An intelligent monitoring system and method for regulating the balance of bovine microbiota.

Abstract

This invention relates to the field of livestock breeding technology and discloses an intelligent monitoring system and method for regulating the microbial community balance in dairy cows. The system includes acquiring high-frequency characteristic data streams and low-frequency true value data streams, converging them to form a multi-source heterogeneous data stream, and centrally processing it by combining an initial intervention influence matrix and physicochemical masking characteristic parameters. A control center calls a projection mapping of high-dimensional fast variables onto low-frequency sequencing parameters, subtracting physicochemical disturbance components, and outputting an inferred microbial abundance vector. A state machine module determines the monitoring state based on the inferred microbial abundance vector, and under the monitoring intervention state, generates an intervention substance regulation ratio based on constraints including a behavioral deviation index, which is then quantitatively fed by a feeding terminal. This invention solves the problem of misalignment between the dimensionality and temporal sequence of monitoring data, eliminates initial non-biological fluctuation interference, incorporates spatial phenotypic anomalies into closed-loop constraints, and achieves safe microbial regulation.

Description

Technical Field

[0001] This invention relates to the field of animal husbandry technology, specifically to an intelligent monitoring system and method for regulating the balance of gut microbiota in dairy cows. Background Technology

[0002] In the monitoring and regulation of rumen microecology in dairy cows, existing monitoring methods typically rely on sensing devices to acquire high-frequency data and combine this with microbial sequencing to obtain low-frequency true values. However, in practical applications, there is an inherent dimensional and temporal misalignment between continuous high-frequency physicochemical monitoring data and lagging low-frequency microbial sequencing data. Due to the lack of centralized coordination of multi-source heterogeneous data streams and the lack of projection mapping processing from high-dimensional rapid variables to low-frequency sequencing parameters, existing technologies cannot continuously extrapolate rumen microecological abundance based on real-time high-frequency characteristic data. This results in a lack of real-time underlying data support when the system determines the monitoring status and executes subsequent regulation.

[0003] When the system is confirmed to be in an intervention state and exogenous intervention substances are administered, the initial dissolution and diffusion of these substances within the rumen generate strong physicochemical fluctuations. These abiotic disturbances mask the true abundance changes dominated by the microbial community. Existing computational models fail to pre-set physicochemical masking characteristic parameters and cannot subtract this physicochemical disturbance component from the high-frequency eigenvectors during projection calculations. This results in abiotic bias in the output projections, leading to a lack of objectivity in the system's projection calculations under intervention conditions.

[0004] Furthermore, existing control programs, when calculating the proportion of intervention substances and administering feeding, often focus monitoring only on the rumen environment, failing to incorporate external phenotypic abnormal responses into the control constraints. When feeding dosage triggers abnormalities, current technology cannot extract spatial residency information entropy and behavioral deviation index based on three-dimensional spatial coordinates, nor can it introduce a penalty decay factor to forcibly lower the upper bound of a single feeding when the behavioral deviation index exceeds the tolerance limit. This quantitative feeding operation, lacking external phenotypic closed-loop constraints, is prone to causing stress overload in dairy cows during continuous execution, and cannot guarantee the safety of the feeding operation. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent monitoring system and method for regulating the balance of dairy cow microbiota. It solves the problem that in existing dairy cow rumen microecological monitoring, there is a misalignment in dimensionality and temporal sequence between high-frequency characteristic data streams and low-frequency true value data streams, which makes it impossible to obtain the microecological status in real time. At the same time, it solves the problem that the physicochemical fluctuations caused by the initial feeding of exogenous intervention substances mask changes in the abundance of internal microbiota, and that single internal environment monitoring is difficult to prevent overload of the body's stress response.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The first aspect of this invention provides an intelligent monitoring system for regulating the balance of dairy cow gut microbiota, comprising:

[0008] The sensing terminals are deployed in the breeding farm environment and output a continuous high-frequency characteristic data stream;

[0009] The analysis unit obtains low-frequency ground truth data streams for Sentinel Bulls through sampling and sequencing operations;

[0010] The calibration end collects simulated intervention response data to generate an initial intervention impact matrix and physicochemical masking characteristic parameters.

[0011] The control center coordinates the convergence of the high-frequency feature data stream and the low-frequency truth data stream to form a multi-source heterogeneous data stream, and performs centralized processing in conjunction with the initial intervention influence matrix and the physicochemical masking feature parameters. The control center utilizes the inference module to invoke the established projection mapping from high-dimensional fast variables to low-frequency sequencing parameters, transforming and outputting the centrally processed multi-source heterogeneous data stream as an inferred microbial abundance vector. The control center uses a state machine module to receive the inferred microbial abundance vector and determine whether the current monitoring state is a free monitoring state or a monitoring intervention state.

[0012] Under the monitoring and intervention state, the optimization module generates the intervention substance regulation ratio according to the constraints, and the feeding terminal performs quantitative feeding operation.

[0013] Preferably, the analysis end includes a flexible guide tube and a pressure-limiting aspiration assembly configured in the sampling device. Within a set low-frequency sampling period, the analysis end controls the pressure-limiting aspiration assembly to initiate negative pressure aspiration to extract the mixed liquid phase, discarding the initially aspirated liquid phase and retaining the rumen liquid phase sample aspirated from the middle section of the flexible guide tube under sealed negative pressure. The analysis end uses a sequencing module to process the rumen liquid phase sample, extracting microecological abundance parameters and converting them into low-frequency ground truth vectors.

[0014] Preferably, the calibration end includes a fermentation module and an extraction module. The fermentation module uses the fermentation broth of the basal diet as a substrate to receive the candidate intervention substances for titration, providing a reaction environment. The extraction module simultaneously collects the reaction environment data of the fermentation module before and after the titration, extracts the difference in microecological abundance parameters and the corresponding mass change, and uses the quotient obtained by dividing the difference in microecological abundance parameters by the mass change as a partial derivative feature. The partial derivative features extracted for all the candidate intervention substances are algebraically recombinated according to a column arrangement to generate the initial intervention influence matrix.

[0015] The extraction module extracts the maximum change range of physicochemical indicators from the titration time based on the reaction environment data. The quotient obtained by dividing the maximum change range by the mass change is used as the maximum change range per unit mass to form a unit masking peak vector. The module also records the time required for the physicochemical indicators to recover from the maximum change range to within the baseline recovery tolerance set by the baseline as the decay time constant. The unit masking peak vector and the decay time constant are encapsulated into paired data tuples to establish a physicochemical masking feature library corresponding to the candidate intervention substances, providing the physicochemical masking feature parameters.

[0016] Preferably, the inference module performs time window alignment operation based on the low-frequency sampling time as the reference time anchor point under the monitoring free state without the application of exogenous intervention substances. The inference module extracts multiple sets of high-frequency feature vectors within the pre-fermentation cumulative time window of the reference time anchor point and calculates the arithmetic mean to output the mean of the high-frequency feature vectors. It then calls the partial least squares regression algorithm, using the mean of the high-frequency feature vectors as the independent variable and the corresponding low-frequency true value vector as the dependent variable, and performs projection dimensionality reduction and regression parameter fitting to establish the projection mapping as the basic steady-state mapping model.

[0017] Preferably, the state machine module calculates the Mahalanobis distance in real time and determines whether the Mahalanobis distance is consistently greater than a set statistical safety threshold within a continuously set monitoring period window. When the Mahalanobis distance is consistently greater than the statistical safety threshold, the state machine module determines that there are precursors to rumen microbiota imbalance, extracts the algebraic difference between the inferred microecological abundance vector and a preset benchmark vector, records it as a microecological deviation vector, and controls the monitoring state to switch from the monitoring free state to the monitoring intervention state. The optimization module uses the microecological deviation vector as the lower limit of the regulation requirement and the physical capacity limit of the single feeding of the candidate intervention substances as the boundary constraint of the feasible region to construct a constrained linear programming model. The constrained linear programming model is mathematically solved to obtain the optimal intervention substance mixing mass ratio vector, and the optimal intervention substance mixing mass ratio vector is converted into a regulation ratio instruction.

[0018] Preferably, the inference module, based on the actual feeding time and the actual feeding quality of various candidate intervention substances fed back by the feeding terminal, retrieves and extracts the unit masking peak vector and the decay time constant corresponding to each candidate intervention substance participating in this feeding from the physicochemical masking feature library based on a preset classification identifier. The inference module uses the extracted unit masking peak vector and the decay time constant, combined with the natural logarithm base, to construct an exponential decay function to calculate the physicochemical perturbation component vector; it subtracts the physicochemical perturbation component vector from the real-time acquired high-frequency feature vector to obtain a corrected high-frequency feature vector; it then calls the basic steady-state mapping model to perform projection dimensionality reduction calculation on the corrected high-frequency feature vector, outputting a net inferred abundance vector.

[0019] Preferably, the sensing terminal includes a behavior node that outputs the spatial phenotypic index. The state machine module divides the activity area of ​​the sentinel bull into multiple two-dimensional spatial grids, extracts the three-dimensional spatial coordinates continuously acquired by the behavior node as the spatial phenotypic index, and calculates the cumulative time length of the sentinel bull within each of the two-dimensional spatial grids. The cumulative time length is divided by the total time length of the monitoring period window to obtain the dwell probability. The state machine module calculates the spatial dwell information entropy based on the dwell probability, calls the pre-calibrated historical health baseline information entropy, calculates the absolute value of the algebraic difference between the current spatial dwell information entropy and the historical health baseline information entropy, and uses the absolute value as the behavior deviation index.

