Mutton sheep quantitative intelligent feeding management system based on Internet of Things

Abstract

The invention relates to the technical field of intelligent animal husbandry, and discloses a mutton sheep quantitative intelligent feeding management system based on the Internet of Things. Comprising a multi-modal sensing module, an edge computing gateway, a metabolic efficiency feedback factor computing unit, an individual nutrition digital twin modeling unit, an active disturbance response verification mechanism execution unit, an intelligent feeding terminal and a group knowledge evolution module. According to the system, an individual nutrition digital twinborn modeling unit is adopted and is combined with an online reinforcement learning algorithm, and an accurate feeding decision is generated according to a metabolic efficiency feedback factor and a nutrition elasticity coefficient matrix; and an active disturbance response verification mechanism execution unit applies controllable tiny disturbance when the model is uncertain, and an elastic coefficient matrix is dynamically updated. According to the method, the physiological state of the individual is sensed in real time by using multi-modal data, dynamic closed-loop control based on metabolic efficiency is realized, the cold start problem of the new individual is solved through knowledge distillation, and the feed conversion efficiency and the culture benefit are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent animal husbandry technology, specifically to a quantitative intelligent feeding management system for meat sheep based on the Internet of Things. Background Technology

[0002] In the field of sheep farming, achieving precision feeding to improve feed conversion ratio, optimize growth performance, and ensure individual health is one of the core goals of modern animal husbandry. Existing sheep feeding and management practices largely rely on feeding standards based on herd averages or growth stages. These methods typically group sheep according to static parameters such as age or average weight and apply a uniform feed formula and feeding amount to the entire herd.

[0003] This group-based management approach is technically difficult to address the significant physiological and metabolic differences among individual sheep. In practice, adopting a uniform standard can lead to underfeeding of some metabolically efficient individuals, limiting their growth potential; while overfeeding can result in feed waste and increased farming costs for others with low metabolic efficiency.

[0004] Although some automated feeding systems have attempted to use electronic tags for individual identification and quantitative feeding, the strategy models upon which their feeding decisions are based are usually static or open-loop. These systems lack an effective closed-loop feedback mechanism and cannot adjust the composition and total amount of feed in real time based on the actual, dynamically changing energy conversion efficiency of each individual.

[0005] Furthermore, existing technologies often fail to effectively address the issue of time-varying individual biological models, meaning that the model's predictive accuracy decreases as the individual's physiological state changes, and they lack online adaptive validation and calibration mechanisms. Additionally, when new individuals are introduced, the lack of historical data presents challenges in quickly and accurately setting initial individualized feeding parameters. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a quantitative intelligent feeding management system for meat sheep based on the Internet of Things, which solves the problems of difficulty in quantifying individual metabolic differences in meat sheep farming, lack of dynamic adaptability of feeding strategies, and difficulty in cold start in existing technologies.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solution: a quantitative intelligent feeding management system for meat sheep based on the Internet of Things, comprising: a multimodal perception module, an edge computing gateway, a metabolic efficiency feedback factor calculation unit, an individual nutrition digital twin modeling unit, an active disturbance response verification mechanism execution unit, an intelligent feeding terminal, and a group knowledge evolution module.

[0008] Preferably, the multimodal sensing module is configured to collect weight data, rumen region thermal imaging data, rumination acoustic signals, activity behavior data, and environmental parameters for each individual sheep. The environmental parameters include ambient temperature, relative humidity, and ammonia concentration. The edge computing gateway is configured to receive the data collected by the multimodal sensing module and perform time synchronization and binding processing on the multi-source data based on individual electronic identifiers. The electronic identifier remains unique throughout the entire process of data acquisition, modeling, decision-making, and execution of the system. Communication between the edge computing gateway, the intelligent feeding terminal, and the multimodal sensing module can utilize a low-power wide area network (LPWAN) or a local area network (LAN).

[0009] Preferably, the metabolic efficiency feedback factor calculation unit is configured to receive synchronous data from the edge computing gateway and periodically calculate the metabolic efficiency feedback factor for each individual sheep. The calculation process utilizes the environmental temperature and humidity index to suppress and correct the original weight gain to obtain a corrected weight gain, aiming to eliminate the interference of environmental heat stress on weight gain. The formula for calculating the environmental temperature and humidity index is as follows:

[0010] ;

[0011] in, Ambient temperature in degrees Celsius. Relative humidity is expressed as a percentage. The total feed nutrient intake is obtained by weighting and calculating the intake of each feed component and its corresponding standardized nutrient equivalent. The calculation method is as follows:

[0012] ;

[0013] in, Indicates the total amount of nutrients ingested from the feed. For the first Intake of similar feed components, For the first Standardized nutritional equivalents of similar feed components This represents the total number of feed components. The metabolic efficiency feedback factor is defined as the ratio of corrected body weight gain to total feed nutrient intake, and its expression is as follows:

[0014] ;

[0015] in, This represents a metabolic efficiency feedback factor. This indicates a correction for weight gain. When the total feed nutrient intake within the observation window is detected to be zero, the system is configured to use the metabolic efficiency feedback factor value from the previous observation window.

