamic prompts.

[0013] Furthermore, the structured natural language report also includes early warning information based on trend prediction, including: The deviation between the current monitoring period and historical data for the same period is calculated. When the deviation shows that the activity of aquatic animals is showing a continuous downward trend, the generated report automatically inserts a water quality deterioration warning and sampling verification suggestion for the future period.

[0014] Furthermore, the system is also configured with a cloud-edge collaborative reverse control mechanism, including: When the report generated by the cloud platform layer determines that the monitored object is in a high-risk state, it automatically generates a hierarchical control instruction and sends it to the edge analysis layer. The hierarchical control instruction is automatically generated based on the decision suggestions generated by the cloud-based big model. The edge analysis layer responds to the hierarchical control commands and adjusts the setting parameters of the visual perception unit in real time.

[0015] On the other hand, the present invention provides an intelligent monitoring and analysis method for aquatic animal behavior based on edge-cloud collaboration, implemented using the system described above, including: The raw video stream is collected in an outdoor water environment through the terminal perception layer and transmitted to the edge analysis layer; In the edge analysis layer, a deep learning model is used to identify key points of aquatic animals' bodies in the video frame by frame, and output time-series coordinate data. The kinematic parameters of key points are calculated in the edge analysis layer, and the behavior state is classified based on the preset threshold to generate analysis results containing only parameters and state labels, while the original video data is stored locally. The analysis results are uploaded to the cloud platform layer, which then calls a large language model to automatically generate a monitoring and analysis report based on the results.

[0016] Compared to existing technologies, the intelligent monitoring and analysis system and method for aquatic animal behavior based on edge-cloud collaboration proposed in this invention achieves efficient, intelligent, and low-cost outdoor aquatic animal monitoring through an edge-cloud collaborative architecture. It boasts significant advantages such as low bandwidth requirements, real-time response, in-depth analysis dimensions, and intuitive output results. Specifically, it includes: (1) Edge computing-driven resource-intensive and real-time interactive architecture: This system adopts a deep decoupling strategy on the edge side, achieving an exponential reduction in communication bandwidth requirements by only transmitting structured processing results back, ensuring highly reliable transmission in extreme network environments in remote waters. At the hardware level, relying on PoE power supply technology and ultra-low power circuit design, it supports a solar energy replenishment system, building true outdoor unattended self-sufficiency. Crucially, the near-source parallel computing on the edge side compresses the identification delay of key stress events to the millisecond level, giving the system crucial pre-intervention time.

[0017] (2) Multidimensional behavior quantification system based on posture estimation and kinematics: The system integrates a deep learning posture estimation model to extract the temporal coordinates of key anatomical sites of aquatic animals. By constructing a kinematic coupling analysis model and combining it with a three-level dynamic velocity threshold algorithm, the system can accurately map fragmented visual pixel features into behavioral semantic indicators such as instantaneous velocity and tail wagging frequency. This deep transformation path of "video-feature-semantics" eliminates the human observation error of traditional monitoring and greatly enhances the scientific research-grade confidence of monitoring data.

[0018] (3) Higher professionalism and accuracy of decision generation: This system abandons simple model calling and innovatively proposes a "dynamic prompt word construction method based on knowledge graph". By injecting aquatic expert knowledge as contextual constraints into the large model, it effectively solves the defect of general large models that are prone to "illusion" in vertical fields, and ensures that the generated environmental attribution (such as "high static state is due to insufficient dissolved oxygen") has a scientific basis.

[0019] (4) Achieved proactive closed-loop collaboration of "cloud-edge-device": Unlike traditional one-way monitoring, this system has constructed a reverse control mechanism. After the cloud brain detects the risk, it can directly drive the edge nerve endings to adjust the acquisition strategy (such as automatically increasing the frame rate for verification), transforming the originally lagging response that required manual intervention into millisecond-level automated intervention, which significantly reduces the risk of aquaculture. Attached Figure Description

[0020] Figure 1 A schematic diagram of the intelligent monitoring and analysis system for aquatic animal behavior based on edge-cloud collaboration provided by the present invention; Figure 2 The system architecture diagram provided for practical application of this invention; Figure 3 This invention provides a schematic diagram of the core module and workflow of the edge analysis layer. Figure 4 The cloud platform report generation logic diagram provided by this invention; Figure 5 This is a flowchart illustrating the intelligent monitoring and analysis method for aquatic animal behavior based on edge-cloud collaboration provided by the present invention. Detailed Implementation

