Method and System for Generating Animal Feeding Plan Based on Feeding Monitoring Data

Abstract

The present application provides a method and system for generating an animal feeding plan based on feeding monitoring data. The method includes: obtaining the feeding time series data of each animal in the breeding space for the feed storage box; obtaining the health index sampling data of each animal; obtaining the nutritional components corresponding to the feed types of each feed storage box to construct a feed nutrition relationship graph; processing the feed nutrition relationship graph based on a graph neural network to obtain the updated node features of each feed type graph node; using the obtained data to define the input state of the feeding plan generation model, and determining the animal feeding plan corresponding to the target action through reinforcement learning. Thus, through data-driven and combined with advanced graph neural network and reinforcement learning technologies, the feeding amount and feeding time of the feed can be accurately calculated, realizing the intelligent management of the animal feeding process.

Description

Technical Field

[0001] The present application relates to the technical field of intelligent breeding, and particularly to a method and system for generating an animal feeding plan based on feeding monitoring data. Background Art

[0002] In animal breeding and experimental research, scientific and reasonable feeding management has an important impact on the growth, health and metabolic status of animals. Taking the cynomolgus monkey as an example, as an experimental animal commonly used in metabolic disease research (such as obesity, diabetes and cardiovascular diseases), accurate monitoring of its food intake and nutritional requirements is the key to ensuring the reliability of experimental data.

[0003] Most traditional feeding plans rely on static settings, that is, fixed feeding amounts are set according to basic parameters such as the weight, age and gender of animals. For example, assume an adult male cynomolgus monkey weighing 6 kg is set to ingest a certain amount of standard feed per day in a study. Although this method can ensure the basic nutritional needs of animals to a certain extent, it ignores the individual differences in animal feeding behaviors. Specifically, even two cynomolgus monkeys with the same weight may have differences in metabolic rate, activity level and feeding preferences.

[0004] In addition, as a primate, the feeding behavior of cynomolgus monkeys is very complex. Different monkeys may choose different types of food at different time periods, and there are even significant differences in the ingestion speed, frequency and manner of food. Traditional static feeding plans usually cannot take these individual behavior differences into account and can only set fixed feeding times and amounts, resulting in an inability to fully adapt to the natural feeding habits and health needs of cynomolgus monkeys.

[0005] In response to the above problems, the industry has not yet proposed a better technical solution. Summary of the Invention

[0006] The present application provides a method, system, storage medium, computer program product and electronic device for generating an animal feeding plan based on feeding monitoring data, so as to at least solve the problems of ignoring individual differences and insufficient adaptability of feeding plans in traditional feeding plans.

[0007] In a first aspect, an embodiment of the present application provides an animal feeding plan generation method based on feeding monitoring data, including: for each feed storage box in the breeding space, obtaining the feeding time series data of each animal in the breeding space for the feed storage box; each of the feed storage boxes is used to dispense corresponding types of feed, and the feeding time series data includes the feeding time and the feeding amount for the corresponding feed type; obtaining the health index sampling data of each animal; obtaining the nutrient components corresponding to the feed types of each of the feed storage boxes to construct a feed nutrition relationship graph; the feed nutrition relationship graph includes feed graph nodes, nutrition graph nodes, and edge connections, the feed graph nodes are used to indicate the corresponding feed types, the nutrition graph nodes are used to indicate the corresponding types of nutrient components, the edge weight of the feed-nutrition edge connection is defined by the association strength between the feed type and the nutrient component, and the edge weight of the nutrition-nutrition edge connection is defined by the dependence relationship between the nutrient components; aggregating and updating the node features of each feed type graph node in the feed nutrition relationship graph and the feature information of adjacent nodes based on a graph neural network to obtain the updated node features of each feed type graph node; defining the input state of the feeding plan generation model with each of the health index sampling data, the feeding time series data corresponding to each feed storage box, the updated node features, and the current feed surplus, so as to determine the animal feeding plan corresponding to the target action through reinforcement learning; the animal feeding plan includes the planned feeding time and the corresponding planned feeding amount of each feed storage box.

[0008] Second aspect, an animal feeding plan generation system based on feeding monitoring data provided by an embodiment of the present application includes: a feeding data acquisition unit configured to acquire, for each feed storage bin in a breeding space, feeding time series data of each animal in the breeding space for the feed storage bin; each of the feed storage bins is respectively configured to dispense a corresponding type of feed, and the feeding time series data includes the feeding time and the feeding amount for the corresponding feed type; a health data acquisition unit configured to acquire health index sampling data of each of the animals; a nutrition graph construction unit configured to acquire the nutritional components corresponding to the feed types of each of the feed storage bins to construct a feed nutrition relationship graph; the feed nutrition relationship graph includes feed graph nodes, nutrition graph nodes, and edge connections, the feed graph nodes are configured to indicate the corresponding feed types, the nutrition graph nodes are configured to indicate the corresponding types of nutritional components, the edge weight of the feed-nutrition edge connection is defined by the association strength between the feed type and the nutritional component, and the edge weight of the nutrition-nutrition edge connection is defined by the dependence relationship between the nutritional components; a graph structure update unit configured to aggregate and update the node features of each feed type graph node in the feed nutrition relationship graph and the feature information of adjacent nodes based on a graph neural network to obtain updated node features of each feed type graph node; a feeding plan generation unit configured to define the input state of a feeding plan generation model with each of the health index sampling data, the feeding time series data corresponding to each feed storage bin, the updated node features, and the current remaining feed amount, so as to determine an animal feeding plan corresponding to a target action through reinforcement learning; the animal feeding plan includes the planned feeding time and the corresponding planned feeding amount of each feed storage bin.

[0009] Third aspect, an electronic device is provided, which includes: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and when the instructions are executed by the at least one processor, the at least one processor is enabled to execute the steps of the method for generating an animal feeding plan based on feeding monitoring data according to any embodiment of the present application.

[0010] Fourth aspect, an embodiment of the present application provides a storage medium, on which a computer program is stored, wherein when the program is executed by a processor, the steps of the method for generating an animal feeding plan based on feeding monitoring data according to any embodiment of the present application are implemented.

[0011] Fifth aspect, an embodiment of the present application provides a computer program product, including a computer program / instructions, and when the computer program / instructions are executed by a processor, the steps of the method for generating an animal feeding plan based on feeding monitoring data according to any embodiment of the present application are implemented.

[0012] A method and system for generating an animal feeding plan based on feeding monitoring data provided by this application can at least produce the following technical effects:

[0013] (1) By obtaining the feeding time-series data and health indicators of each animal, constructing a feed nutrition relationship graph for each feed storage bin, applying a graph neural network to model and analyze the complex relationship between feeds and nutritional components, and using a reinforcement learning model for intelligent agent optimization calculation, comprehensively applying the deep learning capabilities of the graph neural network and reinforcement learning, realizing data-driven intelligent decision-making, and being able to dynamically optimize the feed delivery time and delivery amount. Thus, the feeding plan can respond in real time to the behavioral changes and health status of animals, significantly improving the flexibility and adaptability of feeding management, and avoiding feed waste or animal malnutrition caused by fixed feeding times and delivery amounts.

[0014] (2) By obtaining real-time feeding time-series data, it can reflect the feeding habits of animals at different time periods, and dynamically optimize the feeding plan in combination with the health indicators of the animals. It not only adapts to the changes in the individual feeding habits of animals, but also can adjust the feeding strategy in a timely manner through real-time monitoring results, providing a more personalized feeding plan to ensure that each animal can receive feed delivery that meets its health needs.

