Deep learning-based south day clams breeding equipment group cross-process collaborative control method

Abstract

The present application relates to the field of artificial intelligence, and more particularly to a method for cross-process collaborative control of south China hard clam breeding equipment groups based on deep learning, which comprises comprehensively sensing the state of each process of breeding by collecting and fusing multi-modal data in real time to provide a data basis for accurate control; using a multi-task deep learning model to make parallel predictions and output key decision information, which can handle multiple tasks simultaneously to improve decision efficiency and accuracy and timely respond to complex situations in the breeding process; generating a collaborative work plan based on collaborative rules and global optimization objectives to achieve dynamic coordination across processes, optimize resource allocation, reduce work conflicts, and improve breeding efficiency; converting the work plan into control instructions and distributing them to the breeding equipment while obtaining execution state data to ensure accurate execution of the instructions, timely grasp of equipment operation, and smooth progress of breeding operations; and through incremental training of the model, the model continuously adapts to changes in the breeding environment.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a cross-process collaborative control method for abalone farming equipment based on deep learning. Background Technology

[0002] Against the backdrop of the booming development of smart agriculture, the aquaculture sector is also actively introducing various advanced technologies, moving towards intelligent and refined practices. Nanri abalone farming, as an important component of aquaculture, possesses considerable economic value and market potential.

[0003] However, current abalone farming in Nanri region still faces several challenges, including: First, the lack of effective coordination mechanisms between different processes leads to independent operations, hindering efficient integration based on overall farming conditions. This results in low operational efficiency, potential process conflicts or unreasonable time intervals, impacting the stability of the abalone's growth environment and farming profitability. Second, critical decisions regarding optimal cleaning timing and intensity, optimal hoisting timing, and feeding strategies often rely on experience rather than precise judgment based on real-time data and scientific models. This lack of dynamic adjustment based on the abalone's actual growth status and environmental changes affects abalone growth quality and yield. Third, the absence of scientific scheduling plans for farming equipment can lead to multiple pieces of equipment using the same farming area simultaneously, causing spatial conflicts, disrupting normal equipment operation, and potentially causing unnecessary disturbance and stress to the abalone. Fourth, traditional models for predictive farming decisions may struggle to adapt to the complexity and variability of the Nanri abalone farming environment. As time and the environment change, the models may fail to accurately reflect the actual situation, resulting in significant prediction deviations and failing to provide continuous and effective decision support for the farming process. To address this, the present invention proposes a cross-process collaborative control method for abalone farming equipment based on deep learning. Summary of the Invention

[0004] The purpose of this invention is to solve the problems in the background art by proposing a cross-process collaborative control method for abalone farming equipment based on deep learning.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A deep learning-based cross-process collaborative control method for abalone farming equipment groups includes:

[0007] S1. Through the sensor network and vision system deployed in the breeding area, multimodal data related to each process is collected in real time and uploaded to the central control platform for data fusion to generate a collaborative perception dataset across processes.

[0008] S2. Based on the collaborative sensing dataset, use a multi-task deep learning model to perform parallel prediction and output key decision information;

[0009] S3. The central control platform receives key decision information from the multi-task deep learning model and dynamically generates cross-process collaborative operation plans based on preset collaborative rules and global optimization objectives.

[0010] S4. The central control platform will convert the generated collaborative operation plan into control instructions to drive each breeding equipment to perform the corresponding process, and distribute them to the corresponding breeding equipment, while acquiring the execution status data of the breeding equipment.

[0011] S5. Based on the execution status data of the aquaculture equipment and the newly collected on-site monitoring data, incremental training is performed on the multi-task deep learning model.

[0012] Furthermore, in step S1, the multimodal data includes environmental data, feeding-related data, caisson status data, and aquaculture equipment status data; wherein, the environmental data includes water temperature, salinity, and dissolved oxygen; the feeding-related data includes feeding amount, feeding time, and feed settling and diffusion observed by an underwater camera; the caisson status data includes image or video data of caisson surface attachments collected by an underwater robot or camera; and the aquaculture equipment status data includes the working status and energy consumption information of the feeder, cleaning robot, and crane.

[0013] Furthermore, in step S2, the process of using a multi-task deep learning model to perform parallel prediction and output key decision information includes:

[0014] The multi-task deep learning model consists of three sub-models, including the first sub-model, the second sub-model, and the third sub-model;

[0015] The first sub-model is used for the task of predicting the optimal cleaning time and intensity: Based on the multi-source temporal characteristics of bait residue, water quality dynamics and dirt growth, the first sub-model outputs the optimal cleaning time window and cleaning intensity level for the next cleaning operation in parallel.

[0016] The second sub-model is used for the task of predicting the best hoisting time: Based on the characteristics of the caisson state, process coordination and environmental constraints, the second sub-model aims to minimize the stress level of abalone and efficiently connect the cleaning process to determine the best hoisting time window.

[0017] The third sub-model is used for the task of dynamically adjusting the feeding strategy: The third sub-model determines the basic feeding amount based on the abalone's growth stage, and then performs two-level corrections by combining the residual feed rate and the comprehensive environmental index to output the final feeding amount; at the same time, it uses spatial optimization to generate the coordinates of the feeding point.

[0018] Furthermore, the first sub-model, based on the multi-source temporal characteristics of feed residue, water quality dynamics, and fouling growth, outputs the optimal cleaning time window and cleaning intensity level for the next cleaning operation in parallel.

[0019] In the task of predicting the optimal cleaning time and intensity, multi-source temporal features are acquired, including feed residue features, water quality dynamic features, and fouling growth correlation features. These multi-source temporal features are then input into the first sub-model, which is based on a multi-task deep learning architecture and includes a shared feature extraction network, a cleaning time prediction branch, and a cleaning intensity prediction branch. The cleaning time prediction branch outputs the probability distribution of cleaning demand in future time periods, and the cleaning intensity prediction branch outputs the corresponding cleaning intensity level. The continuous time period in which the probability distribution of cleaning demand exceeds a preset cleaning threshold is determined as the optimal cleaning time window, and the cleaning intensity level is determined based on the output of the cleaning intensity prediction branch.

