A marine ranching intelligent management and control system and method based on a multi-modal AI algorithm

Abstract

The application discloses a kind of based on multi-modal AI algorithm's ocean ranching intelligent management and control system and method, belong to wisdom breeding field.System includes perception layer, algorithm layer, decision-making layer, execution layer and feedback optimization layer, perception layer collects multi-source data, algorithm layer realizes data intelligent analysis by five big AI algorithms (water environment prediction, fry evaluation, accurate feeding, disease early warning, unmanned aerial vehicle path planning), decision-making layer generates standardization instruction, execution layer completes feeding, water quality regulation, fry sorting operation, feedback optimization layer updates model parameter and forms closed loop.The application solves the problems of traditional ocean ranching feed waste, fry sorting difficulty, disease early warning lag, tail water pollution and high artificial dependence, realizes that feed utilization rate is improved by 18%, fry survival rate is improved by 25%, disease loss rate is reduced by 45%, artificial cost is reduced by 80%, and is suitable for large-scale industrial ocean ranching.

Description

Technical Field

[0001] This invention relates to the intersection of marine ranching technology and artificial intelligence. Specifically, it is an intelligent management and control system and method for marine ranches based on multimodal AI algorithms. It is applicable to fish fry quality assessment, precise feed feeding, fish disease early warning, aquaculture water quality control and unmanned operation management in factory-scale marine ranches, and belongs to the category of integrated application of smart agriculture and artificial intelligence technology. Background Technology

[0002] The marine ranching industry currently faces five core technological challenges that hinder its large-scale and green development:

[0003] 1. Conflict between feed utilization and water pollution: Traditional feeding relies on human experience, and the amount of feed given does not match the feeding needs of fish. The uneaten feed rate is as high as 15%-20%. Uneaten feed and fish excrement lead to excessive ammonia nitrogen and nitrite in the aquaculture water, causing intestinal diseases in fish and increasing the mortality rate of aquaculture.

[0004] 2. Lack of standardized fish fry quality sorting: The physical condition of fish fry is affected by many factors such as the quality of fertilized eggs, hatching water temperature, dissolved oxygen, etc. High-quality fish fry are mixed with low-quality fish fry. Manual sorting is inefficient (less than 5,000 fish per person per day) and has a large error (sorting accuracy rate <60%). The survival rate of low-quality fish fry is less than 30%, which reduces the overall aquaculture income.

[0005] 3. Delayed early warning of fish diseases: Vibrio disease, enteritis and other diseases are prone to outbreak during the high temperature season (water temperature > 28℃). Traditional monitoring relies on manual observation of the fish population. By the time the symptoms are discovered, the disease has already spread, resulting in a loss rate of 40%-60% in aquaculture, and in severe cases, no harvest at all.

[0006] 4. Heavy environmental pressure from aquaculture wastewater: Aquaculture wastewater contains pollutants such as uneaten feed, fish medicine residues, and excrement. Direct discharge leads to eutrophication of surrounding sea areas, and the environmental compliance rate is less than 50%. Some farms have been forced to stop production due to non-compliance of wastewater standards.

[0007] 5. High dependence on manual labor and lack of standardization: Large-scale farms require 10-20 people per 10,000 fish. High-intensity operations such as feeding, water quality monitoring, and fry sorting rely on manual operation, resulting in a high error rate (e.g., feeding amount error ±15%). The lack of unified digital management standards restricts the expansion of farm production.

[0008] In existing technologies, some solutions attempt to alleviate the aforementioned pain points through single devices. For example, patent CN202310221567.8 proposes a "marine ranch water quality monitoring device," which can only collect water quality data and cannot correlate it with fish population status for comprehensive decision-making. Patent CN202211456890.3 proposes a "drone feeding system," but lacks a dynamic feeding strategy supported by AI algorithms, and feed waste still exists. Existing technologies generally suffer from defects such as a disconnect between the "perception-decision-execution" process, insufficient multi-source data fusion capabilities, and limited application of AI algorithms, failing to form a fully intelligent management and control system. There is an urgent need for a solution that integrates multimodal AI algorithms and covers the entire aquaculture process. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention provides an intelligent management and control system and method for marine ranches based on multimodal AI algorithms. By constructing a closed-loop system of "perception-analysis-decision-execution-feedback", it solves problems such as feed waste, difficulty in sorting fish fry, delayed disease early warning, tailwater pollution, and high dependence on manual labor, thereby achieving high-quality aquaculture in marine ranches.

