Cowshed intelligent feeding scheduling and dieting control method and system integrated with multi-source sensing

Abstract

The invention provides a multi-source sensing integrated intelligent feeding scheduling and dieting control method and system for a cowshed, and relates to the technical field of intelligent agricultural control, and the method comprises the steps: collecting cattle physiology, cowshed environment, feed state and equipment operation parameters in real time through a distributed sensing network; a machine learning model fusing a time sequence convolutional network and a self-attention mechanism is utilized to predict an individualized feeding demand and a dieting threshold value, and a feeding-dieting risk map for quantifying health and production risks is generated; simulating an extreme scene based on a digital twinborn model to perform dual consistency check, and generating a robust detailed scheduling scheme; converting the scheme into a control instruction to drive execution equipment, and iteratively optimizing the model and the rule; the system comprises a multi-source sensing data acquisition module, an instruction execution module, a feedback optimization module and the like. According to the method, the accurate and self-adaptive feeding management efficiency of individual cattle in a group is improved, the feed utilization efficiency is effectively improved, and the metabolic disease risk is reduced.

Description

Technical Field

[0001] This invention relates to the field of intelligent agricultural control technology, and in particular to a method and system for intelligent feeding scheduling and ration control in cattle sheds that integrates multi-source sensing. Background Technology

[0002] In large-scale dairy farming, traditional feeding management typically employs a uniform standard for the entire herd or a formula based on fixed stages, making it difficult to address the diverse nutritional needs of individual cattle due to their health status, stress responses, and real-time environmental changes. For example, in hot and humid summer weather, heat stress environments easily form in barns, leading to a decrease in the overall feed intake of the herd. However, individual cattle exhibit varying tolerances and physiological responses to heat stress. Current technologies rely on manual observation or simple, phased weight adjustments, failing to accurately perceive and respond to these individual differences in real time. This often results in some cattle with poor tolerance... Insufficient feed intake leads to decreased physical strength and a sharp decline in milk production in cattle, while some more resilient cattle may suffer from feed waste or even folliculitis due to overfeeding caused by failure to adjust feed intake in a timely manner. Although existing automated feeding equipment can achieve timed and quantitative feeding, it lacks the ability to intelligently judge the dynamic needs of individual cattle and the coupled influence of the external environment. Its scheduling strategy is rigid and cannot optimize feed resource allocation under complex and changing conditions, nor can it timely prevent cattle health risks. Therefore, there is an urgent need for a new intelligent feeding scheduling and feed control method and system for cattle sheds that integrates multi-source sensing. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for intelligent feeding scheduling and feed control in cattle sheds that integrates multi-source sensing, in order to solve the problems in the existing technology, such as rigid feeding strategies, low feed utilization efficiency, and difficulty in actively preventing metabolic diseases in cattle due to the inability to accurately sense the real-time physiological needs and environmental changes of individual cattle. The specific technical solution is as follows: This invention provides a method for intelligent feeding scheduling and ration control in cattle sheds that integrates multi-source sensing, including: Step 1: Collect the physiological parameters of cattle, environmental parameters, feed status parameters and feeding equipment operation parameters in the cattle shed through multi-source distributed sensor terminals. The physiological parameters of cattle include body weight, rumination frequency and feeding speed, the environmental parameters include temperature and humidity and ammonia concentration, and the feed status parameters include nutrient composition, remaining amount and freshness. Step 2: The collected multi-source parameters are preprocessed and then input into a pre-trained machine learning model. The machine learning model adopts a fusion structure of temporal convolution and attention mechanism to generate probability prediction results of personalized feeding demand and dieting threshold. Gradient feeding regulation signals are applied within the physiological safety boundary of the cattle, and nutritional supply and demand response characteristics are generated based on the differences in the cattle's feeding response. These characteristics are then fused with the probability prediction results to construct a feeding-dieting risk map. Step 3: Based on the feeding-dieting risk map, use digital twin technology to generate extreme environment and feed shortage scenarios. Verify the consistency of probability prediction results and scheduling schemes in the data domain and physical domain. Based on the verification results, combine the feeding task priority and the feed intake delayability to execute multi-level feeding-dieting collaborative scheduling to form the scheduling result. Step 4: Send the scheduling results to the feeding execution device and transmit back the actual feed intake data, cattle body weight change data and feed utilization data. Iterate and update the parameters of the machine learning model and the dieting threshold judgment rules based on the transmitted data.

[0004] Furthermore, in step 1, the multi-source distributed sensing terminal works in coordination with edge computing nodes deployed at the cattle shed site to perform timestamp alignment, unit unification, and outlier filtering on the collected raw data, forming a structured multi-source parameter set. Among the bovine physiological parameters, weight is dynamically estimated by motion data from the accelerometer and gyroscope integrated in the smart collar using a weight estimation algorithm; rumination frequency is obtained by analyzing sound data collected by the miniature microphone in the smart collar in combination with motion patterns using a pattern recognition algorithm; and feeding speed is calculated by monitoring the time the bovine's head stays on the ground and the feed consumption rate through individual feeding stations in the feeding area or 3D vision sensors. The environmental parameters are collected using integrated temperature and humidity sensors and electrochemical ammonia sensors deployed in the cattle shed ventilation openings, rest areas and feeding areas, with a set sampling period. Among the feed status parameters, the nutrient composition is obtained through online analysis by a near-infrared spectroscopy probe, the remaining amount is monitored by a weighing sensor at the bottom of the feed trough, and the freshness is calculated by combining the feed delivery timestamp with environmental parameters using a spoilage model or determined by camera image recognition. The operating parameters of the feeding equipment include the current and speed of the precision feeding motor, the operating status of the conveyor belt or screw feeder, the on / off status of each discharge port and the preset discharge amount, and the material level height of the hopper.

[0005] Further, in step 2, the specific architecture of the machine learning model is as follows: a temporal convolutional network is used to perform multi-level convolution operations on the input multi-source parameter temporal tensor to extract local and short-term features; a multi-head self-attention mechanism is introduced to model the global dependency of the high-order feature sequence output by the temporal convolutional network, focusing on the long-range correlation between parameters and historical states; the Gaussian distribution mean and variance of the personalized feeding demand, as well as the truncated normal distribution parameters of the dieting threshold, are output through a fully connected layer, where the upper bound of the dieting threshold distribution is the current feeding demand; The gradient feeding regulation signal is generated as follows: based on the cattle breed, current weight, body condition score and growth stage, a pre-set knowledge rule base is queried to obtain the safe range of daily dry matter intake. Using the average of the current predicted feeding demand as a benchmark, several equally spaced feeding levels are taken above and below the safe range to form a gradient feeding regulation signal sequence. The nutritional supply and demand response characteristics include the response delay time of feed intake to the regulatory signal, the sensitivity represented by the slope of the response curve, feed intake at saturation, and the correlation coefficient between changes in rumination activity and changes in feed intake. The feeding-dieting risk map is constructed by splicing and fusing the nutrient supply and demand response feature vector with the probability prediction results, and mapping it to a two-dimensional grid through a neural network or rule scoring system. The horizontal axis of the grid represents the feeding amount level, and the vertical axis represents the dieting intensity level. A comprehensive risk score is calculated for each grid point, taking into account the decline in weight gain benefits, the risk of metabolic diseases, and the reduction in feed conversion rate.

