Cow heat stress online monitoring and early warning method and system based on multi-source data

Abstract

The invention discloses a dairy cow heat stress online monitoring and early warning method and system based on multi-source data. The method comprises the steps that dairy cow living environment temperature and humidity, individual activity frequency, body surface temperature and respiratory frequency data are collected through a multi-source data intelligent pasture pupil early warning platform; inputting the data into a federated learning heterogeneous data fusion model for processing to generate a unified data feature set; inputting the feature set into a dairy cow metabolic behavior multi-modal fusion model, and outputting a real-time metabolic behavior evaluation result; determining a current heat stress risk threshold interval through a heat stress risk dynamic threshold calculation model in combination with historical data; comparing an evaluation result with a threshold interval to judge a risk level; the risk level is transmitted to the display terminal through the platform, and early warning is triggered to the management terminal. The corresponding method of the system comprises six units. According to the method, effective integration and dynamic threshold calculation of multi-source data are realized, the comprehensiveness and accuracy of heat stress monitoring and early warning are improved, and the method is suitable for large-scale cow breeding management.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of dairy cow heat stress monitoring, and in particular to a dairy cow heat stress online monitoring and early warning method and system based on multi-source data. BACKGROUND

[0002] In modern dairy farming, heat stress is one of the key factors affecting the health status and production performance of dairy cows. With the global climate change, high temperature and humidity weather frequently occurs in summer, and dairy cows are prone to problems such as accelerated respiratory rate, elevated body temperature, and metabolic disorder when they are in such environment for a long time, which further leads to reduced milk production, decreased reproductive ability, and even various diseases, resulting in significant economic losses for the breeding enterprises. At the same time, in traditional dairy farming, heat stress monitoring relies mainly on manual observation, which has the limitations of long monitoring cycle, strong subjectivity, and inability to capture real-time changes in dairy cow physiology and environment, and thus cannot meet the needs of timely and accurate heat stress early warning in large-scale and fine breeding. Under such circumstances, the development of a dairy cow heat stress online monitoring and early warning method and system based on multi-source data has become an important direction to solve the problems of traditional monitoring and improve the level of breeding management. By integrating multi-dimensional data such as environment, physiology, and behavior, real-time monitoring and accurate early warning of dairy cow heat stress can be achieved, which is of great significance to the protection of dairy cow welfare and the stability of breeding benefits.

[0003] The existing technology has two obvious shortcomings in dairy cow heat stress monitoring and early warning. On the one hand, the existing technology lacks the ability to integrate and process multi-source data, and mostly only analyzes single type of data (such as environmental temperature and humidity or dairy cow body surface temperature), which fails to effectively integrate multi-dimensional heterogeneous data such as environment, physiology, and behavior, resulting in insufficient utilization of data information and difficulty in fully reflecting the real state of dairy cow heat stress, which further affects the integrity and reliability of the monitoring results. On the other hand, the setting of heat stress risk threshold in the existing technology mostly uses fixed values, without considering the influence of metabolic differences of individual dairy cows, physiological characteristics changes in different growth stages, and dynamic fluctuations of environmental factors on heat stress risk. The fixed threshold cannot adapt to complex breeding scenarios, and is prone to early warning lag or misjudgment, which cannot provide accurate and effective decision support for dairy cow heat stress prevention and control. SUMMARY

[0004] In order to overcome the shortcomings and deficiencies of the existing technology, the present application provides a dairy cow heat stress online monitoring and early warning method and system based on multi-source data.

[0005] The technical scheme adopted by the present application is a dairy cow heat stress online monitoring and early warning method based on multi-source data, comprising the following steps: S1, through the temperature and humidity sensor, activity amount collector and physiological index monitoring module carried by the multi-source data intelligent pasture pupil early warning platform, the temperature and humidity data of the dairy cow living environment, the activity frequency data of the individual dairy cow, the body surface temperature and breathing frequency data of the dairy cow are collected; S2, the multi-type data collected in S1 is input into the federal learning heterogeneous data fusion model, the feature alignment and fusion processing of the heterogeneous data are carried out through the distributed training mechanism between the nodes in the model, and a unified data feature set is generated; S3, the unified data feature set generated in S2 is input into the dairy cow metabolism behavior multi-modal fusion model, the weighted fusion and correlation analysis of the multi-dimensional features related to the metabolism of the dairy cow are carried out through the model, and the real-time metabolism behavior evaluation result of the dairy cow is output; S4, based on the metabolism behavior evaluation result output by S3, the environmental and physiological data at the time of historical heat stress are combined, and input into the heat stress risk dynamic threshold calculation model, the dynamic weight distribution and threshold iteration calculation of the risk influencing factors are carried out through the model, and the heat stress risk threshold interval of the dairy cow under the current environment is determined; S5, the real-time metabolism behavior evaluation result output by S3 is compared and analyzed with the heat stress risk threshold interval determined by S4, and the current heat stress risk grade of the dairy cow is judged; S6, the heat stress risk grade judged by S5 is transmitted to the platform display terminal in the form of data signal through the multi-source data intelligent pasture pupil early warning platform, and the early warning signal of the corresponding risk grade is triggered and transmitted to the specified management terminal.