[0010] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0011] On the one hand, the present invention provides an intelligent management and control system for marine ranches based on multimodal AI algorithms, including a perception layer, an algorithm layer, a decision-making layer, an execution layer and a feedback optimization layer, with each layer interacting with each other via industrial Ethernet and / or 5G networks.

[0012] The specific structure is as follows:

[0013] 1. Sensing Layer: Employs RS485 communication protocol, with a data acquisition frequency of 10 minutes / time. Monitored parameters include:

[0014] Water temperature (measurement range 0-35℃, accuracy ±0.2℃);

[0015] Salinity (measurement range 0-40‰, accuracy ±0.1‰);

[0016] Dissolved oxygen (measurement range 0-20 mg / L, accuracy ±0.1 mg / L);

[0017] pH (measurement range 6.0-9.0, accuracy ±0.1);

[0018] Ammonia nitrogen (measurement range 0-5 mg / L, accuracy ±0.05 mg / L);

[0019] Nitrite (measurement range 0-1 mg / L, accuracy ±0.01 mg / L);

[0020] Underwater HD camera: 1920×1080 resolution, 25fps frame rate, IP68 waterproof rating, supports low-light shooting (minimum illumination 0.01 lux), and captures images and video data of fish schools;

[0021] Meteorological sensor: Collects air pressure (measurement range 800-1100 hPa, accuracy ±1 hPa);

[0022] Illumination (measurement range 0-200000 lux, accuracy ±5%)

[0023] Wind speed (measurement range 0-30m / s, accuracy ±0.3m / s);

[0024] Fish fry sorting sensor: Combining infrared ranging and image recognition, it collects the body length of fish fry (measurement range 1-10cm, accuracy ±0.1cm).

[0025] Activity data such as swimming speed (measurement range 0-5cm / s, accuracy ±0.05cm / s).

[0026] 2. Algorithm Layer: Five core AI algorithms are deployed, and the algorithm running environment is a GPU server (configured with NVIDIA A10, 24GB of video memory), developed using the TensorFlow 2.8 framework, with data processing latency <100ms;

[0027] Multi-factor spatiotemporal prediction model for water environment: a time series model constructed based on LSTM + attention mechanism;

[0028] Fish fry / adult fish intelligent assessment model: constructed based on YOLOv8+ fish body quality scoring neural network;

[0029] Dynamic Precision Feeding Model: Built Based on DQN Reinforcement Learning;

[0030] Early warning model for fish diseases: constructed based on a multimodal fusion diagnostic model;

[0031] Intelligent feeding path planning model for drones: constructed based on AI trajectory planning and dynamic grid method.

[0032] 3. Decision-making level:

[0033] Hardware configuration: Industrial-grade server (CPU Intel Xeon Gold 6330, memory 64GB, hard drive 2TB SSD).

[0034] Functional modules: Data integration module (supports MySQL and MongoDB data storage), decision command generation module (outputs standardized JSON format commands), and human-computer interaction module (web-based visual interface, supporting real-time data viewing and manual command intervention).

[0035] 4. Execution layer:

[0036] Drone feeding system: 5kg payload, 30-minute flight time, positioning accuracy ±0.5m (RTK positioning), feed rate adjustment range 0-500g / min;

[0037] Automatic water quality control equipment includes a variable frequency aerator (oxygenation efficiency 1.5 kg O2 / h) and an intelligent water exchange device (water exchange flow rate 0-5 m³ / h). 3 / h, supports precise traffic control);

[0038] Automatic fish fry sorting device: sorting efficiency 1000 fish / minute, sorting accuracy ±0.5cm (body length screening).

[0039] 5. Feedback Optimization Layer:

[0040] Data collection cycle: 1 hour / time. The collected parameters include feed residue rate (measurement range 0-30%, accuracy ±1%), fish growth rate (measurement range 0-5g / day, accuracy ±0.1g / day), and water quality index changes.

[0041] Model optimization mechanism: Incremental training is adopted, and the algorithm model parameters are updated every 7 days to ensure continuous improvement in decision accuracy.

