Dairy cow perinatal metabolic health intelligent quick detection system and ai assisted decision method

Abstract

This invention provides an intelligent rapid detection system and AI-assisted decision-making method for peripartum metabolic health of dairy cows, relating to the fields of agricultural intelligence and animal health monitoring. The system includes a multi-source data acquisition module, a rapid detection module, a data fusion and processing module, a health risk assessment module, an AI-assisted decision-making module, and a data storage and management module. By collecting various physiological parameters, metabolic indicators, behavioral characteristics, and environmental data of dairy cows during the peripartum period, and combining this with rapid detection results from biological samples such as blood and milk, a comprehensive and dynamic health profile can be constructed, enabling precise quantitative analysis of metabolic status. The fusion processing of multi-source data eliminates redundancy and noise between data, making health characteristics more representative and reliable. Machine learning analysis is then performed using multi-dimensional features.

Description

Technical Field

[0001] This invention relates to the fields of agricultural intelligence and animal health monitoring technology, specifically to an intelligent rapid detection system for peripartum metabolic health of dairy cows and an AI-assisted decision-making method. Background Technology

[0002] The field of intelligent agriculture and animal health monitoring technology refers to the real-time monitoring, assessment, and optimized management of animal health in agricultural production through advanced information technology, sensing technology, artificial intelligence (AI), and big data analysis. This technology involves using sensors, cameras, drones, and other equipment to collect physiological, behavioral, and environmental data of animals, and then providing intelligent decision support systems for disease early warning, nutritional optimization, and behavioral monitoring through data analysis and modeling. Specifically, the intelligent rapid detection system for peripartum metabolic health in dairy cows and the AI-assisted decision-making method refer to a dairy cow health monitoring and management system based on artificial intelligence technology. This system collects various physiological and environmental data from the peripartum period of dairy cows and performs intelligent analysis and health assessment. The main purpose of this system is to assess the health risks of dairy cows during the peripartum period by real-time monitoring and rapid detection of metabolic health indicators, and to provide feeding management suggestions and disease early warnings based on AI algorithms, thereby reducing the incidence of metabolic diseases and improving dairy cow productivity.

[0003] Existing technologies, relying on single data sources or monitoring a limited number of indicators, cannot comprehensively reflect the metabolic status of dairy cows during the peripartum period. This can lead to biased risk assessments and a failure to provide individualized management solutions. For example, depending on routine physiological or behavioral observation data, key metabolic abnormalities such as ketone bodies and non-esterified fatty acids in the blood may be missed, making timely disease warnings impossible. Health assessments lack dynamic and multi-dimensional analysis, resulting in coarse risk level classifications. Feeding management and intervention measures are often based on experience, leading to unstable or delayed intervention effects. Fragmented data records and a lack of traceability systems hinder long-term health trend analysis and scientific decision-making, impacting dairy cow productivity improvement and optimized farming management. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an intelligent rapid detection system for peripartum metabolic health in dairy cows and an AI-assisted decision-making method. This solves the problems of existing technologies, which rely on a single data source or a limited number of indicators, making it difficult to comprehensively reflect the metabolic status of dairy cows during the peripartum period. This can lead to biased risk assessments, failure to provide individualized management solutions, reliance on routine physiological or behavioral observation data, and potential oversight of key metabolic abnormalities such as ketone bodies and non-esterified fatty acids in the blood, thus failing to provide timely disease warnings.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution: a smart rapid detection system for peripartum metabolic health of dairy cows, comprising the following modules: a multi-source data acquisition module, used to collect physiological parameter data, metabolic index data, behavioral data, and breeding environment data of dairy cows during the peripartum period;

[0006] The rapid detection module is used to rapidly detect metabolic indicators in dairy cow blood, biological fluids, or milk samples to obtain data on ketone body concentration, glucose concentration, non-esterified fatty acid concentration, and β-hydroxybutyrate.

[0007] The data fusion processing module is used to preprocess the multi-source data and form a unified health feature vector through a multimodal data fusion algorithm;

[0008] The health risk assessment module analyzes the health feature vector based on a machine learning model to obtain the peripartum metabolic health risk level of dairy cows.

[0009] The AI-assisted decision-making module outputs feeding and management suggestions, nutritional intervention plans, and disease warning information based on the health risk level.

[0010] The data storage and management module is used to store collected data, analysis results, and decision-making information to form traceable health records.

[0011] Preferably, the multi-source data acquisition module includes a body temperature sensor, an activity monitoring sensor, a rumination monitoring device, an environmental temperature and humidity sensor, and a feed intake monitoring device.

[0012] Preferably, the data fusion processing module uses principal component analysis (PCA) and weighted feature fusion methods to perform dimensionality reduction and feature fusion on physiological data, behavioral data, and metabolic detection data.