[0020] Furthermore, the second sub-model, based on the caisson status, process coordination, and environmental constraints, aims to minimize abalone stress levels and efficiently connect cleaning processes. The process for determining the optimal hoisting time window includes:

[0021] In the optimal hoisting timing prediction task, the caisson state characteristics, process coordination characteristics, and environmental constraint characteristics are acquired and input into the second sub-model. The second sub-model includes a stress level prediction module and a multi-objective optimization decision module. The stress level prediction module is used to predict the abalone stress level caused by hoisting operations under different hoisting time windows based on the caisson state characteristics, process coordination characteristics, and environmental constraint characteristics, and outputs a stress response intensity score. The multi-objective optimization decision module is used to receive the stress response intensity score and the caisson state characteristics, process coordination characteristics, and environmental constraint characteristics, and optimizes the solution under the condition of maximizing process connection efficiency and minimizing stress response intensity score, with the goal of maximizing process connection efficiency and minimizing stress response intensity score, and outputs the optimal hoisting time window.

[0022] Furthermore, the third sub-model determines the basic feeding amount based on the abalone's growth stage, and then performs two-level corrections by combining the uneaten feed rate and the comprehensive environmental index to output the final feeding amount; simultaneously, the process of generating the feeding point coordinates using spatial optimization includes:

[0023] In the dynamic adjustment of feeding strategy, the current weight and growth stage of abalone are estimated using an abalone growth model. Based on the growth stage, a preset basic feeding requirement table is consulted to determine the basic feeding amount. The distribution of uneaten feed is analyzed using an underwater camera to calculate the uneaten feed rate. Based on this rate, an uneaten feed correction amount is calculated to make the first correction to the basic feeding amount. Water temperature and dissolved oxygen data are acquired and a comprehensive environmental index is generated. Based on this index, an environmental correction amount is calculated to make a second correction to the feeding amount after the first correction, resulting in the final feeding amount. A spatial optimization algorithm is used to analyze the feeding location: after the aquaculture area is gridded, a predetermined number of grids are selected as feeding points by comprehensively calculating the historical abalone aggregation heat and the current water flow diffusion suitability of each grid unit. The final output feeding strategy includes the feeding amount and the coordinates of the feeding points.

[0024] Furthermore, in step S3, the process of dynamically generating cross-process collaborative work plans based on preset collaboration rules and global optimization objectives includes:

[0025] Receive key decision information output from the first sub-model, the second sub-model, and the third sub-model, respectively. The key decision information includes:

[0026] One or more optimal cleaning time windows and corresponding cleaning intensity levels from the first sub-model;

[0027] One or more optimal hoisting time windows from the second sub-model;

[0028] The final amount of bait and the coordinates of the baiting point from the third sub-model;

[0029] Based on a global collaboration strategy, key decision-making information is integrated and further optimized to generate the final collaborative work plan; the global collaboration strategy includes:

[0030] Based on the optimal cleaning time window and the optimal hoisting time window, a global timing optimization algorithm is used to determine the unique execution time of the cleaning operation, the hoisting operation, and the baiting operation.

[0031] Furthermore, the global time-series optimization algorithm adopts a dynamic weighted multi-objective optimization strategy, constructing an objective function system with total operation time, total energy consumption, and abalone growth interference as optimization dimensions, where the weight of each objective is dynamically allocated according to the real-time operation status.

[0032] Furthermore, in step S4, the central control platform converts the generated collaborative operation plan into control instructions that drive each piece of aquaculture equipment to execute corresponding procedures, and distributes these instructions to the corresponding aquaculture equipment. Simultaneously, the process of acquiring the execution status data of the aquaculture equipment includes:

[0033] Based on the cleaning operation execution time and cleaning intensity level determined in the collaborative operation plan, control instructions are generated and sent to the cleaning equipment. The control instructions include the start time and the cleaning power corresponding to the cleaning intensity level.

[0034] Based on the hoisting operation execution time set in the collaborative operation plan, control instructions are generated and sent to the hoisting equipment. The control instructions include the hoisting time determined from the optimal hoisting time window output from the second sub-model and the target caisson number to be operated.

[0035] Based on the feeding operation execution time obtained from the collaborative operation plan, control instructions are generated and sent to the feeding equipment. The control instructions include the final feeding amount and feeding point coordinates output from the third sub-model.

[0036] The control command distribution adopts an asynchronous communication mechanism based on message queues, and the execution status data of the aquaculture equipment is obtained through a predefined state transition mechanism.

[0037] Furthermore, in step S5, incremental training adopts an online learning strategy, periodically adding the execution status data of the aquaculture equipment and newly collected on-site monitoring data to the training set, and using the stochastic gradient descent algorithm to update the parameters of the multi-task deep learning model.

[0038] Compared with existing technologies, the beneficial effects of this invention are as follows: By utilizing sensor networks and vision systems, multimodal data can be collected and fused in real time and comprehensively, providing rich and accurate information to detect changes in the aquaculture process and provide a reliable basis for precise control; through parallel prediction of multi-task deep learning models, multiple tasks such as the optimal cleaning time can be processed simultaneously, enabling rapid output of key decision information, improving decision-making efficiency, and making aquaculture decisions more timely and accurate; the central control platform generates collaborative operation plans based on collaborative rules and global optimization objectives, realizing dynamic coordination across processes, optimizing resource allocation according to real-time conditions, reducing operational conflicts, and improving overall aquaculture efficiency; by converting operation plans into control instructions and distributing them to aquaculture equipment, while simultaneously acquiring execution status data, accurate execution of instructions is ensured, the operating status of aquaculture equipment is monitored in a timely manner, and aquaculture operations are guaranteed to proceed smoothly according to plan; incremental training of the model based on execution status and on-site monitoring data enables the model to continuously adapt to changes in the aquaculture environment, improving the model's adaptability and prediction accuracy, and providing a more scientific and reliable basis for aquaculture decisions. Attached Figure Description

[0039] Figure 1 This is a flowchart of the cross-process collaborative control method for abalone farming equipment based on deep learning proposed in this invention. Detailed Implementation

[0040] 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.