[0042] The core algorithm implementation method is as follows:

[0043] (1) Multi-factor spatiotemporal prediction model of water environment

[0044] 1) Data preprocessing: Outliers were removed using the 3σ criterion (e.g., dissolved oxygen > 20 mg / L or < 0 mg / L were considered outliers); missing values ​​were imputed using linear interpolation (when the missing rate was < 5%), and LSTM interpolation was used when the missing rate was ≥ 5%; the data were normalized (normalization range [0,1]), using the following formula:

[0045]

[0046] in This is the original data. , These are the historical minimum and maximum values ​​for this parameter, respectively.

[0047] 2) Model building:

[0048] Input layer: 8-dimensional features (water temperature, salinity, dissolved oxygen, pH, ammonia nitrogen, nitrite, light intensity, air pressure), time step 24 (i.e., input data from the previous 24 hours);

[0049] LSTM layer: 2 hidden layers, the first layer has 64 neurons and the second layer has 32 neurons, using ReLU activation function; Attention layer: uses additive attention mechanism, calculated as follows:

[0050]

[0051] Where Q is the query matrix, K is the key matrix, and V is the value matrix. Scaling factor Let be the dimension of the key vector. The Softmax function is applied row-by-row, performing exponential normalization independently on each row of the matrix.

[0052] The detailed calculation process of the above formula can be expressed element-by-element mathematically. Let the score matrix be: Its element in the i-th row and j-th column is ,but:

[0053] a. Attention weight calculation (row-by-row Softmax):

[0054]

[0055] in, Let represent the attention weight of the i-th query on the j-th key, and for any i, .

[0056] b. Context vector output:

[0057] Output matrix The i-th row (i.e., the result of the i-th query) is:

[0058]

[0059] in, It is the j-th row (vector) of the value matrix V.

[0060] Output layer: Fully connected layer, outputs predicted values ​​of 8 water quality parameters for the next 1-24 hours.

[0061] 3) Early Warning Judgment: The formula for calculating the water quality health index is set as follows:

[0062]

[0063] in The weights for each parameter are as follows (dissolved oxygen w=0.3, water temperature w=0.2, and other parameters w=0.1). Value range: 0-100; Warning threshold: A water quality warning is triggered if the water temperature is less than 60 minutes, an oxygen deficiency warning is triggered if the dissolved oxygen level is less than 5 mg / L, and a high temperature warning is triggered if the water temperature is greater than 28℃.

[0064] (2) Dynamic precision feeding model

[0065] 1) Input data dimensions:

[0066] Environmental dimensions: water quality health index, water temperature, dissolved oxygen; biological dimensions: average fish weight (W, in g), fish density (D, in fish / m³). 3 Fish activity level (A, value 0-1, 1 being the highest);

[0067] Historical dimensions: average feed amount over the past 3 days (F_hist, in g), average feed residue rate over the past 3 days (R_hist, in %).

[0068] 2) Reinforcement learning model construction:

[0069] State space: S={HI, T, DO, W, D, A, F_hist, R_hist}, where T is water temperature and DO is dissolved oxygen;

[0070] Action space: A = {feeding time t, feeding amount f, feeding frequency k}, where t ranges from 6:00 to 18:00 (1-hour interval), f ranges from 0 to 500g, and k ranges from 1 to 3 times per day;

[0071] Reward function: R=0.4×(1-R) / 0.3+0.3×ΔW / 5 +0.3×HI, where ΔW is the daily growth rate, R is the current bait residue rate, and the reward value ranges from 0 to 1. The higher the value, the better the action.

[0072] Iterative training: An ε-greedy strategy is used to select actions. The initial value of ε is 0.9, and it decreases by 0.01 every 1000 training steps until ε=0.1.

[0073] 3) Feeding instruction generation:

[0074] Optimal feeding time: Choose a time period when HI > 70 and A > 0.7;

[0075] Single feeding amount: f=K×W×D×V×f(T, DO), where K is the basic coefficient (K=0.05 for juvenile fish, K=0.03 for adult fish), V is the volume of the culture water, and f(T, DO) is the environmental correction coefficient (f=1 when T=25℃ and DO=8mg / L, and is corrected by ±0.1 / ℃ and ±0.05 / mg / L when deviating).

[0076] Conditions for prohibiting feeding: When HI < 60 points or fish disease risk level > 70%, output the command to prohibit feeding. (3) Other key parameters of the model

[0077] 1) Intelligent evaluation model for fry / adult fish: YOLOv8 detection confidence threshold of 0.7. The input features of the fish body quality scoring model include body length / body width ratio, swimming speed, and tail wagging frequency. The output score is 0-10 points (≥8 points for robust, 5-7 points for average, and <5 points for poor quality).