[0006] Furthermore, in step 3, the generation of extreme environment and feed shortage scenarios using digital twin technology includes: Acquire historical extreme weather data and feed inventory safety early warning rules; Historical meteorological extreme value data are injected into the environmental parameter module of the digital twin model to construct an extreme environmental parameter sequence; Based on the feed inventory safety early warning rules and the current feed consumption rate, a feed shortage event is simulated in the feed supply module of the digital twin model. Drive the digital twin model to run under extreme environmental parameter sequences and simulated feed shortage events to generate extreme scenario datasets containing virtual cattle reactions and equipment status; Among them, historical meteorological extreme value data includes records of the highest temperature, lowest temperature and humidity extreme values ​​that have occurred in the cattle shed location over the past few years, and feed inventory safety early warning rules include inventory thresholds that trigger shortage warnings and predicted consumption time thresholds.

[0007] Furthermore, in step 3, the consistency verification of the probability prediction results and scheduling scheme in the data domain and physical domain includes: In the data domain, extreme scenario datasets are input into machine learning models to obtain probability prediction results under extreme scenarios. These results are then compared with the original probability prediction results under normal scenarios to calculate prediction stability deviation. The prediction stability deviation is quantified by comparing the difference between the mean and variance of the probability distribution. In the physical domain, a preliminary scheduling scheme is generated based on the feeding-dieting risk map, which is then converted into an executable instruction sequence for the feeding device. This sequence is then tested in the simulation module of the physical device connected to the digital twin model or during the idle period of the actual device, and the success rate of instruction execution and state deviation are recorded. Consistency verification results are obtained by comprehensively considering the predicted stability deviation, instruction execution success rate, and state deviation.

[0008] Furthermore, in step 3, the multi-level feeding-dieting coordinated scheduling based on the verification results, combined with the feeding task priority and the delayability of feeding includes: Based on the prediction stability bias in the verification results, the feeding-dieting risk map was adjusted with confidence weighting to obtain a robust risk map. Based on the robust risk map and feeding task priority, a preliminary scheduling instruction set containing basic feed amount and feeding restriction instructions is generated for each cow. The feeding task priority is dynamically divided according to cow type, milk production, pregnancy stage and health status. Based on the feed intake delayability, the expected execution time of all instructions in the initial scheduling instruction set is flexibly adjusted to resolve equipment resource conflicts and generate a time-optimized instruction sequence. The feed intake delayability represents the maximum allowable delay time for a single feeding of cattle. The time-optimized instruction sequence is mapped to specific feeding execution devices, and the device capabilities and status are verified. Conflicting instructions are parsed and sorted according to priority, delayability and device reset time to form detailed scheduling results.

[0009] Furthermore, in step 4, the step of sending the scheduling results to the feeding execution device and transmitting the data back includes: The instruction sequence in the detailed scheduling results is encoded into a standard control instruction set that can be recognized by the feeding execution device controller. The standard control instruction set includes the target device address, action type, action parameters and execution timestamp, and is sent through industrial bus or wireless network protocol. The feeding equipment completes the feed delivery operation according to the control command, and the multi-source sensor network synchronously monitors and records the actual feed intake, changes in cattle weight and environmental parameters. The raw data of the monitoring records are cleaned, aligned and aggregated to extract the actual feed intake and related weight trend data of cattle corresponding to each scheduling instruction, and to calculate the feed utilization rate index. The feed utilization rate is calculated by the ratio of output to feed intake, and the output includes weight gain or milk production. The actual feed intake, weight change data, and feed utilization rate indicators are packaged into feedback data packets and sent back to the central processing system.

[0010] Furthermore, in step 4, the iterative updating of the machine learning model parameters and the diet threshold determination rules based on the returned data includes: The multi-dimensional errors between the feeding demand and suggested feeding threshold based on the feedback data packet calculation model and the actual observed cattle responses and production efficiency indicators include prediction errors, feed utilization gaps, and correlations of health events. Using multidimensional error as part of the loss function, incremental learning or periodic batch training methods are used to adjust and update the network weight parameters of the machine learning model; By utilizing historical feedback data within a time sliding window, statistical analysis is conducted on the health and production performance indicators of cattle under different dieting threshold triggering conditions, and the conditional parameters or logical combinations in the dieting threshold determination rules are dynamically adjusted.

[0011] This invention also provides an intelligent feeding scheduling and ration control system for cattle sheds that integrates multi-source sensing, for implementing the method described above. The system includes: The multi-source sensor data acquisition module is used to collect bovine physiological parameters, barn environmental parameters, feed status parameters, and feeding equipment operating parameters in real time or periodically through a multi-source distributed sensor network, and to preprocess and store them through edge computing nodes. The personalized feeding demand prediction and risk map generation module is used to obtain preprocessed multi-source parameter time series data from edge computing nodes, input it into a pre-trained machine learning model, output personalized feeding demand prediction values ​​and dieting threshold suggestions, and generate a feeding-dieting risk map through gradient feeding adjustment and response feature fusion. The digital twin verification and scheduling decision module is used to simulate extreme scenarios based on the feeding-dieting risk map using a digital twin model, verify the consistency between the prediction results and the preliminary scheduling plan, and perform multi-level collaborative scheduling in combination with the feeding task priority and the lag of feeding to generate detailed scheduling results. The feeding control command execution and feedback optimization module is used to convert detailed scheduling results into equipment control commands and send them to the feeding execution equipment, collect actual execution data, and iteratively update machine learning model parameters and dieting threshold determination rules based on feedback.

[0012] Furthermore, the multi-source sensor data acquisition module includes a sensor network coordination unit, a bovine physiological parameter acquisition unit, a cattle shed environmental parameter acquisition unit, a feed status parameter acquisition unit, and a feeding equipment operation parameter acquisition unit; The personalized feeding demand prediction and risk map generation module includes a data preprocessing unit, a machine learning model inference unit, a gradient feeding adjustment unit, a nutrition supply and demand response feature calculation unit, and a risk map construction unit. The digital twin verification and scheduling decision module includes an extreme scenario simulation unit, a consistency verification unit, a robust risk graph generation unit, and a multi-level collaborative scheduling unit. The feeding control command execution and feedback optimization module includes a control command encoding and distribution unit, an actual data acquisition and processing unit, a model parameter update unit, and a diet rule optimization unit.

[0013] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the method.

[0014] The beneficial effects of this invention are as follows: By deeply integrating multi-source sensor networks and machine learning models, this invention achieves accurate prediction of individual cattle nutritional needs and quantitative assessment of health risks; by utilizing digital twin technology for extreme scenario simulation and dual consistency verification, it significantly improves the robustness and reliability of feeding scheduling schemes in real and complex environments; by combining feeding priority and feeding behavior flexibility for multi-level collaborative decision-making, it effectively resolves equipment resource conflicts and optimizes the overall feeding efficiency of the herd; and by continuously adaptively optimizing the model and rules through closed-loop learning of execution feedback data, it ultimately achieves the comprehensive goals of improving feed utilization efficiency, preventing metabolic diseases, and ensuring cattle health and production performance while reducing the intensity of human intervention.

[0015] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the steps of the intelligent feeding scheduling and ration control method for cattle sheds that integrates multi-source sensing according to the present invention; Figure 2 This is a schematic diagram of the structure of the intelligent feeding scheduling and ration control system for cattle sheds that integrates multi-source sensing according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0018] In embodiments of the present invention, a method for intelligent feeding scheduling and ration control in cattle sheds integrating multi-source sensing is provided. Please refer to [link to relevant documentation]. Figure 1 The method includes the following steps: Step 1: Through a multi-source distributed sensor network deployed inside the cattle shed and its associated facilities, the physiological parameters of the target cattle, the environmental parameters of the cattle shed in its local microenvironment, the feed status parameters in the feed trough or feed tower, and the operating parameters of the automated feeding equipment are collected in real time or periodically.