[0015] (3) For each feed storage bin arranged in the feed space, by constructing a feed nutrition relationship graph, systematically modeling the relationship between different feed types and their nutritional components, aggregating and updating the features of feed type nodes and nutritional component nodes through a graph neural network, so as to achieve an in-depth understanding of the complex dependence relationship between feeds and nutritional components. Thus, it is promoted to comprehensively consider the dependence between feed nutritional components when making feed delivery decisions, and intelligently select the most suitable feed type and nutritional ratio according to the health status and feeding behavior of animals, avoiding the negative impact on animal health caused by unbalanced feed nutritional components. In addition, combined with the decision-making mechanism of reinforcement learning, the system intelligently adjusts the strategy in the process of continuously obtaining new feeding and health data to adapt to the growth changes and health conditions of animals, and automatically generates the optimal feeding plan.

[0016] Through this technical solution, by means of data-driven and combined with advanced graph neural network and reinforcement learning technologies, accurately calculating the feed delivery amount and delivery time, realizing the intelligent management of the animal feeding process, being able to effectively meet the personalized nutritional needs of animals, and improving the efficiency and accuracy of feeding management. At the same time, by real-time monitoring the feeding behavior and health indicators of animals, the system can respond and adjust the feeding plan in a timely manner, effectively avoiding the limitations of traditional static feeding plans that cannot adapt to animal health changes in a timely manner. Description of the Drawings

[0017] To more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the accompanying drawings required for the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can also be obtained based on these drawings.

[0018] Figure 1 Fig. 4 shows a flowchart of an example of a method for generating an animal feeding plan based on feeding monitoring data according to an embodiment of the present application;

[0019] Figure 2 Fig. 8 shows a schematic diagram of an external structure design of an example of a feed storage box according to an embodiment of the present application;

[0020] Figure 3 Fig. 12 shows a schematic connection diagram of an example of the internal electronic structure of a feeding trough according to an embodiment of the present application;

[0021] Figure 4 Fig. 16 shows an operation flowchart of an example of obtaining feeding timing data according to an embodiment of the present application;

[0022] Figure 5 Fig. 20 shows an operation flowchart of an example of constructing a feed nutrition relationship diagram according to an embodiment of the present application;

[0023] Figure 6 Fig. 24 shows a schematic connection diagram of an example of the structure of a feeding plan generation model according to an embodiment of the present application;

[0024] Figure 7 Fig. 28 shows a structural block diagram of an example of a system for generating an animal feeding plan based on feeding monitoring data according to an embodiment of the present application;

[0025] Figure 8 Fig. 32 is a schematic structural diagram of an embodiment of an electronic device of the present application. Specific Embodiments

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present application in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are some, but not all, of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present application.

[0027] In the technical solutions of the present application, for the processing of collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved, etc., they all comply with the provisions of relevant laws and regulations and do not violate public order and good customs.

[0028] Figure 1 The flowchart of an example of the method for generating an animal feeding plan based on feeding monitoring data according to an embodiment of the present application is shown.

[0029] Regarding the execution subject of the method according to the embodiment of the present application, it can be any controller or processor with computing or processing capabilities. By integrating advanced data acquisition, graph neural network analysis, and reinforcement learning optimization technologies, it realizes the automated generation of feeding plans, reduces manual intervention and operation errors, and significantly improves the intelligence level and scientific nature of animal feeding management.

[0030] In some examples, it can be integrated and configured in an electronic device, a terminal, or a server in a software, hardware, or software-hardware combination manner, and the types of terminals, electronic devices, or servers can be diverse, such as mobile phones, tablets, or desktop computers, etc.

[0031] As Figure 1 shown, in step S110, for each feed storage bin in the breeding space, obtain the feeding time series data of each animal in the breeding space for the feed storage bin.

[0032] Here, each feed storage bin is respectively used to dispense the corresponding type of feed, and the feeding time series data includes the feeding time and the amount of feed intake for the corresponding feed type.

[0033] In some embodiments, in the breeding space, a feeding monitoring device (such as a feeding image perception system, an RFID tag, an NFC tag, a weight sensor, etc.) is installed for each feed storage bin to record the feeding behavior of each animal for different feed storage bins in real time. In addition, each feed storage bin corresponds to a specific type of feed. For example, a certain storage bin is dedicated to dispensing feed with a high protein content, and another storage bin is used to dispense high-fiber feed. In this way, each time an animal eats, the sensor can accurately record the time stamp of feeding, the duration of each feeding, the type of feed ingested, and the amount of intake, etc.

[0034] In step S120, obtain the health index sampling data of each animal.

[0035] In an example of the embodiment of the present application, the breeder can collect the health index data of the animal by regularly checking the animal, such as measuring the blood glucose and blood lipid of the animal, and input the collected data into the system platform. In another example of the embodiment of the present application, the health index data can be collected by a non-invasive wearable sensing device (for example, a smart collar) to achieve real-time sampling and recording of the animal health index. For example, the heart rate can be monitored in real time through the Photoplethysmography (PPG) technology, the blood glucose and blood lipid of the animal can be monitored in real time through the near-infrared spectroscopy analysis of the spectrum and optical sensors, the weight of the animal can be monitored through the smart weighing scale installed at the bottom of the feeding area, etc., to ensure that the feeding plan can be dynamically adjusted according to the health status of the animal.

[0036] In some embodiments, the feeding behavior data includes at least one of the following: feeding time, feeding frequency, and feeding duration. In addition, the health index sampling data is used to indicate at least one type of health index among the following: weight, heart rate, blood glucose, and blood lipid. By incorporating multi-dimensional feeding behavior data such as feeding time, feeding frequency, and feeding duration into the model, rich behavioral pattern information is provided. In addition, by incorporating various health indexes such as weight, heart rate, blood glucose, and blood lipid during data analysis, the health status of the animal can be more comprehensively reflected.

[0037] In step S130, the nutritional components corresponding to the feed types of each feed storage box are obtained to construct a feed nutrition relationship diagram.

[0038] Here, the feed nutrition relationship diagram includes feed diagram nodes, nutrition diagram nodes, and edge connections. The feed diagram nodes are used to indicate the corresponding feed types, the nutrition diagram nodes are used to indicate the corresponding types of nutritional components, the edge weight of the feed-nutrition edge connection is defined by the association strength between the feed type and the nutritional component, and the edge weight of the nutrition-nutrition edge connection is defined by the dependence relationship between the nutritional components.

[0039] In some embodiments, a corresponding set of nutritional component data, such as protein, fat, carbohydrates, vitamins, and minerals, etc., is obtained by parsing the nutrition table of the feed type. Exemplarily, the user obtains it by uploading the nutrition table to the system platform for parsing or querying the nutrition database based on the feed type.

[0040] Specifically, by constructing a heterogeneous graph structure containing multiple types of graph nodes, feed graph nodes are used to represent different feed types, and nutrient graph nodes are used to represent different nutrient components. There may be edge connections between feed graph nodes and nutrient graph nodes, which can reflect the correlation between feed and nutrient components, such as the proportion of a certain nutrient component in the feed, so as to define the edge weight of the feed-nutrient edge connection. In addition, there may also be edge connections between different nutrient graph nodes, which can reflect the correlation between nutrient components, such as synergistic or antagonistic relationships. Exemplarily, the synergistic relationship between protein and amino acids, the synergistic proportion relationship between calcium and phosphorus, vitamin D affects calcium absorption, and there is a competitive absorption phenomenon between iron and zinc, and so on. Furthermore, by referring to nutritional research knowledge or sample experimental data, such synergistic or antagonistic relationships are quantified as the edge weights of the corresponding nutrient-nutrient edge connections.