[0006] Further, the expression of the dairy cow metabolism behavior multi-modal fusion model is: wherein, represents the real-time metabolism behavior evaluation result of the dairy cow, represents the number of evaluation dimensions of metabolism behavior, represents the weight coefficient of the first evaluation dimension, represents the feature mapping function of the first evaluation dimension, represents the original data corresponding to the first evaluation dimension, represents the correlation adjustment coefficient of the first evaluation dimension, represents the number of correlation data dimensions, represents the correlation coefficient of the first evaluation dimension and the first correlation data dimension, represents the original data corresponding to the first correlation data dimension, represents the model error correction term.

[0007] Further, the expression of the federal learning heterogeneous data fusion model is: wherein, denotes a set of unified data features, denotes the number of federated learning nodes, denotes the weight coefficient of the th node, denotes the local data processing coefficient of the th node, denotes the heterogeneous data matrix collected by the th node, denotes the local data feature weight matrix of the th node, denotes the data interaction coefficient of the th node with other nodes, denotes the data interaction weight matrix of the th node with the th node.

[0008] Further, the expression of the heat stress risk dynamic threshold calculation model is: wherein, denotes the critical value of the heat stress risk threshold interval, denotes the historical threshold reference value, denotes the environmental factor fluctuation adjustment coefficient, denotes the number of environmental impact factors, denotes the weight of the th environmental impact factor, denotes the real-time value of the th environmental impact factor, denotes the historical average value of the th environmental impact factor, denotes the physiological factor deviation adjustment coefficient, denotes the number of physiological impact factors, denotes the weight of the th physiological impact factor, denotes the real-time value of the th physiological impact factor, denotes the historical average value of the th physiological impact factor.

[0009] Further, the data transmission efficiency calculation expression of the multi-source data intelligent pasture pupil early warning platform is: wherein, denotes the platform data transmission efficiency, denotes the number of data transmissions, denotes the data amount of the th transmission, denotes the integrity coefficient of the th transmission data, represents the start time of the first data transmission, represents the end time of the first data transmission, represents the start time of the first data transmission, represents the end time of the first data transmission, represents the platform transmission channel stability coefficient.

[0010] Further, the feature extraction expression of the dairy cow heat stress online monitoring data is: wherein, represents the feature extraction result of the monitoring data, represents the number of feature extraction dimensions, represents the weight of the first feature extraction dimension, represents the feature conversion function of the first feature extraction dimension, represents the original monitoring data corresponding to the first feature extraction dimension, represents the interaction coefficient of the first feature extraction dimension, represents the number of interaction data dimensions, represents the interaction index of the first feature extraction dimension and the first interaction data dimension, represents the original monitoring data corresponding to the first interaction data dimension. represents the interaction coefficient of the first feature extraction dimension, represents the number of interaction data dimensions, represents the interaction index of the first feature extraction dimension and the first interaction data dimension, represents the original monitoring data corresponding to the first interaction data dimension. Further, the S3 comprises the following steps: S31, screening the feature data related to the dairy cow metabolic behavior from the unified data feature set generated in S2, including the dairy cow activity frequency derived feature, the body surface temperature change feature and the respiratory frequency fluctuation feature; S32, classifying the screened feature data according to the preset dimension classification rule, dividing into the metabolic basic feature class and the metabolic correlation feature class; S33, inputting the classified feature data into different input layers of the dairy cow metabolic behavior multi-modal fusion model, and performing nonlinear conversion on the feature data through the activation function of each layer in the model; S34, weighting and fusing the converted feature data of each input layer, combining the feature correlation weight determined in the model training process, and generating the dairy cow real-time metabolic behavior evaluation result.