[0019] It should be noted that, in this application, the bovine physiological parameters refer to a set of quantitative indicators that directly or indirectly reflect the current health status, growth stage, and nutritional requirements of an individual cattle; the barn environmental parameters refer to a set of external microclimate conditions and air quality indicators that affect cattle feeding behavior, feed preservation, and conversion efficiency; the feed state parameters refer to a set of information describing the current physical state, chemical composition, and edible quantity of feed; and the feeding equipment operating parameters refer to a set of data reflecting the current working status, execution accuracy, and remaining supply capacity of automated feeding equipment. Specifically, bovine physiological parameters may include real-time estimated body weight monitored by sensors installed on the cattle's body (such as collars or ear tags), rumination frequency obtained based on acoustic or movement pattern analysis of rumination behavior, and feeding speed characterized by the dry matter mass consumed per unit time calculated by a feeding behavior monitoring unit. Barn environmental parameters may include air temperature and relative humidity collected by temperature and humidity sensors deployed in different zones of the barn, and ammonia concentration detected by a gas detection unit. Feed status parameters may include the content of major nutrients in the feed measured online using technologies such as near-infrared spectroscopy analysis; the amount of feed remaining as sensed by a weighing sensor installed at the bottom of the feed container; and feed freshness indicators calculated based on feed feeding time and environmental parameters or determined by image recognition. Feeding equipment operating parameters may include the current and speed of the precision feeding motor, the operating status of the conveyor belt or screw feeder, the on / off status and preset discharge amount of each discharge port, and the material level in the hopper.

[0020] In practice, an edge computing node deployed at the cattle shed coordinates the operation of the entire multi-source distributed sensor network. This network consists of various types of sensor terminals, which are distributed according to function and physical location. For collecting bovine physiological parameters, a three-axis accelerometer, gyroscope, and miniature microphone are integrated into the smart collar worn by the target cattle. The accelerometer and gyroscope continuously collect the cattle's movement and posture data, which is used to dynamically estimate the cattle's weight using a pre-established weight estimation algorithm. The sound data collected by the miniature microphone, combined with movement patterns, is used to analyze the rumination chewing sounds and rhythms using a pattern recognition algorithm, thereby calculating the rumination frequency. In the cattle feeding area, individual feeding stations based on weight change monitoring are installed, or 3D vision sensors are installed above the public feed troughs to calculate the cattle's feeding speed by monitoring the time the cattle's head stays in front of the feed trough and the rate of feed consumption. For cattle shed environmental parameters, integrated temperature and humidity sensors and electrochemical ammonia sensors are deployed at key locations such as ventilation openings, rest areas, and feeding areas. These sensors collect data at set sampling intervals and transmit the data to edge computing nodes via wireless sensor network protocols such as LoRa or Zigbee. For feed status parameters, near-infrared spectroscopy probes are installed at key nodes of feed mixers or conveyors to analyze the nutritional composition of the mixed feed online. High-precision weighing sensors are installed at the bottom of each feed trough to monitor the remaining feed in real time. By recording the feed delivery timestamp and combining it with current environmental temperature and humidity parameters, a preset spoilage model is used to calculate the feed freshness index. Alternatively, images of the feed are periodically captured by cameras installed above the feed troughs, and image recognition algorithms are used to determine whether there are signs of mold, clumping, or other characteristics indicating a decline in freshness. For feeding equipment operating parameters, the operating current, speed, and cumulative running time of the precision feeding motor are read from its driver. The start / stop status and travel information are obtained through position sensors installed on the feeding mechanism. The total volume of remaining feed is obtained through ultrasonic or pressure level gauges in the feed hopper. Edge computing nodes poll or receive data packets from all the aforementioned sensing terminals through their integrated multi-protocol communication interfaces. They perform timestamp alignment, unit unification, and filtering of invalid and outlier values ​​based on preset rules on the received raw data to form a structured set of multi-source parameters, which is then stored in a local buffer and awaits the call in step 2.

[0021] Step 2 involves retrieving the multi-source parameter temporal data preprocessed in Step 1 from the buffer of the edge computing node, organizing it into a normalized input tensor, and inputting it into a pre-trained machine learning model employing a fusion architecture of temporal convolutional networks and self-attention mechanisms. This model analyzes and calculates the input, outputting a probability distribution of the predicted personalized feeding requirements for the target individual cattle and a suggested range for a rationing threshold. Subsequently, within the physiological safety boundary determined by the cattle's current weight, breed standards, and growth stage, a gradient-varying feeding regulation signal is applied to the cattle, and its feeding behavior response is monitored in real time. Based on the monitored response data, a set of dynamic nutrient supply and demand response features is calculated and generated. Finally, these response features are fused with the probability prediction results initially output by the model at the feature level to construct a feeding-rationing risk map that quantifies health and production risks under different feeding levels.

[0022] It should be noted that, in this application, the personalized feeding requirement refers to the theoretical feed supply calculated to meet the optimal nutritional intake of a specific individual cattle under current physiological and environmental conditions. The feed restriction threshold refers to the critical value set to trigger active reduction of feed amount intervention to control the cattle's body condition, prevent metabolic diseases, or optimize feed efficiency. The probability prediction result refers to the possible range of feed requirement values ​​and their confidence levels, expressed in the form of a probability distribution, as output by the machine learning model, and the possible range and probability of the feed restriction threshold. The gradient feeding regulation signal refers to a series of small, stepwise, experimental feed amount change instructions designed within the physiological safety boundary. The nutritional supply and demand response characteristics refer to the dynamic feature vector extracted by analyzing the lag, sensitivity, and saturation of the cattle's actual feed intake, feeding behavior patterns, and rumination activity relative to the signal changes after receiving the gradient regulation signal. The feeding-dieting risk map is a two-dimensional or higher-dimensional data matrix or probability surface, whose horizontal and vertical axes respectively map the feeding amount level and the dieting intensity level. Each element in the matrix or each point on the surface is associated with a quantified risk score, which is used to characterize the probability of cattle experiencing health risks or decreased production performance under this feeding and dieting combination.

[0023] In practice, the model first reads all temporal parameters of the target cattle and their associated environment, feed, and equipment within the most recent preset time window from the buffer. Missing values ​​are imputed and standardized to form a multi-channel temporal tensor as model input. The core architecture of this machine learning model is as follows: the front end uses a temporal convolutional network to perform multi-level convolution operations on the input tensor to efficiently extract local and short-term periodic and trend features from the parameter sequence; the back end introduces a multi-head self-attention mechanism to perform global dependency modeling on the high-order feature sequences extracted by the temporal convolutional network, paying particular attention to long-range correlations between different parameters and between the current state and historical states, such as associating an abnormal decrease in rumination frequency with the high ammonia concentration environment of the previous few hours. The model ultimately outputs two probabilistic prediction results through a fully connected layer: the first part is the personalized feeding requirement of the cattle in the next feeding cycle, with the model outputting the mean and variance of its Gaussian distribution; the second part is the suggested feeding threshold, with the model outputting a parameter of a truncated normal distribution, with a lower bound of zero and an upper bound of the current feeding requirement. Next, based on the cattle's breed, current weight, body condition score, and growth stage, the system queries a pre-set knowledge rule base for the safe range of their daily dry matter intake. Using this safe range as a boundary, the system generates a gradient-based feeding regulation signal sequence containing several discrete levels. For example, within the safe range, based on the average demand predicted by the current model, several equally spaced feeding levels are taken upwards and downwards. In one or more independent validation cycles, the system adjusts the actual feed delivery amount to the cattle's feeding unit according to this sequence or based on a specific strategy. Simultaneously, the system uses the sensor network from step 1 to frequently monitor and record changes in the cattle's actual feed intake, feeding duration, feeding rate curve, and rumination behavior. Based on this response data, a set of nutrient supply and demand response characteristics is calculated, which may include: the response delay time of feed intake to the regulation signal, the sensitivity represented by the slope of the response curve, the plateau feed intake at saturation, and the correlation coefficient between changes in rumination activity and changes in feed intake. Finally, this set of dynamic response feature vectors is concatenated and fused with the probability distribution of feeding demand and the probability distribution of dieting threshold from the initial output of the model. Through another lightweight neural network or rule-based scoring system, it is mapped onto a two-dimensional grid. The horizontal axis of the grid represents different feeding levels, and the vertical axis represents different dieting intensity levels. Each grid point calculates a comprehensive risk score. The score comprehensively considers factors such as the decline in weight gain efficiency of cattle under this condition, the risk of metabolic diseases such as acidosis or laminitis, and the reduction in feed conversion rate, thus forming an intuitive and quantifiable feeding-dieting risk map.