[0016] Preferably, the individual nutrition digital twin modeling unit is configured to construct a state vector containing static livestock parameters, dynamic physiological characteristics, environmental covariates, and a nutritional elasticity coefficient matrix. The dynamic physiological characteristics specifically include a metabolic efficiency feedback factor, a rumination intensity index extracted from rumination acoustic signals via a hybrid model of a one-dimensional convolutional neural network and a Transformer, the temperature fluctuation variance of rumen region thermal imaging data, the proportion of lying down time, and an estimated value of activity energy expenditure. The nutritional elasticity coefficient matrix is ​​used to characterize the local sensitivity of the metabolic efficiency feedback factor to changes in the proportions of various feed components. The unit utilizes an online reinforcement learning algorithm to generate the total feed amount and feed component proportions for the next feeding cycle based on the state vector and the metabolic efficiency feedback factor. The reward function used in the online reinforcement learning is a weighted average of the metabolic efficiency feedback factor and the unit cost of the current feed formulation, and its expression is:

[0017] ;

[0018] in, For the reward function, and These are configurable weighting coefficients. This represents the unit cost of the current feed formulation. The proportions of the feed components generated by the algorithm strictly satisfy the linear constraint that they are non-negative and sum to 1.

[0019] Preferably, the intelligent feeding terminal is configured to receive instructions on the total feeding amount and feed component ratio, and to identify the electronic identifier of an individual approaching the feeding trough through radio frequency identification or visual recognition. When the corresponding individual is identified, the terminal performs a quantitative mixing and feeding operation specific to that individual and determined by the instructions.

[0020] Preferably, the group knowledge evolution module is configured to aggregate the nutritional digital twin states of all individuals to construct a group metabolic phenotype library for cold start parameter initialization of newly added individuals.

[0021] Specifically, the active perturbation response verification mechanism execution unit is configured to trigger when the uncertainty of the individual nutritional digital twin exceeds a preset threshold. Upon triggering, the execution unit applies a controllable small perturbation to the feed component ratio. The perturbation amplitude is limited to a preset perturbation upper limit and applies only to a single feed component; the ratios of other feed components are scaled proportionally to maintain normalization. Subsequently, the execution unit collects the rate of change of metabolic efficiency feedback factors and the rumination intensity index shift within the observation window after the perturbation as a multimodal response, and updates the nutritional elasticity coefficient matrix based on this multimodal response data. The update of the nutritional elasticity coefficient matrix employs an exponential smoothing algorithm, the mathematical model of which is as follows:

[0022] ;

[0023] In the formula, and They represent the times respectively. and individual For the The elastic coefficient of feed components; Indicates the individual caused by the disturbance The change in metabolic efficiency feedback factors; This represents the actual proportional disturbance amount applied; The learning rate is used to control the weighting of historical and current observation information. This mechanism enables the system to inherently explore and validate its own model online, addressing the technical challenges of complex biological system models drifting over time and the difficulty in consistently and accurately representing individual nutritional elasticity coefficients.

[0024] Preferably, the population knowledge evolution module is configured to execute a cold-start strategy during the new individual's entry into the pen: based on the new individual's initial rumination intensity index and early metabolic efficiency feedback factor trend, similar historical individuals are retrieved and matched from the population metabolic phenotype library. Subsequently, the system uses lightweight knowledge distillation technology to transfer some prior parameters from the nutritional digital twins of similar historical individuals to the new individual model to achieve the initialization of the cold-start parameters. This method can significantly shorten the stabilization time of the new individual model and reduce the initial growth or health risks caused by strategy blind spots.

[0025] Preferably, the specific hardware configuration of the multimodal sensing module includes: a directional microphone array for acquiring acoustic signals of rumination; an infrared thermal imager for acquiring thermal imaging data of the rumen region; an inertial measurement unit collar worn around the neck for acquiring activity behavior data; and an environmental sensor for acquiring ambient temperature, relative humidity, and ammonia concentration.

[0026] This invention provides a quantitative intelligent feeding management system for meat sheep based on the Internet of Things (IoT). It has the following beneficial effects:

[0027] 1. This system quantifies the feed energy conversion efficiency of individual meat sheep through a metabolic efficiency feedback factor calculation unit, and utilizes an online reinforcement learning mechanism to integrate multimodal physiological characteristics as state input through an individual nutrition digital twin modeling unit. This structure enables the system to generate the total feed amount and feed component ratio for the next feeding cycle based on measured individual metabolic performance, rather than the population average standard. This achieves precise matching of feeding strategies with the real-time physiological metabolic state of meat sheep, avoiding nutrient waste or insufficient intake.

[0028] 2. The system's unique active perturbation response verification mechanism execution unit applies controllable micro-perturbations to the feed component ratios when model uncertainty is high, and collects the rate of change of metabolic efficiency feedback factors and the rumination intensity index shift as multimodal responses. This response data is used to update the nutritional elasticity coefficient matrix in real time via exponential smoothing. The system structure inherently possesses the ability to explore and verify its own model online, effectively solving the technical challenge of complex biological system models easily drifting over time and the difficulty in continuously and accurately representing individual nutritional elasticity coefficients.