[0011] Further, the feature extraction expression of the dairy cow heat stress online monitoring data is:

[0012] ​​Further, the S4 comprises the following sub-steps: S41, retrieving the environmental temperature and humidity data, physiological index data and metabolic behavior data of the dairy cow when heat stress occurs in the past three years from the historical database, and constructing a historical heat stress data set; S42, performing feature screening on the historical heat stress data set, retaining feature items consistent with the current monitoring data dimensions to form a standardized historical data set; S43, inputting the real-time metabolic behavior evaluation result output by S3 and the standardized historical data set into a heat stress risk dynamic threshold calculation model, determining the initial weight of the risk influencing factor through the similarity calculation of the historical data and the real-time data by the model; S44, based on the initial weight, adjusting the heat stress risk threshold multiple times through the threshold iteration calculation module of the model, and finally determining the heat stress risk threshold interval under the current environment.

[0013] Further, the S5 comprises the following sub-steps: S51, setting heat stress risk grade division standards, dividing the risk grade into three grades of low risk, medium risk and high risk, each grade corresponding to a different threshold range; S52, comparing the calibration indicators in the real-time metabolic behavior evaluation result output by S3, including metabolic intensity value and physiological index deviation, with the heat stress risk threshold interval determined by S4 one by one; S53, counting the number of calibration indicators falling into each risk threshold range, and calculating the matching degree of each risk grade; S54, determining the current heat stress risk grade of the dairy cow according to the risk grade with the highest matching degree.

[0014] The application discloses a dairy cow heat stress online monitoring and early warning system based on multi-source data, which is applied to a dairy cow heat stress online monitoring and early warning method based on multi-source data, and comprises a multi-source data acquisition unit connected with a multi-source data intelligent pasture pupil early warning platform, which is used for collecting dairy cow living environment temperature and humidity data, dairy cow individual activity frequency data, dairy cow body surface temperature and breathing frequency data through a temperature and humidity sensor, an activity amount collector and a physiological index monitoring module respectively, and transmitting the collected data to a data fusion unit; a federated learning heterogeneous data fusion unit connected with the multi-source data acquisition unit and a metabolic behavior evaluation unit respectively, which is used for receiving the data transmitted by the multi-source data acquisition unit, performing feature alignment and fusion processing on the heterogeneous data through a federated learning heterogeneous data fusion model, generating a unified data feature set, and transmitting the unified data feature set to the metabolic behavior evaluation unit; a dairy cow metabolic behavior multi-modal evaluation unit connected with the federated learning heterogeneous data fusion unit and a risk threshold calculation unit respectively, which is used for receiving the unified data feature set, performing weighted fusion and correlation analysis on dairy cow metabolism related multi-dimensional features through a dairy cow metabolic behavior multi-modal fusion model, outputting a dairy cow real-time metabolic behavior evaluation result, and transmitting the evaluation result to the risk threshold calculation unit; a heat stress risk threshold calculation unit connected with the dairy cow metabolic behavior multi-modal evaluation unit and a risk grade judgment unit respectively, which is used for receiving the real-time metabolic behavior evaluation result, combining historical heat stress data, performing dynamic weight distribution and threshold iteration calculation on risk influence factors through a heat stress risk dynamic threshold calculation model, determining a heat stress risk threshold interval, and transmitting the threshold interval to the risk grade judgment unit; a heat stress risk grade judgment unit connected with the heat stress risk threshold calculation unit and an early warning signal transmission unit respectively, which is used for receiving the real-time metabolic behavior evaluation result and the heat stress risk threshold interval, judging a current heat stress risk grade of the dairy cow through comparison and analysis, and transmitting the risk grade to the early warning signal transmission unit; and an early warning signal processing and transmission unit connected with the risk grade judgment unit and a display terminal and a management terminal of the multi-source data intelligent pasture pupil early warning platform respectively, which is used for transmitting the risk grade to the display terminal in the form of a data signal, generating corresponding early warning signals to the management terminal according to the risk grade, and displaying and delivering early warning information.