[0024] In step 3, based on the feeding-dieting risk map generated in step 2, various extreme environmental conditions and feed shortage scenarios are simulated using a digital twin model corresponding to the cattle shed and herd. In the data domain constructed by the digital twin model and the physical domain associated with the actual feeding equipment, the probability prediction results output in step 2 and the preliminary scheduling scheme generated accordingly are double-checked for consistency. Based on the deviation and feasibility assessment results obtained from the check, combined with the feeding task priorities preset for different individual cattle or groups and the delayability characteristics of their feeding behavior, a multi-stage collaborative scheduling and decision-making process for feeding tasks and diet control instructions is executed, ultimately outputting a detailed scheduling result containing specific execution timing, targets, and quantities.

[0025] It should be noted that, in this application, the extreme environment and feed shortage scenarios are virtual combinations of harsh conditions exceeding current actual monitoring values ​​constructed in the digital space to verify the robustness of the scheduling scheme; the consistency verification is a dual verification and stress test of the prediction model and scheduling logic in both the virtual data space and the physical space connecting the actual equipment; the feeding task priority is a feeding instruction execution priority level pre-divided based on the cattle's physiological state, production value, and health urgency; the feed intake delayability refers to the length of time or flexible range during which a single feeding time can be delayed without affecting the cattle's health and production performance; the multi-level feeding-feeding-restriction coordinated scheduling is a hierarchical decision-making mechanism, which first allocates basic feeding and feeding-restriction instructions at the global level based on priority and risk map, then adjusts the timing of instructions based on delayability at the time resource level to resolve conflicts, and finally performs refined scheduling at the equipment resource level.

[0026] In some embodiments, generating extreme environment and feed shortage scenarios based on feeding-dieting risk maps using digital twin technology can be achieved through the following steps: Acquire historical extreme weather data and feed inventory safety early warning rules; Historical meteorological extreme value data are injected into the environmental parameter module of the digital twin model to construct an extreme environmental parameter sequence; Based on the feed inventory safety early warning rules and the current feed consumption rate, a feed shortage event is simulated in the feed supply module of the digital twin model. The digital twin model is driven to run under injected extreme environmental parameter sequences and simulated feed shortage events, generating an extreme scenario dataset containing virtual cattle reactions and equipment status.

[0027] It should be noted that in this application, the historical meteorological extreme value data are records of the highest temperature, lowest temperature, highest humidity, and other extreme values ​​that have occurred in the past few years at the location of the cattle shed, retrieved from local or cloud-based meteorological databases; the feed inventory safety early warning rule is based on farm management experience and sets the inventory threshold and predicted consumption time threshold for triggering feed shortage early warning.

[0028] In practice, firstly, the system retrieves records from the connected cloud-based agricultural meteorological database or locally stored historical records of any three consecutive days in summer where the average maximum temperature exceeds 35 degrees Celsius, and records of any winter minimum temperature below -10 degrees Celsius, for the location of the cattle shed within the past five years. These extreme temperature values, along with their associated typical humidity extremes, are used as historical meteorological extreme value data. Simultaneously, the system reads preset feed inventory early warning strategies from the farm management system. For example, a level one warning is triggered when the inventory of a certain type of concentrate feed in the feed tower falls below the expected consumption for the next three days, and a level two warning is triggered when it falls below the expected consumption for the next day. This strategy is quantified into specific inventory percentage thresholds and time thresholds, serving as feed inventory safety early warning rules. Then, in the constructed digital twin model of the cattle shed, a submodule simulating environmental parameters is located. The acquired historical high-temperature extreme value data is used to construct a virtual heat stress environmental parameter sequence lasting 72 hours, or the historical low-temperature extreme value data is used to construct a virtual cold environment parameter sequence. This sequence is then injected into the environmental parameter module, replacing the original real-time sensor data stream. Next, in the feed supply and consumption simulation submodule of the digital twin model, based on the acquired feed inventory safety early warning rules, the current feed balance parameter of the virtual silo in the model is modified to a value below the first-level early warning threshold, while keeping the feed consumption rate of the virtual cattle herd unchanged, thus simulating a feed shortage event that is about to occur but has not yet actually happened. Finally, the simulation engine of the digital twin model is started, allowing it to run the simulation for the next 24 to 72 hours in an accelerated manner under the injected extreme environmental parameter sequence and the set feed shortage conditions. Based on its internally integrated cattle physiological response model, feeding behavior model, and equipment control logic model, the model calculates and outputs the predicted feed intake changes of the virtual cattle, stress response indicators, execution status of the virtual feeding equipment, and alarm information under this extreme scenario. The collection of all these output data constitutes the extreme scenario dataset for subsequent verification.

[0029] In some embodiments, the consistency verification of probability prediction results and scheduling schemes in the data domain and physical domain can be achieved by the following steps: In the data domain, the extreme scenario dataset is input into the machine learning model used in step 2 to obtain the probability prediction results under extreme scenarios. These results are then compared with the original probability prediction results of the model under normal scenarios to calculate the prediction stability deviation. In the physical domain, a preliminary scheduling plan is generated based on the feeding-dieting risk map, and the preliminary scheduling plan is converted into an executable instruction sequence for the feeding equipment. The plan is then tested in the physical equipment simulation module connected to the digital twin model or during the idle period of the actual feeding equipment through a security interface. The success rate of instruction execution and state deviation are recorded. Consistency verification results are obtained by comprehensively considering the predicted stability deviation, instruction execution success rate, and state deviation.

[0030] It should be noted that, in this application, prediction stability bias is used to measure the degree of output fluctuation of the probability prediction result when facing extreme input conditions. The smaller the fluctuation, the higher the stability. Instruction execution success rate refers to the proportion of physical devices or high-fidelity simulation modules that can accurately execute the issued instruction sequence. State deviation refers to the difference between the actual state of the physical device after execution and the expected state of the instruction.