[0013] Preferably, the health risk assessment module employs a deep learning prediction model, which includes a convolutional neural network (CNN) or a long short-term memory network (LSTM) to identify abnormal metabolic trends in dairy cows during the peripartum period.

[0014] Preferably, the AI-assisted decision-making module constructs a nutrition intervention decision model based on a reinforcement learning algorithm, and dynamically generates feed nutrition supplementation strategies according to metabolic health status.

[0015] The preferred method for intelligent rapid detection of peripartum metabolic health in dairy cows using AI-assisted decision-making includes the following steps:

[0016] S1: Collect physiological monitoring data, behavioral monitoring data, environmental data, and metabolic detection data of dairy cows during the peripartum period;

[0017] S2: Perform data cleaning and standardization on the collected data and construct a multidimensional health feature dataset;

[0018] S3: A multimodal data fusion algorithm is used to fuse data from different sources to generate a comprehensive health feature vector.

[0019] S4: Input the comprehensive health feature vector into the trained health prediction model for analysis to obtain metabolic health risk assessment results;

[0020] S5: Based on the risk assessment results, generate feeding management suggestions and disease early warning information through an AI decision-making model;

[0021] S6: Store health assessment results and decision-making information and establish a peripartum health management database for dairy cows.

[0022] Preferably, the data cleaning process in S2 includes outlier removal, missing value imputation, and time series smoothing.

[0023] Preferably, the multimodal data fusion algorithm described in S3 uses a combination of a weighted fusion model and a principal component analysis algorithm for feature integration.

[0024] Preferably, the health prediction model described in S4 is trained using historical perinatal data and uses support vector machine (SVM) or deep neural network (DNN) to predict the risk of metabolic diseases.

[0025] Preferably, the AI ​​decision-making model in S5 dynamically optimizes the feed ratio for dairy cows during the peripartum period based on genetic algorithms or particle swarm optimization algorithms, so as to reduce the probability of metabolic diseases and improve the production performance of dairy cows.

[0026] This invention provides an intelligent rapid detection system for peripartum metabolic health in dairy cows and an AI-assisted decision-making method. It offers the following advantages:

[0027] This invention collects various physiological parameters, metabolic indicators, behavioral characteristics, and farming environment data of dairy cows during the peripartum period, and combines this with rapid detection results from biological samples such as blood and milk. This enables the construction of a comprehensive and dynamic health profile, achieving precise quantitative analysis of metabolic status. The fusion processing of multi-source data eliminates redundancy and noise, making health characteristics more representative and reliable. Machine learning analysis using multi-dimensional features allows for the tiered assessment of metabolic health risks in peripartum dairy cows, making risk assessment more personalized and dynamic. Based on this, targeted feeding management and nutritional intervention strategies are generated using health risk results, and disease warning information is output, enabling early intervention and precise management. Data storage and tracking mechanisms ensure the entire health management process is traceable and verifiable, providing a reliable basis for long-term health monitoring and farming optimization. Attached Figure Description

[0028] Figure 1 This is a system block diagram of the present invention;

[0029] Figure 2 This is a schematic diagram of the steps of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Example:

[0032] like Figure 1-2 As shown, this embodiment of the invention provides an intelligent rapid detection system for peripartum metabolic health of dairy cows and an AI-assisted decision-making method, including the following modules: a multi-source data acquisition module, used to collect physiological parameter data, metabolic index data, behavioral data and breeding environment data of dairy cows during the peripartum period;

[0033] First, wearable neck collar sensors, rumination monitors, and environmental temperature and humidity recording terminals are deployed in the cattle shed. The neck collar device records the number of steps, body surface temperature, and duration of rumination at a sampling cycle of approximately every few minutes. The monitor uses a simple counter to count the number of rumination chewings per unit time, forming behavioral record values. ,in This indicates the sampling time sequence number, followed by the measurement of body temperature data using an electronic temperature tag. The environmental terminal simultaneously records the temperature. With humidity The data is aligned using a unified timestamp in the data recording software. The data acquisition system archives behavioral, temperature, and environmental records using spreadsheet software. If there are several dozen sampling points within a monitoring period, each sampling point is recorded as a separate line. For example, at a certain time point, the recorded number of ruminations might be several dozen, the body surface temperature might be around 30 degrees Celsius, the ambient temperature might be around 20 degrees Celsius, and the humidity might be around several tens of percent. Then, a simple data integrity check rule is used to determine the missing data. For example, a missing data threshold might be set as "the proportion of consecutive missing samples exceeds approximately one-tenth." If the missing proportion is below this threshold, the record is retained; if it is above, it is supplemented using the average of adjacent time points. The average calculation formula is written as follows: ,in As a supplementary value, and These represent the preceding and following sample values, respectively. This continuous recording method forms a raw record set containing physiological, behavioral, and environmental items, which is then used in subsequent processing.