[0029] 3. This system, through a group knowledge evolution module, aggregates the nutritional digital twin states of all individuals into a group metabolic phenotype library. Based on the early physiological characteristics of new individuals, it uses lightweight knowledge distillation technology to transfer prior parameters from similar historical individuals. This mechanism enables the new individual model to quickly obtain initial, near-optimal parameters, significantly shortening the model stabilization time and reducing early growth or health risks caused by strategy blind spots. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the process of the present invention.

[0031] Among them, 100 is the multimodal perception module; 200 is the edge computing gateway; 300 is the metabolic efficiency feedback factor calculation unit; 400 is the individual nutrition digital twin modeling unit; 500 is the active disturbance response verification mechanism execution unit; 600 is the intelligent feeding terminal; and 700 is the group knowledge evolution module. Detailed Implementation

[0032] The technical solutions in 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.

[0033] Please see the appendix Figure 1 This invention provides an IoT-based intelligent quantitative feeding management system for meat sheep, comprising: a multimodal sensing module 100, an edge computing gateway 200, a metabolic efficiency feedback factor calculation unit 300, an individual nutritional digital twin modeling unit 400, an active perturbation response verification mechanism execution unit 500, an intelligent feeding terminal 600, and a population knowledge evolution module 700. These modules work together to form a closed-loop data flow and decision optimization system.

[0034] First, the multimodal sensing module 100 continuously collects multi-dimensional data from each individual sheep, including weight data, rumen region thermal imaging data, rumination acoustic signals, activity behavior data, and environmental parameters. Environmental parameters include ambient temperature, relative humidity, and ammonia concentration.

[0035] Subsequently, the collected raw data is transmitted to the edge computing gateway 200. The edge computing gateway 200 is responsible for receiving and synchronizing this heterogeneous data in time, and binding the data according to the electronic identifiers of each individual sheep to form time-aligned individual status data frames, thereby integrating the scattered raw data into structured data suitable for subsequent processing. The edge computing gateway 200 communicates with the multimodal sensing module 100 via a low-power wide area network or a local wireless network.

[0036] The integrated individual status data frame is sent to the metabolic efficiency feedback factor calculation unit 300. Based on the data collected within a preset observation window, the metabolic efficiency feedback factor calculation unit calculates the metabolic efficiency feedback factor for each individual sheep. This calculation process first obtains the corrected weight gain, which is obtained by suppressing and correcting the original weight gain using the environmental temperature and humidity index. The formula for calculating the environmental temperature and humidity index is as follows:

[0037] ;

[0038] in, The ambient temperature is expressed in degrees Celsius. The relative humidity is expressed as a percentage.

[0039] Simultaneously, the metabolic efficiency feedback factor calculation unit 300 calculates the total feed nutrient intake, which is a weighted sum of the intake of each feed component and its corresponding standardized nutrient equivalent. The calculation method is as follows:

[0040] ;

[0041] in, Indicates the total amount of nutrients ingested from the feed. For the first Intake of similar feed components within the observation window For the first Standardized nutritional equivalents of similar feed components This represents the total number of categories of feed components.

[0042] Ultimately, the metabolic efficiency feedback factor is defined as the ratio of corrected body weight gain to total feed nutrient intake, and its expression is as follows:

[0043] ;

[0044] in, This represents the metabolic efficiency feedback factor. Under specific circumstances, when the total feed nutrient intake is zero, to ensure system continuity, the metabolic efficiency feedback factor calculation unit 300 is configured to use the metabolic efficiency feedback factor value from the previous observation window.

[0045] The calculated metabolic efficiency feedback factor and other dynamic physiological characteristics (including rumination intensity index, temperature fluctuation variance of rumen region thermal imaging data, recumbency ratio, and estimated activity energy expenditure) are input into the individual nutrition digital twin modeling unit 400. The individual nutrition digital twin modeling unit constructs and maintains a state vector for each sheep, containing static livestock parameters, dynamic physiological characteristics, environmental covariates, and a nutritional elasticity coefficient matrix. Based on this state vector and the metabolic efficiency feedback factor, the individual nutrition digital twin modeling unit 400 generates the total feed amount for the next feeding cycle using an online reinforcement learning algorithm. Ratio of feed components ( The reward function used in online reinforcement learning ( ) is composed of metabolic efficiency feedback factors ( Compared with the current unit cost of feed formulation ( The weighted composition is expressed as follows:

[0046] ;

[0047] in, and These are configurable weight coefficients. The feed component ratios output by the reinforcement learning algorithm (...). It satisfies the constraints that the coefficients are non-negative and sum to 1. The nutritional elasticity coefficient matrix is ​​used to characterize the local sensitivity of metabolic efficiency feedback factors to the proportions of each feed component.