[0031] In practice, the first step is data domain verification. The currently used machine learning model is obtained from step 2, and the generated extreme scenario dataset is used as new input. The model will then output a new set of probability predictions for personalized feeding requirements and rationing thresholds under these virtual extreme conditions. These extreme scenario predictions are then compared with the original probability predictions obtained from real data in step 2. The comparison can be made by comparing the mean difference and variance change of the two probability distributions, for example, by calculating the absolute value of the difference between the means of two Gaussian distributions, or by calculating the percentage change in variance. These quantified differences are used as indicators of prediction stability deviation. If the deviation exceeds a preset threshold, it indicates that the model's predictions fluctuate significantly under extreme conditions. The second step is physical domain verification. Based on the current feeding-rationing risk map, and with the principle of avoiding high-risk areas and selecting low-risk feeding-rationing combinations for each cow, a preliminary scheduling plan is generated, including the planned feeding amount, feeding time, and possible rationing markers for each cow in the next cycle. This scheme is then converted into a specific sequence of equipment control commands, such as "At time T1, send a pulse to motor M in trough 2, slot 3, to drive the discharge of X kg of material." In a high-fidelity physical equipment simulation module tightly coupled with the digital twin model, or through a secure isolation interface during maintenance or idle periods of the actual feeding equipment, this command sequence is sent to the actuator simulation unit or the actual equipment controller. The simulation module or the actual equipment attempts to execute these commands. The system records whether each command is successfully received and executed, and the difference between the remaining amount in the simulated trough or the feedback value from the actual sensor and the expected value of the command after execution. The proportion of commands successfully executed is used as the command execution success rate, and the average feedback gap is calculated as the state deviation. Finally, the calculated predicted stability deviation, command execution success rate, and state deviation are summarized to form a consistency verification result that includes various quantitative indicators and an overall feasibility assessment.

[0032] In some embodiments, multi-level feeding-dieting coordinated scheduling is performed based on the verification results, combined with the feeding task priority and the deferability of feeding, and the scheduling result can be achieved through the following steps: Based on the prediction stability bias in the verification results, the feeding-dieting risk map was adjusted with confidence weighting to obtain a robust risk map. Based on the robust risk map and feeding task priority, a preliminary scheduling instruction set containing basic feeding amount and feeding restriction instructions is generated for each cow. Based on the lagability of food intake, the expected execution time of all instructions in the initial scheduling instruction set is flexibly adjusted to resolve equipment resource conflicts and generate a time-optimized instruction sequence. The time-optimized instruction sequence is mapped to specific feeding execution devices, device capabilities and status are verified, conflicting instructions are finally parsed and sorted, and a detailed scheduling result that can be issued and executed is formed.

[0033] It should be noted that in this application, the robust risk map is a risk quantification map with higher reliability after being verified by extreme scenarios and adjusted for confidence; the preliminary scheduling instruction set is an instruction set generated only based on the individual needs and priorities of cattle without considering time resource competition; and the time-optimized instruction sequence is an instruction sequence that is rearranged in the time dimension to avoid conflicts after taking into account the flexibility of feeding behavior.

[0034] In practice, the system first reads the consistency verification results. If the prediction stability deviation is small, the original feeding-dieting risk map generated in step 2 is considered to have high confidence, and no adjustment or only minor adjustments are made. If the prediction stability deviation is large, the risk scores of the corresponding areas in the risk map are attenuated and weighted according to the magnitude of the deviation. For example, the risk value of the area corresponding to high deviation is multiplied by a confidence coefficient less than one, thereby reducing the decision weight of these areas with high uncertainty and obtaining a robust risk map. Then, the system reads the feeding task priority of each cow from the breeding management database. This priority can be dynamically calculated based on factors such as cow type, milk production, gestation stage, and health status. For example, dairy cows in their peak lactation period are set to the highest priority, sick cows that need to be isolated for observation are set to high priority, and growing cows are set to ordinary priority. Next, for each cow, in the robust risk map, the point with the lowest risk score that meets its basic nutritional needs is found on the corresponding feed amount-diet intensity plane. The feed amount corresponding to this point is taken as the cow's basic feed amount. If the point corresponds to a non-zero diet intensity, a diet instruction is generated. Combined with priority, cows with higher priority have the right to have their basic feed amount met first during resource competition. This method generates an initial scheduling instruction set for all cows. Next, the system handles time conflicts. From the individual cow profiles or behavioral models, the feed intake deferability parameter for each cow is obtained. This parameter can be represented as the maximum time a cow's planned feeding time can be postponed. The system sorts all initial scheduling instructions according to their original planned time points, forming a timeline. When it is detected that within a short period, the number of instructions pointing to the same physical feeding device, such as the same precision feeder, exceeds its processing capacity, a time adjustment is triggered. The system will adjust the execution time of some instructions based on the cow's priority and feed intake deferability. The adjustment rule prioritizes the timely execution of high-priority instructions, while shifting instructions with greater delayability and relatively lower priority to the next possible timeframe, until the equipment load during that time period decreases to within its capacity. After multiple rounds of iterative adjustments, a time-optimized instruction sequence is generated that is more evenly distributed in time and avoids instantaneous equipment overload. Finally, the system performs equipment-level verification. Each instruction in the time-optimized instruction sequence is mapped to a specific feeding execution device, such as a specific mixer truck, conveyor belt motor, or precision feeding port. The system checks whether the target device is available at the instruction's time point and whether its remaining feed is sufficient. If conflicts are found, such as two instructions requiring the same device to perform different actions at the same time, the conflicting instructions are reordered or a small waiting interval is inserted based on instruction priority, cattle delayability, and the time required for equipment reset. After all conflicts are resolved, a final detailed scheduling result is generated.The result is a complete plan that includes the precise execution timestamp of each instruction, the target cattle identifier, the target equipment identifier, the specific feeding amount or diet control parameters, and the expected execution time.

[0035] In step 4, the detailed scheduling results finalized in step 3 are converted into specific equipment control command sequences via the field control network and sent to the corresponding feeding execution equipment. After the equipment executes the commands, it collects real-time data on the actual feeding process, periodic changes in cattle body weight, and indirectly calculated feed utilization rates through the multi-source sensor network described in step 1, and transmits this data back to the central processing system. Based on the difference between the actual results and the original scheduling expectations, the system constructs model feedback error and rule evaluation indicators, which drive the iterative update of the machine learning model parameters in step 2 and adaptively optimize the embedded dieting threshold determination rules.

[0036] It should be noted that, in this application, the actual feed intake data refers to the set of quantitative information on the actual feed intake of an individual cattle, obtained directly by feeding monitoring equipment or indirectly by feeding monitoring equipment after the feeding instruction is executed; the cattle body weight change data is data obtained through periodic or continuous monitoring, reflecting the trend of cattle weight increase or decrease over a period of time; the feed utilization rate data is a derived indicator used to evaluate feed conversion efficiency, calculated by comprehensively considering production performance indicators such as feed intake, weight change, and milk production; the iterative update refers to the process of retraining and fine-tuning the internal weight parameters of the machine learning model using new input and output data; and the dieting threshold determination rule is a set of logical conditions used to determine under what conditions a dieting intervention should be initiated for a specific cattle.

[0037] In some embodiments, the scheduling results can be sent to the feeding execution device and the actual feed intake data, cattle body weight change data, and feed utilization data can be transmitted back using the following steps: The instruction sequence in the detailed scheduling results is encoded into a standard control instruction set that can be recognized and executed by the feeding execution device controller, and then distributed via industrial bus or wireless network protocol. The feeding execution equipment completes the feed delivery operation according to the control command, and the multi-source sensor network synchronously monitors and records the execution result data and bovine response data related to the command. The raw data from the monitoring records is cleaned, aligned, and aggregated to extract the actual feed intake and related cattle weight trend data for each scheduling instruction. The feed utilization rate index is then calculated and packaged into a feedback data package for transmission back.

[0038] It should be noted that, in this application, the standard control instruction set is a data packet set that conforms to the feeding device communication protocol format and contains the target device address, action type, action parameters and execution timestamp.