[0041] Reference Figure 1 A deep learning-based cross-process collaborative control method for abalone farming equipment in Nanri, including:

[0042] S1. Through the sensor network and vision system deployed in the breeding area, multimodal data related to each process is collected in real time and uploaded to the central control platform for data fusion to generate a collaborative perception dataset across processes.

[0043] S2. Based on the collaborative sensing dataset, use a multi-task deep learning model to perform parallel prediction and output key decision information;

[0044] S3. The central control platform receives key decision information from the multi-task deep learning model and dynamically generates cross-process collaborative operation plans based on preset collaborative rules and global optimization objectives.

[0045] S4. The central control platform will convert the generated collaborative operation plan into control instructions to drive each breeding equipment to perform the corresponding process, and distribute them to the corresponding breeding equipment, while acquiring the execution status data of the breeding equipment.

[0046] S5. Based on the execution status data of the aquaculture equipment and the newly collected on-site monitoring data, incremental training is performed on the multi-task deep learning model.

[0047] It should be further explained that, in the specific implementation process, in step S1, the multimodal data includes environmental data, feeding-related data, caisson status data, and aquaculture equipment status data; among them, environmental data includes water temperature, salinity, and dissolved oxygen; feeding-related data includes feeding amount, feeding time, and feed settling and diffusion observed through underwater cameras; caisson status data includes image or video data of caisson surface attachments collected by underwater robots or cameras; and aquaculture equipment status data includes the working status and energy consumption information of the feeder, cleaning robot, and crane.

[0048] It should be further explained that, in the specific implementation process, in step S2, a multi-task deep learning model is used for parallel prediction to output key decision information, including:

[0049] The multi-task deep learning model consists of three sub-models, including the first sub-model, the second sub-model, and the third sub-model;

[0050] The first sub-model is used for the optimal cleaning timing and intensity prediction task. Based on multi-source temporal features of bait residue, water quality dynamics, and fouling growth, the first sub-model outputs the optimal cleaning time window and cleaning intensity level for the next cleaning operation in parallel. This includes: acquiring multi-source temporal features in the optimal cleaning timing and intensity prediction task, including bait residue features, water quality dynamics features, and fouling growth-related features; inputting these multi-source temporal features into the first sub-model, which is based on a multi-task deep learning architecture and includes a shared feature extraction network, a cleaning time prediction branch, and a cleaning intensity prediction branch; the cleaning time prediction branch outputs the probability distribution of cleaning demand for future periods, and the cleaning intensity prediction branch outputs the corresponding cleaning intensity level; the continuous time period where the probability distribution of cleaning demand exceeds a preset cleaning threshold is determined as the optimal cleaning time window, and the cleaning intensity level is determined based on the output of the cleaning intensity prediction branch.

[0051] Specifically, C1, video data after feeding is collected by an underwater camera and combined with the feeding amount records of the feeder to construct feed residue characteristics; for images or video sequences, a background subtraction algorithm is used to identify the areas where feed has settled, and the feed density per unit area is calculated through pixel statistics; three key indicators are calculated at 15-minute monitoring intervals: feed settling rate = (initial feed density - current frame feed density) / monitoring interval; feed residue rate = current frame feed density / initial feed density × 100%; feed distribution uniformity = 1 - (standard deviation of feed density in each area / average density); simultaneously, water quality sensors continuously monitor water temperature, dissolved oxygen, and salinity data; after processing... Subsequently, dynamic water quality characteristics such as dissolved oxygen change rate, water temperature gradient, and salinity fluctuation index are formed. Dissolved oxygen change rate = (current value - average value of the previous hour) / monitoring time interval; water temperature gradient is the temperature difference between different depth layers; salinity fluctuation index is the coefficient of variation of salinity values ​​over the past 6 hours. Submerged robot periodically collects surface image data of the caisson, which, after color space conversion and threshold segmentation, extracts fouling coverage, fouling growth rate, and algal biomass index to obtain fouling growth-related characteristics. Fouling coverage = number of algal pixels / total number of pixels × 100%; fouling growth rate = (current coverage - coverage rate of the same period last week) / monitoring time interval; algal biomass index = average saturation of algal area × coverage.

[0052] C2. During the training phase of the first sub-model, historical data is collected to construct training samples. Each sample contains time-series data on bait residue characteristics, water quality dynamics, and fouling growth characteristics within a specific (24-hour) time period. These feature data serve as the basis for the first sub-model's learning and are used to predict cleaning timing and intensity-related indicators. The shared feature extraction network of the first sub-model uses a one-dimensional convolutional neural network to extract local temporal patterns, with a kernel size of 3 and a stride of 1. After processing by the one-dimensional convolutional neural network, the extracted local temporal features are input into a long short-term memory (LSM) network layer. The LSM layer contains 64 hidden units to analyze long-term trends and capture long-term dependencies in the feature data. The output layer of the first sub-model uses two fully connected branches: a cleaning time prediction branch and a cleaning intensity prediction branch. The cleaning time prediction branch outputs the probability of cleaning demand at each time point in the next 24 hours, and this probability distribution predicts the urgency of cleaning operations in the future (e.g., the next 24 hours). The cleaning intensity prediction branch outputs the confidence levels for three cleaning intensity levels.