[0015] Beneficial effects: the present application proposes a kind of dairy cow heat stress online monitoring and early warning method and system based on multi-source data, the method and system are integrated environment, physiological, behavior multidimensional data by multi-source data wisdom pasture pupil early warning platform, the effective fusion of heterogeneous data is realized by federated learning heterogeneous data fusion model, in combination with dairy cow metabolic behavior multi-modal fusion model, the metabolic state of dairy cow is accurately evaluated, and the risk threshold value of adaptive real-time scene is determined by heat stress risk dynamic threshold calculation model, finally realize the online monitoring and early warning of heat stress risk, can real-time, comprehensive capture dairy cow heat stress related information, improve the timeliness and accuracy of monitoring, provide fine management support for large-scale breeding, guarantee dairy cow health and breeding benefit.Meanwhile, in view of the problem of insufficient multi-source data fusion capability of prior art, it is characterized by federated learning heterogeneous data fusion model to align and fuse environment temperature and humidity, dairy cow activity frequency, body surface temperature and other multi-type heterogeneous data, make full use of multidimensional data information, fully reflect the real state of dairy cow heat stress, improve the integrity and reliability of monitoring result;in combination with dairy cow real-time metabolic behavior evaluation result and historical heat stress data, the dynamic threshold calculation model is used to dynamically weight and threshold iteration calculation for risk influence factor, so that the threshold adapts to individual differences of dairy cow, growth stage and environmental fluctuation, avoid early warning lag or misjudgment, provide accurate decision support for heat stress prevention and control. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 For the method step flow chart of the present application;

[0017] Figure 2 For the system unit composition diagram of the present application. DETAILED DESCRIPTION

[0018] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict, and the present application will be further described in detail below in combination with the drawings and specific embodiments.

[0019] As Figure 1 shown, a dairy cow heat stress online monitoring and early warning method based on multi-source data includes the following steps:

[0020] S1, the temperature and humidity sensor, activity amount collector and physiological index monitoring module carried by multi-source data wisdom pasture pupil early warning platform are used to collect dairy cow living environment temperature and humidity data, dairy cow individual activity frequency data, dairy cow body surface temperature and respiratory frequency data respectively;

[0021] Specifically, the implementation of step S1 is through a distributed collection network built by the multi-source data intelligent herd eye early warning platform, which includes temperature and humidity sensors deployed in different areas of the cowshed, activity collectors bound to the cow collars, and physiological indicator monitoring modules attached to the cow body surface. The temperature and humidity sensor adopts a high-precision digital sensor, with a sampling frequency set to once every 5 minutes, covering a temperature range of 0-50℃ and a relative humidity range of 0-100%, ensuring real-time capture of temperature and humidity changes in the cow living environment; the activity collector is built-in with a three-axis acceleration sensor, with a sampling frequency of once every 1 minute, by recording the number and duration of standing, walking and lying of the cow per hour, generating individual activity frequency data of the cow, with data accuracy accurate to the duration of a single activity in seconds; the physiological indicator monitoring module uses a non-contact infrared temperature measuring element and a respiration rate sensing element, collecting cow body surface temperature and respiration rate every 3 minutes, with a body surface temperature collection range of 35-42℃ and a respiration rate collection range of 10-80 times / minute, ensuring that the sensor distance from the cow body surface is controlled at 5-10 cm during collection to avoid environmental interference causing data deviation, and all collected data is uploaded in real time to the local database of the multi-source data intelligent herd eye early warning platform through a wireless transmission module, providing complete and accurate raw data support for subsequent data processing.

[0022] S2, inputting the multi-type data collected in S1 into a federated learning heterogeneous data fusion model, performing feature alignment and fusion processing on the heterogeneous data through a distributed training mechanism among nodes in the model, and generating a unified data feature set;

[0023] Specifically, step S2 deploys a federated learning heterogeneous data fusion model on the edge computing node of the multi-source data intelligent pasture pupil early warning platform. First, the three types of data collected in S1 are standardized in format. The temperature and humidity data are converted to decimal numerical format, the activity data are converted to frequency statistical format, and the physiological index data are converted to continuous numerical sequence format, ensuring that different types of data have a unified data interaction interface. The model sets 5-8 distributed training nodes, each node corresponding to a data processing task of a dairy cow breeding unit. The nodes share data features through an encrypted communication protocol to avoid privacy leakage caused by cross-node transmission of raw data. During the training process, each node first extracts features from local data, including the rate of change of environmental temperature and humidity, the time period difference of cow activity, and the fluctuation amplitude of physiological indicators. Then, the federated average algorithm is used to globally fuse the features extracted by each node. When fusing, set the node weight coefficient, adjust the contribution of each node to the global features according to the number of cows in each breeding unit (the weight coefficient is positively correlated with the number of cows, ranging from 0.1 to 0.3), and finally generate a unified data feature set including environmental, behavioral, and physiological features. The dimension of the feature set is controlled within 20-30 dimensions to ensure that it includes key information and avoids data redundancy, providing high-quality feature input for subsequent metabolic behavior evaluation.