[0020] Preferably, the state machine module compares the behavioral deviation index with a preset stress tolerance limit. When the behavioral deviation index remains greater than the stress tolerance limit within the stress determination time window, the state machine module determines that the current feeding dose triggers an abnormal response, calculates the difference between the behavioral deviation index and the stress tolerance limit, and divides the difference by the stress tolerance limit to obtain the relative exceedance magnitude. The state machine module limits the relative exceedance magnitude to the maximum decay saturation range, and calculates a penalty decay factor by subtracting the limited relative exceedance magnitude from a constant. The penalty decay factor is multiplied into the upper bound vector of the single intervention substance feeding, and the values ​​of each dimension of the upper bound vector are forcibly lowered as new constraints and fed back to the constrained linear programming model.

[0021] Preferably, the state machine module determines whether the updated Mahalanobis distance has fallen back to within the statistical safety threshold, and checks whether the latest behavioral deviation index has fallen back to within half of the stress tolerance limit. When the Mahalanobis distance is less than or equal to the statistical safety threshold, and the behavioral deviation index is less than or equal to half of the stress tolerance limit, the state machine module determines that the fermentation environment has stabilized and recovers, and switches the control level from the monitoring intervention state back to the monitoring free state.

[0022] Preferably, the sensing terminal is divided into physicochemical nodes and behavioral nodes, which are respectively deployed inside the rumen and on the surface of the sentinel cattle to continuously acquire physicochemical indicators and spatial phenotypic indicators. The sensing terminal uses a microprocessor to synchronize and align the physicochemical indicators and spatial phenotypic indicators on the time axis. Based on a set reference timestamp, linear interpolation or a zero-order hold algorithm is used to resample and compensate for misaligned sampling data points. The resampled and compensated physicochemical indicators and spatial phenotypic indicators are concatenated to form a high-frequency feature data stream corresponding to the current high-frequency sampling time, and the high-frequency feature data stream is transformed into a high-frequency feature vector in an algebraic structure.

[0023] Preferably, when the net extrapolated abundance vector and the preset benchmark vector do not show a convergence trend, the analysis terminal obtains the low-frequency truth vector corresponding to the current moment and transmits it to the optimization module. The optimization module calculates the abundance algebraic difference between the low-frequency truth vector corresponding to the current moment and the previously output low-frequency truth vector, and uses the abundance algebraic difference as the actual abundance increment vector; it establishes a correspondence between the actual abundance increment vector and the actual feeding ratio vector when the feeding terminal actually performed feeding when the low-frequency truth vector was output, and uses the gradient descent error feedback learning law to perform posterior online correction on the intervention influence matrix of the current control cycle.

[0024] A second aspect of this invention provides an intelligent monitoring method for regulating the balance of bovine microbiota, comprising the following steps:

[0025] Sensing terminals are deployed in the farm environment to acquire continuous high-frequency feature data streams, and low-frequency truth data streams are acquired for the sentinel cattle in the farm environment.

[0026] Collect simulated intervention response data to generate an initial intervention impact matrix and physicochemical masking characteristic parameters;

[0027] The high-frequency feature data stream and the low-frequency truth data stream are integrated to form a multi-source heterogeneous data stream, which is then centrally processed in conjunction with the initial intervention influence matrix and the physicochemical masking feature parameters.

[0028] The multi-source heterogeneous data stream after centralized processing is invoked to establish a projection mapping from high-dimensional fast variables to low-frequency sequencing parameters. The projection mapping is used to output an inferred microbial abundance vector, and the current monitoring state is determined to be either a monitoring free state or a monitoring intervention state based on the inferred microbial abundance vector.

[0029] Once the monitoring intervention state is determined, an intervention substance regulation ratio is generated based on the constraints, and a quantitative feeding operation is performed according to the intervention substance regulation ratio to complete the intelligent monitoring of the microbial community balance regulation execution state.

[0030] This invention provides an intelligent monitoring system and method for regulating the bacterial flora balance in dairy cows. It has the following beneficial effects:

[0031] 1. This invention solves the problem of dimensional and temporal misalignment between continuous physicochemical monitoring and delayed microbial sequencing by centrally processing continuous high-frequency feature data streams and low-frequency true value data streams, and by calling the established projection mapping of high-dimensional fast variables to low-frequency sequencing parameters. It realizes the continuous inference of rumen microecological abundance based on real-time high-frequency data, and provides real-time data support for monitoring status determination and subsequent regulation execution.

[0032] 2. This invention calculates the physicochemical perturbation component by pre-generated physicochemical masking characteristic parameters at the calibration end, and subtracts the perturbation component from the high-frequency feature vector when the system performs projection calculation. This eliminates the non-biological physicochemical fluctuations caused by the dissolution and diffusion of exogenous intervention substances in the rumen during the initial feeding stage, so that the output net inferred abundance vector can truly reflect the actual changes dominated by the microbial community, and improves the objectivity of the system's inference calculation under intervention conditions.

[0033] 3. This invention calculates the spatial residence information entropy and behavior deviation index by extracting the three-dimensional spatial coordinates of sentinel cattle, and calculates the penalty decay factor when the behavior deviation index continuously exceeds the stress tolerance limit, forcibly lowering the upper bound of the quality of a single feeding in the constrained linear programming model, incorporating the abnormal response of external phenotypic behavior into the closed-loop constraint of intervention and regulation, avoiding stress overload caused by continuous feeding, and ensuring the safety of quantitative feeding operation. Attached Figure Description

[0034] Figure 1 This is a diagram of the intelligent monitoring system architecture for regulating the dairy cow microbiome balance according to the present invention.

[0035] Figure 2 This is a flowchart illustrating the overall workflow of the intelligent monitoring method of the present invention.

[0036] Figure 3 This is a flowchart of the multi-source heterogeneous data stream separation and acquisition method of the present invention;

[0037] Figure 4 This is a flowchart of the online micro-ecological state simulation and targeted formulation generation method of the present invention;

[0038] Figure 5 This is a flowchart of the intervention-state dual-scale collaborative monitoring and masking compensation deduction method of the present invention;

[0039] Figure 6 This is a comparison diagram of the transient fluctuations in pH and the masking interference stripping effect of the present invention;

[0040] Figure 7 This is a graph showing the online extrapolation and true value correction of microbial abundance in this invention. Detailed Implementation

[0041] 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.

[0042] See attached document Figure 1 This invention provides an intelligent monitoring system for regulating the balance of dairy cow microbiota, including a sensing terminal, an analysis terminal, a calibration terminal, a control center, and a feeding terminal.

[0043] Sensing terminals are deployed in the dairy farm environment to acquire continuous high-frequency characteristic data streams. These high-frequency characteristic data streams include physicochemical indicators and spatial phenotypic indicators. The sensing terminals are divided into physicochemical nodes and behavioral nodes. Physicochemical nodes are deployed inside the rumen of the dairy cows, continuously acquiring physicochemical indicators and sending them to the control center. Behavioral nodes are attached to the surface of the dairy cows, continuously acquiring spatial phenotypic indicators and sending them to the control center.

[0044] The analysis unit acquires low-frequency ground truth data streams containing microbial abundance parameters from specific sentinel cattle within the population. The analysis unit comprises a sampling device and a sequencing module. The sampling device, equipped with a flexible guide tube and pressure-limiting aspiration assembly, aims to obtain a representative liquid phase from the abdominal sac via the esophagus under anti-reflux operating procedures. It then probes deep into the rumen to extract a representative liquid phase sample and transfers it to the sequencing module. The sequencing module processes the liquid phase sample, extracts the microbial abundance parameters to form a low-frequency ground truth data stream, and sends it to the control center.

[0045] The calibration unit comprises a fermentation module and an extraction module. The fermentation module provides the reaction environment for titrating the intervention substance using the fermentation broth of the basal diet as the substrate. The extraction module collects the titration reaction process data from the fermentation module, generates an initial intervention effect matrix and related physicochemical masking characteristic parameters of the intervention substance, and sends them to the control center.

[0046] The control center establishes data communication links with the sensing terminal, analysis terminal, and calibration terminal respectively. It coordinates the reception of high-frequency feature data streams from the sensing terminal and low-frequency truth data streams from the analysis terminal, converging them into a multi-source heterogeneous data stream. This data stream is then centrally processed in conjunction with the initial intervention influence matrix and physicochemical masking characteristic parameters sent by the calibration terminal. The control center comprises a deduction module, a state machine module, and an optimization module. The deduction module establishes a projection mapping from high-dimensional fast variables to low-frequency sequencing parameters based on the centrally processed multi-source heterogeneous data stream and outputs an inferred microbial abundance vector. The state machine module receives the inferred microbial abundance vector and determines whether the system is in a monitoring free state or a monitoring intervention state. The optimization module generates the intervention substance regulation ratio based on constraints in the monitoring intervention state.

[0047] The control center generates and sends the regulation ratio command to the feeding terminal. The feeding terminal receives the regulation ratio command, executes the quantitative feeding operation, and feeds back the actual execution data to the control center, realizing intelligent monitoring of the execution status of microbial balance regulation.

[0048] See attached document Figure 2 This invention provides an intelligent monitoring method for regulating the balance of dairy cow gut microbiota, comprising the following steps:

[0049] S10. Separate Acquisition of Multi-Source Heterogeneous Data Streams. High-frequency feature data streams are continuously acquired using sensing terminals. Low-frequency safety sampling is performed on selected sentinel cattle at specific intervals using the analysis terminal to acquire low-frequency ground truth data streams containing microbial abundance parameters. The high-frequency feature data streams and low-frequency ground truth data streams together constitute a multi-source heterogeneous data stream.