[0078] The core of this model is a nonlinear regression model based on supervised learning, and its principle is as follows:

[0079] Model Structure: A fully connected deep neural network with multiple inputs and a single output is used. Alternatively, a random forest or gradient boosting regression tree can be employed. In this embodiment, a deep neural network with three hidden layers is preferred, with 16, 8, and 4 neurons in each layer, respectively. The activation function is ReLU, and the output layer is a linear unit.

[0080] Input features: Three normalized continuous features—body length / body width ratio (reflecting the robustness of the fry), swimming speed (cm / s, reflecting vitality), and tail wagging frequency (times / second, reflecting stress response and respiratory rate).

[0081] Output: A continuous value, ranging from 0 to 10, representing the overall quality score.

[0082] Training Mechanism: Thousands of fish fry samples were pre-scored manually by aquaculture experts to create a "feature-score" labeled dataset. The model uses mean squared error as the loss function and updates the weights through backpropagation using the Adam optimizer, learning the complex mapping relationship from physical features to quality scores. After training, the model can consistently output objective scores for unknown fish fry that highly match expert evaluations, with a scoring error within ±0.5 points.

[0083] The advantage of this model is that it transforms the traditional qualitative judgment that relies on human experience into an objective quantitative evaluation based on multi-dimensional quantifiable features, which is the key to achieving automated and accurate sorting of fish fry.

[0084] 2) Fish disease early warning model: Abnormal behavior judgment threshold (proportion of fish surfacing > 10%, proportion of fish that stop feeding > 15%), disease risk level = 0.5 × proportion of abnormal behavior + 0.3 × water temperature rise rate (°C / 24h) + 0.2 × dissolved oxygen variation coefficient;

[0085] 3) Drone path model: The shortest path is planned using the A* algorithm, and the amount of feed is allocated to the feeding area based on the fish density heat map.

[0086] The A* (A-Star) algorithm is a highly efficient heuristic search algorithm for finding the shortest path in a static road network. It combines the advantages of Dijkstra's algorithm (guaranteed to find the shortest path) and greedy best-first search (high search efficiency), using an evaluation function to determine the search direction.

[0087] Its core evaluation function is: f(n) = g(n) + h(n)

[0088] in:

[0089] g(n) represents the actual cost of moving from the starting node to the current node n (in this scheme, it is the cumulative path distance).

[0090] h(n) is a heuristic function that estimates the expected cost of moving from node n to the target node. h(n) must satisfy the "admissibility" property, meaning it will not overestimate the actual cost. Euclidean distance or Manhattan distance is typically used.

[0091] The A* algorithm selects the node with the smallest f(n) value for expansion each time until the target node is found.

[0092] In this system, the specific application of the UAV intelligent delivery path planning model is as follows:

[0093] Environmental modeling: The aquaculture area and its surrounding airspace are mapped as a dynamic raster map with a resolution of 1m×1m. Each raster is divided into passable (normal airspace), obstacles (no-fly zones, facility towers, etc.), and target points (feeding waypoints).

[0094] Pathfinding: Starting from the drone's current location, the farthest feed point is the endpoint. The A* algorithm calculates f(n) for adjacent grid cells from the starting point. Since the drone can fly in any direction, g(n) uses accumulated Euclidean distances, and h(n) uses the Euclidean distance from the node to the endpoint as a heuristic to ensure the shortest path and no collisions.

[0095] Dynamic cost fusion: To optimize flight safety, the algorithm adds an additional thermal cost based on fish density to g(n). If flying over a grid cell would disturb a high-density fish population, the cost of passing over it is increased, guiding the path to avoid the airspace above the feeding target and only descending upon reaching the feeding point.

[0096] Real-time replanning: When temporary obstacles occur during the feeding process (such as personnel entering), the grid map is dynamically updated, and the A* algorithm can complete local path replanning in milliseconds.

[0097] Through the A* algorithm, drones can quickly (total planning time <10 seconds) plan a three-dimensional feeding route that takes into account safety, avoidance of no-fly zones, minimal disturbance to fish, and the shortest total flight distance in complex coastal environments. Combined with heat maps, the drones can allocate differentiated feeding amounts to each waypoint, achieving truly intelligent and precise feeding.