[0034] The rapid detection module is used to rapidly detect metabolic indicators in dairy cow blood, biological fluids, or milk samples to obtain data on ketone body concentration, glucose concentration, non-esterified fatty acid concentration, and β-hydroxybutyrate.

[0035] During the sampling process, staff use disposable lancets to obtain a small blood sample and insert it into a test strip. The test strip is then inserted into a portable biochemical detection device, which generates an electrochemical response signal. The device internally converts changes in current into metabolite concentration values ​​using a simple linear conversion relationship. The concentration records were obtained, among which This represents an estimated value of metabolite concentration. Indicates the equipment calibration coefficient. The baseline compensation value is represented by the calibration coefficient, which is determined through standard solution experiments. For example, a calibration table is established using multiple concentration samples. When the electrical signal is in a certain current range, the corresponding coefficient is approximately on the order of a few tenths of a unit. Subsequently, the detection device provides ketone body-related index values, blood glucose record values, and non-esterified fatty acid index record values, with β-hydroxybutyrate data as a key recording item. In the example detection, if the device output signal is approximately in a certain current range, the β-hydroxybutyrate value is calculated to be in the range of approximately a few tenths to several units through coefficient conversion. The system sets the range division rules. For example, when the β-hydroxybutyrate value is in the range of approximately a few tenths to one unit, it is considered to be in the normal range, and when it is in the range of approximately one to several units, it is marked as being out of range. The detection time, cow number, and record values ​​of each index are written into a data table file through the electronic recording terminal to form metabolic index record data.

[0036] The data fusion processing module is used to preprocess multi-source data and form a unified health feature vector through a multimodal data fusion algorithm;

[0037] The data processing script reads behavioral records, environmental records, and metabolic index records, and performs a unified format conversion on data from different sources. For example, it converts all time units to the same time scale and temperature units to the same Celsius representation. Then, it performs normalization calculations on each type of data. The normalization expression is written as follows: ,in This represents the standardized value. Represents the original sampled value. This represents the average value of the same indicator over the monitoring period. The standard deviation and the mean are expressed as follows: , This indicates the number of samples. For example, if a dairy cow's rumination samples during a monitoring period average within a range of approximately several dozen times, with a standard deviation of several times, then when the rumination value at a certain point in time is slightly higher than the average, the standardized result is a positive interval value. Subsequently, the system combines multiple indicators according to preset weights to form a unified feature vector. Calculation method writing ,in This represents the weight of each indicator. The weight is set with reference to the contribution ratio of historical variance. For example, the contribution ratio of behavioral data is approximately a certain percentage, and the contribution ratio of metabolic indicators is approximately a certain percentage. Therefore, the weight is set to a numerical range close to this ratio. A unified health feature record vector is generated by a simple weighted summation method.

[0038] The health risk assessment module analyzes health feature vectors based on machine learning models to obtain the peripartum metabolic health risk level of dairy cows.

[0039] After obtaining the unified feature vector, it is imported into the machine learning modeling software interface. Historical dairy cow sample data is input in tabular format, and a classification model is trained. The model first reads the feature vector. ,in The number of indicators is represented, and then the risk probability value is calculated using logistic regression. Expression writing ,in Represents the coefficient of the constant term. Indicates the first Each indicator corresponds to a coefficient. The coefficient range is obtained through repeated iterations using training samples. In the example, some metabolic indicator coefficients are in the positive range while behavioral indicator coefficients are in the negative range, indicating that they have different directional effects on the probability value. After calculating the probability value, the system classifies it according to the interval division rules. For example, the probability value in the range of approximately zero to a few tenths of a percent is classified as the normal range, and the probability value in the range of approximately a few tenths of a percent to a higher range is classified as the intermediate level. If the probability value enters a higher range, it is marked as a risk level record. Through this calculation process, the metabolic health level of an individual dairy cow during the peripartum period is obtained.

[0040] The AI-assisted decision-making module outputs feeding and management suggestions, nutritional intervention plans, and disease warning information based on the health risk level.

[0041] Upon receiving a health level record, the system calls the management rules database for matching. The database pre-stores multiple sets of feeding parameter records, such as dietary energy density, protein ratio, and dry matter intake range. When the system reads a dairy cow's risk level, it recalculates the nutritional parameters using rule expressions. In the example calculation, the energy supplementation coefficient is set to... The basic daily energy value is recorded as The adjusted energy value calculation formula is written as follows: ,in The coefficient is set based on historical feeding records. If the risk level is in the middle range, the coefficient is taken as a value close to a few tenths of a percent; if the risk level is in a higher range, the coefficient is taken as a higher value. The system also records feed intake. With cow weight Establish proportional relationships When the ratio value is outside a certain reference range, an alert record is generated on the management terminal, and corresponding feeding adjustment suggestions and health warning record entries are generated through the electronic terminal.