[0024] S3, input the unified data feature set generated in S2 into the dairy cow metabolic behavior multi-modal fusion model. Through the weighted fusion and correlation analysis of the model on the multi-dimensional features related to cow metabolism, the evaluation results of the real-time metabolic behavior of the cow are output.

[0025] Specifically, step S3 inputs the unified data feature set generated by S2 into the dairy cow metabolic behavior multi-modal fusion model deployed on the cloud server of the multi-source data intelligent pasture platform. The model adopts a multi-layer neural network structure, with 20-30 neurons in the input layer corresponding to each dimension of the unified data feature set. The hidden layer is set to 3-5 layers, with 40-60 neurons in each layer. The ReLU activation function is used for nonlinear conversion of the features, with emphasis on strengthening features related to dairy cow metabolism, such as the coordinated change of body temperature and respiratory rate, and the correlation between activity and temperature and humidity. During model training, historical metabolic behavior data (including 10,000+ sets of dairy cow metabolic data under healthy and heat stress conditions) are used for supervised training. The training iteration number is set to 500-800, and the learning rate is controlled at 0.001-0.005 to ensure that the model converges to the optimal state. In the prediction process, the model performs weighted fusion on the input feature data, with weights determined according to the correlation between the features and metabolic behavior. For example, the respiratory rate feature weight is set to 0.2-0.3, the body temperature feature weight is set to 0.15-0.25, and the activity feature weight is set to 0.1-0.2. The weighted calculation outputs the dairy cow real-time metabolic behavior evaluation results, including metabolic intensity level (divided into low, medium, and high) and metabolic stability index (range 0-1, with a value closer to 1 indicating more stable metabolism), providing direct basis for subsequent heat stress risk judgment.

[0026] S4, based on the metabolic behavior evaluation results output by S3, combine the historical environmental and physiological data when heat stress occurs, input the heat stress risk dynamic threshold calculation model, through the model to dynamically allocate the weights of risk influencing factors and iteratively calculate the threshold, determine the heat stress risk threshold interval of dairy cows under the current environment;

[0027] Specifically, step S4 is based on the metabolic behavior evaluation results output by S3, combined with the heat stress data stored in the wisdom pasture pupil early warning platform historical database (including 2000+ times of environmental data, physiological data and metabolic behavior data of heat stress events in the past 3-5 years) to carry out calculation. First, the historical data consistent with the current dairy cow group breed and growth stage (such as lactation period, dry period) are screened from the historical database, and the data amount after screening reaches 500+ groups, ensuring the representativeness of the data; the screened historical data and the real-time metabolic behavior evaluation results (metabolic intensity level, metabolic stability index) output by S3 are input into the heat stress risk dynamic threshold calculation model. The model first statistically analyzes the heat stress occurrence critical value in the historical data to determine the critical range of environmental temperature and humidity, physiological indicators and metabolic behavior; then adjusts the critical range according to the current real-time metabolic behavior evaluation results, such as when the metabolic intensity level is "high", the environmental temperature critical value is lowered by 1-2℃, and when the metabolic stability index is lower than 0.5, the respiratory rate critical value is lowered by 5-8 times / minute; after adjustment, the critical value is iteratively optimized by sliding window algorithm (window size is set to 30-50 groups of data), and the iteration number is 10-15 times, finally the heat stress risk threshold interval of dairy cows under the current environment is determined, including temperature threshold (such as 28-32℃), humidity threshold (such as 65%-80%), respiratory rate threshold (such as 40-60 times / minute), metabolic intensity threshold (such as "medium-high" level), which provides clear standards for subsequent risk level judgment.