[0041] In step S140, based on the graph neural network, the node features of each feed type graph node in the feed-nutrient relationship graph and the feature information of adjacent nodes are aggregated and updated to obtain the updated node features of each feed type graph node.

[0042] Here, in the Graph Neural Network (GNN), through the message passing mechanism, each graph node (feed type and nutrient component) aggregates and updates features based on the information of its adjacent nodes to achieve node feature update. For example, the features of the feed type node will consider the feature information of the adjacent nutrient component nodes, thereby updating its own features, so that these updated node features can reflect the overall nutritional value of the feed.

[0043] In step S150, each health indicator sampling data, the feeding time series data corresponding to each feed storage bin, the updated node features, and the current feed surplus are used to define the input state of the feeding plan generation model, so as to determine the animal feeding plan corresponding to the target action through reinforcement learning.

[0044] Here, the animal feeding plan includes the planned feeding time and the corresponding planned feeding amount of each feed storage bin.

[0045] It should be noted that the feeding plan generation model can adopt various types of reinforcement learning models, such as Q-learning, deep Q-network, etc., which can generate the optimal target actions (planned feeding time and the corresponding planned feeding amount) according to the current input state (such as animal health indicators, the feeding time series data corresponding to each feed storage bin, the updated node features, and the current feed surplus).

[0046] In the embodiments of the present application, the Agent is the part responsible for generating the feeding plan, which can make decisions based on the input status (i.e., the animal feeding plan). The Agent operates in the Environment and obtains feedback from the Environment. Here, the Environment contains a state space defined by information such as health index sampling data, feeding time series data, node features, and feed surplus.

[0047] It should be understood that the design of the Reward function can be diversified and can be set or adjusted according to business requirements. In some examples, the Reward function can be comprehensively set according to factors such as the health status of the animal or the effective utilization rate of the feed. Exemplarily, if the health index of the animal improves after the execution of the feeding plan, the Agent will obtain a positive reward or a greater reward feedback; if the feeding plan is unreasonable, resulting in feed waste or a decline in the health status of the animal, the Agent will obtain a negative reward or a smaller reward feedback.

[0048] In some embodiments, as factors such as the health status of the animal, feeding behavior, or feed surplus change in real time, the system can adjust the output of the reinforcement learning model, automatically update the feeding time and corresponding feeding amount of different feed storage bins, so as to adaptively meet the personalized feeding needs of the animals.

[0049] Figure 2 The figure shows a schematic diagram of the external structure design of an example of a feed storage bin according to an embodiment of the present application.

[0050] As Figure 2 shown, the feed storage bin adopts a multi-layer box structure. The feed storage bin includes an upper box body 21 and a lower box body 22. The upper box body 21 is a feeding area provided with a feeding trough 211, and the lower box body is used to receive the animal feces from the feeding area through the grid through holes at the bottom of the upper box body. Thus, the food and urine feces of the animal can be effectively separated, while ensuring the diet hygiene of the animal, the food intake can be calculated more accurately. In addition, generally, the breeder only needs to clean the feces in the lower box body 22, which can also reduce the workload burden on the breeder to a certain extent. Preferably, as Figure 2 the multiple feed storage bins in

[0051] Figure 3 The figure shows a schematic connection diagram of an example of the electronic structure inside the feeding trough according to an embodiment of the present application.

[0052] As Figure 3As shown, the internal electronic structure of the feeding trough 300 includes an RFID reader 310, a weighing sensor 320, and a signal transmitter 330. In some embodiments, the RFID reader 310 can be designed on the side close to the animal activity area to sense the unique RFID tag worn by each animal and identify the target animal in the feeding area in real time. In addition, each RFID reader equipped with a feeding trough has a uniquely corresponding reader device number, so as to realize the inductive identification of the feed type of animal feeding.

[0053] The weighing sensor 320 is designed as the bottom tray of the feeding trough, which is used to sense the weight change of the contained feed in real time and record the corresponding feed consumption and consumption time.

[0054] The signal transmitter 330 is used to upload the data collected by the RFID reader 310 and the weighing sensor 320 to the animal feeding management platform. For example, the MQTT protocol or other lightweight protocols suitable for Internet of Things devices can be used to achieve efficient data transmission.

[0055] Figure 4 The operation flowchart of an example for obtaining feeding timing data according to an embodiment of the present application is shown.

[0056] As Figure 4 shown, in step S410, based on the RFID reader disposed inside the feeding trough, the RFID tags of each animal are sensed to monitor the target animals in the feeding area in real time.

[0057] Specifically, taking the cynomolgus monkey as an example, each cynomolgus monkey is equipped with a unique RFID tag, which contains the unique identification code (Unique Identification Data, UID) of the animal and related basic information (such as number, gender, age, health status, etc.). The RFID tag is securely fixed on the collar or other parts of each cynomolgus monkey to ensure that the tag is not easily detached or damaged during the animal's feeding process. When the cynomolgus monkey is in the feeding area close to the feeding trough, the RFID reader detects and records the UID, entry time, and exit time of the animal entering the feeding area in real time. By matching the read UID with the animal information in the database, the accuracy of the data and the reliability of individual identification are ensured.

[0058] In step S420, based on the weighing sensor disposed inside the feeding trough, the feed consumption and feed consumption time of the corresponding feed type are monitored in real time to determine the feeding timing data of the corresponding target animal.

[0059] Here, a strain gauge sensor or other electronic weighing sensors that support high precision and fast response can be selected, and the weighing sensor can be installed on the bottom tray of the feeding trough to ensure accurate measurement of the feed consumption.

[0060] Thus, the individual feeding behavior monitoring is realized through the RFID sensing technology, ensuring the independent collection and analysis of the feeding data of each animal. Through the high-precision weighing sensor, the high accuracy and reliability of the feeding amount and feeding time data are ensured, providing a reliable data basis for realizing personalized feeding management.

[0061] Figure 5 The operation flowchart showing an example of constructing a feed nutrition relationship diagram according to an embodiment of the present application is shown.

[0062] As Figure 5 shown, in step S510, based on the feed nutrition component database, the nutrient content of the feed types in each of the feed storage bins and the bioavailability and nutrient metabolism experimental data corresponding to each nutrient are determined.

[0063] Here, the feed nutrition component database can be a self-built database or a third-party database, and can be constructed by integrating scientific literature, laboratory test results, data provided by feed manufacturers, etc. By collecting the nutrient content data of various feeds, including the specific contents of protein, fat, carbohydrates, vitamins, minerals, etc. The bioavailability data of each nutrient is investigated, that is, the proportion actually absorbed and utilized in the animal body. The nutrient metabolism experimental data related to animal health and metabolic functions are collected. For example, through animal metabolism experiments, the bioavailability of each nutrient is measured, which can also include indicators such as metabolic rate, energy consumption, and nutrient absorption efficiency. In some examples, the nutrient metabolism experimental data can also include experimental samples considering the impact of nutrient intake on various animal health indicators, which helps to reveal the relationship between nutrient intake and animal health.

[0064] In step S520, based on the nutrient content of each feed type and the bioavailability and nutrient metabolism experimental data corresponding to each nutrient, a feed nutrition relationship diagram is constructed.

[0065] Here, the node features of the feed diagram nodes are defined by the feed name, feed classification, and feed nutrient content of the corresponding feed type, and the node features of the nutrient diagram nodes are defined by the nutrient index name, nutrient classification, nutrient metabolism experimental data, and bioavailability of the corresponding nutrient.