[0050] S20. Construction of the basic mapping model and parameter benchmark. The intervention process is simulated using the calibration terminal to generate an initial intervention impact matrix and a physicochemical masking feature library. The derivation module aligns the high-frequency feature data stream and the low-frequency true value data stream within the synchronization time window to establish the initial basic steady-state mapping model of the system;

[0051] S30, Online Microbial Ecosystem Status Simulation, Anomaly Identification, and Targeted Formula Generation. Under routine monitoring in a free state, the simulation module uses a basic steady-state mapping model to calculate the high-frequency characteristic data stream acquired in real time, outputting an inferred microbial ecosystem abundance vector. The state machine module calculates the deviation between the inferred microbial ecosystem abundance vector and the health baseline. When the deviation exceeds a safety threshold, it identifies an impending imbalance and controls the system to switch to monitoring intervention mode. After switching, the optimization module constructs a linear programming objective function, introduces regulatory requirements and feeding amount boundary constraints, solves for the optimal intervention ratio vector, and the feeding terminal receives the optimal intervention ratio vector to feed the target object. The control center synchronously monitors the actual feeding execution status of the feeding terminal.

[0052] S40, Dual-scale Cooperative Monitoring and Physicochemical Masking Compensation. In the monitoring intervention state, the state machine module extracts changes in spatial dwell probability to generate a behavioral deviation index. The inference module extracts attenuation parameters of corresponding substances in the physicochemical masking feature library to construct a compensation matrix, strips physicochemical perturbation components from the high-frequency feature data stream, and outputs a net inference abundance vector after interference isolation.

[0053] S50, Dynamic Constraint Adaptive Correction and Truth-Anchored Posterior Loop Closure. Under the condition that the net inferred abundance vector does not show a convergence trend and the behavior deviation exponent does not converge, the optimization module compresses the upper bound boundary of the feed and regenerates the formula. After the sentinel cattle's real sequencing results are returned in the next sampling cycle, the control center uses the ratio of the real microecological change to the cumulative feeding amount to perform a posterior update of the influence matrix, achieving adaptive updates of the monitoring model and ensuring continuous output of intelligent monitoring results that closely match the real situation.

[0054] See attached document Figure 3 After the system is deployed and a communication link is established, in order to synchronously acquire the high-frequency physical boundary features and low-frequency biological truth values ​​of the target object, and to solve the technical defect that a single high-frequency sensor cannot directly characterize the abundance of the core bacterial community, the control center coordinates the sensing terminal and the analysis terminal to execute step S10, multi-source heterogeneous data stream separation and acquisition, which specifically includes the following sub-steps:

[0055] S11. The sensing terminal continuously acquires high-frequency characteristic data streams. To obtain the physical boundary conditions reflecting the transient physiological fluctuations of the target object, in this embodiment, physicochemical nodes are deployed in the rumen and reticulum regions of the target object. Physicochemical indicators, including pH, rumen fluid temperature, and redox potential, are continuously collected according to a preset high-frequency sampling time step to monitor the transient fermentation environment inside the rumen. The high-frequency sampling time step is a discrete time interval set based on the upper limit of the physiological change frequency of the target object and according to the Nyquist sampling theorem to ensure that the collected physicochemical indicators do not suffer from aliasing distortion. Specifically, pH is used to directly characterize the rate equilibrium state of carbohydrate fermentation and acid production, rumen fluid temperature fluctuations reflect the dynamic changes in metabolic heat production by the microbial community, and redox potential is used to assess the stability of the rumen anaerobic environment.

[0056] The behavioral node is fixed to the neck or ear of the target object, and synchronously collects spatial phenotypic indicators including three-dimensional spatial coordinates and triaxial motion acceleration at the same high-frequency sampling time step. The purpose is to analyze the feeding frequency of the target object through the change of three-dimensional spatial coordinates and to analyze its rumination activity rhythm using triaxial motion acceleration, thereby providing a macroscopic physiological homeostasis reference. Regarding the analog-to-digital conversion and signal sensing principles of the underlying probes within the physicochemical and behavioral nodes, those skilled in the art can use conventional electrochemical sensing probes and microelectromechanical inertial measurement units. The underlying hardware sensing mechanism is well-known technology in this field and will not be elaborated here.

[0057] The microprocessor inside the sensing terminal synchronizes and aligns physicochemical indicators with spatial phenotypic indicators on the time axis. Addressing the sampling clock offset and transmission delay physical phenomena generated during the operation of physicochemical and behavioral nodes, the microprocessor, based on a set reference timestamp, uses linear interpolation or a zero-order hold algorithm to resample and compensate for misaligned sampling data points, ensuring the consistency of operating conditions across all sensing channels at the same time point. For interpolation and resampling compensation of multi-source discrete signals, those skilled in the art can utilize conventional digital signal processing algorithms; the basic principles of these algorithms are well-known in the field and will not be elaborated upon here.

[0058] After data synchronization is complete, the microprocessor concatenates the physicochemical and spatial phenotypic indices to form a high-frequency feature data stream corresponding to the current high-frequency sampling time. For ease of subsequent processing, this high-frequency feature data stream is algebraically transformed into a stream with the following dimensions: High-frequency eigenvectors of ×1 .in, Represents the high-frequency sampling time; The total number of parameter dimensions representing physicochemical and spatial phenotypic indicators. The sensing terminal continuously outputs high-frequency feature vectors during operation. And send it to the control center.

[0059] S12. The analysis end performs low-frequency sampling on the sentinel cattle to obtain low-frequency ground truth data streams. During the parallel phase of acquiring high-frequency feature data streams, to obtain low-frequency real biological targets for subsequent model calibration and derivation, and to avoid increased system operating costs and population stress caused by invasive rumen fluid sampling of all individuals in the large-scale farming group, the system pre-selects a specific proportion of representative individuals from the large-scale farming group as sentinel cattle based on parity, lactation days, and historical health scores. The specific proportion is the minimum sample confidence ratio calculated using a finite population sampling formula after setting a significance level and tolerance limit based on the total population size of the large-scale farming group, to ensure the statistical representativeness of the sentinel cattle subset to the overall population.

[0060] Within a set low-frequency sampling period, the analysis unit controls the sampling device to extract representative rumen fluid samples from sentinel cattle. The low-frequency sampling period is a discrete number of days based on the biological cycle of natural rumen microbiota succession, filtering out high-frequency transient noise and extracting long-term steady-state drift characteristics. The system limits the operating time window of the sampling device to 2 to 4 hours after the target sentinel cattle finish feeding, eliminating dilution interference from feeding and drinking behaviors on the physicochemical properties of the rumen fluid samples. Based on the steady-state characteristics of carbohydrate fermentation substrate concentration and microbial metabolic activity within the rumen during this time period, the extracted rumen fluid samples have a more accurate characterization reference value for the core microecological state.

[0061] The flexible guide tube of the sampling device is inserted through the esophagus of a sentinel cow. Based on the scale markings on the outer side of the flexible guide tube, the device targets a representative fluid phase from the abdominal sac, locating it to a deep rumen fluid phase region 180 cm to 200 cm below the incisors. This depth positioning threshold range is based on the average physical distance measured from the oral-esophageal inlet to the deep rumen fluid retention area, obtained from the standard anatomical characteristics of adult target subjects. After positioning, the pressure-limiting aspiration component of the sampling device initiates negative pressure aspiration to extract the mixed fluid phase.

[0062] The sampling device employs anti-backflow operation during operation, discarding the initial 50-100 mL of liquid phase aspirated via a pressure-limiting aspiration component controlled by the system. This minimizes the impact of slightly alkaline saliva carried during rumination on the microecological abundance parameter sequencing results. The emptying volume threshold is calculated based on the total dead volume of the internal tubing of the flexible guide tube and the estimated flushing volume of contaminants adhering to the tube wall. After the initial emptying, the system retains only the rumen liquid phase sample aspirated from the middle section of the flexible guide tube under sealed negative pressure.

[0063] The sampling device seals and transfers the rumen liquid phase sample to the sequencing module. The sequencing module extracts microecological abundance parameters reflecting the rumen fermentation state from the rumen liquid phase sample. For the nucleic acid extraction, library construction, and biochemical processing of the rumen liquid phase sample for high-throughput sequencing, those skilled in the art can perform these procedures according to standard metagenomic analysis protocols. The biochemical reactions and instrument operations involved are well-known techniques in the field and will not be elaborated upon here.

[0064] The sequencing module packages the extracted microbial abundance parameters into a low-frequency ground truth data stream according to a predefined data protocol. This low-frequency ground truth data stream is also algebraically transformed into a dimensionality of [missing information]. ×1 low-frequency truth vector .in, Representing the A low-frequency sampling moment; The total number of dimensions representing the abundance of microbial ecosystem parameters.

[0065] The analysis end completes the first Low-frequency sampling time After data acquisition, the low-frequency truth vector The data is sent to the control center. High-frequency feature data streams and low-frequency truth data streams converge within the control center, forming a multi-source heterogeneous data stream system. These multi-source heterogeneous data streams are paired and recorded in the control center's data storage based on timestamp mapping relationships, supporting the modeling and deduction tasks of subsequent computing units.

[0066] After acquiring multi-source heterogeneous data streams, to address the lack of theoretical parameter support in the initial stage of the control algorithm, the control center collaborative calibration terminal executes step S20, which involves constructing the basic mapping model and parameter benchmark. In this embodiment, step S20 specifically includes the following sub-steps:

[0067] S21. The initial intervention impact matrix is ​​generated by simulating the intervention process using the calibration terminal. To obtain the quantitative correlation between the effects of candidate intervention substances on microecological abundance parameters, the system performs offline experimental calibration using the calibration terminal. The fermentation module of the calibration terminal is equipped with an in vitro biomimetic fermentation system. The fermentation module uses the fermentation broth of the basic diet used in the farm as the substrate and provides a reaction environment that simulates the actual internal temperature, pH and anaerobic conditions of the rumen.

[0068] The fermentation module contains a variety of alternative intervention substances, including specific strains of probiotic complexes and plant extracts. The specific strains of probiotic complexes provide targeted colonization substrates for the core microbial community, while the plant extracts inhibit the metabolic activity of competing harmful microorganisms. In this embodiment, the total number of alternative intervention substances is set to [number missing]. As a preferred method, the mass change of the candidate intervention substance is controlled by a precision micro-pump. The fermentation module independently titrates a known mass of a single candidate intervention substance into the reaction environment.