[0048] The intelligent feeding terminal 600 receives total feeding amount and feed component ratio instructions from the individual nutrition digital twin modeling unit 400. The intelligent feeding terminal 600 identifies the electronic identifier of the individual approaching the feeding trough via RFID or visual recognition, and only performs its own instruction-determined quantitative mixing and feeding operations on the identified individual. The intelligent feeding terminal 600 communicates with the edge computing gateway 200 via a low-power wide area network or local wireless network. The actual feeding execution record of the intelligent feeding terminal 600 is fed back to the metabolic efficiency feedback factor calculation unit 300 for subsequent calculation of total feed nutrient intake.

[0049] To ensure the continuous accuracy and adaptability of the individual nutrition model, the system includes an active perturbation response verification mechanism execution unit 500. When the uncertainty of the individual nutrition digital twin output by the individual nutrition digital twin modeling unit 400 exceeds a preset threshold, the active perturbation response verification mechanism execution unit 500 is triggered. Upon triggering, the active perturbation response verification mechanism execution unit 500 applies a controllable small perturbation to the feed component ratios. The perturbation amplitude does not exceed a preset perturbation upper limit and only applies to a single feed component; the ratios of other feed components are scaled proportionally to maintain a normalization sum of 1. Subsequently, the active perturbation response verification mechanism execution unit 500 collects the rate of change of metabolic efficiency feedback factors and the rumination intensity index shift within the observation window after the perturbation as multimodal responses. Based on this multimodal response, the active perturbation response verification mechanism execution unit 500 updates the nutritional elasticity coefficient matrix using an exponential smoothing method. The update formula for the nutritional elasticity coefficient matrix is ​​as follows:

[0050] ;

[0051] in, Indicates at time individual For the The elastic coefficient of similar feed components, Indicates at time Updated elasticity coefficient; Indicates the individual caused by the disturbance The change in metabolic efficiency feedback factors; This represents the actual proportional disturbance amount applied; This represents the learning rate, used to balance the influence of historical information and current observation information.

[0052] In addition, this system includes a population knowledge evolution module 700. This module aggregates the nutritional digital twin states of all individuals to construct a population metabolic phenotype library, supporting the initialization of cold-start parameters for newly enrolled individuals. Upon enrolling a new individual, the module matches similar historical individuals from the population metabolic phenotype library based on its initial rumination intensity index and early metabolic efficiency feedback factor trends. It then transfers some prior parameters from the nutritional digital twins of these similar historical individuals using lightweight knowledge distillation technology to complete the cold-start parameter initialization, thereby accelerating the convergence speed of the new individual model. The electronic identifiers of all individuals remain unique throughout the system and are used throughout the entire process of data acquisition, modeling, decision-making, and execution.

[0053] Please see the appendix Figure 1The multimodal sensing module 100 is configured to collect data on each individual sheep and its breeding environment. The multimodal sensing module 100 includes, but is not limited to, the following hardware sub-units: an individual identification-linked weighing device for acquiring the sheep's weight data and identifying its electronic identifier as it passes by; an infrared thermal imager for acquiring surface temperature distribution data of the sheep's rumen region, forming rumen region thermal imaging data; a directional microphone array for collecting acoustic signals from the sheep's rumination; and an inertial measurement unit (IMU) worn around the sheep's neck. The collar is used to continuously collect data on the sheep's activity and behavior, such as acceleration and angular velocity. Environmental sensors are used to collect environmental parameters from the breeding environment, including ambient temperature, relative humidity, and ammonia concentration. All sensors are equipped with independent timestamp functions to ensure the initial time-series information of the raw data.

[0054] Edge computing gateway 200 receives heterogeneous raw data streams from multimodal sensing module 100. This gateway is configured with a high-precision clock synchronization mechanism (e.g., synchronization with a central clock server via Network Time Protocol (NTP) or Precision Time Protocol (PTP)) to ensure that all received data is aligned in the time dimension. The core function of edge computing gateway 200 is to bind data from different sensors within similar timestamps based on the individual electronic identifier of the sheep. For example, individual ID and weight data from a weighing device, rumen region thermal images from an infrared thermal imager, rumination acoustic signals from a directional microphone array, activity behavior data from the collar, and corresponding environmental parameters are all integrated into a time-aligned individual status data frame. This process also includes preliminary data cleaning, format conversion, and compression of the raw data to improve the efficiency of data transmission and subsequent processing. The processed time-aligned individual status data frame is then transmitted to metabolic efficiency feedback factor calculation unit 300. The edge computing gateway 200 communicates with the smart feeding terminal 600 and the multimodal sensing module 100 via a low-power wide area network or a local wireless network to ensure the reliability and real-time performance of data transmission.

[0055] The metabolic efficiency feedback factor calculation unit 300 is configured to periodically calculate the metabolic efficiency feedback factor of each individual sheep based on the data of the individual within a preset observation window.

[0056] First, the unit 300 receives time-aligned individual status data frames from the edge computing gateway 200 and obtains actual feeding execution records from the smart feeding terminal 600.