[0021] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0022] Please see Figure 1This embodiment provides an intelligent monitoring and analysis system 100 for aquatic animal behavior based on edge-cloud collaboration, including a terminal perception layer 101, an edge analysis layer 102 and a cloud platform layer 103 connected in sequence. Terminal sensing layer 101, deployed at the monitoring site, is equipped with POE single-wire transmission power supply and data acquisition equipment, used to collect raw video streams of aquatic animals and execute control commands; Edge analysis layer 102 is equipped with a deep learning inference engine, which is used to perform localized frame-by-frame analysis on the original video stream, extract the temporal coordinates of key points on the body of aquatic animals and calculate kinematic parameters, determine the behavioral state of aquatic animals based on kinematic parameters, and upload structured text data containing kinematic parameters and behavioral state labels to the cloud platform layer only when a change in behavioral state is detected or at preset periodic time points. The cloud platform layer 103 is used to receive lightweight analysis results from multiple time periods and use a large language model to generate structured reports containing environmental problem attributions and decision-making recommendations based on the analysis results.

[0023] The system provided in this embodiment adopts a three-tier architecture of edge-cloud, enabling complex deep learning inference and behavior quantification to be completed at the edge. Only the processed lightweight data (such as velocity values ​​and status labels) is uploaded to the cloud, effectively reducing bandwidth consumption and improving real-time response speed. Utilizing deep learning pose estimation technology, the system extracts the temporal coordinates of key points in aquatic animals. Combined with professional kinematic formulas and multi-level velocity thresholds, it can accurately quantify the instantaneous velocity and behavioral state of animals. Simultaneously, it integrates a large-scale language model to automatically transform abstract monitoring data into structured reports containing environmental attribution and intervention suggestions, lowering the technical application threshold. This system achieves end-to-end data mining from "video → coordinates → behavioral semantics → environmental association," outputting directly applicable decision-making suggestions.

[0024] In a preferred embodiment, the terminal perception layer includes: a visual perception unit, a PoE switch, and an instruction execution unit; A visual sensing unit is used to collect behavioral data of aquatic animals at the monitoring site; A PoE switch can simultaneously power the acquisition device and transmit data via a single network cable; The instruction execution unit is used to adjust the monitoring angle, focal length, and aperture parameters of the visual perception unit in real time according to the control instructions of the edge analysis layer.

[0025] As a specific example, such as Figure 2 As shown, Figure 2This is a schematic diagram of the system architecture for practical applications. In practical applications, the visual perception unit uses a Hikvision DS-2CD3T47SWDV3-L series POE camera, or hardware devices with equivalent sampling capabilities (4 megapixels, 2.8mm focal length, IP67 waterproof, infrared night vision), which is adaptable to underwater and humid environments, supports high resolution of 1080P@30fps to ensure that animal behavior details are clearly discernible, and reduces the system's dependence on network bandwidth and hardware power consumption, supporting long-term (≥6 months) unattended monitoring in outdoor waters.

[0026] The PoE switch uses Hikvision DS-3E0510P-E (8 PoE ports), which can simultaneously power the camera (30W / port) and transmit data with a single network cable, eliminating the need for separate power cables and simplifying outdoor cabling.

[0027] The execution unit is used to receive instructions from the edge analysis layer and supports 360° rotation of the gimbal (to adjust the monitoring range), electric focus adjustment (to focus on a single animal), and aperture control (to adapt to changes in day and night lighting).

[0028] In some embodiments, the execution unit further includes a field oxygen control or alarm device for performing oxygen control and alarm operations at the monitoring site in response to control commands from the edge analysis layer.

[0029] In a preferred embodiment, the edge computing device of the edge analysis layer is equipped with a GPU acceleration unit to support real-time inference of the deep learning model; the edge analysis layer includes a pose estimation unit and a behavior analysis unit. The attitude estimation unit is used to identify the coordinates of specific anatomical key points of aquatic animals at different time points, and to align the temporal coordinate data of the key points with environmental parameters using a unified timestamp. The behavior analysis module is equipped with a BioGrid Analyzer algorithm unit, which is used to calculate kinematic parameters based on the temporal coordinate data of key points, and to determine the behavioral state of aquatic animals based on the kinematic parameters.