[0098] On the other hand, the present invention also provides an intelligent management and control method for marine ranches based on multimodal AI algorithms. This method, based on the aforementioned system, specifically includes the following steps:

[0099] 1. Data Acquisition Phase (S1): The sensing layer devices collect water quality, fish population, weather, and fish fry data at a set frequency and transmit them to the algorithm layer via the 5G network. The data transmission latency is <500ms.

[0100] 2. Algorithm Calculation Stage (S2): After receiving the data, the algorithm layer performs parallel calculations using five AI algorithms, outputting water quality prediction results (including early warnings), fry scoring results, feeding plans, disease risks, and drone paths. The calculation time is <100ms.

[0101] 3. Decision Instruction Generation Stage (S3): The decision layer integrates the algorithm output results to generate standardized instructions. The instructions are displayed in real time on the web interface and manual intervention is supported (intervention response time < 10 seconds). After the instructions are generated, they are automatically pushed to the execution layer.

[0102] 4. Execution Operation Phase (S4): After receiving the instruction, the execution layer equipment executes the feeding according to the path, the water quality control equipment adjusts the oxygenation and water exchange volume according to the parameters, the fish fry sorting device selects fish fry according to the score, and the execution status is fed back to the decision-making layer in real time.

[0103] 5. Feedback Optimization Phase (S5): The feedback optimization layer collects data after execution, calculates indicators such as residual bait rate and growth rate, and updates the algorithm model parameters every 7 days to form a closed-loop optimization.

[0104] 6. Cyclic Operation: The system runs continuously 24 hours a day, repeating the S1-S5 process to achieve intelligent control of the entire process.

[0105] Compared with the prior art, the present invention has the following beneficial effects:

[0106] 1. Economic benefits: Feed utilization rate increases by 18%±2%, fish fry survival rate increases by 25%±3%, and fish disease loss rate decreases by 45%±5%. For a farm with a scale of 100,000 fish, the annual comprehensive income can be increased by 500,000 to 800,000 yuan.

[0107] 2. Environmental benefits: The ammonia nitrogen content in the aquaculture wastewater is reduced by 25%±3%, and the nitrite content is reduced by 20%±2%, which enables the wastewater quality to stably meet the discharge requirements of the "Fishery Water Quality Standard" (GB 11607-2000) and reduce pollution to the surrounding sea areas.

[0108] 3. Management benefits: Labor demand is reduced by 80% ± 5%, the average management scale per person increases from 10,000 fish / person to 50,000 fish / person, the standardization rate of the breeding process reaches over 90%, reducing human operation errors and providing a technological foundation for a large-scale, replicable smart ranch model. Attached Figure Description

[0109] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention.

[0110] Figure 2 This is a schematic diagram of the workflow of the dynamic precision feeding model in this invention.

[0111] Figure 3 This is a schematic diagram of the UAV path planning and bait distribution in this invention. Detailed Implementation

[0112] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0113] This embodiment selects a factory-style marine ranch as the implementation object. The main species farmed on this ranch is red sea bream, and the total aquaculture water volume is approximately 1000 m³. 3 The stocking density was set at 50 fish / m². 3 The fish fry introduced into the pond are approximately 5cm in size per fish. The system deployment period is 15 days, and the operation and testing period is 30 days.

[0114] I. System Deployment

[0115] 1. Sensing Layer Deployment: Six multi-parameter water quality sensors are deployed at five symmetrical points around the perimeter of each aquaculture pond and one point at the center of the pond. These sensors are connected to a data acquisition unit via an RS485 bus, with a data acquisition frequency set to 10 minutes per acquisition. The sensor measurement range and accuracy strictly conform to those described in the claims.

[0116] Two underwater high-definition cameras (1920×1080, 25fps, IP68, 0.01lux) are installed diagonally above each breeding pond to ensure that the main activity area of ​​the fish can be covered without blind spots.

[0117] An integrated meteorological sensor was installed at the highest point of the ranch to collect micro-meteorological data such as air pressure, light intensity, and wind speed.

[0118] A fish fry sorting sensor unit is deployed on the conveyor line before the fish fry are released into the pond. It is equipped with an infrared rangefinder and a high-speed industrial camera to measure the body length of each fish fry and estimate its swimming speed.

[0119] 2. Algorithm layer deployment:

[0120] Deploy a GPU server (configured with an NVIDIA A10 GPU and 24GB of video memory) in the ranch control room and install the TensorFlow 2.8 deep learning framework and the required dependency libraries.