[0066] Specifically, a dual graph structure is constructed, including feed graph nodes and nutrient graph nodes. The graph nodes are connected by edges, representing the relationship between feeds and nutrient components or the relationship between nutrient components. Each feed type corresponds to a node, and the node features include feed name, feed classification, content of each nutrient component, etc. Each nutrient component corresponds to a node, and the node features include nutrient index name (such as protein, fat, etc.), nutrient classification, relevant nutrient metabolism experiment data (such as the impact on metabolic rate or health indicators), biological utilization rate, etc.

[0067] In some examples of the embodiments of the present application, the feed classification includes at least one of the following: plant-based feed, animal-based feed, energy feed, fiber feed, and protein feed. By introducing more refined feed classification and nutrient classification in the definition of graph nodes, the features of each feed type node not only reflect the nutrient content but also characterize its nutrient function orientation and biological role through category identification. The nutrient classification includes at least one of the following: macronutrients, micronutrients, electrolytes, and essential amino acids. Macronutrients provide a large amount of energy required by animals daily, including protein, fat, and carbohydrates. The demand for micronutrients by animals is relatively small, but they are crucial for health and physiological functions, including vitamins and minerals. Electrolytes such as sodium, potassium, calcium, and magnesium maintain fluid balance and nerve conduction. Essential amino acids such as lysine and tryptophan cannot be synthesized by animals themselves and must be ingested through diet. In this way, the nutrient node features are more systematic and interpretable, making the feature representation of graph nodes no longer a simple nutrient content vector but a multi-dimensional feature set organically integrated by category, functional attributes, and physiological significance, greatly enriching the available information for the model to process the same type of nodes in the heterogeneous graph.

[0068] Furthermore, the edge weight connected by the feed-nutrient edge is calculated by the following formula:

[0069] w(F i ,N j )=α1·[C(F i ,N j )] γ ·Imp(N j )·B(N j ), Equation (1)

[0070]

[0071] In the formula, F i represents the feed type corresponding to the i-th feed graph node, N j represents the nutrient component corresponding to the j-th nutrient graph node, w(F i ,N j ) represents the edge used to connect F i and N jThe edge weight of the edge connected to the corresponding graph node; C(F i ,N j ) represents F i in N j 's normalized content, Imp(N j ) represents N j 's importance factor for the health and metabolic functions of animals, B(N j ) represents N j 's bioavailability factor, γ represents the non-linear parameter used to control the sensitivity of nutrient content, α1 is the first metric global scaling factor; H = {H1, H2, …, H m} represents an array containing m animal health indicators; H a represents the ath health indicator, whose value is normalized to [0, 1], and the larger the value, the better the health condition; Corr(N j ,H a ) represents the correlation metric between the intake level of N j and H a , whose value range is [-1, 1], a positive value indicates a positive correlation between intake and health indicators, and a negative value indicates a negative correlation between intake and health indicators; Q is the total number of experimental samples in the nutrient metabolism experimental data, where the measured intake of nutrient component N j in the qth sample is N j (q), and the measured performance value of H a in the qth sample is H a (q), is the mean value of all experimental samples corresponding to N j (q), is the mean value of all experimental samples corresponding to H a (q); Pos(Corr(N j ,H a )) represents mapping the correlation metric Corr(N j ,H a ) to the [0, 1] interval of the corresponding positive score; is the weight of the health indicator H a , indicating the relative importance of this indicator to overall health.

[0072] It should be noted that the importance factor of a nutrient component reflects the degree of association between the change in the intake of this nutrient component and multi-dimensional health indicators (such as weight stability, heart rate healthy range, blood sugar stability, normal blood lipid level, etc.). Specifically, for each nutrient component N j , by analyzing the experimental sample data, calculate its correlation with each health indicator H a respectively, and map it to a positive score. Furthermore, use the weight w aPerform weighted summation to obtain the corresponding importance factor Imp(N j ). If N j has a positive impact on multiple key health indicators (the correlation coefficient is positive and large), then Imp(N j ) will be relatively high; if there is a negative correlation, this value will decrease.

[0073] Here, the Pearson Correlation Coefficient is used to measure the correlation, which can measure the strength and direction of the linear relationship between two variables. When Corr(N j ,H a ) is close to 1, it indicates a significant positive correlation between the nutrient intake and this health indicator (the health indicator tends to improve when the nutrient intake increases). When Corr(N j ,H a ) is close to -1, it indicates a significant negative correlation (the health indicator tends to deteriorate when the nutrient intake increases). When Corr(N j ,H a ) is close to 0, it indicates that there is no significant linear relationship between the two.

[0074] Furthermore, to avoid the over-suppression effect of negative correlation on the calculated importance factor, the Pos(·) function is used to map the correlation to the [0,1] interval. By analyzing the historical sample experimental data, calculate the linear correlation degree between the nutrient intake level and the health indicator, so that the model can comprehensively weigh the response degree of multiple health dimensions to the intake of a single nutrient. The system can dynamically allocate feed components according to different health focuses, which helps to generate more targeted and personalized nutrition plans.

[0075] The edge weight of the nutrient-nutrient edge connection is calculated by the following formula:

[0076] w(N j ,N k ) = α2·exp(λ·Rel(N j ,N k ))·[B(N j )·B(N k )] δ , Equation (5)

[0077] Rel(N j ,N k ) = tanh(BaseRel(N j ,N k ))), Equation (6)

[0078]

[0079] In the formula, Nk represents the nutrient component corresponding to the k-th nutrient graph node, w(N j ,N k ) represents the edge weight of the edge connection used to connect the graph nodes corresponding to N j and N k ; Rel(N j ,N k ) represents the metabolic relatedness between N j and N k ; B(N k ) is the bioavailability factor of N k ; δ is a parameter that controls the sensitivity of the relationship between nutrients to bioavailability; λ is a non-linear mapping parameter used to map Rel(N j ,N k ) to the positive value space; α2 is the second global scaling factor; exp(·) represents the exponential function; BaseRel(N j ,N k ) represents the quantified value of the synergistic or antagonistic effect of the nutrient components of N j and N k ; Corr((N j ,N k ),H a ) represents the correlation measure considering the combined intake level of two nutrient components N j and N k and the health indicator H a ; Corr(N j ,H a ) and Corr(N k ,H a ) represent the correlation measures of the individual intake of N j and N k with the health indicator H a ; is the weight of the health indicator H a when determining the relationship between nutrients.

[0080] It should be noted that the relationship between nutrient components can be manifested as a synergistic or antagonistic effect, which is determined by analyzing historical experimental data to determine the difference in the impact on health indicators when two nutrient components are ingested simultaneously compared to individual ingestion.

[0081] Specifically, calculate the correlation of the health indicator when two nutrients are ingested simultaneously, Corr((N j ,N k ),H a ), and then calculate the average contribution value relative to H a when ingested individually If Corr((N j ,Nk ), H i ), if it is significantly higher than this average value, it indicates a synergistic effect; if it is lower than this average value, it indicates an antagonistic effect. Since this value can be negative (antagonistic) or positive (synergistic). Further, it is necessary to map it to the interval [-1, 1] and distinguish the positive and negative relationships, and then use the hyperbolic tangent function (tanh) to smoothly map the value. If BaseRel(N j , N k ) >> 0, then Rel(N j , N k ) is close to 1, indicating a strong synergy; if BaseRel(N j , N k ) << 0, then Rel(N j , N k ) is close to -1, indicating a strong antagonism.

[0082] Thus, when two nutrient components are ingested simultaneously and are significantly superior to single ingestion in multi-dimensional health indicators, BaseRel(N j , N k ) is positive and large, and after being mapped by the hyperbolic tangent function, it is close to 1, indicating an obvious synergy. When the simultaneous ingestion of two nutrient components shows significantly worse performance than single ingestion, the value is negative and large, and after mapping, it is close to -1, indicating a strong antagonism. If the relationship between the two is not significant, the value will hover around 0. Thus, it is possible to identify and quantify the synergistic or antagonistic effects between nutrient components from historical experimental data, enabling the system to more precisely consider the complex interactions between nutrient components when formulating feed plans, thereby avoiding misjudgments caused by simple linear summation and ultimately improving animal health and feeding efficiency.