[0039] In practice, firstly, the central processing system converts the detailed scheduling results generated in step 3—the plan containing precise timestamps, target devices, and feeding amounts—into specific device drive instructions. For example, for a record stating "At time T, send an instruction to device D, dispensing amount M," the system encodes it into a data frame that the device controller can directly parse. This data frame typically contains a unique network identifier for the target device, an instruction action code, a numerical parameter representing the dispensing amount, and a time-synchronized absolute execution timestamp. These data frames are multicast or unicast to the corresponding feeding execution device controller via a pre-deployed field industrial Ethernet, CAN bus, or low-power wide-area wireless network. Then, the embedded controller of the feeding execution device, such as a precision feeder or automatic feeding cart, executes the corresponding drive operation when the instruction timestamp is reached, such as controlling a stepper motor to rotate a specific number of revolutions to dispense a quantitative amount of feed. Meanwhile, the sensor network deployed in the cattle shed is triggered or operates at a fixed frequency, collecting execution result data: weighing sensors or RFID readers installed at the target cattle's feeding position or individual identification channel record the actual feed intake of the cattle during the instruction period; the weight estimate continuously monitored by the smart collars worn by the cattle is reported periodically; environmental sensors continue to record environmental parameters. This data is tagged with the corresponding scheduling instruction ID. Finally, edge computing nodes or the central server process the collected raw data. The cleaning operation includes removing obviously erroneous sensor readings; the alignment operation associates data from different sensors within the same time window based on the instruction ID and timestamp; the aggregation operation summarizes the corresponding actual feed intake for each executed scheduling instruction and associates it with the cattle's weight change curve before and after the instruction execution. Based on the actual feed intake and weight change data, combined with the cattle's milk production and other production data, the feed utilization rate index for that period is calculated using formulas such as feed utilization rate = output (e.g., weight gain, milk production) / feed intake. All processed data, including instruction ID, planned feed amount, actual feed intake, weight change trend, and calculated feed utilization rate, are encapsulated into a structured feedback data packet and sent back to the data storage module of the central processing system.

[0040] In some embodiments, iteratively updating the parameters of the machine learning model and the diet threshold determination rules based on the returned data can be achieved using the following steps: Based on the returned feedback data packets, the multi-dimensional errors between the feeding demand and feeding threshold suggestions predicted by the model and the actual observed cattle responses and production efficiency indicators are calculated. Using the aforementioned multidimensional errors as part of the loss function, incremental learning or periodic batch training methods are employed to adjust and update the network weight parameters of the machine learning model. By utilizing historical feedback data within a sliding time window, we can statistically analyze the health and production performance indicators of cattle under different dieting threshold triggering conditions, and dynamically adjust the conditional parameters or logical combinations in the dieting threshold determination rules.

[0041] It should be noted that, in this application, multidimensional error is a comprehensive indicator that measures the difference between the model's predicted value and the actual observed value across multiple relevant evaluation dimensions; incremental learning is a training mode that uses new sample data to continuously and online fine-tune an existing model; periodic batch training is a mode that periodically collects a batch of new data to retrain the model; and time sliding window refers to a fixed recent historical period that is shifted forward over time for analysis.

[0042] In practice, firstly, the model management module of the central processing system reads feedback data packets related to recently executed scheduling instructions from the database. For each scheduling cycle of each cow, the module extracts the mean of the probability distribution of the personalized feeding demand initially predicted by the model in step 2, and the actual feed intake returned in step 4. It calculates the absolute or relative error between the predicted demand and the actual feed intake. Simultaneously, the system evaluates the production efficiency of that cycle, such as calculating the feed utilization rate. If the model has provided a suggested feeding threshold, it observes whether the cows experience unexpected health events under that threshold suggestion. The prediction error, the gap between the actual feed utilization rate and industry standards or historical best levels, and the correlation with health events are combined into a multi-dimensional error vector. Then, the model management module constructs a new loss function, using minimizing a norm of this multi-dimensional error vector, such as the L2 norm, as one of the optimization objectives, combined with the original training data. The system employs an incremental learning algorithm, using feedback data packets as new training samples and inputting them into the machine learning model in step 2 with a small learning rate. The weight parameters of the temporal convolutional layers, attention mechanism layers, and fully connected layers in the model are fine-tuned through backpropagation. Alternatively, every fixed period, such as a week, the system merges all feedback data generated in the past week with historical data into a new training batch, performing a complete retraining of the model. In this way, the model parameters are iteratively updated, making its predictions more closely reflect the actual dynamics of the cattle herd. Finally, for updating the dieting threshold determination rules, the rule management module sets a 30-day sliding window, continuously analyzing all cases within the window that trigger the dieting threshold determination rules and execute dieting operations. The system statistically analyzes the changes in weight change rate, health event incidence rate, and subsequent feed utilization rate of the cattle after dieting in these cases. By analyzing the statistical relationship between these indicators and the applied dieting threshold, the rules are dynamically adjusted. For example, if analysis reveals that when the model-suggested diet threshold falls below a new critical value X, the probability of subsequent health problems in cattle increases while feed efficiency does not improve significantly, the system automatically modifies the "start dieting" condition in the rule base from the original "predicted threshold below Y" to "predicted threshold below min(X, Y)". Similarly, the logical parameters in the rules regarding dieting intensity and duration are adaptively optimized based on the statistical analysis results of historical feedback data, making dieting control more precise and safer.

[0043] This method, implemented in this embodiment, comprehensively perceives multi-dimensional information such as individual cattle physiology, environment, feed status, and equipment operation through a multi-source sensor network. It then utilizes an advanced machine learning model integrating temporal convolution and self-attention for deep analysis, achieving a paradigm shift from standardized group feeding to precise individual feeding. Its core benefits lie in its ability to dynamically predict the personalized nutritional needs of each cattle and, innovatively, proactively detect the physiological responses of cattle to changes in feed intake through gradient adjustment signals, thereby constructing a quantitative "feeding-dieting risk map." This method not only improves feed utilization efficiency and reduces the risk of metabolic diseases but also uses digital twin technology for robustness verification and multi-level collaborative scheduling in extreme scenarios, ensuring the safety, adaptability, and feasibility of feeding strategies in complex and ever-changing real-world environments. Ultimately, it achieves dual optimization of cattle health and production efficiency.

[0044] like Figure 2 As shown, this embodiment of the invention also provides an intelligent feeding scheduling and ration control system for cattle sheds that integrates multi-source sensors, used to implement the above-mentioned intelligent feeding scheduling and ration control method for cattle sheds that integrates multi-source sensors. The system includes: The multi-source sensor data acquisition module is used to collect, in real time or periodically, the physiological parameters of the target cattle, the environmental parameters of the cattle shed, the feed status parameters, and the operating parameters of the feeding equipment through a multi-source distributed sensor network deployed inside the cattle shed and its related facilities, and to preprocess and store the data through edge computing nodes. The personalized feeding demand prediction and risk map generation module is used to obtain preprocessed multi-source parameter time series data from edge computing nodes, input it into a pre-trained machine learning model, output personalized feeding demand prediction values ​​and dieting threshold suggestions, and generate a feeding-dieting risk map through gradient feeding adjustment and response feature fusion. The digital twin verification and scheduling decision module is used to simulate extreme scenarios based on the feeding-dieting risk map using a digital twin model, verify the consistency between the prediction results and the preliminary scheduling plan, and perform multi-level collaborative scheduling in combination with the feeding task priority and the lag of feeding to generate detailed scheduling results. The feeding control command execution and feedback optimization module is used to convert detailed scheduling results into equipment control commands and send them to the feeding execution equipment, collect actual execution data, and iteratively update machine learning model parameters and dieting threshold determination rules based on feedback.