[0121] Historical water quality data, meteorological data, and fish growth records from the past three months of the ranch (approximately 100,000 valid data points in total) were imported to pre-train various AI models. The DQN reinforcement learning model underwent 10,000 epochs of trial-and-error training in a simulation environment. The initial prediction accuracy of all models reached over 85%.

[0122] 3. Deployment of the decision-making and feedback optimization layers:

[0123] The data integration and instruction generation module is deployed on an industrial-grade server, and the database adopts a hybrid architecture of MySQL and MongoDB to accommodate both structured and unstructured data.

[0124] Build a web-based remote visual management and control platform with functions such as real-time data dashboard, command preview and manual intervention, and historical data traceability.

[0125] Configure an automatic model update script. The script monitors the feedback database. When the number of new valid feedback data exceeds 5,000 or the 7-day time threshold is reached, it automatically starts the incremental training task for each model and uses the Elastic Weight Consolidation (EWC) algorithm to ensure that the model does not forget old knowledge when learning new knowledge.

[0126] 4. Execution layer deployment:

[0127] The drone feeding system's docking point and automatic charging station are located in the center of the ranch, while the RTK differential reference station is set up in an unobstructed open area to ensure that the drones achieve centimeter-level positioning accuracy.

[0128] Automatic water quality control equipment, including variable frequency aerators and intelligent water exchange devices, is connected to a decision-making server via industrial Ethernet to receive PID control parameters.

[0129] The automatic fish fry sorting device is located at the downstream station of the sensing unit. Its high-speed diversion baffle can, within 100ms, divert inferior fry with a score of <5 and qualified fry with a score of ≥5 to different temporary holding ponds based on the scoring results issued by the algorithm layer.

[0130] II. Detailed Explanation of the Core Methodology Implementation Process

[0131] During the operational testing phase, the system operates continuously according to the following closed-loop process:

[0132] S1 Data Acquisition: All devices in the perception layer collect data according to a preset cycle, and send the data packets to the algorithm layer server with a delay of less than 500ms via 5G CPE (customer terminal equipment).

[0133] S2 multi-task parallel computing:

[0134] Water environment prediction: The LSTM-Attention model receives 8-dimensional data from the past 24 hours and predicts water quality trends for the next 1-24 hours. Calculation results show that the prediction error for water temperature in the next 24 hours is ±0.3℃, and the prediction error for dissolved oxygen is ±0.2mg / L. A water quality warning signal is immediately generated when the calculated HI index falls below 60.

[0135] Fish fry assessment: The YOLOv8 model performs real-time inference on images of fish fry passing by at high speeds, with a confidence threshold set at 0.7. For identified fish fry, their length-to-width ratio, instantaneous swimming speed, and tail wagging frequency are extracted and input into the scoring network. The measured sorting accuracy reached 91%, and the rate of removing inferior fish fry increased from 40% by manual sorting to 85%.

[0136] Dynamic feeding decision-making: The DQN model agent makes decisions based on the current state (e.g., HI=85 points, T=25℃, DO=8mg / L, W=15g, D=50 tails / m). 3 With an exploration probability of ε=0.1 (A=0.8, etc.), the following actions were ultimately selected: feeding time at 8:00 AM, single feeding amount of 280g, and feeding frequency of 2 times / day. The single feeding amount was further fine-tuned according to the formula f=K×W×D×V×f(T,DO).

[0137] Disease Warning: The visual analysis module detected that only 2% of the fish are currently surfacing, and 5% have stopped feeding. The model, considering a 1.5℃ increase in water temperature over 24 hours and a dissolved oxygen coefficient of variation of 0.1, calculates the disease risk level as: 0.5 × (2% + 5%) + 0.3 × 1.5 + 0.2 × 0.1 = 0.035 + 0.45 + 0.02 = 0.505 = 50.5%. The risk level is less than 70%, which is within the normal monitoring range. If the risk level exceeds 70%, the system will issue the highest-level red warning and forcibly output a "Do Not Feed" command.

[0138] Path planning: The path planning model receives feeding decisions and a real-time fish density heatmap generated by an underwater camera, dividing the feeding area into 1m×1m grids. After avoiding no-fly zone grids, the A* algorithm plans a flight path approximately 150 meters long. Based on the heatmap, the feeding amount for the central high-density grids is increased by 20%. The entire planning calculation takes only 8.5 seconds.