[0083] Through the embodiments of the present application, by using the constructed feed-nutrient relationship graph and corresponding parameter calculation formulas (such as feed-nutrient edge weight, nutrient-nutrient relationship metric, nutrient component importance factor, etc.), the originally static, experience-dominated feeding strategy is transformed into a data-driven decision-making process. By performing feature aggregation and update on the information extracted from the feeding time series data, physiological index sampling data, and feed nutrient component database, the ability to automatically extract core nutrient relationship features from large-scale complex data is technically realized, thereby providing a basis for subsequent feeding plan decision analysis.

[0084] In some examples of the embodiments of the present application, the graph neural network adopts a graph attention network (Graph Attention Network, GAT), which dynamically assigns weights to neighbor nodes through an attention mechanism and integrates category information and edge weights to aggregate and update node features.

[0085] Specifically, for each pair of adjacent graph nodes, calculate the attention score:

[0086] e uv = LeakyReLU(Q T · [WX u ||WX v ||C uv )), Equation (8)

[0087] In the formula, u and v represent adjacent graph nodes u and v, W and Q represent the learnable linear transformation matrix and the attention parameter vector respectively, || represents the vector concatenation operation, X u and X v represent the node features of graph nodes u and graph node v respectively; C uv is the information vector corresponding to the feed classification or nutrition classification, which uses one-hot encoding; LeakyReLU(·) represents the LeakyReLU activation function.

[0088] Here, the feed classification or nutrition classification information of the graph nodes is incorporated into the attention calculation in the form of one-hot encoding, enabling the relationship between different classification nodes to be distinguished when the model calculates the attention weights, enhancing the expression ability of the node features, thereby improving the pertinence and effectiveness of feature aggregation, and helping the system to better identify and utilize the characteristics of different feeds.

[0089] The edge weight is incorporated into the attention score to form a weighted attention score:

[0090] e' uv = e uv · w uv ), Equation (9)

[0091] In the formula, e' uv represents the weighted attention score, and w uv represents the edge weight of the edge connection between graph node u and graph node v.

[0092] Here, the pre-computed edge weight w uv of the edge connection is incorporated into the attention score calculation formula, so that in the feature aggregation process, not only the similarity of node features is considered, but also the association strength between feeds and nutritional components and the metabolic relevance between nutritional components are reflected.

[0093] Normalize the attention score to obtain the weighted category attention weight:

[0094]

[0095] In the formula, α uv represents the weighted category attention weight between graph node u and graph node v, p ∈ N(u) represents any graph node p in the set N(u) of adjacent graph nodes of graph node u, e' upRepresents the weighted category attention weight between graph node u and graph node p.

[0096] Weighted aggregation of the node features of neighbor nodes is performed based on the weighted category attention weight to update the node features:

[0097]

[0098] In the formula, σ represents a non-linear activation function (e.g., RELU activation function, etc.), and H’ u Represents the updated node feature of graph node u.

[0099] As a preferred implementation manner of the embodiment of the present application, multiple layers of GATs with the above structure can also be stacked to capture deeper node relationships and feature representations.

[0100] Through the embodiment of the present application, the weights of neighbor nodes are adaptively allocated based on the attention mechanism, and category information and pre-computed edge weights are incorporated to organically integrate the multi-dimensional features of feed and nutrition, improving the expressiveness and scientific nature of node features, being able to generate personalized features for different feed types, and supporting the formulation of individualized feeding strategies. In addition, based on the attention weight mechanism, interpretability of the feed-nutrition relationship is provided, helping managers understand the feature update process and enhancing the transparency and trust of the system.

[0101] Figure 6 Shows a schematic structural connection diagram of an example of a feeding plan generation model according to an embodiment of the present application.

[0102] As Figure 6 shown, the feeding plan generation model 600 adopts a hierarchical reinforcement learning model, which includes an attention fusion layer 610, a high-level policy network 620, and a low-level policy network 630.

[0103] The attention fusion layer 610 is used to fuse each of the health indicator sampling data with the feeding time series data corresponding to each of the feed storage bins, the updated node features, and the current feed surplus through a self-attention mechanism to obtain an attention fusion feature.

[0104] F attended = Attention(MLP(H t ), MLP(F t ), MLP(G t ), MLP(S t ))), Equation (12)

[0105] In the formula, F attended represents the attention fusion feature, Attention(·) represents the self-attention mechanism, H t represents the health indicator at time step t, Ft Represents the feeding timing data at time step t, G t Represents the updated node features at time step t, S t Represents the current feed surplus at time step t, and MLP(·) represents the multi-layer perceptron processing function.

[0106] Here, a multi-layer perceptron (Multilayer Perceptron, MLP) is used to fuse features from different sources to extract high-dimensional feature representations. Specifically, the model projects health indicator data (H t ), feeding timing data (F t ), graph node features (G t ), current feed surplus (S t ), etc. into a unified feature space through a multi-layer perceptron.

[0107] Furthermore, using the self-attention mechanism (Attention), according to the relevance of each feature to the current decision, more important information is automatically selected and given higher weights, and then they are weighted and fused. The resulting weighted fusion feature F attended is a comprehensive representation that adaptively weights and integrates multiple heterogeneous data to ensure that decisions are made based on the most relevant and useful information.

[0108] Define the input state based on the attention fusion feature to trigger the high-level policy network 620 to determine the corresponding feed plan delivery time.

[0109] T t = Softmax(W m ·F attended + b m ), Equation (13)

[0110] In the formula, T t represents the time step decision vector of feed delivery output by the high-level policy network, which is a vector of length K, and each element represents the probability of feeding at the corresponding time step; Softmax(·) represents the softmax activation function; W m and b m represent the weight matrix and bias vector of the high-level policy network respectively.

[0111] Here, the high-level policy network takes the attention fusion feature as the input state and generates a policy output T t through a series of linear and non-linear transformations, which expresses a probability distribution of "when to replenish feed to each feed storage bin next", and is used to decide when to execute the delivery action.

[0112] Input the attention fusion feature and the feed plan delivery time into the low-level policy network 630 to determine the corresponding feed plan delivery quantity.

[0113] F worker_input = [F attended ||T t , Equation (14)

[0114]

[0115]

[0116] A t = N(μ, σ 2 ), Equation (19)

[0117] Where, F worker_input represents the input feature vector of the low-level policy network, and respectively represent the weight matrix, bias vector and output feature of the first hidden layer of the low-level policy network, and respectively represent the weight matrix, bias vector and output feature of the second hidden layer of the low-level policy network, W μ and b μ respectively represent the weight matrix and bias vector of the output mean layer of the low-level policy network, W σ and b σ respectively represent the weight matrix and bias vector of the output standard deviation layer of the low-level policy network; μ represents the mean vector of the delivery quantity of each feed storage bin, σ represents the standard deviation vector of the delivery quantity of each feed storage bin, Softplus(·) represents the Softplus activation function; N(μ, σ 2 ) represents the normal distribution, which is used to sample continuous actions; A t represents the feed delivery quantity vector of each feed storage bin at time step t.