[0045] The multi-source sensor data acquisition module specifically includes: The sensor network coordination unit is used to coordinate the operation of a multi-source distributed sensor network through edge computing nodes. It polls or receives data packets from various sensor terminals and performs timestamp alignment, unit unification, and outlier filtering to form a structured set of multi-source parameters. The bovine physiological parameter acquisition unit is used to collect bovine movement, posture and sound data through an accelerometer, gyroscope and miniature microphone in the smart collar, and to calculate estimated weight, rumination frequency and feeding speed. The cattle shed environmental parameter acquisition unit is used to deploy temperature and humidity sensors and ammonia sensors at key locations in the cattle shed to collect data at set intervals and send it to the edge computing node via a wireless network protocol. The feed status parameter acquisition unit is used to analyze the nutritional components of feed through a near-infrared spectroscopy probe, monitor the remaining amount of feed through a weighing sensor, and determine the freshness of feed through a spoilage model or image recognition. The feeding equipment operating parameter acquisition unit is used to read the current, speed and running time through the motor driver, obtain the status of the feeding mechanism through the position sensor, and obtain the feed inventory through the level gauge.

[0046] The personalized feeding demand prediction and risk map generation module specifically includes: a data preprocessing unit, used to read multi-source parameter time series data from the buffer, perform missing value imputation and standardization processing, and form a multi-channel time series tensor as model input; The machine learning model inference unit is used to extract local features using a temporal convolutional network and model global dependencies through a self-attention mechanism, outputting a probability distribution of feeding demand and suggestions for dieting thresholds. A gradient feeding regulation unit is used to generate a gradient feeding regulation signal sequence based on the physiological safety boundary of cattle and to monitor the feeding behavior response of cattle. The nutrient supply and demand response characteristic calculation unit is used to analyze changes in feed intake, feed intake rate and rumination behavior, and extract dynamic feature vectors such as response delay, sensitivity and saturation period; The risk map construction unit is used to fuse response features with model prediction results and map them onto a two-dimensional grid through a neural network or rule-based scoring system to generate a feeding-dieting risk map with a quantitative risk score.

[0047] The digital twin verification and scheduling decision module specifically includes: an extreme scenario simulation unit, used to acquire historical extreme meteorological data and feed inventory early warning rules, inject extreme environmental parameters and simulate feed shortage events into the digital twin model, and generate a virtual dataset; The consistency verification unit is used to compare the prediction stability deviation between extreme scenarios and normal scenarios in the data domain, and to calculate the success rate and state deviation of the instructions through device simulation or trial execution in the physical domain, and generate consistency verification results. The robust risk map generation unit is used to perform confidence-weighted correction on the risk map based on the prediction stability bias to obtain a reliable risk map. A multi-level collaborative scheduling unit is used to generate a preliminary scheduling instruction set based on a robust risk map and feeding task priorities, and to perform time optimization and equipment verification in conjunction with the delayability of feeding, and output detailed scheduling results.

[0048] The feeding control command execution and feedback optimization module specifically includes: a control command encoding and distribution unit, used to encode detailed scheduling results into a standard control command set and distribute it to the feeding execution device via an industrial bus or wireless network; The actual data acquisition and processing unit is used to monitor the equipment execution results and cattle response data, perform cleaning, alignment and aggregation, and extract actual feed intake, weight change and feed utilization rate. The model parameter update unit is used to calculate multi-dimensional errors based on feedback data and adjust the weights of the machine learning model using incremental learning or periodic batch training. The diet rule optimization unit is used to analyze historical data through a time sliding window and dynamically adjust the condition parameters and logical combinations of the diet threshold determination rules.

[0049] This system achieves intelligent scheduling and feed control in cattle sheds through the collaborative work of multiple modules, ensuring consistent methodology and efficient operation.

[0050] This system embodiment constructs a closed-loop automated management system integrating real-time perception, intelligent decision-making, precise execution, and continuous learning. Through the collaborative work of multiple modules, it realizes intelligent scheduling and ration control of cattle shed feeding, ensuring consistent methodological logic and efficient operation. Modular design enables efficient execution and stable operation and maintenance of complex logic. The combination of edge computing and central processing ensures real-time performance. The deployment of this system can effectively free up manpower, reduce the uncertainty of human intervention, and realize automated, refined, and intelligent management of personalized feeding and ration control for large-scale cattle herds. This is conducive to cost reduction, efficiency improvement, and sustainable development in modern animal husbandry.

[0051] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0052] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for intelligent feeding scheduling and diet control of cowshed by fusing multi-source sensing, characterized in that, The method comprises the following steps: Step 1: Collecting the physiological parameters of the cattle, environmental parameters, feed state parameters and feeding equipment operation parameters in the cowshed through multi-source distributed sensing terminals. The physiological parameters of the cattle include body weight, rumination frequency and feeding speed. The environmental parameters include temperature and humidity, ammonia concentration. The feed state parameters include nutritional composition, residual amount and freshness. Step 2: Inputting the preprocessed multi-source parameters into the pre-trained machine learning model to generate the probability prediction results of individualized feeding demand and fasting threshold. Applying gradient feeding adjustment signals within the safety boundary of the cattle physiological parameters, generating the nutritional supply-demand response characteristics according to the differences in the feeding response of the cattle, and fusing the probability prediction results to construct the feeding-fasting risk map. Step 3: Generating the extreme environment and feed shortage scenarios based on the feeding-fasting risk map using digital twin technology, and performing consistency checking of the probability prediction results and scheduling scheme in the data domain and the physical domain. According to the checking results, multi-level feeding-fasting collaborative scheduling is performed combined with the feeding task priority and the feeding delay to form the scheduling results. Step 4: Distributing the scheduling results to the feeding execution equipment and returning the actual feeding data, cattle body weight change data and feed utilization rate data. According to the returned data, the parameters of the machine learning model and the fasting threshold determination rules are iteratively updated.

2. The method of claim 1, wherein, In step 1, the multi-source distributed sensing terminals work in coordination with the edge computing nodes deployed on the spot of the cowshed. The raw data collected by the terminals are time-stamped, unitized and filtered for abnormal values to form a structured multi-source parameter set. In the physiological parameters of the cattle, the body weight is dynamically estimated by the motion data of the acceleration sensor and the gyroscope integrated in the intelligent collar through a body weight estimation algorithm. The rumination frequency is obtained by analyzing the sound data collected by the microphone in the intelligent collar combined with the motion pattern through a pattern recognition algorithm. The feeding speed is calculated by the individual feeding station or the 3D visual sensor in the feeding area based on the head residence time and the feed consumption rate. The environmental parameters are collected by the temperature and humidity integrated sensor and the electrochemical ammonia sensor deployed at the ventilation port, the resting area and the feeding area of the cowshed at a set sampling period. In the feed state parameters, the nutritional composition is obtained by online analysis through a near-infrared spectrum probe. The residual amount is monitored by the weighing sensor at the bottom of the feed trough. The freshness is determined by the feed delivery timestamp combined with the environmental parameters through a spoilage model or by camera image recognition. The feeding equipment operation parameters include the current and speed of the precision feeding motor, the running state of the conveyor belt or screw feeder, the switch state of each discharge port, the preset discharge amount, and the material level of the silo.