[0139] S3 Command Generation and Issuance: The decision-making level integrates the above parallel outputs into a single JSON command and issues it. The web interface displays the command synchronously. If the administrator does not intervene within 10 seconds, the system will execute it automatically.

[0140] S4 command execution: The drone takes off automatically and delivers food precisely to the predetermined waypoint; the oxygenator speed is automatically adjusted; the sorting device's flow divider operates on time.

[0141] S5 Feedback Optimization: After the operation, the feedback system calculated the uneaten bait rate to be 4.8% and the daily growth rate to be 3.9g based on underwater uneaten bait monitoring sensors and the next day's fish sampling and weighing. Seven days later, the incremental training script ran automatically, using the newly generated tens of thousands of interaction data as the training set to fine-tune the Q-network of the DQN model for 100 rounds. The EWC algorithm effectively protected the generalization ability learned from historical data, and the model's decision reward value in the new scenario improved by 2.5%.

[0142] III. Implementation Results

[0143] After 30 days of continuous operation and testing, the following results were achieved:

[0144] (1) Water quality prediction accuracy: water temperature prediction error ±0.3℃, dissolved oxygen prediction error ±0.2mg / L, water quality health index calculation accuracy 92%, early warning response time <1 minute;

[0145] (2) Fish fry sorting effect: the sorting accuracy rate was 91%, the rate of removing inferior fish fry increased from 40% in manual sorting to 85%, and the fish fry survival rate increased from 65% to 90%;

[0146] (3) Feeding optimization effect: The uneaten feed rate decreased from 18% to 5%, the feed utilization rate increased by 17%, and the daily growth rate of the fish increased from 2.5g to 3.8g;

[0147] (4) Disease early warning effect: During the high-temperature season (water temperature 29℃), disease early warning is issued 72 hours in advance, and the disease outbreak rate drops from 50% to 12%;

[0148] (5) Effect of labor substitution: The number of workers in the ranch was reduced from 15 to 3, the labor cost was reduced by 80%, and the standardization rate of the breeding process was increased from 30% to 92%.

[0149] (6) The aquaculture wastewater was tested and found that the ammonia nitrogen and nitrite content were consistently lower than the national standard limits, with a compliance rate of 100%.

[0150] It should be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0151] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A marine ranching intelligent management and control system based on multimodal AI algorithms, characterized in that, include: The sensing layer includes multi-parameter water quality sensors, underwater high-definition cameras, meteorological sensors, and fish fry sorting sensors, which are used to collect data on water quality, fish populations, weather, and fish fry. At the algorithm layer, a multi-factor aquatic environment spatiotemporal prediction model, a fish fry / adult fish intelligent assessment model, a dynamic precision feeding model, a fish disease early warning model, and a drone feeding path planning model are deployed. The decision-making layer includes a data integration module and a decision instruction generation module, which are used to integrate algorithm results and generate standardized instructions. The execution layer, including the drone feeding system, automatic water quality control equipment, and automatic fish fry sorting device, is used to execute decision-making instructions; The feedback optimization layer is configured to perform incremental training and parameter updates on each model in the algorithm layer periodically based on the result data after the execution layer's operation. The perception layer, algorithm layer, decision-making layer, execution layer, and feedback optimization layer interact with each other via a network, forming a closed-loop control system for the entire process.

2. The system according to claim 1, characterized in that, The multi-parameter water quality sensor collects parameters including water temperature, salinity, dissolved oxygen, pH, ammonia nitrogen, and nitrite, with a collection frequency of 10 minutes per time.

3. The system according to claim 2, characterized in that, The multi-factor spatiotemporal prediction model for the water environment is a time-series prediction model based on a Long Short-Term Memory (LSTM) network with an attention mechanism. Its input includes eight temporal features: water temperature, salinity, dissolved oxygen, pH, ammonia nitrogen, nitrite, light intensity, and air pressure. The time step is 24, and the output is the predicted values ​​of eight water quality parameters for the next 1-24 hours. Based on these predicted values, the water quality health index HI is calculated using the following formula: in, Let be the weights of the i-th type of parameter, with dissolved oxygen having a weight of 0.3, water temperature having a weight of 0.2, and all other parameters having a weight of 0.

1. For the predicted values ​​of the parameters, and These are the historical minimum and maximum values ​​of the parameter, respectively.