[0042] The data storage and management module is used to store collected data, analysis results, and decision-making information to form traceable health records.

[0043] After data generation, all records are written to the database system, and an individual cow profile table is created. The database structure includes fields for cow ID, time, behavior record, metabolic indicators, and risk level. Before writing, the system performs duplicate record detection by comparing the cow ID with the time key. If the same key is found, it is recorded as a duplicate entry, and the most recent record is retained using a time sorting function. Subsequently, the system generates a historical record sequence for each cow. ,in Indicates the first The monitoring results are used to periodically count the frequency of a particular dairy cow's risk level over a period of time. For example, the statistical expression is written as follows: ,in Indicates the number of times a certain risk level occurs. This indicates the total number of monitoring sessions. When the frequency value enters a higher range, continuous monitoring records are marked in the archive. At the same time, all data are arranged in chronological order to form a searchable historical archive record set.

[0044] The multi-source data acquisition module includes a body temperature sensor, an activity monitoring sensor, a rumination monitoring device, an environmental temperature and humidity sensor, and a feed intake monitoring device.

[0045] The operation of each sensor in the data acquisition unit, which consists of multiple types of monitoring devices, is broken down into three consecutive steps: continuous sampling, time-synchronized recording, and unified format packaging. Among them, the body temperature detection element samples the temperature at the base of the ear or rumen of the target dairy cow at a set sampling period and generates the raw temperature sequence. Indices in the sequence Indicates the first The sampling record and The motion monitoring element records motion signals and calculates motion amplitude through a triaxial acceleration module. ,in The device represents acceleration values ​​in three directions with units of the same dimension. The rumination detection device identifies the rumination chewing cycle through neck vibration signals and records the number of chewing cycles identified per unit time as a variable. ,in Indicates the first Within a time window, the environmental module synchronously records temperature variables. With humidity variables The feed intake monitoring device uses an electronic weighing structure to record the difference in feed weight after each feeding. ,in For the purpose of feeding, the feed trough is heavy. To determine the weight of the feed tank after feeding, the data generated by each device is first sorted according to a unified timestamp, and the time deviation is calculated using a formula. Calculate and Records are retained when the data is within the allowed range; otherwise, the timeline is realigned. In the example scenario, when a dairy cow wears a collar-type activity device and completes a feeding behavior in the feeding trough area, the system records the continuous motion amplitude sequence and the difference in feed weight, and simultaneously obtains the body temperature and environmental variables for the same time period. This results in a unified structured monitoring record set containing the body temperature sequence, motion amplitude sequence, number of rumination cycles, environmental temperature and humidity, and the difference in feed weight.

[0046] The data fusion processing module uses principal component analysis (PCA) and weighted feature fusion to perform dimensionality reduction and feature fusion on physiological data, behavioral data, and metabolic detection data.

[0047] After multi-source monitoring information enters the fusion processing unit, it first performs standardization on various variables, then dimensionality reduction and feature combination, where each type of original variable... Indicates the first The sample at the th The observed values ​​for each indicator are standardized using a formula. Processing to obtain standardized variables , in the formula For the first The index sample mean and The two are calculated by summing the sample standard deviation. and To obtain the covariance matrix, it is necessary to first construct the covariance matrix when performing principal component decomposition. Matrix elements Indicates the first With the The covariance relationship between indicators is determined, and then eigenvectors are calculated using eigenvalue decomposition. With corresponding eigenvalues When the cumulative contribution rate When a feature falls within a set proportion range, the corresponding principal component is retained; during the weight fusion stage, coefficients are assigned to features of different categories. The weights are calculated based on the historical sample variance percentage. In the example, when the variance percentage of behavioral indicators is in a high range, the weight range is assigned to a moderate proportion. Physiological indicators have slightly higher weights than environmental indicators, and the weights are determined using a fusion formula. Generate comprehensive features, where These represent physiological, behavioral, and metabolic-related feature vectors, respectively. Within a certain monitoring period, the system inputs variables such as body temperature, rumination frequency, and exercise amplitude into a matrix for calculation, obtains several principal components, and forms a new set of comprehensive feature variables through weighted combination.

[0048] The health risk assessment module uses a deep learning prediction model, including a convolutional neural network (CNN) or a long short-term memory network (LSTM), to identify abnormal metabolic trends in dairy cows during the peripartum period.