[0030] Please see Figure 2 As the "nerve center" of the system, the edge analytics layer is responsible for video parsing, behavior analysis, and event recognition, enabling lightweight data processing. Therefore, edge computing devices need to be equipped with GPU acceleration units. In specific applications, a high-performance AI computing module (NVIDIA 4060ti 16G or other CUDA-enabled graphics cards are recommended) is built-in to run the DeepLabCut model.

[0031] In a preferred embodiment, the attitude estimation unit includes an inference model unit and a time-scale synchronization unit; The inference model unit adopts a customized model trained based on the DeepLabCut framework; The time-stamp synchronization unit is used to align the extracted key point time-series coordinate data with environmental parameters using a unified timestamp.

[0032] As a specific implementation, the inference model unit runs a custom deep learning model trained using the DeepLabCut open-source toolkit. This module receives raw video streams from a camera and manually annotates approximately 100 clear crayfish images for training the DeepLabCut model. It detects seven key points of the crayfish: mouth, head, telson, left and right claws, and claws. During inference, it outputs the two-dimensional coordinates of the key points frame by frame, using ResNet-50 or higher (ResNet-101, ResNet-152) neural networks. The model is trained for 50,000 epochs with a learning rate of 0.001, and the measured inference accuracy reaches 85% or higher.

[0033] In a preferred embodiment, the step of calculating kinematic parameters based on the temporal coordinate data of key points and determining the behavioral state of aquatic animals based on the kinematic parameters includes: Calculate the instantaneous velocity V based on the time-series coordinate data of key points. i The calculation formula is: , Among them, (Xi, Yi) and ( , ) represents the coordinates of the same key point in adjacent frames, and ΔT is the length of the interval between the two frames; The methods for classifying the behavioral states of aquatic organisms based on preset speed thresholds include: When V0>V i When V1 is greater than V1, it is determined to be a state of stress; when V i When V<V2, it is considered to be in a static state; when V2≤V i A value ≤ V1 is considered normal. Where V2 < V1 < V0, V 0、 V1 and V2 are preset speed thresholds.

[0034] As a specific implementation, the BioGrid Analyzer parameters in the behavior analysis module are set as follows: based on the camera's frame rate of 30fps, the time interval ΔT=0.033s is set; based on crayfish behavioral experimental data, the speed thresholds are set as follows: V0=15cm / s (to avoid extreme noise), V1=5cm / s (critical value for stress state), and V2=0.1cm / s (critical value for stationary state).

[0035] As a specific embodiment, the workflow of the behavior analysis module includes: First, align the coordinate data with the environmental parameters to create a three-dimensional dataset of "time-behavior-environment"; Secondly, it receives the temporal keypoint coordinate data generated by the DeepLabCut module, and calculates the kinematic parameters based on the displacement of one or more keypoints.

[0036] Finally, the target state is classified in real time and a state label is generated by using built-in configurable thresholds (based on the target animal's behavioral experimental data, with preset stress threshold V1 and rest threshold V2).

[0037] In some embodiments, the edge computing device uses the UGREEN DH2100 (8TB dual-bay) as the local storage and management unit, which distributes the storage of original video (retained for 7 days), coordinate data (retained for 30 days), and behavior analysis results (retained permanently), and supports fast retrieval by time / event type.

[0038] like Figure 3 As shown, Figure 3 The architecture diagram of the edge analytics layer in a practical application is shown.

[0039] In a preferred embodiment, the cloud platform layer includes a macro trend analysis engine, a report generation module, and a model training module; The macro trend analysis engine is used to aggregate, mine, and analyze data across time and space scales based on the lightweight analysis results uploaded by the edge analysis layer, and to identify macro trends in population behavior patterns and environmental relevance. The report generation module is used to call a large language model and, based on a preset prompt word template containing aquatic animal behavioral terms, perform semantic analysis and reasoning on the lightweight analysis result data uploaded by the edge analysis layer to generate a structured natural language report containing data trend interpretation, environmental anomaly attribution, and aquaculture intervention suggestions. The model training module is used to retrain the deep learning model using historical monitoring data, remove redundant parameters through model distillation, compress the model volume, and send it to the edge analysis layer to update the deep learning pose estimation module.