4. The system according to claim 2, characterized in that, The dynamic precision feeding model is a reinforcement learning model based on a deep Q-network (DQN). Its state space includes the water quality health index HI, water temperature T, dissolved oxygen DO, fish body weight W, fish density D, fish activity A, historical feeding amount, and historical uneaten feed rate; its action space is a discrete combination of feeding time t, single feeding amount f, and feeding frequency k; and its reward function R is: R=0.4×(1-R) / 0.3+0.3×ΔW / 5+0.3×HI Where R is the uneaten feed rate, ΔW is the daily growth rate of the fish population, and HI is the water quality health index.

5. The system according to claim 3, characterized in that, The fish disease early warning model is a multimodal fusion diagnostic model, and its output disease risk level is determined by the following formula: Disease risk level = 0.5 × percentage of abnormal behavior + 0.3 × rate of water temperature rise + 0.2 × dissolved oxygen variation coefficient; When the disease risk level is greater than 70%, an emergency warning signal is triggered, and the feeding decision is forcibly overridden, outputting a command to prohibit feeding.

6. The system according to claim 2, characterized in that, The intelligent evaluation model for fry / adult fish includes a target detection network based on YOLOv8 and a fish body quality scoring model. The target detection network is used to identify fish targets from underwater images (confidence ≥ 0.7) and extract body length / body width ratio, swimming speed, and tail wagging frequency as input features, which are then input into the quality scoring model to output a quality score of 0 to 10. Among them, a score ≥ 8 indicates a robust fry, 5 to 7 indicates a general fry, and < 5 indicates a poor-quality fry.

7. The system according to claim 2, characterized in that, The drone feeding path planning model uses the A* algorithm to plan the shortest path and allocates the amount of feed to the feeding area according to the fish density heat map: based on the basic feeding amount, the feeding amount in the area with the highest density is increased by 20%, and the feeding amount in the no-fly zone and the fishless zone is set to zero. The total time for path planning is less than 10 seconds.

8. The system according to claim 2, characterized in that, The feedback optimization layer is configured to collect the uneaten feed rate, daily growth rate of fish, and changes in water quality parameters every 24 hours after the operation, and trigger the incremental training process of each AI model based on a set cumulative data threshold or a 7-day time period. The incremental training process, while retaining the original model architecture, only uses newly collected data to fine-tune the network weights, and uses the elastic weight consolidation EWC algorithm to prevent catastrophic forgetting.

9. A method for intelligent management and control of marine ranches based on multimodal AI algorithms, applied to the system described in any one of claims 1 to 8, characterized in that, Includes the following steps: S1. Data Acquisition and Transmission: The perception layer collects water quality, fish population, weather, and fry data in the aquaculture environment at a preset frequency and transmits them to the algorithm layer in real time via the network. S2. The algorithm layer calls the following five AI models in parallel to perform computations: A multi-factor spatiotemporal prediction model for water environment outputs predicted values ​​of water quality parameters and early warning signals. Intelligent assessment model for fry / adult fish, outputting quality scores and grading results for fry or adult fish; The dynamic precision feeding model outputs the optimal feeding plan, which includes feeding time, feeding amount, and feeding frequency. The fish disease early warning model outputs the disease risk level and corresponding warning signals. The drone feeding path planning model outputs feeding waypoints and the amount of feed at each point based on fish density heatmaps and dynamic grid methods. S3. The decision-making layer integrates all the outputs of step S2 to generate a unified structured instruction set, which is then distributed to the execution layer. At the same time, it is visualized through a human-computer interaction interface, allowing manual intervention or confirmation within a preset time window. S4: The execution layer devices parse the instruction set, the drones carry out precise feeding according to the planned path, the automatic water quality control equipment adjusts the oxygenation and water exchange, and the automatic fry sorting device selects fry according to the scoring results. S5. The feedback optimization layer collects the effect data after execution, and performs incremental training on the five AI models in step S2 every 7 days to update the model parameters and achieve closed-loop management.

10. The method according to claim 9, characterized in that, In step S4, the automatic fry sorting device performs real-time screening of continuously passing fry based on the quality score output in step S2, and automatically removes inferior fry with a score of <5 through a diversion device; the automatic water quality control equipment includes a variable frequency aerator and an intelligent water exchange device, and its control logic is as follows: based on the predicted and current values ​​of water quality parameters, the speed of the variable frequency aerator and the valve opening of the intelligent water exchange device are precisely adjusted by a PID controller to maintain the dissolved oxygen and ammonia nitrogen concentrations within the preset optimal range.

Images