[0049] When the health assessment module performs deep model training and prediction processing on the fused feature sequences, it first represents the feature matrix formed within the continuous monitoring period as follows: subscript The length of the time series is represented by the convolution kernel in convolutional processing. Slide along the timeline and compute the convolution result. , in the formula Indicates the first convolution kernel Each weight parameter This represents the kernel length. The initial weights are randomly set within a small range and then adjusted during training based on the error function. Iterative updates are performed; if a sequence network structure is used, the state update formula is applied. Calculate the hidden state, where For the weight matrix As bias parameters, all parameters are adjusted round by round through gradient descent. During the training phase, historical samples are divided into training and validation sets, and the change of loss function in each round is recorded. When the magnitude of loss change falls into the stable range, the model parameters are retained. In the actual ranch example, the system inputs the comprehensive feature vector sequence of a dairy cow for multiple consecutive daily cycles and obtains the output predicted value sequence through convolution operation or sequence memory calculation. Then, the metabolic state change record of that time period is recorded according to the trend of the predicted sequence change.

[0050] The AI-assisted decision-making module constructs a nutrition intervention decision model based on reinforcement learning algorithms, and dynamically generates feed nutrition supplementation strategies according to metabolic health status.

[0051] After receiving the health assessment output, the AI-assisted decision-making module constructs a reinforcement learning process based on a state-action structure, where the state variables... Indicates the first The combination of comprehensive health indicators of dairy cows over a time period and action variables This represents the corresponding nutritional supplement combination, processed through a reward function. Each decision outcome is numerically evaluated, and rewards are calculated using a formula. ,in Indicates the change in health indicators Indicates the deviation from the benchmark index, coefficient The policy is set within a proportional range and gradually adjusted by fitting historical data, using a value function during the policy update process. Indicates the state Select action Expected returns and in accordance with updated rules Iterative adjustments, among which For the learning rate parameter As a discount factor, both sample values ​​are selected within the empirical range and tend to stabilize after multiple rounds of training. In the example farm scenario, the system reads the current state vector of a dairy cow and selects a feed supplement combination from the action set. Then, it calculates the reward value and updates it based on the changes in body temperature, rumination, and feed intake variables monitored in the next cycle. This will allow for the gradual formation of a set of nutritional supplementation strategies corresponding to the current monitoring status.

[0052] A smart rapid detection method for peripartum metabolic health in dairy cows using AI-assisted decision-making includes the following steps:

[0053] S1: Collect physiological monitoring data, behavioral monitoring data, environmental data, and metabolic detection data of dairy cows during the peripartum period;

[0054] Technicians deployed neck ring sensors, rumination monitors, and environmental monitoring nodes above the cattle shed passage and initiated the data recording program. First, the step sequence recorded by the neck ring sensors was divided into time windows, for example, dividing a monitoring cycle into several equal-length segments and recording the cumulative step count for each time segment. Then, the rumination duration recorded by the rumination monitor was read and correlated with the number of feeding actions in the behavior recording module. When processing the environmental monitoring node data, the raw records of temperature, humidity, and gas concentration were read first, and outlier removal was performed. For example, if the temperature value in the monitoring data sequence deviated from the set range outside the historical average, then… The data was completed using linear interpolation. Subsequently, blood ketone body test results were extracted from the metabolic monitoring records and their units were standardized. For example, when the units given by different testing devices were inconsistent, a conversion factor was used for standardization. The conversion factor was obtained by comparing historical test samples. A time-stamped correspondence was established between behavioral monitoring and metabolic monitoring results. For example, when the rumination time series and the blood ketone body test time point differed by no more than one monitoring cycle, a marker association was established. Then, the physiological monitoring records, behavioral monitoring records, environmental records, and metabolic monitoring records of each dairy cow were merged and stored according to the dairy cow number to form a set of original peripartum monitoring records.

[0055] S2: Perform data cleaning and standardization on the collected data and construct a multidimensional health feature dataset;

[0056] First, the monitoring record set is read and anomaly cleanup is performed. During the cleanup process, a statistical distribution interval is established for the physiological data sequence. For example, a reference interval is established based on historical samples, and values ​​deviating from this interval are marked as outliers. Then, outliers are replaced using the nearest time-slice averaging method, where the nearest time-slice averaging is calculated according to the formula... Calculation, where This represents the monitoring record value for a certain time slice. This indicates the number of time slices used in the average calculation. The cleaned data is then standardized using the formula... Completed processing, among which Indicates the first The monitoring records the values. Represents the sample mean. This represents the sample standard deviation. In the actual calculation example, we first collect body temperature records for a dairy cow over multiple monitoring periods and calculate the mean and standard deviation. Then, we substitute these values ​​into the formula to obtain the standardized result. Subsequently, we convert physiological monitoring data, behavioral monitoring data, environmental data, and metabolic detection data into characteristic variables of a unified format. For example, we convert rumination duration into rumination rate per unit time using the formula. Calculation, where This represents the recorded value indicating the duration of rumination. This represents the total duration of the monitoring period, ultimately forming a health characteristic data set containing multiple characteristic variables.