[0069] The extraction module synchronously collects reaction environment data from the fermentation module before and after titration. To quantify the intervention effect, the extraction module calculates the ratio of the difference in microbial abundance parameter changes to the corresponding change in the mass of the candidate intervention substance. To prevent division by zero overflow due to minor mass fluctuations, when the change in the mass of the candidate intervention substance approaches zero, a preset ratio stability constant is added to the denominator in the ratio calculation. The value of the ratio stability constant depends on the minimum effective precision limit of the floating-point operation of the microprocessor inside the control center. The extraction module defines the above ratio calculation result with anti-divergence constraints as the partial derivative feature. The partial derivative feature, at the physical level, characterizes the targeted microbial colonization gain brought by a unit intervention substance under a specific reaction system.

[0070] Subsequently, the extraction module will target all The partial derivative features of the extracted candidate intervention substances are algebraically recombinated based on the column arrangement, generating a dimension of... OK Initial intervention effect matrix of the real field of the column The initial intervention impact matrix reflects the static response gain of the microecology to the regulation of material input under ideal fermentation conditions. Among them, The total number of dimensions representing the abundance of microbial ecosystem parameters; This represents the total number of candidate intervention substances; This represents the initial intervention impact matrix. Regarding the environmental temperature control and gas path regulation logic of the in vitro biomimetic fermentation system operating in the fermentation module, those skilled in the art can set it based on the working principles of conventional microbial in vitro culture devices. The mechanical structure and operating method of its reaction device are well-known technologies in this field and will not be elaborated upon here.

[0071] S22. Extract physicochemical masking features using the calibration end and construct a physicochemical masking feature library. Exogenous candidate intervention substances possess specific acid-base buffering properties or thermodynamic properties. The introduction of candidate intervention substances instantaneously alters the local ion activity and heat distribution of the reaction liquid phase, thereby causing transient disturbances in the physicochemical indicators of the reaction environment, masking the true biological metabolic state signals. To eliminate the aliasing distortion between the physicochemical disturbance signals and the true microbial succession signals, the extraction module extracts transient disturbance features in parallel during the titration process in the fermentation module.

[0072] Regarding the first A number of alternative intervention substances, among which The range of values ​​is The extraction module monitors the maximum change in physicochemical indicators from the moment of titration, and divides the maximum change by the mass change of the corresponding candidate intervention substance for normalization. The maximum change per unit mass is then extracted to form the dimension. ×1 unit masking peak vector .

[0073] In forming unit masking peak vector Subsequently, the extraction module continues to monitor the time required for the physicochemical indicators to recover from their maximum variation to within the baseline recovery tolerance set by the initial stable baseline. In this embodiment, the baseline recovery tolerance is preferably 5% of the initial stable baseline, and the baseline recovery tolerance is a system steady-state recovery limit set based on the sensor's background noise tolerance and conventional measurement error experience. The extraction module records the recorded time as a decay time constant. .

[0074] Furthermore, the extraction module extracts the unit masking peak vector. With decay time constant Encapsulate into paired data tuples, and establish corresponding first... The physical and chemical masking characteristics of the candidate intervention substances are determined, and after performing a traversal calibration on all candidate intervention substances, the physical and chemical masking characteristic library is constructed and the data is sent to the control center.

[0075] S23. Establish the initial basic steady-state mapping model of the system using the deduction module. Under the monitoring free state without external intervention, the deduction module calls upon the multi-source heterogeneous data stream records aggregated within the data storage. Considering the asynchronous characteristics of high-frequency feature data streams and low-frequency true value data streams in terms of sampling frequency and physical time scale, the deduction module uses the low-frequency sampling time... Perform time window alignment operation for the reference time anchor point.

[0076] The inference module extracts low-frequency sampling times. Multiple high-frequency feature vectors within the fermentation accumulation time window corresponding to the sampling time. In this embodiment, the fermentation cumulative time window is preferably 2 hours. The duration of the fermentation cumulative time window is set based on the average circulation cycle of the rumen liquid chyme and the hysteresis constant of microbial attachment fermentation, to ensure that the extracted high-frequency feature data stream accurately reflects the cumulative fermentation environment during the sequencing sample generation stage. The extrapolation module analyzes the high-frequency feature vectors within the fermentation cumulative time window. Perform an arithmetic mean operation and output the mean of the high-frequency feature vectors after eliminating high-frequency measurement noise.

[0077] Based on the physical coupling and multicollinearity among the parameters of high-frequency feature vectors, the deduction module calls the partial least squares regression algorithm, using the mean of the high-frequency feature vectors as the independent variable and the corresponding low-frequency truth vectors as the independent variable. Using the variable as the dependent variable, projection dimensionality reduction and regression parameter fitting are performed to establish a basic steady-state mapping model reflecting the projection relationship from high-dimensional sensing signals to low-frequency sequencing parameters. The structure of the basic steady-state mapping model follows the formula:

[0078] ;

[0079] in, Represents the low-frequency sampling time The extracted low-frequency truth vector; Represents the low-frequency sampling time The mean of high-frequency feature vectors obtained by arithmetic averaging within the corresponding fermentation cumulative time window; This represents the symbol for mathematical mean operations; The transpose of the partial least squares regression coefficient matrix; The linear mapping baseline component, which represents the model prediction obtained by operating on the mean of high-frequency eigenvectors and the partial least squares regression coefficient matrix, represents the basic steady-state mapping model. This represents the residual vector during the fitting calculation process, used to absorb random error components of the system that are not explained by linear projection.

[0080] For the factor extraction rules and matrix cross-validation fitting process of the partial least squares regression algorithm, those skilled in the art can deploy and implement them using conventional multivariate statistical software or algorithm libraries. The algebraic decomposition derivation process at the underlying level of the algorithm is a well-known technique in the field and will not be elaborated here. The derivation module saves the basic steady-state mapping model as a benchmark for state inference in the subsequent online operating environment.

[0081] See attached document Figure 4 After completing the construction of the basic mapping model and parameter benchmark, the control center performs high-frequency online calculations based on the calibrated correlation parameters, executing step S30, online micro-ecological state simulation, anomaly identification, and targeted formula generation. In this embodiment, step S30 specifically includes the following sub-steps:

[0082] S31. The inference module uses a basic steady-state mapping model to calculate the high-frequency feature data stream acquired in real time, outputting an inferred microbial abundance vector. Under normal monitoring conditions, the system relies solely on the sensing terminal to continuously acquire the high-frequency feature data stream, ceasing low-frequency true value sampling and offline analysis. The sensing terminal continuously sends the high-frequency feature vector to the control center. Because the fluctuations in rumen physicochemical indicators and host phenotypic characteristics have a direct correlation with the metabolic products and activity rhythms of the microbial community, the inference module uses this multi-dimensional sensor data as an input source, receives the high-frequency feature vector, and establishes a real-time projection relationship using a partial least squares regression coefficient matrix. This allows it to infer the rumen microbial abundance, which is difficult to measure directly, using high-frequency physicochemical and spatial phenotypic indicators. The specific calculation logic of the inference module follows the formula:

[0083] ;

[0084] in, Represents the high-frequency sampling time The inferred microbial abundance vector output by the system is given by the values ​​of each dimension of the vector, which represent the inferred relative abundance of the corresponding core bacterial community in the current environment. The transpose of the partial least squares regression coefficient matrix; Represents the high-frequency sampling time The high-frequency feature vectors obtained are used to infer the abundance vector of the microbial ecosystem. Online soft measurement and estimation of the succession trend of the core microbial community has been achieved.

[0085] S32. The state machine module calculates and infers the deviation between the microecological abundance vector and the health baseline, identifies early signs of imbalance, and controls the system to switch to a monitoring and intervention state. The system pre-stores a baseline vector representing the stable state of the healthy microecological environment and a covariance matrix reflecting the distribution characteristics of historical healthy samples. To eliminate differences in numerical dimensions between different features and consider the correlation between variables, the state machine module uses the Mahalanobis distance algorithm to quantify the degree of deviation. As a preferred approach, to avoid multicollinearity of variables leading to singularity in the covariance matrix and thus making it impossible to invert, the state machine module adds a small anti-singularity compensation constant to the main diagonal elements of the covariance matrix to ensure the stability of the matrix calculation. The anti-singularity compensation constant is preferably 10. -6 Up to 10 -4 The constant is determined based on the lower limit of the precision of microprocessor floating-point operations to prevent excessively large compensation constants from damaging the covariance structure of the original data.

[0086] The logic of the state machine module in real-time calculation of Mahalanobis distance follows the formula:

[0087] ;

[0088] in, Represents the high-frequency sampling time The Mahalanobis distance calculated below is used to quantify the deviation of the inferred abundance from the health baseline; Represents the high-frequency sampling time The inferred microbial abundance vector output by the system; The value represents a preset baseline vector representing the homeostasis of a healthy microbiome, and its value is set based on the long-term sequencing history average of the healthy group. The transpose of the vector representing the algebraic difference between the inferred microbial abundance vector and the baseline vector; This represents the inverse matrix of the historical healthy sample covariance matrix with an added anti-singularity compensation constant.

[0089] The state machine module continuously monitors the Mahalanobis distance over time. To avoid system malfunctions caused by occasional data spikes, a sliding time window mechanism is introduced. The state machine module further determines whether the Mahalanobis distance consistently exceeds a set statistical safety threshold within a continuously set monitoring period window. The preferred monitoring period window length is 2 to 4 hours to match the basic physiological lag cycle of rumen emptying and fermentation, thereby filtering out transient data spikes caused by drinking water or short-term feeding. The statistical safety threshold is determined based on the chi-square distribution followed by the normal healthy sample set data, selecting the critical value corresponding to the 95% confidence level. The degrees of freedom of this chi-square distribution are set to be equal to the total number of dimensions of the inferred microbial abundance vector, thus ensuring the statistical significance of anomaly detection. When the conditions are met, the state machine module determines that there are precursors to rumen microbiota imbalance.