[0053] C3. The first sub-model ultimately outputs the probability distribution of cleaning demand and the confidence level of the cleaning intensity level. The optimal cleaning time window is determined as the continuous time period during which the probability distribution exceeds a preset cleaning threshold (e.g., 0.7). Simultaneously, for this time window, the level with the highest confidence level output by the cleaning intensity prediction branch is determined as the final cleaning intensity level (mild, moderate, strong). Assume the output structure of the cleaning intensity prediction branch is as follows: it outputs a list containing three values, representing the model's confidence levels for the mild, moderate, and strong levels. If the model outputs [0.15, 0.70, 0.15], it means the model considers the probability of moderate cleaning to be 70%, and the probabilities of mild and strong cleaning to be 15% each. Among these three values ​​(0.15, 0.70, 0.15), the maximum value is 0.70, and 0.70 corresponds to a moderate cleaning level. Therefore, the final automatically determined cleaning intensity is moderate.

[0054] The second sub-model is used for the optimal hoisting timing prediction task. Based on the caisson status, process coordination, and environmental constraints, the second sub-model aims to minimize abalone stress levels and efficiently connect cleaning processes to determine the optimal hoisting time window. This includes: acquiring caisson status characteristics, process coordination characteristics, and environmental constraints in the optimal hoisting timing prediction task; and inputting these into the second sub-model. The second sub-model contains a stress level prediction module and a multi-objective optimization decision module. The stress level prediction module predicts the abalone stress level caused by hoisting operations under different hoisting time windows based on caisson status characteristics, process coordination characteristics, and environmental constraints, and outputs a stress response intensity score. The multi-objective optimization decision module receives the stress response intensity score, caisson status characteristics, process coordination characteristics, and environmental constraints, and optimizes the solution to maximize process connection efficiency and minimize the stress response intensity score under the condition of satisfying preset operational constraints, outputting the optimal hoisting time window. Simultaneously, when there are conflicting hoisting demands for multiple caissons, a game theory model is used to generate the optimal hoisting scheduling scheme by combining the optimal hoisting time windows of each caisson.

[0055] Specifically, D1 involves acquiring the caisson status characteristics, process coordination characteristics, and environmental constraint characteristics: obtaining the fouling coverage rate and determining the current degree of fouling of the caisson by assessing the fouling coverage rate; using underwater video analysis technology, statistically analyzing the abalone activity frequency as a basic indicator of stress response to obtain the abalone activity status, and combining the caisson fouling degree and abalone activity status to obtain the caisson status characteristics; from the predicted cleaning plan output by the cleaning prediction task (including the estimated cleaning completion time), and calculating the process time interval between cleaning and hoisting (the process time interval between cleaning and hoisting is the difference between the candidate hoisting time and the estimated cleaning operation completion time), thereby obtaining the process coordination characteristics; and obtaining future (e.g., 12-hour) weather and sea state information, including wind speed, wave height, and current speed, by accessing the meteorological department's API, calculating the comprehensive sea state index to obtain the environmental constraint characteristics, where the comprehensive sea state index = 0.4 × wind speed + 0.3 × wave height + 0.3 × current speed.

[0056] D2. Input the caisson state characteristics, process coordination characteristics, and environmental constraint characteristics into the stress level prediction module. This module is a three-layer fully connected network structure with hidden layer nodes set to 32, 16, and 8 respectively, using the ReLU activation function. The module's output is a stress response intensity score, used to quantify and predict the abalone stress level that may be triggered by hoisting operations under given characteristics. Input the stress response intensity score output by the stress level prediction module, along with the original caisson state characteristics, process coordination characteristics, and environmental constraint characteristics, into the multi-objective optimization decision module. The multi-objective optimization decision module establishes the objective functions: Objective function 1 is to maximize the connection efficiency, and objective function 2 is to minimize the stress level. Maximizing the connection efficiency = 1 / (time interval between cleaning and hoisting processes + 0.1), by maximizing this function, the hoisting time and cleaning process are connected as efficiently as possible. Minimizing the stress level = stress response intensity score, by minimizing this function... To reduce stress on abalone during hoisting, the following constraints were set: the time interval between cleaning and hoisting processes must be ≥30 minutes to ensure hoisting occurs after cleaning and to guarantee a reasonable sequence of processes; the overall sea state index must be ≤ a safe operating threshold, which is pre-set based on actual operating conditions to ensure hoisting operations are conducted under suitable sea conditions; a hoisting suitability score is calculated using the formula: Hoisting Suitability Score = 0.6 × Connection Efficiency + 0.4 × (1 - Stress Level). This formula integrates the impact of connection efficiency and stress level on hoisting suitability, with weight allocation reflecting their relative importance in decision-making; a sea state penalty is added: when the overall sea state index exceeds the safe operating threshold, the hoisting suitability score decreases by 0.1 for every 10% exceeding the threshold; for example, if the overall sea state index exceeds the safe operating threshold by 20%, the score decreases by 0.2. This penalty further considers the impact of sea conditions on hoisting suitability.

[0057] D3. The second sub-model outputs hoisting suitability scores for multiple candidate time windows. The highest-scoring consecutive 3-hour period is selected as the initial candidate optimal hoisting time window, facilitating completion of the hoisting task within a relatively stable timeframe and reducing operational costs and risks caused by time dispersion. Simultaneously, this period ensures the following hard constraints: the hoisting and cleaning processes are guaranteed to proceed in the correct order after the predicted cleaning operation is completed; the future comprehensive sea state index is below the safe operation threshold, ensuring the safe conduct of the hoisting operation; this period includes at least two slack tide periods, during which sea states are relatively stable and conducive to hoisting operations. Finally, the consecutive period with the highest hoisting suitability score that meets all conditions is determined as the optimal hoisting time window. When multiple caissons need to be hoisted, a game theory model is used to resolve resource conflicts.

[0058] D31. Treat each caisson that needs to be hoisted as an agent. Each agent represents an independent decision-making unit. The agent's policy set includes its optional hoisting time window (e.g., selecting one from multiple candidate time windows output by the second sub-model).