[0057] S3: A multimodal data fusion algorithm is used to fuse data from different sources to generate a comprehensive health feature vector.

[0058] First, feature weight sets are established for different data sources, and weight allocation is performed. During the weight setting process, statistical calculations are performed based on the correlation between each feature variable and metabolic abnormality records in historical monitoring samples. For example, the correlation coefficient between a certain feature variable and metabolic abnormality records is calculated, and weight ranges are set according to interval division. When the correlation coefficient falls into the medium interval, a medium weight is assigned; when it falls into the strong interval, a larger weight is assigned. Subsequently, a weighted fusion formula is used. Calculate the comprehensive feature vector, where Indicates the first Values ​​of each feature variable, This represents the corresponding feature weight coefficient. This indicates the number of features involved in the fusion. In the example operation, behavioral activity index, rumination ratio, and metabolic detection indicators are first selected as input features and substituted with their corresponding weight values. Then, the weighted result is calculated to obtain the comprehensive feature value at a single time point. Subsequently, a time series is established for the comprehensive feature values ​​of continuous monitoring periods, and a moving average formula is used. Smoothing is performed, where Indicates the first The comprehensive characteristic value of each time point, This represents the number of time points involved in the smoothing calculation. After processing, a set of comprehensive health feature vectors for dairy cows during the peripartum period is obtained.

[0059] S4: Input the comprehensive health feature vector into the trained health prediction model for analysis to obtain metabolic health risk assessment results;

[0060] First, the comprehensive feature vector set is read and input into the trained prediction calculation module. During prediction calculation, a model input matrix is ​​built for each feature variable; for example, the first feature variable is... The feature vectors of each time node are represented as follows: ,in Indicates the first The first time point Numerical values ​​of each feature variable This indicates the number of feature variables, which are then used in the model computation according to the linear combination formula. Calculate the risk score, where This represents the weight coefficients obtained during the model training phase. This represents the bias coefficient. In the example calculation, the comprehensive feature vector of a dairy cow is first read and substituted into the formula to calculate the score value. Then, the risk level is divided according to the score range. For example, when the score falls into the basic range, it is marked as a low-risk range; when the score falls into the middle range, it is marked as a medium-risk range; and when the score enters the upper range, it is marked as a high-risk range. The range division is set by statistically analyzing the historical sample score distribution and dividing it into equal segments. Finally, the metabolic health risk assessment record of the corresponding dairy cow is generated.

[0061] S5: Based on the risk assessment results, generate feeding management suggestions and disease early warning information through an AI decision-making model;

[0062] First, risk assessment records are read, and a set of rule parameters is established in the decision calculation module. During the rule parameter setting process, corresponding management parameter thresholds are established for different risk ranges. For example, the rate of change in feed intake is statistically analyzed based on historical feeding records, and a reference rate of change is calculated using a formula. Calculation, where This represents the recorded feed intake value for the current monitoring period. This represents the feed intake record value from the previous cycle. In the example calculation, the average rate of change is calculated using feed intake records from multiple cycles, and a range is set. Subsequently, in the decision-making module, the risk score and the feed intake change rate are jointly judged. For example, when the risk score enters the medium-risk range and the feed intake change rate falls into the abnormal range, a feeding adjustment record is generated. At the same time, the change ratio of rumination time is read and adjusted according to the formula. Calculation, where This indicates the rumination record value for the current monitoring period. This represents the historical reference rumination record value. If the calculation result falls within the preset abnormal range, a corresponding early warning record is generated. Subsequently, the corresponding management suggestion text record and early warning label information are output according to the rule table.

[0063] S6: Store health assessment results and decision-making information and establish a peripartum health management database for dairy cows.

[0064] First, risk assessment records and decision output records are read and data structure is organized. During this process, a unique index is created for each cow, and a monitoring record sequence is generated chronologically. Then, association fields are established for health assessment records, behavior monitoring records, and environmental records, such as a joint index using the cow's ID and monitoring time identifier. When writing data to the database, the data table structure is constructed and field parameters are set, such as characteristic variable fields, risk score fields, and management record fields. Data integrity is verified during the writing process, for example, using formulas. Calculate record completeness, where Indicates the number of valid fields. This indicates the total number of fields to be recorded. The record writing operation is completed when the calculation result is within the set range. In the example record processing, the health assessment results and management records of a dairy cow for multiple consecutive monitoring cycles are written into the database table in chronological order and a corresponding index is generated, thereby forming a set of perinatal health management data records.