[0090] After identifying early signs of imbalance, the state machine module extracts the algebraic difference between the inferred microbial abundance vector and the baseline vector representing a healthy microbial homeostasis, recording this algebraic difference as the microbial deviation vector. Subsequently, the state machine module switches the system control level from the routine monitoring free state to the monitoring intervention state.

[0091] S33. The optimization module constructs a linear programming objective function, solves for the optimal intervention ratio vector, and executes feeding via the feeding terminal. After switching to the monitoring and intervention state, the optimization module activates the targeted formulation generation mechanism. The optimization module extracts the pre-set economic cost parameters of intervention materials within the system and constructs a linear programming objective function with the goal of minimizing the total intervention cost. Simultaneously, the optimization module uses the acquired microecological deviation vector as the lower limit of the regulation requirement and the single-feeding physical capacity limit of the candidate intervention materials as the boundary constraint of the feasible region. The constrained linear programming model constructed by the optimization module follows the formula:

[0092] ;

[0093] The constraints are:

[0094] ;

[0095] ;

[0096] ;

[0097] in, This represents the objective function value of the intervention cost; A mathematical operation identifier representing the search for the minimum value; The transpose of the intervention cost vector reflects the unit mass price of various alternative intervention substances; This represents the vector representing the optimal intervention ratio to be solved. The process of calculating the inner product representing the total cost of intervention; This represents the formulation requirement relaxation margin vector, used to provide a feasible solution space when constraints conflict; The transpose of the preset penalty coefficient vector; This represents the algebraic penalty term introduced when slack margin is enabled; This represents the intervention impact matrix used in the current regulatory cycle. When the intervention is first triggered, the intervention impact matrix used in the current regulatory cycle is equivalent to the initial intervention impact matrix. ; This represents the expected increase in the abundance of the microbial ecosystem after implementing a specific ratio; This represents the micro-ecological deviation vector extracted by the state machine module;

[0098] This represents a logical constraint inequality that indicates the regulatory effect must meet a lower limit requirement. The lower bound vector representing the mass of a single dose of intervention substance is used to prevent the controlled dose from being too small and losing its physiological significance. It is calibrated based on the lowest effective physiological activation dose of various candidate intervention substances in the body. The upper bound vector representing the mass of a single feeding of intervention substances is used to prevent digestive and metabolic stress caused by overfeeding. It is determined based on the toxicological safety thresholds of various candidate intervention substances and the maximum physical feeding capacity of the host animal in a single feeding. Physical boundary constraint inequalities representing the amount of intervention substance fed; This represents a logical boundary constraint with non-negative relaxation margin.

[0099] The optimization module mathematically solves the constrained linear programming model to obtain the optimal mass ratio vector of intervention materials that satisfies the regulation requirement constraint and whose cost is within the lower bound of the feasible region. Under specific severe imbalance conditions, if the compensation required by the microecological deviation vector is too large, causing an algebraic conflict between the regulation requirement constraint and the upper bound constraint of the feeding quality, making it unsolvable, the optimization module automatically introduces a formula requirement relaxation margin to reduce the lower limit of the regulation requirement. By obtaining a suboptimal feasible approximate solution, the uninterrupted operation of the control link is ensured. The weight of the formula requirement relaxation margin is controlled by a preset penalty coefficient. This penalty coefficient is empirically preset based on the economic tolerance and imbalance risk level of the farm. As an optimal method, the penalty coefficient is set to be on the order of 10 to 100 times the unit mass price in the intervention cost vector, thus forming an algebraic penalty tendency during the iterative process of mathematical solution, ensuring that the relaxation margin is activated when there is no solution.

[0100] For the solution process of linear programming models using the simplex method or interior point method, those skilled in the art can call conventional operations research algorithm solvers for calculation. The underlying iterative search principle is a well-known technology in this field and will not be elaborated here.

[0101] The optimization module transforms the obtained optimal intervention substance mixture mass ratio vector into a control ratio command. The control center sends the control ratio command to the corresponding communication interface of the feeding terminal. Based on the control ratio command, the feeding terminal controls the micro-discharge valves of each feed hopper inside, precisely dispensing the corresponding mass of intervention substance into the target feed trough, completing the targeted feeding action. The mechanical execution structure and electrical drive principle of the micro-discharge valves controlled by the feeding terminal can be implemented using conventional agricultural automated feeding devices or industrial weighing and batching scales. The related mechanical transmission and electrical control logic are well-known technologies in this field and will not be elaborated here. The control center simultaneously listens to and records the actual dispensing mass confirmation signal fed back by the feeding terminal, forming a closed loop for monitoring the execution status of the intervention action.

[0102] See attached document Figure 5 After the system switches to the monitoring and intervention state and the feeding terminal administers the alternative intervention substances, in order to address the problem of distortion in the microecological inference model caused by transient physicochemical interference from exogenous substance intake, and to introduce macroscopic signs as a physical verification aid for microscopic inference, the control center executes step S40, dual-scale collaborative monitoring, and physicochemical masking compensation. In this embodiment, step S40 specifically includes the following sub-steps:

[0103] S41. The deduction module extracts the attenuation parameters of corresponding substances in the physicochemical masking feature library to construct a compensation matrix, removes the physicochemical perturbation components from the high-frequency feature data stream, and outputs the net deduced abundance vector after isolating interference. In this embodiment, after the alternative intervention substance is fed into the rumen by the feeding terminal, the acid-base buffering properties and thermodynamic solubility properties of the alternative intervention substance itself will cause ion activity fluctuations and heat release in the liquid phase environment. This phenomenon is a physical response caused by the introduction of exogenous substances. In order to avoid the deduction module misjudging this transient fluctuation as a signal of microbial community metabolic succession and thus causing distortion of abundance deduction, it is necessary to remove the physicochemical perturbation components from the high-frequency feature data stream.

[0104] To this end, the simulation module obtains the actual feeding time and the actual feeding quality of various candidate intervention substances from the feeding terminal feedback. After data synchronization is completed, the module retrieves and extracts the unit masking peak vector and decay time constant corresponding to each candidate intervention substance participating in this feeding from the physicochemical masking feature library based on the preset classification labels. As a supplementary explanation, the unit masking peak vector and decay time constant in the physicochemical masking feature library are obtained in advance through isothermal in vitro digestion simulation experiments, by pulsely injecting a single dose of the candidate intervention substance into a standard volume of artificial rumen fluid, and calibrating the transient response curve of the physicochemical sensor by exponential decay fitting.

[0105] Based on the response principle of the damping attenuation characteristics of a first-order system, the derivation module constructs a formula for calculating the time-varying physicochemical perturbation component vector, specifically following the expression below:

[0106] ;

[0107] in, Represents the high-frequency sampling time The physicochemical perturbation component vector obtained by the cumulative calculation; Represents all Mathematical identifiers for algebraic summation of candidate intervention substances; Represents the obtained number The actual feeding quality of the candidate intervention substances; The corresponding number in the physical and chemical masking feature library The unit masking peak vector of the candidate intervention substances; Representing the The maximum theoretical masking peak response vector elicited by the actual feeding mass of the candidate intervention substances; The base of the natural logarithm; This represents the actual timing of the administration of the candidate intervention substance; It represents the elapsed time span between the current moment and the actual feeding action. The corresponding number in the physical and chemical masking feature library The decay time constant of the candidate intervention substances; Represents the dimensionless decay rate over time; This represents the decay coefficient of transient physicochemical disturbances over time, calculated based on the exponential decay law.

[0108] To ensure the rationality of the physical logic, the above calculations require time to meet certain conditions. And decay time constant The value of is constrained by the system's physical inertia, with a minimum physical limit greater than zero to prevent the exponential term from diverging by division by zero. After obtaining the physicochemical perturbation component vector, the derivation module subtracts the physicochemical perturbation component vector from the real-time acquired high-frequency eigenvector vector to obtain the corrected high-frequency eigenvector vector after isolating and masking interference. The calculation logic follows the formula:

[0109] ;

[0110] in, This represents the corrected high-frequency feature vector after isolating and masking interference. Represents the high-frequency sampling time The high-frequency feature vectors obtained through the sensing terminal; Represents the high-frequency sampling time The physicochemical perturbation component vector obtained by the cumulative calculation; This represents the environmental state characterization component of bio-fermentation after the elimination of external physicochemical shocks.

[0111] Furthermore, the inference module calls the established basic steady-state mapping model to perform projection dimensionality reduction calculation on the corrected high-frequency feature vector, and outputs the net inferred abundance vector. The calculation logic follows the formula:

[0112] ;

[0113] in, Represents the net inferred abundance vector after isolating physicochemical interferences; The transpose of the partial least squares regression coefficient matrix; This represents the corrected high-frequency feature vector after isolating and masking interference. This represents the algebraic transformation process of projecting the corrected high-frequency sensing signal onto the microbial abundance feature space through linear combination. This net inferred abundance vector eliminates short-term physical blind zone interference caused by feeding actions, ensuring the continuity and state fidelity of the microbial ecosystem inference results under intervention conditions.

[0114] S42. The state machine module extracts changes in spatial dwell probability to generate a behavioral deviation index. Beyond the micro-ecological scale of deduction, the system introduces behavioral features in parallel to establish a second monitoring scale. During the rumen microbiota imbalance and recovery period following intervention, the target object's feeding, drinking, and recumbent rumination patterns undergo consistent changes at a macroscopic level. Based on this, the state machine module divides the target object's activity area into multiple two-dimensional spatial grids with predetermined physical dimensions. As a preferred approach, the physical dimensions of the two-dimensional spatial grids are set based on the empirical ratio of the target object's average body length to its stride length; in this embodiment, it is set to 1m × 1m to ensure the capture of local activity aggregation features. For the coordinate mapping transformation of the two-dimensional spatial grids, those skilled in the art can use conventional raster map modeling methods; its spatial discretization modeling is a well-known technique in the field and will not be elaborated upon here.