[0028] S5, compare and analyze the real-time metabolic behavior evaluation results output by S3 with the heat stress risk threshold interval determined by S4 to determine the current heat stress risk level of dairy cows;

[0029] Specifically, step S5 performs comparison and analysis process in the data analysis module of the multi-source data intelligent pasture pupil early warning platform. First, the key indicators are extracted from the real-time metabolic behavior evaluation results output in S3, including metabolic intensity level, metabolic stability index, hourly average body surface temperature, and hourly average respiratory frequency, a total of 4 core indicators. Then, the threshold range of the corresponding indicators is extracted from the heat stress risk threshold interval determined in S4 to establish the index-threshold correspondence table. Subsequently, each core indicator is matched with the corresponding threshold range in a step-by-step comparison manner. When matching, the judgment rule is set: if the metabolic intensity level is higher than the upper limit of the threshold interval, or the metabolic stability index is lower than the lower limit of the threshold interval, or the average body surface temperature is higher than the upper limit of the threshold interval, or the average respiratory frequency is higher than the upper limit of the threshold interval, it is determined that the indicator triggers risk warning. The number of indicators triggering risk warning is counted, and the risk level determination standard is set: triggering 0 indicators is low risk, triggering 1-2 indicators is medium risk, and triggering 3-4 indicators is high risk. Finally, the current heat stress risk level of the dairy cow is determined according to the statistical result, and the determination process is completed within 1-2 minutes to ensure the timeliness of risk level judgment and gain time for subsequent early warning signal transmission.

[0030] S6, the heat stress risk level determined in S5 is transmitted to the platform display terminal in the form of data signal through the multi-source data intelligent pasture pupil early warning platform, and the early warning signal of the corresponding risk level is triggered to the specified management terminal.

[0031] Specifically, step S6 realizes data transmission and early warning triggering through the communication and control module of the multi-source data intelligent pasture pupil early warning platform. First, the display terminal of the platform (large screen display deployed in the breeding management center and mobile terminal APP of the management personnel) adopts real-time data pushing mechanism to display the heat stress risk level (low, medium, and high) determined in S5 in the form of text and icon combination. The text color is set to distinguish the risk level (green for low risk, yellow for medium risk, and red for high risk), and the icon adopts different shapes (circle for low risk, triangle for medium risk, and square for high risk) to ensure that the management personnel can intuitively identify the risk level. At the same time, the early warning signal transmission module of the platform generates corresponding early warning instructions according to the risk level. For low risk, only text reminder instruction is sent (pushed through APP), for medium risk, text reminder + sound and light alarm instruction is sent (sound and light alarm of the breeding management center is started), and for high risk, text reminder + sound and light alarm + device control instruction is sent (cooling devices such as fans and sprinkler system in the dairy cow house are started). Early warning signal transmission adopts 5G or LoRa wireless communication technology, and transmission delay is controlled within 100-300 milliseconds to ensure that the instruction is delivered quickly. In addition, the platform automatically records the early warning time, risk level, triggered early warning instruction, and device response state, and stores them in the early warning log database. The log preservation period is set to 1-3 years to facilitate subsequent breeding management analysis and traceability.

[0032] Preferably, the expression of the dairy cow metabolic behavior multi-modal fusion model is: wherein, represents the real-time metabolic behavior evaluation result of the dairy cow, represents the number of metabolic behavior evaluation dimensions, represents the weight coefficient of the th evaluation dimension, represents the feature mapping function of the th evaluation dimension, represents the original data corresponding to the th evaluation dimension, represents the correlation adjustment coefficient of the th evaluation dimension, represents the number of correlation data dimensions, represents the correlation coefficient of the th evaluation dimension and the th correlation data dimension, represents the original data corresponding to the th correlation data dimension, represents the model error correction term.

[0033] Specifically, the dairy cow metabolic behavior multi-modal fusion model, in implementation, first determines the number of evaluation dimensions of the model input, which is usually set to 8-12 in combination with the dairy cow metabolic behavior monitoring requirements, including key dimensions such as activity frequency, body surface temperature, and respiratory rate; the weight coefficient of each evaluation dimension is determined through historical data training, the training sample uses metabolic data of dairy cows in different seasons and different growth stages in the past 2-3 years, and the sample size needs to reach 5000+ groups; in the training process, the weight is iteratively optimized through the gradient descent algorithm, finally making the weight coefficient of each dimension in a reasonable interval of 0.05-0.3, wherein the weight of dimensions closely related to metabolism such as respiratory rate and body surface temperature is set to 0.2-0.3, and the weight of auxiliary dimensions such as activity frequency is set to 0.05-0.15. The correlation adjustment coefficient is calculated according to the correlation between each evaluation dimension, the correlation is determined through Pearson correlation coefficient analysis, and the value range is 0.1-0.5; the number of correlation data dimensions is consistent with the number of evaluation dimensions, and the correlation coefficient is also calibrated through historical data to ensure that the model can effectively integrate multi-dimensional features. The model error correction term is dynamically adjusted according to the deviation between the actual monitoring data and the model prediction result, the deviation is calculated using the mean square error method, and the correction term value range is -0.05-0.05; through continuous iteration, the model prediction error is controlled within 5%, and the implementation of the model can realize the accurate evaluation of the real-time metabolic behavior of dairy cows, provide reliable metabolic state basis for subsequent heat stress risk judgment, and solve the problem that single-dimensional evaluation cannot fully reflect the metabolic condition of dairy cows.