3. The method of claim 1, wherein, The specific architecture of the machine learning model in step 2 is as follows: a multi-layer convolution operation is performed on the input multi-source parameter time series tensor using a time series convolution network to extract local and short-term features; a multi-head self-attention mechanism is introduced to model the global dependency of the high-order feature sequence output by the time series convolution network, focusing on the long-range correlation between parameters and historical states; the Gaussian distribution mean and variance of the individualized feeding demand, as well as the truncated normal distribution parameters of the fasting threshold, are output through a fully connected layer, where the upper limit of the fasting threshold distribution is the current feeding demand; The generation method of the gradient feeding adjustment signal is as follows: based on the breed of the cattle, the current body weight, the body condition score, and the growth stage, the safe range of daily dry matter intake is obtained by querying the pre-set knowledge rule base, and the current predicted feeding demand mean is taken as the benchmark, a number of equally spaced feeding levels are taken upward and downward within the safe range to form a gradient feeding adjustment signal sequence; The nutritional supply and demand response characteristics include the response delay time of intake to the adjustment signal, the sensitivity represented by the response curve slope, the saturated intake, and the correlation coefficient between the change of rumination activity and the change of intake; The construction method of the feeding-fasting risk map is as follows: the nutritional supply and demand response characteristic vector and the probability prediction result are spliced and fused, and are mapped to a two-dimensional grid through a neural network or a rule scoring system, the grid horizontal axis is the feeding level, the vertical axis is the fasting intensity level, the comprehensive risk score is calculated for each grid point, and the score considers the decline of weight gain benefit, the risk of metabolic disease, and the reduction of feed conversion rate.

4. The method of claim 1, wherein, In step 3, the extreme environment and feed shortage scenarios are generated using digital twin technology, including: Obtain historical meteorological extreme value data and feed inventory safety warning rules; Inject historical meteorological extreme value data into the environmental parameter module of the digital twin model to construct an extreme environmental parameter sequence; Simulate a feed shortage event in the feed supply module of the digital twin model based on the feed inventory safety warning rules and the current feed consumption rate; Drive the digital twin model to run under the extreme environmental parameter sequence and the simulated feed shortage event to generate an extreme scenario data set containing virtual cattle responses and equipment states; The historical meteorological extreme value data includes the maximum temperature, minimum temperature, and humidity extreme value records in the past years at the location of the cattle shed, and the feed inventory safety warning rules include the inventory threshold and the predicted consumption time threshold that trigger the shortage warning.

5. The method of claim 1, wherein, In step 3, the consistency check of the probability prediction result and the scheduling scheme in the data domain and the physical domain includes: In the data domain, input the extreme scenario data set into the machine learning model to obtain the probability prediction result under the extreme scenario, compare it with the original probability prediction result under the normal scenario, calculate the prediction stability deviation, and quantify the prediction stability deviation by comparing the probability distribution mean difference and the variance change; In the physical domain, generate a preliminary scheduling scheme based on the feeding-fasting risk map, convert it into a feed equipment executable instruction sequence, and execute it in the idle period of the digital twin model connected physical device simulation module or actual device, record the instruction execution success rate and state deviation; The consistency check result is obtained by integrating the prediction stability deviation, the instruction execution success rate, and the state deviation.

6. The method of claim 1, wherein, The multi-level feeding-fasting collaborative scheduling in Step 3 includes: According to the prediction stability deviation in the verification result, the confidence weighted correction is performed on the feeding-fasting risk map to obtain a robust risk map; Based on the robust risk map and the feeding task priority, a preliminary scheduling instruction set including the basic feeding amount and the fasting instruction is generated for each cow, wherein the feeding task priority is dynamically divided according to the cow type, milk yield, pregnancy stage and health condition; Based on the feeding delay, the expected execution time of all instructions in the preliminary scheduling instruction set is flexibly adjusted to resolve device resource conflicts, and a time-optimized instruction sequence is generated, wherein the feeding delay represents the maximum time allowed for postponing the single feeding time of the cow; The time-optimized instruction sequence is mapped to the specific feeding execution device, the device capability and state are checked, the conflicting instructions are analyzed and sorted according to the priority, the delay and the device reset time, and a detailed scheduling result is formed.

7. The method of claim 1, wherein, In Step 4, the scheduling result is sent to the feeding execution device and the data is returned, including: The instruction sequence in the detailed scheduling result is encoded into a standard control instruction set recognizable by the feeding execution device controller, which includes the target device address, action type, action parameter and execution timestamp, and is sent through industrial bus or wireless network protocol; The feeding execution device completes the feed delivery operation according to the control instruction, and the multi-source sensing network synchronously monitors and records the actual feeding amount, the weight change of the cow and the environmental parameters; The raw data of the monitoring record is processed by cleaning, aligning and aggregating, the actual feeding amount corresponding to each scheduling instruction, the weight trend data of the associated cow, and the feed utilization rate index are extracted, and the feed utilization rate is calculated by the ratio of output to feed intake, including weight gain or milk yield; The actual feeding amount, weight change data and feed utilization rate index are packaged as feedback data packets and returned to the central processing system.

8. The method of claim 1, wherein, In Step 4, the parameters of the machine learning model and the fasting threshold determination rule are iteratively updated according to the returned data, including: Based on the feedback data packet, the multi-dimensional error between the predicted feeding demand, the fasting threshold suggestion and the actual observed cow reaction and production efficiency index is calculated, including the prediction error, the feed utilization rate gap and the health event correlation; Taking the multi-dimensional error as part of the loss function, the network weight parameters of the machine learning model are adjusted and updated by using incremental learning or periodic batch training; Using the historical feedback data in the time sliding window, the health and production performance indicators of the cow under different fasting threshold trigger conditions are statistically analyzed, and the condition parameters or logical combinations in the fasting threshold determination rule are dynamically adjusted.

9. A system for intelligent feeding scheduling and diet control in a cowshed by fusing multi-source sensing, for implementing the method of any one of claims 1-8, characterized in that, It includes: A multi-source sensing data acquisition module for real-time or periodic acquisition of cow body physiological parameters, cow house environmental parameters, feed state parameters and feeding device operation parameters through a multi-source distributed sensing network, and pre-processing and storage through an edge computing node; The personalized feeding demand prediction and risk map generation module is configured to obtain preprocessed multi-source parameter time series data from the edge computing node, input the data into a pre-trained machine learning model, output a personalized feeding demand prediction value and a diet threshold suggestion, and generate a feeding-diet risk map through gradient feeding adjustment and response feature fusion. The digital twin verification and scheduling decision module is configured to perform consistency verification on the prediction result and a preliminary scheduling scheme based on the feeding-diet risk map, perform multi-level collaborative scheduling in combination with a feeding task priority and a forage delay, and generate a detailed scheduling result. The feeding control instruction execution and feedback optimization module is configured to convert the detailed scheduling result into a device control command, issue the command to a feeding execution device, collect actual execution data, and update machine learning model parameters and diet threshold determination rules based on feedback iteration.

10. The system of claim 9, wherein, The multi-source sensing data acquisition module includes a sensing network coordination unit, a cow body physiological parameter acquisition unit, a cow house environment parameter acquisition unit, a feed state parameter acquisition unit, and a feeding device operation parameter acquisition unit. The personalized feeding demand prediction and risk map generation module includes a data preprocessing unit, a machine learning model inference unit, a gradient feeding adjustment unit, a nutritional supply-demand response feature calculation unit, and a risk map construction unit. The digital twin verification and scheduling decision module includes an extreme scenario simulation unit, a consistency verification unit, a robust risk map generation unit, and a multi-level collaborative scheduling unit. The feeding control instruction execution and feedback optimization module includes a control instruction encoding and issuing unit, an actual data acquisition and processing unit, a model parameter updating unit, and a diet rule optimization unit.

Images