[0057] Secondly, calculate the total feed nutrient intake of the sheep within the observation window. The total amount is the intake of each feed component and its corresponding standardized nutritional equivalent (SCE). The weighted sum is calculated as follows:

[0058] ;

[0059] in, This indicates the total amount of feed nutrients ingested by an individual within the observation window. For the first The actual intake of similar feed components within the observation window. For the first Standardized nutritional equivalents of similar feed components This represents the total number of categories of feed components.

[0060] At the same time, the corrected weight gain of the sheep within the observation window was calculated. Corrected body weight gain was obtained by suppressing the original body weight gain through environmental temperature and humidity index, aiming to reduce the impact of environmental factors on growth performance assessment. The formula for calculating the environmental temperature and humidity index is:

[0061] ;

[0062] in, The ambient temperature is expressed in degrees Celsius. Relative humidity expressed as a percentage. Correction function. Piecewise linear functions or sigmoid functions based on empirical data can be used for quantification. Inhibitory effect on weight gain.

[0063] Ultimately, metabolic efficiency feedback factor ( Defined as the ratio of corrected body weight gain to total feed nutrient intake, its expression is as follows:

[0064] ;

[0065] in, This represents a metabolic efficiency feedback factor. This represents the corrected weight gain of an individual within the observation window. Under specific circumstances, when the total feed nutrient intake ( When the calculation result is zero, the metabolic efficacy feedback factor calculation unit 300 is configured to use the metabolic efficacy feedback factor value of the previous observation window for that individual, so as to avoid the mathematical problem of division by zero and maintain system stability.

[0066] Each individual nutritional digital twin modeling unit (400) constructs and continuously updates a high-dimensional state vector for each meat sheep. This state vector includes: static livestock parameters (e.g., breed, sex, age, initial body size); dynamic physiological characteristics (including metabolic efficiency feedback factors, rumination intensity index extracted from rumination acoustic signals via a hybrid model of a one-dimensional convolutional neural network and a Transformer, temperature fluctuation variance of rumen region thermal imaging data, proportion of lying down time, and estimated energy expenditure during activity); and environmental covariates (e.g., current...). Values, ammonia concentration, etc.); and a nutrient elasticity coefficient matrix ( Nutritional elasticity coefficient matrix ( It is used to characterize the local sensitivity of metabolic efficiency feedback factors to the proportion of each feed component.

[0067] Based on this state vector, unit 400 uses an online reinforcement learning algorithm to generate the total feeding amount for the next feeding cycle. Ratio of feed components ( The reinforcement learning algorithm uses the state of a digital twin of a sheep as observations and outputs a feeding strategy as an action. The reward function used in online reinforcement learning (…) ) is composed of metabolic efficiency feedback factors ( ) and the current unit cost of feed formulation ( The weighted composition is expressed as follows:

[0068] ;

[0069] in, and These are configurable weighting coefficients used to balance growth performance and feeding cost objectives. The reinforcement learning algorithm generates feed component ratios (...). At that time, it will ensure that it meets the constraint that it is non-negative and the sum of the proportions of all components is 1.

[0070] The active perturbation response verification mechanism execution unit 500 is designed to ensure the continuous accuracy and adaptability of the individual nutrition model. The active perturbation response verification mechanism execution unit 500 is triggered when the uncertainty of the individual nutrition digital twin output by the individual nutrition digital twin modeling unit 400 exceeds a preset threshold. Uncertainty can be quantified using metrics such as the variance of the model prediction error, the width of the confidence interval for the model parameters, or the consistency of predictions from the ensemble model.

[0071] Upon triggering, the active disturbance response verification mechanism execution unit 500 will check the current feed component ratio ( Apply controlled, small perturbations. The perturbation amplitude is limited to a preset perturbation upper limit (e.g., no more than 1% adjustment to the proportion of a single component) to avoid significant negative impacts on sheep growth. The perturbation only affects a single feed component, for example, changing only the proportion of a certain energy feed or protein feed. To maintain the normalization of the total feed volume, the proportions of the remaining undisturbed feed components are scaled proportionally.

[0072] After the perturbation was applied, the system entered the observation period and collected multimodal responses from meat sheep within the observation window following the perturbation. The multimodal responses included the rate of change of metabolic efficiency feedback factors and the shift in the rumination intensity index. These response data were used to update the nutritional elasticity coefficient matrix in the individual nutritional digital twin modeling unit 400. The nutrient elasticity coefficient matrix is ​​updated using exponential smoothing, and its mathematical model is as follows:

[0073] ;

[0074] in, Indicates at time individual For the The elastic coefficient of similar feed components, Indicates at time Updated elasticity coefficient; Indicates the individual caused by the disturbance The change in metabolic efficiency feedback factors; This represents the actual proportional disturbance amount applied; The learning rate, which ranges from 0 to 1, is used to control the weighting of historical elasticity coefficients and current observed responses, thereby enabling online adaptive adjustment of model parameters.