[0065] The data cleaning process in S2 includes outlier removal, missing value imputation, and time series smoothing.

[0066] First, import the continuous record tables from the body temperature sensor, rumination monitoring collar, and blood metabolism test records into the data processing software environment, for example, by creating a set of fields in a tabular data processing program. Where the subscript i represents the time sampling number, and then in the anomaly identification stage, the statistical dispersion is calculated for each time series, first calculating the series mean. Then calculate the discrete scale. If a certain sample value satisfies These are then classified into the abnormal interval. The coefficient k is set as an empirical proportion between one and two times the discrete scale, based on the empirical range of perinatal body temperature fluctuations. The intermediate proportion is used as an example computational resource through manual sample statistics. For instance, when the average body temperature sequence in the historical sample is close to the median of the normal range, and a certain record deviates significantly from the boundary of the average range, it is marked and removed. Subsequently, the missing data completion process begins, where missing records are interpolated using the nearest time series to establish a linear relationship. Estimation is performed, for example, when rumination activity records are missing for a certain time period, intermediate estimates are calculated using adjacent sample values ​​to fill the record table. Then, a moving average smoothing process is applied to the complete sequence, and calculations are performed within the time window set W. The window length is set to a short-cycle interval value based on the perinatal behavioral fluctuation cycle, and smooth time series data is obtained through continuous calculation.

[0067] The multimodal data fusion algorithm in S3 combines a weighted fusion model with principal component analysis to integrate features.

[0068] First, the physical condition score data Blood metabolic indicators and behavioral activity indicators Organize into a feature matrix Then, in the weighted fusion step, weight parameters are set for each type of indicator. The weighting was set with reference to the correlation statistics of historical samples, and the correlation between each indicator and the target health label H was calculated. Obtain the relative correlation degree, and then perform normalization calculation based on the correlation degree ratio. For example, when blood indicators have a high correlation with metabolic state, their weight accounts for a large proportion of the overall weight, while behavioral indicators, which have a moderate correlation, account for a smaller proportion. Then, a weighted composite calculation is performed based on these weights. Then, the principal component calculation stage begins, where the covariance matrix of the fusion matrix is ​​obtained. Then, the set of principal eigenvectors is obtained through eigenvalue decomposition. When a certain feature value accounts for a significant proportion of all feature values, it is selected as the principal component. For example, when the contribution proportion of the first feature is close to most of the overall contribution, it is used as the comprehensive feature expression, thus obtaining the dimensionality-reduced feature vector sequence.

[0069] The health prediction model in S4 is trained using historical perinatal data and uses support vector machines (SVM) or deep neural networks (DNN) to predict the risk of metabolic diseases.

[0070] First, the historical records of dairy cows were compiled into a training set. ,in This represents the fused multidimensional feature vector. The metabolic state label variable is used to construct a classification function during the support vector machine modeling stage. Where w is the feature weight vector and b is the bias, the optimization objective is solved by... Where C is the penalty coefficient, As a slack variable, the penalty coefficient is set to a moderate constraint ratio based on the sample error tolerance range. For example, when the sample classification error is within the acceptable range, C is adjusted to the middle ratio range to maintain a stable classification margin. Subsequently, a kernel function is used. Calculate the mapping relationship between samples, for example, using the radial function form. Where γ is the kernel width parameter, the parameter size is set in the medium diffusion range according to the dispersion of the sample distribution. During training, the decision boundary is calculated using the sample feature vectors. When a new sample is input, it is processed by a function. To obtain the classification results, if a deep neural network is used, first construct an input layer with the number of nodes equal to the feature dimension, then set several hidden layers with the number of nodes in each layer to be a multiple of the input dimension, and finally compute the classification results through forward propagation. Obtain the output probability value and record the corresponding risk label.

[0071] In S5, the AI ​​decision-making model dynamically optimizes the feed ratio for dairy cows during the peripartum period based on genetic algorithms or particle swarm optimization algorithms, in order to reduce the probability of metabolic diseases and improve dairy cow production performance.