[0059] D32. The utility function of each agent is defined based on its hoisting suitability score, job urgency, and resource competition factors: In the formula, Indicates the first The utility function of an agent is the numerical result calculated by the function, which is the utility value obtained by the agent under a certain strategy (i.e., selecting a certain hoisting time window). Indexing for intelligent agents; The hoisting suitability score (range 0-1) is calculated for the second sub-model. The higher the score, the better the hoisting effect. The priority weight of the caisson is dynamically calculated based on the degree of soiling (soil coverage) and the abalone activity status (for example, the higher the soil coverage and the more abnormal the abalone activity, the greater the priority weight). This is a waiting time penalty, meaning that the penalty increases if the hoisting time is delayed. Equal to the number of hours of delay; Preset weighting coefficients are used to balance hoisting suitability, priority, and waiting time;

[0060] D33. Using a non-cooperative game model, agents independently choose strategies (lifting time windows) to maximize their own utility. The Nash equilibrium state of the game indicates that no agent can improve its own utility by unilaterally changing its strategy.

[0061] D34. Solve the Nash equilibrium using the iterative optimal response dynamic algorithm: Each agent randomly selects an initial policy (lifting time window); each agent takes turns calculating the utility value of all possible policies and selects the policy that maximizes its own utility (optimal response); after updating the policy, the utility of other agents is recalculated; when the policies of all agents no longer change, or the change is less than the preset iteration threshold, it is considered that the Nash equilibrium has been reached; the policy combination corresponding to the Nash equilibrium is the final lifting scheduling scheme, ensuring that the lifting time windows of each caisson are conflict-free and globally optimized;

[0062] D35. During the game, a hard constraint check is added: if multiple agents choose the same time window and the hoisting equipment resources are insufficient (for example, a maximum of 2 caissons can be hoisted in the same time period), the agent with the higher utility value is given priority, and other agents reselect their strategies; all selected hoisting time windows must satisfy the comprehensive sea state index ≤ safe operation threshold; the game model finally outputs a Pareto optimal scheduling scheme that minimizes the total operation time and total stress level while satisfying individual utilities.

[0063] The third sub-model is used for the dynamic adjustment of feeding strategies: Based on the abalone's growth stage, the third sub-model determines the basic feeding amount and performs two-stage corrections using the residual feed rate and comprehensive environmental index, outputting the final feeding amount. Simultaneously, spatial optimization is used to generate feeding point coordinates, including: in the dynamic adjustment of feeding strategies, the current abalone weight and growth stage are estimated using the abalone growth model; the basic feeding amount is determined by querying a preset basic feeding requirement table based on the growth stage; residual feed distribution is analyzed using underwater cameras, the residual feed rate is calculated, and a residual feed correction amount is calculated based on this rate, performing the first correction to the basic feeding amount; water temperature and dissolved oxygen data are acquired and a comprehensive environmental index is generated; an environmental correction amount is calculated based on the comprehensive environmental index, performing a second correction to the feeding amount after the first correction, resulting in the final feeding amount; a spatial optimization algorithm is used for feeding location analysis: after gridding the aquaculture area, a predetermined number of grids are selected as feeding points by comprehensively calculating the historical abalone aggregation heat and current water flow diffusion suitability of each grid unit; the final output feeding strategy includes the feeding amount and feeding point coordinates.

[0064] Specifically, G1, determine the growth stage of abalone by establishing an abalone growth model (weight = initial weight × exp(growth coefficient × time)), obtaining the growth coefficient based on historical data regression, and dividing abalone into three stages: juvenile abalone stage, growth stage, and adult abalone stage. Juvenile abalone weight <50g, growth stage weight 50g-100g, and adult abalone weight >100g.

[0065] Visual analysis of uneaten bait quantifies the remaining bait after baiting by image preprocessing (Gaussian filtering for noise reduction and histogram equalization for contrast enhancement), bait region segmentation (establishing a bait pixel classifier based on color features), and uneaten bait rate calculation (uneaten bait rate = current number of bait pixels / initial number of pixels after baiting).

[0066] Water temperature and dissolved oxygen were extracted from environmental data and standardized to form water temperature suitability and dissolved oxygen influence coefficients. Water temperature suitability = 1 - |current water temperature - optimal water temperature| / temperature tolerance range; the optimal water temperature was set based on the growth physiological characteristics of *Abalone spp.* This value is determined based on the median of the optimal temperature range for abalone growth in local aquaculture manuals; the temperature tolerance range is the span of temperature range in which abalone can grow normally but not optimally, and is set as follows: (For example, from) to The dissolved oxygen influence coefficient is calculated as follows: sigmoid((dissolved oxygen - critical dissolved oxygen concentration) × slope), where the critical dissolved oxygen concentration is set to 5 mg / L, and the slope is used to control the sharpness of the change in the dissolved oxygen influence coefficient. Its value is determined by fitting historical data and is set between 1.5 and 3.0. Based on water temperature suitability and the dissolved oxygen influence coefficient, a comprehensive environmental index is obtained through fusion processing, where the comprehensive environmental index = (water temperature suitability + dissolved oxygen influence coefficient) / 2.