[0034] Preferably, the expression of the federated learning heterogeneous data fusion model is: , wherein, represents a unified data feature set, represents the number of federated learning nodes, represents the weight coefficient of the th node, represents the local data processing coefficient of the th node, represents the heterogeneous data matrix collected by the th node, represents the local data feature weight matrix of the th node, represents the data interaction coefficient of the th node with other nodes, represents the data interaction weight matrix of the th node and the th node.

[0035] Specifically, the federated learning heterogeneous data fusion model first determines the number of federated learning nodes during implementation, and the number of nodes matches the scale of dairy cow breeding. A small and medium-sized breeding farm sets 5-8 nodes, and a large-scale breeding farm sets 8-12 nodes. Each node corresponds to one breeding unit. The node weight coefficient is distributed according to the number of dairy cows in each breeding unit. The weight coefficient of the unit with a number of 50-100 is set to 0.1-0.15, the weight coefficient of the unit with a number of 100-200 is set to 0.15-0.2, and the weight coefficient of the unit with a number of more than 200 is set to 0.2-0.3. It is ensured that the node weight matches the data contribution. The local data processing coefficient is set in combination with the data quality of each node. The coefficient of the node with a data integrity rate of more than 95% is set to 0.8-0.9, the coefficient of the node with a data integrity rate of 90%-95% is set to 0.7-0.8, and the coefficient of the node with a data integrity rate of less than 90% is set to 0.6-0.7. The local data processing priority is optimized through the coefficient. The data interaction coefficient is determined according to the similarity between nodes. The similarity is calculated by cosine similarity, and the value range is 0.3-0.7. The interaction coefficient between nodes with high similarity is set to 0.5-0.7 to promote data feature sharing. The data interaction weight matrix is generated through the feature data trained by encrypted communication transmission between nodes. The number of training iterations is set to 300-500 times, and the learning rate is controlled at 0.001-0.003 to ensure that the matrix can accurately reflect the data association between nodes. The model uses homomorphic encryption technology to protect data privacy during implementation. The encryption key length is set to 2048 bits to avoid leakage of original data. The dimension of the unified data feature set generated finally is controlled at 20-30 dimensions, and the feature integrity reaches more than 90%, providing a high-quality data basis for subsequent metabolic behavior evaluation.

[0036] Preferably, the expression of the heat stress risk dynamic threshold calculation model is: , wherein, This represents the critical value within the heat stress risk threshold range. Indicates the historical threshold baseline value. This represents the adjustment coefficient for fluctuations in environmental factors. Indicates the number of environmental impact factors. Indicates the first The weights of each environmental impact factor, Indicates the first Real-time values ​​of each environmental impact factor. Indicates the first Historical average values ​​of environmental impact factors This represents the physiological factor deviation adjustment coefficient. Indicates the number of physiological influencing factors. Indicates the first The weights of each physiological influencing factor, Indicates the first Real-time values ​​of several physiological influencing factors Indicates the first The historical average value of each physiological influencing factor.