[0040] As a specific embodiment, the report generation module is used to receive structured data (such as speed curves, status statistics, and event logs) and environmental data uploaded from the edge layer after being processed by BioGridAnalyzer. By calling a large language model API (such as DeepSeek) and combining it with predefined prompt words, it automatically generates a multi-page professional report containing package summary, methods, calculation results, data charts, professional conclusions, and practical suggestions.

[0041] In some embodiments, the model training module is built on the TensorFlow / PyTorch framework, and uses historical data from multiple monitoring points (≥100,000 frames of labeled images) to optimize the DeepLabCut model. The model size is compressed from 500MB to 50MB through model distillation (removing redundant parameters) to adapt to the computing power of edge devices.

[0042] In a preferred embodiment, semantic analysis and reasoning are performed on the lightweight analysis results data to generate a structured natural language report containing data trend interpretation, environmental anomaly attribution, and aquaculture intervention recommendations, including: Input system role definitions into a large language model; The key point average velocity, static state time ratio, stress state time ratio, and abnormal event logs in the lightweight analysis results are converted into natural language descriptive text and embedded as contextual information prompts. Preset analysis rules are embedded in the prompts, and these rules include the correspondence between different behavioral states and specific environmental factors. The model generates text according to a four-part structure: monitoring purpose, method summary, results analysis, and conclusions and recommendations.

[0043] In a preferred embodiment, the preset analysis rule constraints include: If the input data shows that the proportion of time spent in a static state exceeds a preset safety threshold, the report generation module configures the model to attribute the anomaly to insufficient dissolved oxygen or heavy metal pollution, and generates a suggestion to turn on the aerator or test the content of heavy metal pollutants in the water. If the input data shows that the instantaneous speed fluctuates drastically at high frequency and reaches the stress state threshold, the report generation module configures the model to attribute the anomaly to pesticide contamination or sudden changes in water temperature, and generates suggestions to investigate drainage from surrounding farmland or detect the rate of change in water temperature.

[0044] As a specific example, the report template is configured with a fixed structure: "Monitoring purpose (assessing crayfish activity and water quality status) - Method summary (edge-cloud collaborative monitoring, DeepLabCut + BioGrid Analyzer analysis) - Result analysis (including velocity curves and status percentages) - Conclusions and recommendations"; Professional associations are: preset association rules such as "surge in stress state → suspected pesticide pollution / sudden temperature change" and "surge in quiescent state → suspected insufficient dissolved oxygen / heavy metal pollution".

[0045] In this preferred embodiment, the report generation module does not directly pass data to a large language model, but instead employs a technical architecture that combines "dynamic prompt word engineering" with "knowledge base retrieval enhancement." The specific execution flow is as follows: Semantic mapping of structured data: The report generation module has a semantic transformation unit for receiving lightweight JSON-formatted data (including average speed V) uploaded by the edge analysis layer. avg The proportion of static time T static Stress frequency F stress (etc.). This unit uses a pre-defined mapping rule table to transform numerical data into semantic descriptions of intermediate states. For example: When T is detected static When the activity level is greater than 30% and its month-on-month growth rate is greater than 15%, the system generates the semantic label "Status: Activity level has decreased significantly; Trend: Deterioration".

[0046] When the maximum speed V is detected max Exceeding the stress threshold V stress threshold When the frequency of the stress exceeds 5 times within 10 minutes, the semantic label "State: Intermittent severe stress; Feature: Paroxysmal" is generated.

[0047] Domain knowledge context embedding: The system has a built-in "aquatic animal behavior-environmental attribution knowledge graph". When constructing prompt words, the system automatically retrieves potential related nodes in the knowledge graph based on the aforementioned semantic tags. For example, for "intermittent severe stress", the system can retrieve related environmental factors such as "organophosphorus pesticide residues" or "sudden pH changes", and inject these professional attribution logics as "expert context constraints" into the System Message of the prompt word. This step effectively limits the "illusion" risk of large language models, ensuring that they are strictly based on aquatic behavioral logic for reasoning.