[0075] The intelligent feeding terminal 600 is a module that physically executes feeding commands. This terminal is equipped with multiple independent feed bins, each storing one type of feed component. Internally, the terminal includes high-precision weighing sensors, a mixing device, and a discharging mechanism. Its workflow is as follows: First, the terminal uses radio frequency identification (RFID)... A reader or integrated visual recognition system identifies the electronic identifier of an individual sheep approaching the feeding trough. Once the individual is identified, the terminal receives specific feeding instructions (including total feed amount) from the individual's nutritional digital twin modeling unit 400. Ratio of feed components Subsequently, the terminal controls each feed bin to precisely measure the corresponding amount of feed according to the required proportions, mixes it in the mixing device, and finally dispenses the mixed feed quantitatively into the feeding trough through the discharging mechanism, ensuring that only the identified individual receives their specific mixed feed. Simultaneously, the intelligent feeding terminal 600 records the actual feeding execution amount and feeds these records back to the metabolic efficiency feedback factor calculation unit 300 for accurate calculation of the actual total feed nutrient intake.

[0076] The Group Knowledge Evolution Module 700 is responsible for aggregating the nutritional digital twin status of all sheep in the flock, constructing and continuously updating a large-scale group metabolic phenotype database. This database stores historical physiological data, feeding strategies, and metabolic efficiency feedback for sheep of different breeds, ages, and growth stages. One of the core functions of this module is to support the initialization of cold-start parameters for newly admitted sheep. When a new sheep enters the flock, the Group Knowledge Evolution Module 700 collects its initial rumination intensity index and early metabolic efficiency feedback factor trends. Based on these initial characteristics, the module retrieves and matches historical individuals with similar initial characteristics from the group metabolic phenotype database. Subsequently, through lightweight knowledge distillation technology, some prior parameters (e.g., the initial range of the nutritional elasticity coefficient matrix, the initial weights of the reinforcement learning strategy network, etc.) from the nutritional digital twins of the matched similar historical individuals are transferred to the model of the newly admitted sheep, thus completing the cold-start parameter initialization and significantly shortening the time required for the new individual's model to reach stability and optimize decision-making.

[0077] See attached document Figure 1 , Figure 1 This is a schematic diagram of the overall system workflow according to an embodiment of the present invention. This embodiment describes the complete life cycle management process of a newly admitted sheep in the system, including its cold start, daily feeding optimization, and model adaptive validation process.

[0078] Step S1: New Individual Entry and Cold Start

[0079] When a new sheep enters the pen, its basic information (such as breed, age, and initial weight) is entered into the system. The multimodal perception module 100 begins collecting initial data on the new individual, including weight, rumination acoustic signals, activity behavior data, and environmental parameters. This data is synchronized and bound in time by the edge computing gateway 200 before being transmitted to the population knowledge evolution module 700. Based on the new individual's initial rumination intensity index (obtained from preliminary analysis of rumination acoustic signals) and early metabolic efficiency feedback factor trend (calculated from preliminary feeding data), the population knowledge evolution module 700 matches historical individuals with similar characteristics in a pre-built population metabolic phenotype library. Subsequently, through lightweight knowledge distillation technology, some prior parameters from the nutritional digital twins of the matched historical individuals, such as the initial distribution of the nutritional elasticity coefficient matrix and the initial weights of the reinforcement learning policy network, are transferred and initialized into the individual nutritional digital twin modeling unit 400 of the new individual. This process aims to provide a reasonable initial model for the new individual, avoiding learning from scratch and thus accelerating its model convergence speed.

[0080] Step S2: Daily Data Collection and Edge Processing

[0081] In daily management, the multimodal sensing module 100 continuously collects data on each individual sheep. For example, individual identification is linked to a weighing device to obtain weight data; an infrared thermal imager obtains thermal imaging data of the rumen region; a directional microphone array collects acoustic signals of rumination; and an inertial measurement unit collar collects activity behavior data. Simultaneously, environmental sensors collect parameters such as ambient temperature, relative humidity, and ammonia concentration in the breeding environment. All collected raw data is transmitted in real time to the edge computing gateway 200. The edge computing gateway 200 performs high-precision time synchronization on these heterogeneous data streams and binds them according to the electronic identifier of each sheep, forming structured, time-aligned individual status data frames. These data frames are then sent to downstream processing units.

[0082] Step S3: Calculation of metabolic efficiency feedback factor

[0083] Individual status data frames and actual feeding execution amounts recorded by the intelligent feeding terminal 600 are transmitted to the metabolic efficiency feedback factor calculation unit 300. This unit periodically (e.g., every 24 hours) calculates the metabolic efficiency feedback factor for each individual sheep.

[0084] First, calculate the total feed nutrient intake of the sheep within the observation window. The total nutrient intake of feed is based on the intake of each feed component and its standardized nutritional equivalent (SEE) in the actual feeding records. The weighted sum is calculated using the following formula:

[0085] ;

[0086] in, For the first The actual intake of similar feed components within the observation window. For the first Standardized nutritional equivalents of similar feed components This represents the total number of categories of feed components.

[0087] At the same time, the corrected weight gain of the sheep within the observation window was calculated. This increase was obtained by suppressing and correcting the increase in original body weight using the environmental temperature and humidity index. The formula for calculating the environmental temperature and humidity index is:

[0088] ;

[0089] in, The ambient temperature is expressed in degrees Celsius. This represents relative humidity as a percentage. A correction function is used for quantization. Inhibitory effect on weight gain.