[0067] G2. The calculation of feeding amount is a multi-level correction process. The final feeding amount is composed of three parts: the basic feeding amount, the residual feed correction amount, and the environmental correction amount. The basic feeding requirement is obtained by referring to a preset basic feeding requirement table based on the growth stage. This requirement table is defined as follows: 3-5% of body weight for juvenile abalone, 2-3% for growth stage, and 1-2% for adult abalone. Feedback adjustment is based on the residual feed rate: if the residual feed rate is >20%, then the residual feed correction amount = -basic feeding amount × (residual feed rate - 20%), indicating overfeeding, and the amount should be reduced next time; if the residual feed rate is <1%, then the amount should be adjusted accordingly. If the residual feed rate is 0%, then the residual feed correction amount = + base feed amount × (10% - residual feed rate), indicating insufficient feeding, and the amount will be increased next time; if 10% ≤ residual feed rate ≤ 20%, then the residual feed correction amount = 0; adjustment is based on the comprehensive environmental index: environmental correction amount = base feed amount × (comprehensive environmental index - 1); the final feed amount is calculated as: final feed amount = base feed amount + residual feed correction amount + environmental correction amount, with the constraint that the final feed amount ∈ [50% of the base feed amount, 150% of the base feed amount], to prevent calculation errors that lead to overfeeding or underfeeding;

[0068] G3. For feeding locations, the abalone aggregation areas over the past 7 days are statistically analyzed, and feed distribution is predicted based on the current water flow diffusion pattern. The aquaculture area is uniformly divided into 10m × 10m grid units. The feed settling effectiveness index of each grid unit is related to historical aggregation and water flow diffusion, and the calculation formula is: Feed Settling Effectiveness Index = α × Historical Aggregation Heat + β × Water Flow Diffusion Suitability. In the formula, historical aggregation heat represents the frequency of abalone appearance in the grid based on underwater video analysis over the past 7 days, and is normalized; water flow diffusion suitability represents the predicted feed distribution from candidate feeding points within the grid based on the current water flow direction and speed. The settling ratio is calculated, and the higher the value, the more effectively the bait can cover the target area. α and β are weighting coefficients, and α+β=1, used to balance the influence of biological behavior and environmental physics. After calculating the bait settling effectiveness index of all grids, the 3-5 grids with the highest bait settling effectiveness index are selected as the feeding points for this time, ensuring that the points are evenly distributed and avoiding excessive concentration. The Euclidean distance between any two selected points should not be less than 15 meters. Finally, the feeding point coordinate sequence is output, including the amount of bait (accurate to 0.1kg) and the feeding point coordinates (a list of latitude and longitude in the GIS coordinate system, used to identify the center of the recommended feeding area).

[0069] It should be further explained that, in the specific implementation process, in step S3, based on preset collaboration rules and global optimization objectives, a cross-process collaborative work plan is dynamically generated, specifically including:

[0070] Receive key decision information output from the first sub-model, the second sub-model, and the third sub-model, respectively. The key decision information includes:

[0071] One or more optimal cleaning time windows and corresponding cleaning intensity levels from the first sub-model;

[0072] One or more optimal hoisting time windows from the second sub-model;

[0073] The final amount of bait and the coordinates of the baiting point from the third sub-model;

[0074] Based on a global collaboration strategy, key decision-making information is integrated and further optimized to generate the final collaborative work plan; the global collaboration strategy includes:

[0075] Based on the optimal cleaning time window and the optimal hoisting time window, a global time-series optimization algorithm is used to determine the unique execution times for cleaning, hoisting, and feeding operations, ensuring that each operation is optimally time-series at the global level. The global time-series optimization algorithm employs a dynamically weighted multi-objective optimization strategy, constructing an objective function system with total operation time, total energy consumption, and abalone growth interference as optimization dimensions. The weights of each objective are dynamically allocated based on the real-time operation status. The dynamically weighted multi-objective optimization function used in the global time-series optimization... Specifically:

[0076] ;

[0077] In the formula, Estimate the total operation time for all processes (cleaning, hoisting, feeding). Estimate the total energy consumption of all equipment during operation. This represents the sum of quantitative assessments of all abalone growth disturbances. Abalone growth disturbance is obtained by quantifying and weighting the stress responses that may be triggered by various operations (cleaning, hoisting, and feeding). The specific calculation formula is: Abalone growth disturbance = Σ(Stress intensity score of operation type × Duration of operation impact × Weight of impact range). The stress intensity score of the operation type is the stress response intensity score (normalized) taken from the output of the second sub-model for hoisting operations. The duration of operation impact represents the expected duration of a single operation of this type. The weight of the impact range is determined according to the proportion of the number of caissons affected by this operation to the total number. These are dynamic weighting coefficients, and ;

[0078] A particle swarm optimization algorithm with inertial weights is used to solve the multi-objective optimization function to determine the optimal start time of each task and generate a collaborative task time sequence diagram. Each particle is a vector representing the start time sequence of all processes. For example, for three processes (cleaning, hoisting, and baiting), the particle position is... ,in This represents the start time of a process, with each time variable being a continuous value (in hours, starting from 00:00 on the current day); the number of particles is set to 50 to balance computational efficiency and search capability; a linear decreasing strategy is adopted, with initial weight values... Weighted final value The inertia weights are updated according to the following formula during the iteration process: In the formula, Indicates the first Inertia weights in the next iteration This represents the current iteration number. This represents the maximum number of iterations (set to 100).

[0079] Check the space occupancy of all aquaculture equipment at the final execution time. When multiple aquaculture equipment are detected to be planning to use the same aquaculture area in the same time period, dynamically adjust their execution sequence based on the urgency and importance of the aquaculture equipment operations to ensure that only one aquaculture equipment is operating in the same aquaculture area at any given time. For example, the overlapping area of ​​the operation areas of two aquaculture equipment exceeds 20% of their total area, or the safe distance is less than 5m.

[0080] It should be further explained that, in the specific implementation process, in step S4, the central control platform converts the generated collaborative operation plan into control instructions that drive each piece of aquaculture equipment to execute the corresponding procedures, and distributes them to the corresponding aquaculture equipment. Simultaneously, it acquires the execution status data of the aquaculture equipment, specifically including:

[0081] Based on the cleaning operation execution time and cleaning intensity level determined in the collaborative operation plan, control instructions are generated and sent to the cleaning equipment. The control instructions include the start time and the cleaning power corresponding to the cleaning intensity level.

[0082] Based on the hoisting operation execution time set in the collaborative operation plan, control instructions are generated and sent to the hoisting equipment. The control instructions include the hoisting time determined from the optimal hoisting time window output from the second sub-model and the target caisson number to be operated.