[0048] Structured Report Generation: The report generation module combines "role definition + semantic tags of monitoring data + retrieved expert attribution logic + standardized output template" into complete prompt words, which are then input into a large-scale language model. The model outputs structured text containing the following four dimensions: (1) Data trend interpretation: describing numerical changes in natural language; (2) Anomaly attribution diagnosis: probabilistic inference based on knowledge base; (3) Risk warning: Based on trend forecasts for the same period in history; (4) Recommendations for aquaculture intervention: such as “turn on aerator No. 2” or “take a sample to test for nitrite.”

[0049] like Figure 4 As shown, Figure 4 It shows the architecture diagram of the cloud platform layer in a practical application.

[0050] In a preferred embodiment, the structured natural language report further includes trend-prediction-based early warning information, including: The deviation between the current monitoring period and historical data for the same period is calculated. When the deviation shows that the activity of aquatic animals is showing a continuous downward trend, the generated report automatically inserts a water quality deterioration warning and sampling verification suggestion for the future period.

[0051] like Figure 5 As shown, an intelligent monitoring and analysis method for aquatic animal behavior based on edge-cloud collaboration is implemented using the system described above, including: Step S101: Collect the raw video stream in the outdoor water environment through the terminal perception layer and transmit it to the edge analysis layer; Step S102: Using a deep learning model, the edge analysis layer is used to identify key points of aquatic animals' bodies frame by frame in the video, and output temporal coordinate data. Step S103: Calculate the kinematic parameters of key points through the edge analysis layer, classify the behavior state based on the preset threshold, generate analysis results containing only parameters and state labels, and save the original video data locally; Step S104: Upload the analysis results to the cloud platform layer, and use the cloud platform layer to call a large language model to automatically generate a monitoring and analysis report based on the analysis results.

[0052] In some embodiments, the cloud platform can use the accumulated result data to retrain or optimize the DeepLabCut model and distribute it to the edge layer for updates, thereby updating the deep learning pose estimation module.

[0053] The method in this embodiment achieves resource optimization and deep intelligent decision-making in a collaborative "edge-cloud" architecture. It utilizes the deep computing capabilities of the edge to perform high-precision pose recognition and lightweight data processing of video, transforming massive amounts of raw video into semantic data containing only parameters and labels. This significantly reduces network bandwidth consumption by over 90%, solving the real-time transmission challenge in outdoor weak network environments. Simultaneously, through secondary mining of the lightweight data using a large cloud-based model, it achieves an automated transition from data to professional decision reports, allowing users to directly obtain intuitive feedback including environmental attribution and intervention suggestions without requiring a biological background.

[0054] 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 changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A smart monitoring and analysis system for aquatic animal behavior based on edge-cloud collaboration, characterized in that, It includes the terminal perception layer, the edge analysis layer, and the cloud platform layer, which are connected in sequence; The terminal sensing layer, deployed at the monitoring site, is equipped with POE single-wire transmission power supply and data acquisition equipment, used to collect raw video streams of aquatic animals and execute control commands; The edge analysis layer is equipped with a deep learning inference engine to perform localized frame-by-frame analysis of the original video stream, extract the temporal coordinates of key points on the body of aquatic animals and calculate kinematic parameters, determine the behavioral state of aquatic animals based on kinematic parameters, and upload structured text data containing kinematic parameters and behavioral state labels to the cloud platform layer only when a change in behavioral state is detected or at preset periodic time points. The cloud platform layer is equipped with a natural language processing module, which receives the structured text data and uses a pre-built domain knowledge base and prompt word templates to map the structured text data into a natural language report containing environmental problem attribution and decision-making suggestions.

2. The system according to claim 1, characterized in that, The terminal perception layer includes: a visual perception unit, a PoE switch, and an instruction execution unit; A visual sensing unit is used to collect behavioral data of aquatic animals at the monitoring site; A PoE switch is used to simultaneously power and transmit data to a visual sensing unit using a single network cable. The instruction execution unit is used to adjust the setting parameters of the visual perception unit in real time according to the control instructions of the edge analysis layer.

3. The system according to claim 1, characterized in that, The edge computing device of the edge analysis layer is equipped with a parallel computing acceleration unit to support real-time inference of the deep learning model; the edge analysis layer includes a pose estimation unit and a behavior analysis unit. The attitude estimation unit is used to identify the coordinates of specific anatomical key points of aquatic animals at different time points, and to align the temporal coordinate data of the key points with environmental parameters using a unified timestamp. The behavior analysis module is equipped with a behavior analysis algorithm unit, which is used to calculate kinematic parameters based on the temporal coordinate data of key points, and to determine the behavioral state of aquatic animals based on the kinematic parameters.