[0115] After constructing the spatial grid, the state machine module extracts the behavior nodes of the sensing terminal at high-frequency sampling times. The three-dimensional spatial coordinates are continuously acquired within a sliding monitoring cycle ending at the target object. The total duration of the sliding monitoring cycle is preferably set to 4 to 6 hours, based on the typical feeding-to-rumination behavior switching cycle of the target object. The system calculates the cumulative time the target object spends within each two-dimensional spatial grid and divides this cumulative time by the total duration of the sliding monitoring cycle to obtain the probability of the target object's presence within each two-dimensional spatial grid.

[0116] Based on Shannon's information entropy principle, the state machine module calculates the spatial residency information entropy, which characterizes the distribution of spatial activities. This is to avoid situations where the residency probability within a certain two-dimensional spatial grid... When the value is zero, the natural logarithm operation causes mathematical divergence, resulting in the appending of a dwell probability lower bound constant to the probability numerical term. The specific calculation logic follows the formula:

[0117] ;

[0118] in, Represents the high-frequency sampling time The spatial dwelling information entropy, updated based on the sliding monitoring cycle, is used to quantify the degree of disorder or concentration in spatial activity patterns. Represents the total number of two-dimensional spatial grids; Represents the target object being in the first position. The dwell probability of a two-dimensional spatial grid; This represents the minimum probability of residence, and its preferred value is 10. -6 Up to 10 -5 The specific setting depends on the lower limit precision of the system processor's floating-point operations; This represents the corrected dwell probability after additional anti-divergence smoothing processing; The logarithm of the modified dwell probability reflects the amount of information in a single spatial grid. Representing the The entropy of the resident information of a two-dimensional spatial grid is calculated. This is a mathematical notation representing the summation of the dwell probability components over all two-dimensional spatial grids.

[0119] Subsequently, the state machine module calls the pre-calibrated and extracted historical health baseline information entropy within the system to calculate the absolute value of the algebraic difference between the currently output spatial residency information entropy and the historical health baseline information entropy. This absolute value is used as the generated behavioral deviation index. The historical health baseline information entropy is calculated by continuously statistically analyzing the spatial residency information entropy of target groups in a normal health state over multiple days under daily monitoring in a free state, and then determining its mathematical expectation. This value represents the behavioral dispersion benchmark under normal, uninterrupted conditions. The behavioral deviation index, as an independent macroscopic physical scale monitoring indicator, forms a cross-scale cross-validation system with the net inferred abundance vector at the micro-ecological scale, jointly supporting the adaptive evaluation and iterative decision-making of subsequent control links.

[0120] In this embodiment, after the intervention action is performed and the dual-scale co-monitoring of the intervention state is completed, the metabolic effect of the exogenous intervention substance in vivo is affected by changes in the individual animal's physiological state. Based on this time-varying metabolic characteristic, in order to achieve online evolution and control loop of system control parameters, the control center executes step S50, adaptive constraint boundary correction, and truth-anchored posterior loop closure. In this embodiment, step S50 specifically includes the following sub-steps:

[0121] S51. The state machine module performs adaptive constraint boundary correction based on the dual-scale collaborative monitoring results. Under monitoring intervention conditions, although the optimal intervention ratio vector output by the optimization module satisfies the regulatory requirements at the algebraic level, large doses of exogenous intervention substances may cause physical overload of the individual's digestive tract or local metabolic stress during actual animal ingestion. Therefore, the system control state machine module continuously receives the behavioral deviation index generated in step S40.

[0122] As a preferred approach, the state machine module compares the real-time calculated behavioral deviation index with a preset stress tolerance threshold. The stress tolerance threshold is the maximum recoverable deviation in animal behavior, set based on veterinary experience in farms; its specific value is determined by the statistical mean of the information entropy deviation of historical abnormal feeding samples. To filter out transient noise interference, the state machine module utilizes a stress determination time window, preferably set to 1 to 2 hours in length. When the behavioral deviation index consistently exceeds the stress tolerance threshold within the stress determination time window, the system determines that the current feeding dose has triggered adverse stress in macroscopic physiological behavior.

[0123] To ensure the physiological safety of the target population, the state machine module activates a boundary adaptive correction mechanism, using the difference between the behavioral deviation index and the stress tolerance limit as a proportional base to dynamically calculate the penalty decay factor. The specific calculation logic is as follows: the relative exceedance magnitude is obtained by dividing the aforementioned numerical difference by the stress tolerance limit. Simultaneously, to prevent excessively large relative exceedance magnitudes from causing system shutdown, the relative exceedance magnitude is limited to the maximum decay saturation range. The maximum decay saturation value is preferably between 0.3 and 0.5, a value calibrated based on the farm's tolerance for regulatory failure caused by a sudden drop in intervention dosage.

[0124] Subsequently, the state machine module calculates the penalty attenuation factor by subtracting the restricted relative exceedance magnitude from a constant of 1. Then, the penalty attenuation factor is multiplied into the upper bound vector of the mass feed for a single intervention, forcibly reducing the values ​​of each dimension of the upper bound vector. The reduced upper bound vector serves as a new boundary constraint condition and is fed back into the constrained linear programming model of the optimization module. Through numerical scaling, the feasible region boundary is dynamically contracted, avoiding the risk of overfeeding in subsequent control cycles.

[0125] S52. The analysis module acquires low-frequency true data streams and updates the intervention impact matrix using the optimization module. The initial intervention impact matrix obtained in the offline in vitro biomimetic fermentation system only represents the static microbial response gain under ideal physical conditions. Under the influence of complex enzyme systems and dynamic flow rates in vivo, the actual intervention efficiency will inevitably undergo systematic shifts. Therefore, the system needs to introduce real posterior data into the control link for parameter iteration.

[0126] Upon reaching the end of the set low-frequency sampling cycle, the system control and analysis terminal restarts the invasive physical sampling of the target object. The specific physical actions and biochemical processes for rumen fluid sample extraction, impurity filtering, and microbial abundance parameter sequencing reuse the low-frequency truth acquisition logic, and will not be elaborated further here.

[0127] The analysis module acquires the latest low-frequency truth vector and transmits it to the optimization module. Considering the physiological metabolic lag between the intake of exogenous intervention substances and the evolution of core microbial abundance, the optimization module extracts the low-frequency truth vectors corresponding to two adjacent low-frequency sampling periods, calculates the abundance algebraic difference between the two vectors, and defines this difference as the actual abundance increment vector. Simultaneously, the optimization module backtracks along the timeline to find intervention feeding records within the metabolic lag period preceding the current low-frequency truth sampling time, and extracts the actual feeding ratio vector of the feeding terminal during this period.

[0128] To achieve parameter iteration, the optimization module employs the gradient descent error feedback learning law. Based on the deviation between the expected theoretical response increment and the actual abundance increment vector, it performs posterior online correction of the intervention impact matrix used in the current regulation cycle. The specific calculation logic follows the formula:

[0129] ;

[0130] in, This represents the updated intervention impact matrix prepared for the next regulatory cycle; The matrix representing the impact of interventions used in the current regulatory cycle; The learning rate coefficient is used to control the step size of parameter iteration updates. It is preferably set between 0.01 and 0.1. The setting of this range is based on balancing the parameter convergence speed and preventing matrix abrupt changes caused by error surface oscillation. This represents the actual abundance increment vector calculated through two low-frequency real sampling operations. This represents the actual feeding ratio vector issued during the actual matching metabolic lag cycle; This represents the theoretical microecological response increment vector calculated based on the unupdated matrix; This represents the prediction error vector between real biological feedback and theoretical prediction models. The transpose of the actual feeding ratio vector; The inner product (i.e., the square of the magnitude of the vector) represents the actual feeding ratio vector. This represents the preset lower bound of the feeding norm, used to prevent divergence when divided by zero; by dividing by Normalization can prevent gradient explosion and matrix iteration divergence caused by excessively large actual feeding quality values. The gradient correction direction matrix representing the model bias with respect to the input weights, combined with the normalization coefficients, drives the intervention influence matrix to stably approximate the true in vivo response law through reverse compensation in the negative gradient direction.

[0131] S53. The derivation module anchors to the low-frequency true value and performs posterior closed-loop correction of the steady-state mapping model. Because the steady-state mapping model experiences sensor baseline drift and micro-ecological natural substrate succession over time, the derivation module needs to update this model synchronously. The derivation module uses the latest acquired low-frequency true value vector as the dependent variable data anchor point and the extracted mean of the high-frequency feature vectors within the corresponding fermentation cumulative time window as the independent variable data anchor point. It then uses an incremental partial least squares algorithm to retrain and update the partial least squares regression coefficient matrix.

[0132] For the online recursive update process of the partial least squares regression coefficient matrix, those skilled in the art can use conventional adaptive filtering or incremental machine learning algorithms for the underlying implementation. The covariance matrix iterative update and eigenvalue decomposition convergence logic involved are well-known technologies in this field and will not be elaborated here.

[0133] After completing the posterior closed-loop anchoring of all parameters, the state machine module uses the updated covariance matrix and the benchmark vector representing the steady state of the healthy micro-ecosystem to recalculate the Mahalanobis distance corresponding to the latest low-frequency truth vector. To ensure the mathematical stability of matrix inversion, anti-singularity compensation constants are also added to the main diagonal elements of the updated covariance matrix.