[0083] Based on the feeding operation execution time obtained from the collaborative operation plan, control instructions are generated and sent to the feeding equipment. The control instructions include the final feeding amount and feeding point coordinates output from the third sub-model.

[0084] The control command distribution adopts an asynchronous communication mechanism based on message queues to ensure reliable transmission and real-time performance of commands. Simultaneously, it acquires the execution status data of the aquaculture equipment through a predefined state transition mechanism. This state transition mechanism is essentially a finite state machine defined for each piece of aquaculture equipment to monitor its lifecycle. The states include: standby, command receiving, execution, completion, or anomaly. The transition conditions and monitoring actions are as follows: From standby to command receiving: The aquaculture equipment correctly receives and verifies the control command; From command receiving to execution: The aquaculture equipment confirms the command is correct and begins execution; From execution to completion: The aquaculture equipment returns a preset completion signal; From execution to anomaly: The system does not receive a heartbeat signal from the aquaculture equipment within a preset time, the aquaculture equipment reports an error code, or sensor data indicates that the operation effect deviates significantly from expectations (e.g., the dirt coverage does not decrease after cleaning). The anomaly handling process is as follows: If the state transitions to an anomaly, the central control platform immediately initiates a recovery process, including: attempting to resend the command, notifying maintenance personnel, or dynamically adjusting the collaborative operation plan.

[0085] It should be further explained that, in the specific implementation process, in step S5, incremental training adopts an online learning strategy, periodically adding the execution status data of the aquaculture equipment and newly collected on-site monitoring data (subsequent environmental and caisson status data) to the training set, and using the stochastic gradient descent algorithm to update the parameters of the multi-task deep learning model; the training process uses a loss function to weightedly balance historical data and new data to prevent the model from forgetting, wherein the loss function is composed of the weighted sum of the losses of the output layers of each task response, and the weights are dynamically adjusted according to the uncertainty of the prediction results of each task; wherein, the parameters of the multi-task deep learning model specifically refer to: the parameters of the shared feature extraction network include the convolutional kernel weights of the one-dimensional convolutional layer, the input weights, recurrent weights, and bias terms of the LSTM layer; the parameters of the output layers of each task include the cleaning time prediction branch and cleaning intensity prediction branch of the first sub-model, the stress level prediction module of the second sub-model, and the weight matrix and bias vector of all fully connected layers in the feeding amount output layer of the third sub-model.

[0086] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. The focus of each embodiment is on its differences from other embodiments. In particular, the apparatus embodiments are described simply because they are fundamentally based on the method embodiments; relevant details can be found in the descriptions of the method embodiments.

[0087] For ease of description, the above devices are described separately by function as various units. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware.

[0088] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0089] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0090] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0091] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0092] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0093] In conclusion, 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. A cross-process collaborative control method for abalone farming equipment based on deep learning, characterized by: S1. Through the sensor network and vision system deployed in the breeding area, multimodal data related to each process is collected in real time and uploaded to the central control platform for data fusion to generate a collaborative perception dataset across processes. S2. Based on the collaborative sensing dataset, use a multi-task deep learning model to perform parallel prediction and output key decision information; The key decision-making information includes at least the optimal cleaning time window output by the first sub-model, the optimal hoisting time window output by the second sub-model, and the final feeding amount output by the third sub-model. The second sub-model, based on the caisson status, process coordination, and environmental constraints, aims to minimize abalone stress levels and efficiently connect the cleaning process. The process for determining the optimal hoisting time window includes: In the optimal hoisting timing prediction task, the caisson state characteristics, process coordination characteristics, and environmental constraint characteristics are acquired and input into the second sub-model. The second sub-model includes a stress level prediction module and a multi-objective optimization decision module. The stress level prediction module is used to predict the abalone stress level caused by hoisting operations under different hoisting time windows based on the caisson state characteristics, process coordination characteristics, and environmental constraint characteristics, and outputs a stress response intensity score. The multi-objective optimization decision module is used to receive the stress response intensity score and the caisson state characteristics, process coordination characteristics, and environmental constraint characteristics, and optimizes the solution under the condition of maximizing process connection efficiency and minimizing stress response intensity score, with the goal of maximizing process connection efficiency and minimizing stress response intensity score, and outputs the optimal hoisting time window. S3. The central control platform receives key decision information from the multi-task deep learning model and dynamically generates cross-process collaborative operation plans based on preset collaborative rules and global optimization objectives. The generation of cross-process collaborative operation plans includes receiving the optimal cleaning time window and optimal hoisting time window output by the first and second sub-models, using a global time series optimization algorithm to determine the execution sequence of each process, constructing an objective function system, and solving to obtain the collaborative operation plan. S4. The central control platform will convert the generated collaborative operation plan into control instructions to drive each breeding equipment to perform the corresponding process, and distribute them to the corresponding breeding equipment, while acquiring the execution status data of the breeding equipment. S5. Based on the execution status data of the aquaculture equipment and the newly collected on-site monitoring data, the multi-task deep learning model is incrementally trained. The incremental training adopts an online learning strategy, regularly adding the execution status data of the aquaculture equipment and the newly collected on-site monitoring data to the training set, and using the stochastic gradient descent algorithm to update the parameters of the multi-task deep learning model.

2. The cross-process collaborative control method for abalone farming equipment based on deep learning according to claim 1, characterized in that, In step S1, the multimodal data includes environmental data, feeding-related data, caisson status data, and aquaculture equipment status data. Among them, the environmental data includes water temperature, salinity, and dissolved oxygen; the feeding-related data includes feeding amount, feeding time, and the sinking and diffusion of feed observed by an underwater camera; the caisson status data includes image or video data of the surface attachments of the caisson collected by an underwater robot or camera; and the aquaculture equipment status data includes the working status and energy consumption information of the feeder, cleaning robot, and crane.