4. The system according to claim 3, characterized in that, The attitude estimation unit includes an inference model unit and a time-scale synchronization unit; The inference model unit adopts a pose estimation model based on deep learning; The time-stamp synchronization unit is used to align the extracted key point time-series coordinate data with environmental parameters using a unified timestamp.

5. The system according to claim 3, characterized in that, The calculation of kinematic parameters based on the temporal coordinate data of key points, and the determination of the behavioral state of aquatic animals based on the kinematic parameters, includes: Calculate the instantaneous velocity V based on the time-series coordinate data of key points. i The calculation formula is: ， Among them, (Xi, Yi) and ( , ) represents the coordinates of the same key point in adjacent frames, and ΔT is the length of the interval between the two frames; The methods for classifying the behavioral states of aquatic organisms based on preset speed thresholds include: When V0>V i When V1 is greater than V1, it is determined to be a state of stress; when V i When V<V2, it is considered to be in a static state; when V2≤V i A value ≤ V1 is considered a normal state; Where V2 < V1 < V0, V 0、 V1 and V2 are preset speed thresholds.

6. The system according to claim 1, characterized in that, The cloud platform layer includes a macro trend analysis engine, a report generation module, and a model training module; The macro trend analysis engine is used to aggregate, mine, and analyze data across time and space scales based on the lightweight analysis results uploaded by the edge analysis layer, and to identify macro trends in population behavior patterns and environmental relevance. The report generation module is used to construct dynamic prompt words based on the data uploaded by the edge analysis layer through semantic mapping and context embedding technology, and call a large language model to generate a structured natural language report. The model training module is used to retrain the deep learning model using historical monitoring data, remove redundant parameters through model distillation, compress the model volume, and send it to the edge analysis layer to update the deep learning pose estimation module.

7. The system according to claim 6, characterized in that, Dynamic prompts are constructed using semantic mapping and context embedding techniques to drive large-scale language models to generate structured natural language reports, including: The structured text data uploaded by the edge analysis layer is transformed into intermediate semantic tags for natural language description through preset mapping rules; Based on the intermediate semantic tags, the associated environmental factors are retrieved from the pre-set behavior-environment attribution knowledge graph and used as expert knowledge constraints. Define system roles, combine system role definitions with intermediate semantic tags, expert knowledge constraints and preset output templates to assemble dynamic prompt words, and input them into a large language model; A large-scale language model is used to generate structured text containing data interpretation, anomaly attribution, and intervention suggestions based on the dynamic prompts.

8. The system according to claim 7, characterized in that, The structured natural language report also includes trend-based early warning information, including: The deviation between the current monitoring period and historical data for the same period is calculated. When the deviation shows that the activity of aquatic animals is showing a continuous downward trend, the generated report automatically inserts a water quality deterioration warning and sampling verification suggestion for the future period.

9. The system according to claim 2, characterized in that, The system is also equipped with a cloud-edge collaborative reverse control mechanism, including: When the report generated by the cloud platform layer determines that the monitored object is in a high-risk state, it automatically generates a hierarchical control instruction and sends it to the edge analysis layer. The hierarchical control instruction is automatically generated based on the decision suggestions generated by the cloud-based big model. The edge analysis layer responds to the hierarchical control commands and adjusts the setting parameters of the visual perception unit in real time.

10. A method for intelligent monitoring and analysis of aquatic animal behavior based on edge-cloud collaboration, characterized in that, The system is implemented using any one of claims 1-9, comprising: The raw video stream is collected in an outdoor water environment through the terminal perception layer and transmitted to the edge analysis layer; In the edge analysis layer, a deep learning model is used to identify key points of aquatic animals' bodies in the video frame by frame, and output time-series coordinate data. The kinematic parameters of key points are calculated in the edge analysis layer, and the behavior state is classified based on a preset threshold to generate analysis results containing only parameters and state labels, while the original video data is stored locally. The analysis results are uploaded to the cloud platform layer, which then calls a large language model to automatically generate a monitoring and analysis report based on the results.

Images