[0118] Here, the model concatenates the time decision result output by the high-level with the attention fusion feature of the input state to form the input of the low-level policy network, which performs a non-linear transformation and multi-layer processing on this input. Specifically, through two hidden layers activated by ReLU and finally outputs two distribution parameters for the "feed delivery quantity": μ and σ, where μ and σ respectively represent the mean and standard deviation of the delivery quantity. Finally, sample A 2 from the normal distribution N(μ, σ t ), use the normal distribution to parameterize the policy, ensure the continuity and differentiability of the output action, and the low-level policy network makes the final decision output, that is, finally determines how much feed should be delivered to each feed storage bin at this moment.

[0119] Through the embodiments of the present application, the high-level policy network and the low-level policy network hierarchically process decision-making tasks. The high-level policy first determines when to deliver (i.e., time decision), and on this basis, the low-level policy further makes a refined decision on the feed delivery amount. This hierarchical structure reduces the complexity of the decision space, enabling the model to maintain efficient learning and adaptation capabilities in high-dimensional decision-making problems (dual variables of time and delivery amount). In addition, the reinforcement learning model learns through continuous interaction and trial and error, and dynamically adapts the decision to the real-time behavior and health status changes of animals by updating the policy parameters. When external conditions (such as animal activity level, fluctuations in health indicators, changes in feed composition) change, the model can automatically adjust the feeding timing and feeding amount, and has good environmental adaptability and real-time feedback capabilities.

[0120] In some examples, the feeding plan generation model uses the PPO (Proximal Policy Optimization) algorithm as the policy optimization algorithm. Specifically, by interacting the current policy network with the environment, data on states, actions, rewards, and the next state are collected, and training is carried out according to the training process of the PPO algorithm. The reward changes and the stability of policy updates during the training process are monitored to ensure the convergence of the model.

[0121] Compared with traditional policy gradient methods, the PPO algorithm restricts the difference between the old and new policies when updating the policy parameters (i.e., the "proximal" constraint), so that the policy will not deviate from the optimal direction due to overly radical single updates during the training process, thereby reducing the instability during the training process and the risk of policy collapse. In addition, in the generation of feeding plans, the experimental and sampling costs are relatively high. PPO can improve the data usage efficiency by repeatedly using the same batch of sample data for updates, which helps to obtain better policy performance under the condition of less actual test resources.

[0122] In some examples of the embodiments of the present application, the reward function of the feeding plan generation model is:

[0123]

[0124] In the formula, R t represents the reward value at time step t, which is used to evaluate the pros and cons of taking action A t at time step t; ΔH t represents the improvement degree of the health indicator after time step t, and a g (t) represents the feed delivery amount of the gth feed storage bin at time step t, and S g(t) represents the current remaining feed in the g-th feed storage bin at time step t, G represents the total number of feed storage bins in the feeding space, and β1, β2, and β3 represent the weight coefficients of the health index improvement term, the total feed input penalty term, and the feed remaining change penalty term, respectively.

[0125] Here, the reward function adopts a multi-objective design. Through multi-objective optimization, it ensures that the model can maintain good decision-making performance under various conditions, and the model shows stronger robustness and stability when facing different feeding environments and data changes.

[0126] Specifically, through the health index reward term (β1·ΔH t ) in the reward function, by real-time monitoring and quantifying the changes in the health index, the feeding strategy is dynamically adjusted, encouraging the model to preferentially select feeding plans that can significantly improve the animal's health index, providing personalized nutrition plans for animals in different health states, ensuring that animals obtain the required nutritional support, and promoting healthy growth.

[0127] Through the feed input reward term in the reward function By quantifying the impact of feed input on the reward, the model is guided to find the most economical feeding plan, encouraging the model to minimize the feed input as much as possible on the premise of meeting the health needs, effectively reducing the feed cost and improving the feed utilization efficiency.

[0128] Through the feed remaining change control reward term in the reward function By controlling the fluctuation of the feed remaining, punishing the drastic change of the feed remaining, encouraging the model to maintain a stable feed inventory, supporting the realization of more accurate feed management, and ensuring the stability of the feed inventory remaining.

[0129] Through the reward function provided by the embodiments of the present application, through quantifiable multi-dimensional reward terms, it clearly reflects key factors such as health index improvement, feed input, and feed remaining change, making the decision-making basis of the system clear and transparent, enhancing the user's trust and acceptance of the system, and also helping the user understand and trust the decision-making process of the system. In addition, β1, β2, and β3 are weight coefficients used to balance different reward terms, which can balance the importance of different reward terms and can be adjusted according to the actual feeding needs. For example, when paying more attention to animal health, the value of β1 can be appropriately increased; when emphasizing feed cost more, the value of β2 can be appropriately increased; and when hoping to maintain the balance of feed reserves more, the value of β3 can be appropriately increased. Thus, through the adjustment of the weight coefficients, the system can flexibly switch the focus according to the actual feeding needs to meet the needs of diverse feeding scenarios.

[0130] It should be noted that, for the foregoing method embodiments, for the sake of simple description, they are all expressed as a series of combined actions. However, those skilled in the art should be aware that the present application is not limited by the described action sequence, because according to the present application, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the present application. In the above embodiments, each embodiment is described with its own emphasis. For parts not detailed in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0131] Figure 7 FIG. 4 shows a structural block diagram of an example of an animal feeding plan generation system based on feeding monitoring data according to an embodiment of the present application.

[0132] As Figure 7 shown, the animal feeding plan generation system 700 based on feeding monitoring data includes a feeding data acquisition unit 710, a health data acquisition unit 720, a nutrition graph construction unit 730, a graph structure update unit 740, and a feeding plan generation unit 750.

[0133] The feeding data acquisition unit 710 is configured to obtain, for each feed storage box in the breeding space, the feeding time series data of each animal in the breeding space for the feed storage box; each of the feed storage boxes is used to dispense a corresponding type of feed, and the feeding time series data includes the feeding time and the feeding amount for the corresponding feed type.

[0134] The health data acquisition unit 720 is configured to obtain the health index sampling data of each of the animals.

[0135] The nutrition graph construction unit 730 is configured to obtain the nutritional components corresponding to the feed types of each of the feed storage boxes to construct a feed nutrition relationship graph; the feed nutrition relationship graph includes feed graph nodes, nutrition graph nodes, and edge connections. The feed graph nodes are used to indicate the corresponding feed types, the nutrition graph nodes are used to indicate the corresponding types of nutritional components, the edge weight of the feed-nutrition edge connection is defined by the association strength between the feed type and the nutritional component, and the edge weight of the nutrition-nutrition edge connection is defined by the dependence relationship between the nutritional components.

[0136] The graph structure update unit 740 is configured to aggregate and update the node features of each feed type graph node in the feed nutrition relationship graph and the feature information of adjacent nodes based on a graph neural network to obtain the updated node features of each feed type graph node.

[0137] The feeding plan generation unit 750 is configured to define the input state of the feeding plan generation model by using each of the health index sampling data, the feeding time series data corresponding to each of the feed storage bins, the updated node features, and the current remaining feed amount, so as to determine the animal feeding plan corresponding to the target action through reinforcement learning; the animal feeding plan includes the planned feeding time and the corresponding planned feeding amount of each feed storage bin.

[0138] In some embodiments, the embodiments of the present application provide a non-volatile computer-readable storage medium, in which one or more programs including execution instructions are stored, and the execution instructions can be read and executed by an electronic device (including but not limited to a computer, a server, or a network device, etc.) to execute the steps of any one of the above-mentioned animal feeding plan generation methods based on feeding monitoring data of the present application.

[0139] In some embodiments, the embodiments of the present application further provide a computer program product, which includes a computer program stored on a non-volatile computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer is caused to execute the steps of any one of the above-mentioned animal feeding plan generation methods based on feeding monitoring data.