[0134] Meanwhile, to ensure the reliability of the judgment and avoid biased judgments due to reliance on calculation results from a single dimension, the state machine module adopts multi-dimensional joint judgment logic. Specifically, the state machine module not only determines whether the updated Mahalanobis distance has fallen back to within the set statistical safety threshold, but also checks in parallel whether the latest behavior deviation index has fallen back to within half of the stress tolerance limit.

[0135] Only when both of the above conditions are met simultaneously does the state machine module determine that the rumen microecological fermentation environment has been stably restored to a healthy and balanced steady state. Subsequently, the state machine module switches the system control level from the monitoring intervention state back to the daily monitoring free state, thus completing a complete closed loop of the entire life cycle of monitoring, deduction, regulation, and truth-anchored adaptive optimization. Specific Implementation

[0136] See attached document Figure 6 and attached Figure 7 This paper provides a specific application example of rumen microecological monitoring and regulation in dairy cows, which corresponds to the intelligent monitoring and intervention process when rumen acidosis precursors occur in a farm.

[0137] To acquire underlying data, during the implementation phase, the physicochemical nodes of the sensing terminal are specifically configured with an ion-selective electrode pH sensor with a glass breathable membrane, a PT1000 platinum resistance temperature sensor, and a redox potential probe with a solid-state reference electrode. The behavioral nodes are specifically configured with a spatial positioning module based on UWB ultra-wideband technology and a microelectromechanical triaxial inertial sensor. The internal microprocessor uses a microcontroller based on the ARM Cortex-M4 core and communicates with the control center via a LoRa radio frequency module. The sequencing module at the analysis end uses a 16S rRNA gene high-throughput sequencer based on the Illumina sequencing platform. The feeding terminal is specifically an automatic precision feeding machine equipped with a stepper motor-driven micro-feeding spiral.

[0138] Multi-source heterogeneous data stream separation and acquisition. The system sets the high-frequency sampling time step of the sensing terminal to 10 seconds. For a herd of 500 lactating cows, based on the finite population sampling formula, a significance level of 0.05 and a tolerance of 5% are set, and 30 cows are selected as sentinel cows. The system sets the low-frequency sampling period to 14 days. Three hours after the sentinel cows finish feeding, the sampling device inserts a flexible guide tube through the esophagus into the deep rumen to extract a mixed liquid phase. The pressure-limiting aspiration component empties the first 75 mL of liquid phase to prevent saliva interference. The sequencing module sequences the intercepted mid-segment sample, extracting microecological abundance parameters including the relative proportions of lactic acid-utilizing bacteria (such as *Macrococcus*) and lactic acid-producing bacteria (such as *Streptococcus*), forming a low-frequency ground truth data stream.

[0139] Basic mapping model and parameter benchmark construction. The system utilizes the fermentation module at the calibration end to perform offline calibration in an artificial rumen fluid environment at a constant temperature of 39℃. The candidate intervention substances were set as Saccharomyces cerevisiae extract, Bacillus subtilis preparation, and sodium bicarbonate buffer. The extraction module recorded parameter changes before and after titration to generate an initial intervention effect matrix. When sodium bicarbonate buffer was injected, the extraction module recorded the unit masking peak vector corresponding to the maximum change in liquid phase pH, and simultaneously recorded the 30-minute recovery time from this transient fluctuation to the baseline recovery tolerance, denoted as the decay time constant, to construct a physicochemical masking feature library. The extrapolation module extracted high-frequency feature data streams within the cumulative fermentation time window 2 hours prior to the low-frequency sampling time, and used partial least squares regression to establish a basic steady-state mapping model.

[0140] The inference module uses a basic steady-state mapping model to calculate the high-frequency feature data stream acquired in real time. When the system enters a monitoring free state, the inference module continuously receives and calculates high-frequency feature vectors, outputting a real-time inferred microbial ecosystem abundance vector.

[0141] The state machine module calculates and infers the deviation between the microbial abundance vector and the health baseline. The state machine module calculates the Mahalanobis distance, with a 95% confidence critical statistical safety threshold of 5.99 for the chi-square distribution. During a certain monitoring period, due to high concentrate intake, the target individual's Mahalanobis distance remained consistently between 6.5 and 7.2 within a 2-hour monitoring cycle. The state machine module determined that there were early signs of imbalance, and the system switched to monitoring intervention mode.

[0142] The optimization module constructs a constrained linear programming model to solve for the intervention formulation. The optimization module aims to minimize intervention costs, importing the unit price parameters of sodium bicarbonate and Bacillus subtilis preparations as the intervention cost vector. It sets a lower limit for regulatory requirements based on the microecological deviation vector, a lower mass bound vector based on the minimum activation dose, and an upper mass bound vector based on the toxicological threshold. The optimization module uses the interior-point method to solve the constrained linear programming model, outputting the optimal mass ratio vector of the intervention substance mixture, which is then administered by the feeding terminal. For the iterative optimization of the constrained linear programming model using the interior-point method, those skilled in the art can utilize general optimization solvers in numerical computing libraries; the Hess matrix calculation and direction search mechanisms are well-known techniques in the field and will not be elaborated upon here.

[0143] The simulation module extracts the physicochemical perturbation component from the high-frequency feature data stream. After the feeding terminal administers sodium bicarbonate buffer, a transient increase in rumen pH occurs due to physicochemical perturbation. Based on the masking compensation formula, the simulation module extracts the corresponding unit masking peak vector and the 30-minute decay time constant to calculate the physicochemical perturbation component vector. The physicochemical perturbation component vector is then subtracted from the high-frequency feature vector to output a corrected high-frequency feature vector (this process is reflected in the appended...). Figure 6 The corrected data after stripping the physical mask in the data is obtained, and the net inferred abundance vector after isolating interference is further calculated.

[0144] The state machine module extracts the behavior deviation index. Based on the three-dimensional spatial coordinates obtained by the UWB module, the state machine module divides the space into a two-dimensional spatial grid with a physical size of 1m × 1m. Within a 6-hour sliding monitoring cycle, the dwell probability of the target object is statistically analyzed, with an additional 10... -5 After calculating the minimum residency probability constant, the spatial residency information entropy is calculated, and the absolute value of the algebraic difference between the entropy and the historical health baseline information entropy is used as the behavioral deviation index.

[0145] The state machine module performs adaptive constraint boundary correction. The system continuously compares the behavioral deviation index with the stress tolerance limit. If the behavioral deviation index is greater than the stress tolerance limit for 1 hour, it indicates that the high-dose intervention has triggered local metabolic stress. The system calculates the relative exceedance magnitude and truncates it within the maximum decay saturation of 0.3, calculates the penalty decay factor, and multiplies it into the upper bound vector of the constrained linear programming model to shrink the constraint boundary.

[0146] The analysis unit obtains low-frequency ground truth values ​​to update the intervention impact matrix. After reaching the 14-day low-frequency sampling cycle node, the sampling device re-acquires liquid-phase sequencing data. The optimization module extracts the actual abundance increment vector and the corresponding actual feeding ratio vector. The learning rate coefficient is set to 0.05, and a feeding norm lower limit is added to prevent divergence. Using the product of the prediction error and the actual feeding data, the intervention impact matrix used in the current regulation cycle is updated along the negative gradient direction.

[0147] The extrapolation module updates the basic steady-state mapping model using a posteriori data points as anchors. It updates the partial least squares regression coefficient matrix and covariance matrix using the latest high-frequency eigenvector mean and low-frequency ground truth vector as data anchors. The state machine module recalculates the Mahalanobis distance and behavioral deviation index, determining that both have fallen back to within half of the statistical safety threshold and stress tolerance limit, respectively, and then switches the control level back to the daily monitoring free state.

[0148] The deduction and verification conclusions of this application embodiment are as follows:

[0149] Combined with appendix Figure 6 The data distribution and physical phenomena shown in the comparison of physicochemical index masking and stripping under the intervention state reveal that after the target object was fed an intervention substance containing sodium bicarbonate (at the 1.5h node of the horizontal axis monitoring time (h), i.e., the time of intervention substance feeding marked by the black dotted line), the original data (red solid line) of the sensing end corresponding to the rumen fluid phase pH on the vertical axis in the traditional uncompensated sensing link exhibits a sharp upward pulse-like fluctuation. Since the partial least squares mapping model is directly built on these physical quantities, this transient disturbance is incorrectly amplified into a drastic mutation in the microecology in the inference model. The system uses a masking compensation formula to strip this physical response from the original signal. The output corrected data after stripping the physical mask (blue dashed line) smoothly preserves the biological evolution trend of the fermentation substrate, confirming the effective role of the physicochemical masking feature library in avoiding false triggering mechanisms in the control loop.

[0150] Combined with appendix Figure 7The global process curves of the dual-scale extrapolation and anomaly detection of microbial abundance demonstrate that the projection mapping of high-frequency features onto microbial parameters possesses reliable early warning capabilities. Before the intervention trigger point (i.e., day 5 of the horizontal axis monitoring time), although the high-frequency extrapolation abundance curve (solid line) corresponding to the relative abundance (%) of lactic acid-utilizing bacteria on the left vertical axis was in a gradual decline phase, the Mahalanobis distance state curve (dotted line) corresponding to the Mahalanobis distance judgment index on the right vertical axis, through the combined multidimensional covariance features, had already exceeded the set statistical safety judgment threshold (dotted line) before the abundance reached a dangerous trough, establishing a preliminary judgment of microbial imbalance and providing the system with a valuable intervention and regulation time window. After implementing targeted formula feeding, the Mahalanobis distance index gradually declined, and the extrapolation abundance curve showed a rebound inflection point. The low-frequency true value sampling point (marked with squares) obtained when reaching the low-frequency sampling node on day 14 fell within the reasonable prediction error range of the extrapolation curve. The introduction of this real data anchor point drove the gradient descent correction of the control center. This closed-loop mechanism eliminates the cumulative prediction errors caused by sensor zero-point drift and individual metabolic differences during long-term operation, confirming that the dual-scale monitoring combined with the posterior update model has long-term accurate inference performance in the face of complex in vivo fermentation environments.