3. The deep learning-based cross-process collaborative control method for a south- China hard clam (Mercenaria mercenaria) aquaculture equipment group according to claim 1, characterized in that, In step S2, the process of using a multi-task deep learning model to perform parallel prediction and output key decision information includes: The multi-task deep learning model consists of three sub-models, including the first sub-model, the second sub-model, and the third sub-model; The first sub-model is used for the task of predicting the optimal cleaning time and intensity: Based on the multi-source temporal characteristics of bait residue, water quality dynamics and dirt growth, the first sub-model outputs the optimal cleaning time window and cleaning intensity level for the next cleaning operation in parallel. The second sub-model is used for the task of predicting the best hoisting time: Based on the characteristics of the caisson state, process coordination and environmental constraints, the second sub-model aims to minimize the stress level of abalone and efficiently connect the cleaning process to determine the best hoisting time window. The third sub-model is used for the task of dynamically adjusting the feeding strategy: The third sub-model determines the basic feeding amount based on the abalone's growth stage, and then performs two-level corrections by combining the residual feed rate and the comprehensive environmental index to output the final feeding amount; at the same time, it uses spatial optimization to generate the coordinates of the feeding point.

4. The cross-process collaborative control method for abalone farming equipment based on deep learning according to claim 3, characterized in that, The first sub-model, based on the multi-source temporal characteristics of feed residue, water quality dynamics, and fouling growth, outputs the optimal cleaning time window and cleaning intensity level for the next cleaning operation in parallel. In the task of predicting the optimal cleaning time and intensity, multi-source temporal features are acquired, including feed residue features, water quality dynamic features, and fouling growth correlation features. These multi-source temporal features are then input into the first sub-model, which is based on a multi-task deep learning architecture and includes a shared feature extraction network, a cleaning time prediction branch, and a cleaning intensity prediction branch. The cleaning time prediction branch outputs the probability distribution of cleaning demand in future time periods, and the cleaning intensity prediction branch outputs the corresponding cleaning intensity level. The continuous time period in which the probability distribution of cleaning demand exceeds a preset cleaning threshold is determined as the optimal cleaning time window, and the cleaning intensity level is determined based on the output of the cleaning intensity prediction branch.

5. The deep learning-based cross-process collaborative control method for the south China hard clam (Mercenaria mercenaria) aquaculture equipment group according to claim 3, characterized in that, The third sub-model determines the basic feeding amount based on the abalone's growth stage, and then performs two-level corrections by combining the uneaten feed rate and the comprehensive environmental index to output the final feeding amount; simultaneously, the process of generating the feeding point coordinates using spatial optimization includes: In the dynamic adjustment of feeding strategy, the current weight and growth stage of abalone are estimated using an abalone growth model. Based on the growth stage, a preset basic feeding requirement table is consulted to determine the basic feeding amount. The distribution of uneaten feed is analyzed using an underwater camera to calculate the uneaten feed rate. Based on this rate, an uneaten feed correction amount is calculated to make the first correction to the basic feeding amount. Water temperature and dissolved oxygen data are acquired and a comprehensive environmental index is generated. Based on this index, an environmental correction amount is calculated to make a second correction to the feeding amount after the first correction, resulting in the final feeding amount. A spatial optimization algorithm is used to analyze the feeding location: after the aquaculture area is gridded, a predetermined number of grids are selected as feeding points by comprehensively calculating the historical abalone aggregation heat and the current water flow diffusion suitability of each grid unit. The final output feeding strategy includes the feeding amount and the coordinates of the feeding points.

6. The deep learning-based cross-process collaborative control method for southern mud clam (Manila clam) breeding equipment groups according to claim 1, characterized in that, In step S3, the process of dynamically generating a cross-process collaborative work plan based on preset collaboration rules and global optimization objectives includes: Receive key decision information output from the first sub-model, the second sub-model, and the third sub-model, respectively. The key decision information includes: One or more optimal cleaning time windows and corresponding cleaning intensity levels from the first sub-model; One or more optimal hoisting time windows from the second sub-model; The final amount of bait and the coordinates of the baiting point from the third sub-model; Based on a global collaboration strategy, key decision-making information is integrated and further optimized to generate the final collaborative work plan; the global collaboration strategy includes: Based on the optimal cleaning time window and the optimal hoisting time window, a global timing optimization algorithm is used to determine the unique execution time of the cleaning operation, the hoisting operation, and the baiting operation.

7. The method for cross-process collaborative control of abalone farming equipment based on deep learning according to claim 6, characterized in that, The global time-series optimization algorithm adopts a dynamic weighted multi-objective optimization strategy, constructing an objective function system with total operation time, total energy consumption, and abalone growth interference as optimization dimensions, where the weight of each objective is dynamically allocated according to the real-time operation status.

8. The deep learning-based cross-process collaborative control method for southern mud clam (Manila clam) breeding equipment groups according to claim 1, characterized in that, In step S4, the central control platform converts the generated collaborative operation plan into control instructions that drive each piece of aquaculture equipment to execute corresponding procedures, and distributes these instructions to the corresponding aquaculture equipment. Simultaneously, the process of acquiring the execution status data of the aquaculture equipment includes: Based on the cleaning operation execution time and cleaning intensity level determined in the collaborative operation plan, control instructions are generated and sent to the cleaning equipment. The control instructions include the start time and the cleaning power corresponding to the cleaning intensity level. Based on the hoisting operation execution time set in the collaborative operation plan, control instructions are generated and sent to the hoisting equipment. The control instructions include the hoisting time determined from the optimal hoisting time window output from the second sub-model and the target caisson number to be operated. Based on the feeding operation execution time obtained from the collaborative operation plan, control instructions are generated and sent to the feeding equipment. The control instructions include the final feeding amount and feeding point coordinates output from the third sub-model. The control command distribution adopts an asynchronous communication mechanism based on message queues, and the execution status data of the aquaculture equipment is obtained through a predefined state transition mechanism.

Images