[0090] Finally, the formula for calculating the metabolic efficiency feedback factor is:

[0091] ;

[0092] when When the value is zero, the system uses the previous observation window value for that individual. Value. Calculated The value is fed back to the individual nutrition digital twin modeling unit 400.

[0093] Step S4: Individual Nutritional Digital Twin Modeling and Feeding Strategy Generation

[0094] The individual nutrition digital twin modeling unit 400 receives the latest data from the metabolic efficiency feedback factor calculation unit 300. The individual nutrition digital twin modeling unit 400 integrates this information, along with other individual state data frames from the edge computing gateway 200, to update the individual nutrition digital twin state vector for each sheep. This state vector includes static livestock parameters, dynamic physiological characteristics (such as...) The data included rumination intensity index, rumen temperature fluctuation variance, recumbent time ratio and estimated activity energy expenditure, environmental covariates and current nutritional elasticity coefficient matrix.

[0095] Based on the updated state vector, the individual nutrition digital twin modeling unit 400 uses an online reinforcement learning algorithm to generate the optimal total feed amount for the next feeding cycle. Ratio of feed components ( The reward function of reinforcement learning algorithms ( ) is defined as:

[0096] ;

[0097] in, and These are the weighting coefficients. This represents the unit cost of the current feed formulation. The feed component ratios output by reinforcement learning ( It satisfies the constraints that are non-negative and sum to 1. The generated feeding instruction ( and The signal is sent to the intelligent feeding terminal 600.

[0098] Step S5: Execution of intelligent feeding instructions

[0099] The intelligent feeding terminal 600 receives feeding instructions from the individual nutrition digital twin modeling unit 400. When an individual sheep approaches the feeding trough, the intelligent feeding terminal 600 confirms the individual's electronic identifier via radio frequency identification or visual recognition system. Once the individual is identified, the terminal accurately measures and mixes the appropriate type and proportion of feed according to the received instructions, and quantitatively feeds it into the feeding trough, ensuring that each sheep receives a customized feed formula. The intelligent feeding terminal 600 simultaneously records the actual feeding amount and feeds this information back to the metabolic efficiency feedback factor calculation unit 300 for subsequent use. calculate.

[0100] Step S6: Active perturbation response verification and model adaptation

[0101] During the continuous feeding optimization process, the individual nutrition digital twin modeling unit 400 monitors the uncertainty of its model. When the prediction uncertainty of a certain meat sheep individual model (e.g., the variance of the prediction error or the confidence interval width of the elasticity coefficient) exceeds a preset threshold, the active perturbation response verification mechanism execution unit 500 is triggered.

[0102] Once triggered, the active perturbation response verification mechanism execution unit 500 applies a controllable, minute perturbation to the feed component ratios for the next feeding cycle generated by the individual nutrition digital twin modeling unit 400. For example, it fine-tunes the proportion of a certain feed. ( (Not exceeding the preset upper limit), while adjusting other feed components proportionally to maintain a total ratio of 1. This disturbance instruction is then executed via the intelligent feeding terminal 600.

[0103] In the subsequent observation window, the multimodal sensing module 100 continues to collect response data from individual sheep, which is then processed by the edge computing gateway 200 and the metabolic efficiency feedback factor calculation unit 300 to generate multimodal responses such as the rate of change of the perturbed metabolic efficiency feedback factor and the shift of the rumination intensity index.

[0104] The active perturbation response verification mechanism execution unit 500 receives these response data and updates the nutritional elasticity coefficient matrix in the individual nutritional digital twin modeling unit 400 using an exponential smoothing algorithm. The updated formula is:

[0105] ;

[0106] in, Indicates at time individual For the The elastic coefficient of similar feed components, This represents the updated elasticity coefficient; Indicates the individual caused by the disturbance The change in metabolic efficiency feedback factors; This represents the actual proportional disturbance amount applied; The learning rate is represented by this value. In this way, the system can explore the physiological responses of sheep to changes in feed formulation online and correct its internal model, achieving continuous adaptive learning of the nutritional elasticity characteristics of sheep.

[0107] The steps S2 to S5 described above constitute a continuous, closed-loop feeding management and optimization cycle. When the model uncertainty increases, the perturbation verification mechanism in step S6 will be activated to enhance the accuracy and stability of the model.