Claims

1. An intelligent monitoring system for regulating the balance of bovine microbiota, characterized in that, include: Sensing terminals are deployed in the breeding farm environment to acquire continuous high-frequency feature data streams; The analysis end acquires low-frequency ground truth data streams for the sentinel cattle in the farm environment; At the calibration end, simulated intervention response data are collected to generate an initial intervention impact matrix and physicochemical masking characteristic parameters. The control center coordinates the convergence of the high-frequency feature data stream and the low-frequency truth data stream to form a multi-source heterogeneous data stream, and performs centralized processing in conjunction with the initial intervention influence matrix and the physicochemical masking feature parameters; The control center uses the inference module to establish a projection mapping from high-dimensional fast variables to low-frequency sequencing parameters based on the centrally processed multi-source heterogeneous data stream, and outputs an inferred microecological abundance vector. The control center uses a state machine module to receive the inferred micro-ecological abundance vector and determine whether the current monitoring state is a free monitoring state or a monitoring intervention state. The control center uses an optimization module to generate an intervention substance regulation ratio based on constraints under the monitoring intervention state; The feeding terminal receives the intervention substance regulation ratio issued by the control center and performs quantitative feeding operation to complete the intelligent monitoring of the microbial community balance regulation status.

2. The intelligent monitoring system for regulating the balance of dairy cow microbiota according to claim 1, characterized in that, The low-frequency truth data stream contains microbial abundance parameters; The analysis end uses a sampling device equipped with a flexible guide tube and a pressure-limiting suction component to obtain liquid samples by probing into the deep liquid phase region of the rumen through the esophagus. Within the set low-frequency sampling period, the analysis terminal controls the pressure-limiting suction component to start the negative pressure suction action to extract the mixed liquid phase, and discards the initial liquid phase initially aspirated, while retaining the rumen liquid phase sample aspirated in the middle section of the flexible guide tube under sealed negative pressure. The analysis unit uses a sequencing module to process the rumen liquid sample to extract microecological abundance parameters, and then converts the extracted microecological abundance parameters into low-frequency ground truth vectors.

3. The intelligent monitoring system for regulating the microbial community balance in dairy cows according to claim 1, characterized in that, The calibration end utilizes a fermentation module to receive the basal diet fermentation broth as a substrate for titration of alternative intervention substances, providing a reaction environment. The calibration end utilizes the extraction module to simultaneously collect reaction environment data of the fermentation module before and after the titration; The extraction module extracts the difference in microbial abundance parameter changes and the corresponding mass change of the titrated candidate intervention substances from the reaction environment data. The quotient obtained by dividing the difference in microbial abundance parameter changes by the mass change is used as a partial derivative feature. The partial derivative features extracted for all candidate intervention substances are algebraically reorganized according to the column arrangement to generate the initial intervention influence matrix. For the candidate intervention substance, the extraction module extracts the maximum change range of physicochemical indicators from the titration time based on the reaction environment data, and divides the maximum change range by the mass change as the quotient value as the maximum change range per unit mass, and extracts the maximum change range per unit mass to form a unit masking peak vector; The extraction module continues to monitor the time required for the physicochemical index to recover from the maximum change to within the baseline recovery tolerance set by the baseline. The time is recorded as a decay time constant, and the unit masking peak vector and the decay time constant are encapsulated into paired data tuples to establish a physicochemical masking feature library corresponding to the candidate intervention substances. The physicochemical masking feature library provides the physicochemical masking feature parameters.

4. The intelligent monitoring system for regulating the microbial balance in dairy cows according to claim 1, characterized in that, In the monitoring free state without the application of exogenous intervention substances, the inference module performs time window alignment operation based on the low-frequency sampling time as the reference time anchor point; The inference module extracts multiple sets of high-frequency feature vectors within the fermentation cumulative time window before the benchmark time anchor point and calculates the arithmetic mean, outputting the mean of the high-frequency feature vectors. The deduction module calls the partial least squares regression algorithm, takes the mean of the high-frequency feature vector as the independent variable and the corresponding low-frequency true value vector as the dependent variable, performs projection dimensionality reduction and regression parameter fitting, and establishes the projection mapping as the basic steady-state mapping model.

5. The intelligent monitoring system for regulating the microbial balance in dairy cows according to claim 1, characterized in that, The state machine module calculates the Mahalanobis distance in real time and determines whether the Mahalanobis distance is always greater than the set statistical safety threshold within a continuously set monitoring period time window. When the Mahalanobis distance is always greater than the statistical safety threshold, the state machine module determines that there are signs of rumen microbiota imbalance in the internal environment, extracts the algebraic difference between the inferred microbiota abundance vector and the preset benchmark vector, records the extracted algebraic difference as a microbiota deviation vector, and controls the monitoring state to switch from the monitoring free state to the monitoring intervention state. The optimization module uses the obtained microecological deviation vector as the lower limit of the regulation requirement, and uses the physical capacity limit of the single feeding of the candidate intervention substances as the boundary constraint of the feasible region to construct a constrained linear programming model. The constrained linear programming model is then mathematically solved to obtain the optimal intervention substance mixing mass ratio vector. The optimal intervention substance mixing mass ratio vector is then converted into a regulation ratio command and sent to the feeding terminal.

6. The intelligent monitoring system for regulating the microbial community balance in dairy cows according to claim 3, characterized in that, The simulation module obtains the actual feeding time and the actual feeding quality of various alternative intervention substances fed back by the feeding terminal. The deduction module, based on the actual feeding time and the actual feeding quality, retrieves and extracts the unit masking peak vector and decay time constant corresponding to each of the candidate intervention substances participating in this feeding from the physicochemical masking feature library based on the preset classification identifier. Then, using the extracted unit masking peak vector and decay time constant, and combining the base of the natural logarithm, it constructs an exponential decay function to calculate the physicochemical perturbation component vector. The inference module subtracts the physicochemical perturbation component vector from the high-frequency feature vector acquired in real time to obtain the corrected high-frequency feature vector, and calls the established projection mapping to perform projection dimensionality reduction calculation on the corrected high-frequency feature vector to output the net inference abundance vector.

7. The intelligent monitoring system for regulating the balance of bovine microbiota according to claim 1, characterized in that, The high-frequency feature data stream includes spatial phenotypic indicators, and the sensing terminal includes behavioral nodes; The state machine module divides the activity area of ​​the sentinel bull into multiple two-dimensional spatial grids and extracts the three-dimensional spatial coordinates of the behavior node as the spatial phenotypic index, which are continuously acquired within the monitoring period time window. The state machine module counts the cumulative time the sentinel bull spends within each two-dimensional spatial grid. It divides the cumulative time by the total time of the monitoring period window to obtain the dwell probability of the sentinel bull in each two-dimensional spatial grid. Based on the dwell probability, it calculates the spatial dwell information entropy and calls the pre-calibrated historical health baseline information entropy to calculate the absolute value of the algebraic difference between the currently output spatial dwell information entropy and the historical health baseline information entropy. The absolute value is used as the behavior deviation index.

8. The intelligent monitoring system for regulating the balance of dairy cow microbiota according to claim 7, characterized in that, The state machine module compares the behavior deviation index with a preset stress tolerance limit. When the behavior deviation index is continuously greater than the stress tolerance limit within the stress determination time window, the state machine module determines that the current feeding dose triggers an abnormal response. The state machine module calculates the difference between the behavior deviation index and the stress tolerance limit, and divides the difference by the stress tolerance limit to obtain the relative exceedance range. The state machine module limits the relative exceedance range to the maximum decay saturation range, and calculates the penalty decay factor by subtracting the limited relative exceedance range from a constant. The penalty decay factor is multiplied into the upper bound vector of the mass of a single feeding of the intervention substance, and the values ​​of each dimension of the upper bound vector of the mass are forcibly lowered as new constraints and fed back to the optimization module.

9. The intelligent monitoring system for regulating the balance of dairy cow microbiota according to claim 8, characterized in that, The state machine module determines whether the updated Mahalanobis distance has fallen back to within the preset statistical safety threshold, and checks in parallel whether the latest behavior deviation index has fallen back to within half of the stress tolerance limit; When the Mahalanobis distance is less than or equal to the statistical safety threshold, and the behavior deviation index is less than or equal to half of the stress tolerance limit, the state machine module determines that the fermentation environment has been stably restored, and switches the control level from the monitoring intervention state to the monitoring free state.

10. An intelligent monitoring method for regulating the balance of bovine microbiota, characterized in that, An intelligent monitoring system for regulating the microbial balance of dairy cows as described in any one of claims 1-9 includes the following steps: Sensing terminals are deployed in the farm environment to acquire continuous high-frequency feature data streams, and low-frequency truth data streams are acquired for the sentinel cattle in the farm environment. Collect simulated intervention response data to generate an initial intervention impact matrix and physicochemical masking characteristic parameters; The high-frequency feature data stream and the low-frequency truth data stream are converged to form a multi-source heterogeneous data stream, and then centrally processed in conjunction with the initial intervention influence matrix and the physicochemical masking feature parameters; The multi-source heterogeneous data stream after centralized processing is invoked to establish a projection mapping from high-dimensional fast variables to low-frequency sequencing parameters. The projection mapping is used to output an inferred microbial abundance vector, and the current monitoring state is determined to be either a monitoring free state or a monitoring intervention state based on the inferred microbial abundance vector. Once the monitoring intervention state is determined, an intervention substance regulation ratio is generated based on the constraints, and a quantitative feeding operation is performed according to the intervention substance regulation ratio to complete the intelligent monitoring of the microbial community balance regulation execution state.

Images