[0072] First, the composition of the perinatal diet is represented as a set of variables. Each variable represents a proportion of feed ingredients, such as the proportion of energy ingredients, crude fiber ingredients, and protein ingredients. Subsequently, in the genetic algorithm calculation step, each combination of proportions is encoded into a chromosome vector. And establish the fitness function Where H represents the metabolic risk assessment quantity and G represents the production performance evaluation quantity, the coefficients α and β are set as complementary ratios based on management experience, and are normalized to satisfy α + β = overall weight ratio. Several ratio combinations are randomly generated during the population initialization phase to ensure... Individuals within the total diet range are ranked by calculating a fitness function. Individuals with higher fitness values ​​are retained for the next iteration. A crossover operation is then performed, for example, randomly selecting split points on two chromosomes and exchanging some proportion parameters to obtain new combinations. Finally, a minor mutation operation is performed on a given chromosome. By slightly increasing or decreasing the proportion and recalculating the fitness value, a stable set of proportion vectors is obtained after multiple iterations. If a particle swarm optimization algorithm is used, each proportion combination is treated as a particle position vector. Through the speed update formula Adjust the proportions, where w is the inertia coefficient. and The learning coefficients are set to a medium weight ratio based on the historical experimental intervals, and the final diet ratio vector is obtained through continuous iterative updates.

[0073] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart rapid detection system for peripartum metabolic health of dairy cows, characterized in that, It includes the following modules: a multi-source data acquisition module, used to collect physiological parameter data, metabolic index data, behavioral data, and breeding environment data of dairy cows during the peripartum period; The rapid detection module is used to rapidly detect metabolic indicators in dairy cow blood, biological fluids, or milk samples to obtain data on ketone body concentration, glucose concentration, non-esterified fatty acid concentration, and β-hydroxybutyrate. The data fusion processing module is used to preprocess the multi-source data and form a unified health feature vector through a multimodal data fusion algorithm; The health risk assessment module analyzes the health feature vector based on a machine learning model to obtain the peripartum metabolic health risk level of dairy cows. The AI-assisted decision-making module outputs feeding and management suggestions, nutritional intervention plans, and disease warning information based on the health risk level. The data storage and management module is used to store collected data, analysis results, and decision-making information to form traceable health records.

2. The intelligent rapid detection system for peripartum metabolic health of dairy cows according to claim 1, characterized in that: The multi-source data acquisition module includes a body temperature sensor, an activity monitoring sensor, a rumination monitoring device, an environmental temperature and humidity sensor, and a feed intake monitoring device.

3. The intelligent rapid detection system for peripartum metabolic health of dairy cows according to claim 1, characterized in that: The data fusion processing module uses principal component analysis (PCA) and weighted feature fusion methods to perform dimensionality reduction and feature fusion on physiological data, behavioral data, and metabolic detection data.

4. The intelligent rapid detection system for peripartum metabolic health of dairy cows according to claim 1, characterized in that: The health risk assessment module employs a deep learning prediction model, which includes a convolutional neural network (CNN) or a long short-term memory network (LSTM), to identify abnormal metabolic trends in dairy cows during the peripartum period.

5. The intelligent rapid detection system for peripartum metabolic health of dairy cows according to claim 1, characterized in that: The AI-assisted decision-making module constructs a nutrition intervention decision model based on reinforcement learning algorithms, and dynamically generates feed nutrition supplementation strategies according to metabolic health status.

6. An AI-assisted decision-making method for intelligent rapid detection of peripartum metabolic health in dairy cows, characterized in that: Includes the following steps: S1: Collect physiological monitoring data, behavioral monitoring data, environmental data, and metabolic detection data of dairy cows during the peripartum period; S2: Perform data cleaning and standardization on the collected data and construct a multidimensional health feature dataset; S3: A multimodal data fusion algorithm is used to fuse data from different sources to generate a comprehensive health feature vector. S4: Input the comprehensive health feature vector into the trained health prediction model for analysis to obtain metabolic health risk assessment results; S5: Based on the risk assessment results, generate feeding management suggestions and disease early warning information through an AI decision-making model; S6: Store health assessment results and decision-making information and establish a peripartum health management database for dairy cows.

7. The AI-assisted decision-making method for intelligent rapid detection of peripartum metabolic health in dairy cows according to claim 6, characterized in that: The data cleaning process in S2 includes outlier removal, missing value imputation, and time series smoothing.

8. The AI-assisted decision-making method for intelligent rapid detection of peripartum metabolic health in dairy cows according to claim 6, characterized in that: The multimodal data fusion algorithm described in S3 uses a combination of a weighted fusion model and a principal component analysis algorithm for feature integration.

9. The AI-assisted decision-making method for intelligent rapid detection of peripartum metabolic health in dairy cows according to claim 6, characterized in that: The health prediction model described in S4 is trained using historical perinatal data and uses support vector machines (SVM) or deep neural networks (DNN) to predict the risk of metabolic diseases.

10. The AI-assisted decision-making method for intelligent rapid detection of peripartum metabolic health in dairy cows according to claim 6, characterized in that: In S5, the AI ​​decision-making model dynamically optimizes the feed ratio for dairy cows during the peripartum period based on genetic algorithms or particle swarm optimization algorithms, in order to reduce the probability of metabolic diseases and improve dairy cow production performance.

Images