[0140] In some embodiments, the embodiments of the present application further provide an electronic device, which includes: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the steps of the animal feeding plan generation method based on feeding monitoring data.

[0141] Figure 8 is a schematic hardware structure diagram of an electronic device for executing the animal feeding plan generation method based on feeding monitoring data provided by another embodiment of the present application, as Figure 8 shown, the device includes:

[0142] one or more processors 810 and a memory 820, Figure 8 Taking one processor 810 as an example.

[0143] The device for executing the animal feeding plan generation method based on feeding monitoring data may further include: an input device 830 and an output device 840.

[0144] The processor 810, the memory 820, the input device 830, and the output device 840 may be connected through a bus or other means, Figure 8 Taking connection through a bus as an example.

[0145] The memory 820, being a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the program instructions / modules corresponding to the animal feeding plan generation method based on feeding monitoring data in the embodiments of the present application. The processor 810 executes various functional applications and data processing of the server by running the non-volatile software programs, instructions, and modules stored in the memory 820, that is, implements the animal feeding plan generation method based on feeding monitoring data in the above method embodiments.

[0146] The memory 820 may include a program storage area and a data storage area. Among them, the program storage area can store an operating system and application programs required for at least one function; the data storage area can store data created according to the use of the electronic device, etc. In addition, the memory 820 may include high-speed random access memory, and may also include non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other non-volatile solid-state storage devices. In some embodiments, the memory 820 may optionally include memories remotely provided with respect to the processor 810, and these remote memories can be connected to the electronic device through a network. Examples of the above network include but are not limited to the Internet, an enterprise intranet, a local area network, a mobile communication network, and combinations thereof.

[0147] The input device 830 can receive input digital or character information, and generate signals related to the user settings and function control of the electronic device. The output device 840 may include a display device such as a display screen.

[0148] The one or more modules are stored in the memory 820, and when executed by the one or more processors 810, execute the animal feeding plan generation method based on feeding monitoring data in any of the above method embodiments.

[0149] The above product can execute the method provided in the embodiments of the present application, and has the corresponding functional modules and beneficial effects of the executed method. For technical details not described in detail in this embodiment, reference can be made to the method provided in the embodiments of the present application.

[0150] The electronic device in the embodiments of the present application exists in various forms, including but not limited to:

[0151] (1) Mobile communication devices: These devices are characterized by having mobile communication functions and mainly aim to provide voice and data communication. Such terminals include: smart phones, multimedia phones, functional phones, and low-end phones, etc.

[0152] (2) Ultra-mobile personal computer devices: These devices belong to the category of personal computers, have computing and processing functions, and generally also have the characteristic of mobile Internet access. Such terminals include: PDAs, MIDs, and UMPC devices, etc.

[0153] (3) Portable entertainment devices: Such devices can display and play multimedia content. This type of device includes: audio and video players, handheld game consoles, e-books, as well as smart toys and portable in-vehicle navigation devices.

[0154] (4) Other on-board electronic devices with data interaction functions, such as in-vehicle device installed on a vehicle.

[0155] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0156] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the above technical solutions in essence or the part that makes contributions to the related technologies can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0157] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of each embodiment of the present application.

Claims

1. A method for generating an animal feeding plan based on feeding monitoring data, comprising: For each feed storage bin in the breeding space, obtaining the feeding time series data of each animal in the breeding space for the feed storage bin; each of the feed storage bins is used to dispense corresponding types of feed, and the feeding time series data includes the feeding time and the feeding amount for the corresponding feed type; Obtaining the health index sampling data of each of the animals; Obtaining the nutritional components corresponding to the feed types of each of the feed storage bins respectively to construct a feed nutrition relationship graph; The feed nutrition relationship graph includes feed graph nodes, nutrition graph nodes and edge connections. The feed graph nodes are used to indicate the corresponding feed types, the nutrition graph nodes are used to indicate the corresponding types of nutritional components, the edge weight of the feed-nutrition edge connection is defined by the association strength between the feed type and the nutritional component, and the edge weight of the nutrition-nutrition edge connection is defined by the dependence relationship between the nutritional components; Aggregating and updating the node features of each feed type graph node in the feed nutrition relationship graph and the feature information of adjacent nodes based on a graph neural network to obtain the updated node features of each feed type graph node; Defining the input state of the feeding plan generation model with each of the health index sampling data, the feeding time series data corresponding to each of the feed storage bins, the updated node features and the current feed remaining amount, so as to determine the animal feeding plan corresponding to the target action through reinforcement learning; the animal feeding plan includes the planned feeding time and the corresponding planned feeding amount of each feed storage bin.

2. The method according to claim 1, wherein, The feeding behavior data includes at least one of the following: feeding time, feeding frequency and feeding duration; the health index sampling data is used to indicate at least one type of health index among the following: body weight, heart rate, blood sugar and blood lipid.

3. The method according to claim 1, wherein The feed storage bin includes an upper box body and a lower box body, wherein the upper box body is a feeding area provided with a feeding trough, and the lower box body is used to receive the animal feces from the feeding area through the grid through holes at the bottom of the upper box body.

4. The method according to claim 3, wherein, The obtaining the feeding time series data of each animal in the breeding space for the feed storage bin includes: Based on the RFID reader arranged inside the feeding trough, sensing the RFID tags of each animal and real-time monitoring the target animals in the feeding area; Based on the weighing sensor arranged inside the feeding trough, real-time monitoring the feed consumption amount and the feed consumption time of the corresponding feed type to determine the feeding time series data corresponding to the target animal.

5. The method according to claim 1, wherein The obtaining the nutritional components corresponding to the feed types of each of the feed storage bins respectively to construct a feed nutrition relationship graph includes: Based on the feed nutrition component database, determining the nutritional component content of the feed types of each of the feed storage bins, the biological utilization rate and the nutritional metabolism experimental data corresponding to each nutritional component; the nutritional metabolism experimental data includes the experimental data of animal health and metabolic functions related to nutrition. Construct a feed nutrition relationship graph based on the nutrient content of each of the feed types, the bioavailability corresponding to each nutrient, and the nutrient metabolism experimental data; wherein, the node characteristics of the feed graph nodes are defined by the feed name, feed classification, and feed nutrient content of the corresponding feed type, and the node characteristics of the nutrient graph nodes are defined by the nutrient index name, nutrient classification, nutrient metabolism experimental data, and bioavailability of the corresponding nutrient; The edge weight connected by the feed-nutrient edge is calculated by the following formula: w(F i ,N j ) = α1·[C(F i ,N j )] γ ·Imp(N j )·B(N j ), Wherein, F i represents the feed type corresponding to the i-th feed graph node, N j represents the nutrient component corresponding to the j-th nutrient graph node, w(F i , N j ) represents the edge weight of the edge connection connecting the graph nodes corresponding to F i and N j ; C(F i , N j ) represents the standardized content of N i in F j , Imp(N j ) represents the importance factor of N j to the health and metabolic functions of animals, B(N j ) represents the bioavailability factor of N j , γ represents the non-linear parameter used to control the sensitivity of nutrient content, α1 is the first metric global scaling factor; H = {H1, H2, …, H m} represents an array containing m animal health indicators; H a represents the a-th health indicator, and its value is normalized to [0, 1], and the larger the value, the better the health condition; Corr(N j , H a ) represents the correlation metric between the intake level of N j and H a , and its value range is [-1, 1]. A positive value indicates a positive correlation between intake and health indicators, and a negative value indicates a negative correlation between intake and health indicators; Q is the total number of experimental samples in the nutrient metabolism experimental data. Among them, the intake measured for the nutrient component N j in the q-th sample is N j (q), and the measured performance value for H a in the q-th sample is H a (q), is the mean value of N j (q) corresponding to all experimental samples, is the mean value of H a (q) corresponding to all experimental samples; Pos(Corr(N j , H a )) represents mapping the correlation metric Corr(N j , H a ) to the [0, 1] interval corresponding to the positive score; is the weight of the health indicator H a , indicating the relative importance of this indicator to the overall health; The edge weight connected by the nutrient-nutrient edge is calculated by the following formula: w(N j ,N k ) = α2·exp(λ·Rel(N j ,N k ))·[B(N j )·B(N k )] δ , Rel(N j ,N k ) = tanh(BaseRel(N j ,N k )) Wherein, N k represents the nutrient component corresponding to the k-th nutrient graph node, w(N j , N k ) represents the edge weight of the edge connection for connecting the graph nodes corresponding to N j and N k ; Rel(N j , N k ) represents the metabolic relatedness between N j and N k , B(N k ) is the bioavailability factor of N k , δ is a parameter for controlling the sensitivity of the relationship between nutrients to bioavailability, λ is a non-linear mapping parameter for mapping Rel(N j , N k ) to the positive value space, α2 is the second global scaling factor, exp(·) represents the exponential function, BaseRel(N j , N k ) represents the quantification value of the synergistic or antagonistic effect of the nutrient components of N j and N k ; Corr((N j , N k ), H a ) represents the correlation measure considering the combined intake level of the two nutrient components N j and N k and the health index H a , Corr(N j , H a ) and Corr(N k , H a ) represent the correlation measures of the individual intake of N j and N k with the health index H a , is the weight of the health index H a when determining the relationship between nutrients.