Claims

1. A quantitative intelligent feeding management system for meat sheep based on the Internet of Things, characterized in that, include: The multimodal sensing module (100) is used to collect the weight data, rumen region thermal imaging data, rumination acoustic signals, activity behavior data and environmental parameters of each individual sheep. An edge computing gateway (200) is used to synchronize and bind the data collected by the multimodal sensing module (100) according to individual electronic identifiers in time; The metabolic efficiency feedback factor calculation unit (300) is used to calculate the metabolic efficiency feedback factor of an individual based on the corrected weight gain and total feed nutrient intake within a preset observation window. Individual nutrition digital twin modeling unit (400) is used to construct a state vector for each meat sheep, which includes static livestock parameters, dynamic physiological characteristics, environmental covariates and nutritional elasticity coefficient matrix, and to generate the total feeding amount and feed composition ratio for the next feeding cycle based on metabolic efficiency feedback factor through online reinforcement learning; The active perturbation response verification mechanism execution unit (500) is used to apply a controllable small perturbation to the feed component ratio when the uncertainty of the individual nutrition digital twin output by the individual nutrition digital twin modeling unit (400) exceeds a preset threshold, and update the nutrition elasticity coefficient matrix according to the multimodal response collected after the perturbation. The intelligent feeding terminal (600) is used to receive the total amount of feed and the proportion of feed components, and to perform quantitative mixing and feeding when it detects that a corresponding individual is approaching the feeding trough. The Group Knowledge Evolution Module (700) is used to aggregate the nutritional digital twin status of all individuals, build a group metabolic phenotype library, and support the initialization of cold start parameters for newly added individuals.

2. The IoT-based intelligent feeding management system for meat sheep according to claim 1, characterized in that, In the metabolic efficiency feedback factor calculation unit (300): the corrected weight gain is obtained by suppressing and correcting the original weight gain through the environmental temperature and humidity index; the total feed nutrient intake is the weighted sum of the intake of each feed component and the corresponding standardized nutrient equivalent; the metabolic efficiency feedback factor is defined as the ratio of the corrected weight gain to the total feed nutrient intake; when the total feed nutrient intake is zero, the metabolic efficiency feedback factor of the previous observation window is used.

3. The IoT-based intelligent feeding management system for meat sheep according to claim 1, characterized in that, In the individual nutrition digital twin modeling unit (400): dynamic physiological features include metabolic efficiency feedback factors, rumination intensity index extracted from rumination acoustic signals through a one-dimensional convolutional neural network and a Transformer hybrid model, temperature fluctuation variance of rumen region thermal imaging data, lying time ratio and estimated activity energy consumption; the nutritional elasticity coefficient matrix is ​​used to characterize the local sensitivity of metabolic efficiency feedback factors to the proportion of each feed component.

4. The IoT-based intelligent feeding management system for meat sheep according to claim 1, characterized in that, In the individual nutrition digital twin modeling unit (400): the reward function used in online reinforcement learning is composed of a weighted average of the metabolic efficiency feedback factor and the unit cost of the current feed formula; the feed component ratio output by online reinforcement learning satisfies the constraint that it is non-negative and the sum is 1.

5. The IoT-based intelligent feeding management system for meat sheep according to claim 1, characterized in that, In the active disturbance response verification mechanism execution unit (500): the amplitude of the controllable micro-disturbance does not exceed the preset disturbance upper limit; Controllable micro-perturbations act only on a single feed component, while the proportions of other feed components are scaled proportionally to maintain normalization; the multimodal response includes the rate of change of metabolic efficiency feedback factors and the shift of rumination intensity index collected within the observation window after the perturbation.

6. The IoT-based intelligent feeding management system for meat sheep according to claim 5, characterized in that, In the active perturbation response verification mechanism execution unit (500): the nutritional elasticity coefficient matrix is ​​updated using an exponential smoothing method; the elasticity coefficient value at the next moment is calculated by weighting the elasticity coefficient value at the previous moment with the ratio of the change in metabolic efficiency feedback factor caused by the perturbation to the perturbation amount, and the weighted combination is controlled by a preset learning rate.

7. The IoT-based intelligent feeding management system for meat sheep according to claim 1, characterized in that, The multimodal sensing module (100) includes: a directional microphone array for collecting ruminant acoustic signals; an infrared thermal imager for acquiring thermal imaging data of the rumen region; an inertial measurement unit collar worn around an individual's neck for collecting activity behavior data; and an environmental sensor for collecting environmental parameters, including ambient temperature, relative humidity, and ammonia concentration.

8. The IoT-based intelligent feeding management system for meat sheep according to claim 1, characterized in that, The intelligent feeding terminal (600) identifies the electronic identifier of an individual approaching the feeding trough through radio frequency identification or visual recognition; and releases a unique mixed feed, corresponding to the total feeding amount and feed component ratio, only to the identified individual.

9. The IoT-based intelligent feeding management system for meat sheep according to claim 1, characterized in that, The group knowledge evolution module (700) initializes cold start parameters by: when a new individual enters the pen, based on the individual's initial rumination intensity index and early metabolic efficiency feedback factor trend; matching similar historical individuals from the group metabolic phenotype library; and transferring some prior parameters of the nutritional digital twins of similar historical individuals through lightweight knowledge distillation.

10. The IoT-based intelligent feeding management system for meat sheep according to claim 1, characterized in that, The edge computing gateway (200) communicates with the intelligent feeding terminal (600) and the multimodal sensing module (100) via a low-power wide area network or a local wireless network; the electronic identifier of each individual remains unique throughout the entire system and is used throughout the entire process of data acquisition, modeling, decision-making and execution.

Images