6. The method according to claim 5, wherein, The feed classification includes at least one of the following: plant-based feed, animal-based feed, energy feed, fiber feed, and protein feed; the nutrient classification includes at least one of the following: macronutrients, micronutrients, electrolytes, and essential amino acids.

7. The method according to claim 5, wherein, The graph neural network adopts a graph attention network, which dynamically assigns weights to neighbor nodes through an attention mechanism and incorporates category information and edge weights to aggregate and update node characteristics: For each pair of adjacent graph nodes, calculate the attention score: e uv = LeakyReLU(Q T ·[WX u ||WX v ||C uv ), where u and v denote adjacent graph nodes u and v, W and Q denote learnable linear transformation matrices and attention parameter vectors respectively, || denotes the vector concatenation operation, X u and X v denote the node features of graph node u and graph node v respectively; C uv is the information vector corresponding to the feed classification or nutritional classification, which uses one-hot encoding; LeakyReLU(·) represents the LeakyReLU activation function; Incorporate the edge weight into the attention score to form a weighted attention score: In the formula, represents the weighted attention score, w uv represents the edge weight of the edge connection between graph node u and graph node v; Normalize the attention score to obtain a weighted category attention weight: where α uv represents the weighted class attention weight between graph node u and graph node v, p ∈ N(u) represents any graph node p in the set N(u) of neighboring graph nodes of graph node u, and e' up represents the weighted class attention weight between graph node u and graph node p; Based on the weighted category attention weight, perform weighted aggregation on the node characteristics of neighbor nodes to update the node characteristics: where, σ represents the non-linear activation function, represents the updated node feature of graph node u.

8. The method according to claim 7, wherein The feeding plan generation model adopts a hierarchical reinforcement learning model, which includes an attention fusion layer, a high-level policy network, and a low-level policy network; The attention fusion layer is used to fuse each of the health indicator sampling data with the feeding time series data corresponding to each of the feed storage bins, the updated node characteristics, and the current feed surplus through a self-attention mechanism to obtain attention fusion characteristics: F attended = Attention(MLP(H t ), MLP(F t ), MLP(G t ), MLP(S t )), Where, F attended represents the attention fusion feature, Attention(·) represents the self-attention mechanism, H t represents the health index at time step t, F t represents each feeding time series data at time step t, G t represents each updated node feature at time step t, S t represents each current feed surplus at time step t, and MLP(·) represents the multi-layer perceptron processing function; Define an input state based on the attention fusion characteristics to trigger the high-level policy network to determine the corresponding feed plan delivery time; T t = Softmax(W m ·F attended + b m ) where, T t represents the time-step decision vector of feed delivery output by the high-level policy network, which is a vector of length K, and each element represents the probability of feeding at the corresponding time step; Softmax(·) represents the softmax activation function; W m and b m represent the weight matrix and bias vector of the high-level policy network respectively; Input the attention fusion characteristics and the feed plan delivery time into the low-level policy network to determine the corresponding feed plan delivery amount: F worker_input = [F attended ||T t ​ A t = N(μ, σ 2 ) where F worker_input represents the input feature vector of the low-level policy network, and represent the weight matrix, bias vector, and output feature of the first hidden layer of the low-level policy network, respectively, and represent the weight matrix, bias vector, and output feature of the second hidden layer of the low-level policy network, respectively. W μ and b μ represent the weight matrix and bias vector of the output mean layer of the low-level policy network, respectively. W σ and b σ represent the weight matrix and bias vector of the output standard deviation layer of the low-level policy network, respectively; μ represents the mean vector of the feeding amounts of each feed storage bin, σ represents the standard deviation vector of the feeding amounts of each feed storage bin, Softplus(·) represents the Softplus activation function; N(μ, σ 2 ) represents the normal distribution, which is used to sample continuous actions; A t represents the feeding amount vector of each feed storage bin at time step t.

9. The method according to claim 8, wherein The reward function of the feeding plan generation model is: where R t represents the reward value at time step t, which is used to evaluate the quality of taking action A t at time step t; ΔH t represents the improvement degree of the health index after time step t, a g (t) represents the feed delivery amount of the g-th feed storage bin at time step t, S g (t) represents the current feed remaining amount of the g-th feed storage bin at time step t, G represents the total number of feed storage bins in the breeding space, and β1, β2, and β3 respectively represent the weight coefficients of the health index improvement term, the total feed delivery amount penalty term, and the feed remaining amount change penalty term.

10. An animal feeding plan generation system based on feeding monitoring data, comprising: A feeding data acquisition unit, configured to acquire, for each feed storage bin in the breeding space, the feeding time series data of each animal in the breeding space for the feed storage bin; each of the feed storage bins is respectively used to dispense feeds of corresponding types, and the feeding time series data includes the feeding time and the feeding amount for the corresponding feed type; A health data acquisition unit, configured to acquire the health indicator sampling data of each of the animals; A nutrition graph construction unit is used to obtain the nutritional components corresponding to the feed types of each of the feed storage bins respectively to construct a feed nutrition relationship graph. The feed nutrition relationship graph includes feed graph nodes, nutrition graph nodes, and edge connections. The feed graph nodes are used to indicate the corresponding feed types, the nutrition graph nodes are used to indicate the corresponding types of nutritional components, the edge weights of the feed-nutrition edge connections are defined by the association strength between the feed type and the nutritional component, and the edge weights of the nutrition-nutrition edge connections are defined by the dependence relationship between the nutritional components. A graph structure update unit is used to aggregate and update the node features of each feed type graph node in the feed nutrition relationship graph and the feature information of adjacent nodes based on a graph neural network to obtain the updated node features of each feed type graph node. A feeding plan generation unit is used to define the input state of a feeding plan generation model with each of the health index sampling data, the feeding time series data corresponding to each of the feed storage bins, the updated node features, and the current feed surplus, so as to determine the animal feeding plan corresponding to the target action through reinforcement learning. The animal feeding plan includes the planned feeding time and the corresponding planned feeding amount of the feed in each feed storage bin.

Images