[0037] Specifically, the dynamic threshold calculation model for heat stress risk first determines the historical threshold baseline value. This value is determined by statistically analyzing the average environmental and physiological data from the past 3-5 years when no heat stress occurred. The statistical sample needs to include data from different seasons and weather conditions, with a sample size of 3000+ groups. The temperature baseline value is typically set at 25-27℃, the humidity baseline value at 55%-65%, and the respiratory rate baseline value at 25-35 breaths / minute. The environmental factor fluctuation adjustment coefficient is set according to local climate fluctuations: 1.2-1.5 for summer (high temperature and humidity), 1.0-1.2 for spring and autumn, and 0.8-1.0 for winter. The number of environmental impact factors is set to 6-8, including temperature, humidity, light intensity, and wind speed. The weight of each factor is determined using the analytic hierarchy process (AHP): temperature and humidity weights are set at 0.25-0.35, and light intensity and wind speed weights are set at 0.05-0.15. The physiological factor deviation adjustment coefficient is dynamically adjusted according to the physiological state of dairy cows. The coefficient is set to 1.1-1.3 for lactating cows and 0.9-1.1 for dry cows. The number of physiological influencing factors is set to 4-6, including body surface temperature, respiratory rate, and heart rate. The weights are also determined by the analytic hierarchy process, with respiratory rate and body surface temperature weighted at 0.3-0.4, and heart rate weighted at 0.1-0.2. During model calculation, the real-time values ​​and historical averages of each factor need to be obtained in real time. The historical averages are calculated using data from the same period within the past month. This model allows the heat stress risk threshold range to be dynamically adjusted according to environmental and physiological changes, with threshold accuracy controlled within ±1℃ (temperature), ±5% (humidity), and ±2 breaths / minute (respiratory rate), avoiding warning deviations caused by fixed thresholds.

[0038] Preferably, the data transmission efficiency calculation expression of the multi-source data intelligent pasture pupil early warning platform is: , wherein, represents the platform data transmission efficiency, represents the data transmission times, represents the data amount of the nth transmission, represents the integrity coefficient of the data of the nth transmission, represents the start time of the nth data transmission, represents the end time of the nth data transmission, represents the platform transmission channel stability coefficient.

[0039] Specifically, the data transmission efficiency calculation of the multi-source data intelligent pasture pupil early warning platform is implemented by first determining the data transmission times statistical period. To ensure real-time performance, the statistical period is set to 1 hour. The transmission times in each period are determined according to the data acquisition frequency. The temperature and humidity data are transmitted once every 5 minutes, and the transmission times in 1 hour are 12 times. The activity data are transmitted once every 1 minute, and the transmission times in 1 hour are 60 times. The physiological index data are transmitted once every 3 minutes, and the transmission times in 1 hour are 20 times. The total transmission times in each period are calculated to be about 92 times. The data amount of each transmission is determined according to the data type. The temperature and humidity data are 8 bytes per time, the activity data are 16 bytes per time, and the physiological index data are 12 bytes per time. The data integrity coefficient is determined through data verification. The CRC32 verification algorithm is used. The data packet integrity coefficient is set to 1.0 when the verification is passed, to 0.5-0.9 when the verification fails but can be repaired, and to 0.0 when it cannot be repaired. The integrity coefficient mean is maintained above 0.95 through continuous optimization of the transmission protocol. The data transmission start and end times are recorded through the platform timestamp. The timestamp accuracy is controlled at the millisecond level. The transmission channel stability coefficient is calculated according to the transmission success rate in the past 24 hours. The coefficient is set to 0.9-1.0 when the success rate is above 99%, to 0.8-0.9 when the success rate is between 95% and 99%, and to 0.7-0.8 when the success rate is below 95%. Through the efficiency calculation, the platform data transmission state can be monitored in real time. When the efficiency is below 0.8 MB / s, the transmission optimization instruction is triggered to adjust the transmission frequency band and compress the data format, ensuring the stability of the platform data transmission and providing timely data support for subsequent monitoring and early warning.

[0040] Preferably, the feature extraction expression of the dairy cow heat stress online monitoring data is: , wherein, represents the feature extraction result of the monitoring data, represents the number of feature extraction dimensions, represents the feature extraction dimension of the nth​​​​ Weights for each feature extraction dimension Indicates the first Feature transformation function for each feature extraction dimension Indicates the first The original monitoring data corresponding to each feature extraction dimension Indicates the first Interaction coefficients of each feature extraction dimension Indicates the number of dimensions in the interactive data. Indicates the first The feature extraction dimension and the first Interaction index of each interaction data dimension Indicates the first The original monitoring data corresponding to each interactive data dimension.

[0041] Specifically, for feature extraction from online monitoring data of heat stress in dairy cows, the first step is to determine the number of feature dimensions. Based on the needs of heat stress monitoring, this number is set to 10-14, including key dimensions such as temperature change rate, humidity fluctuation range, peak activity level, and respiratory rate coefficient of variation. The weight of each dimension is determined through feature importance analysis. A random forest algorithm is used to evaluate the features of historical monitoring data, with an evaluation sample size of over 4000 groups. Ultimately, the weights of core dimensions such as temperature change rate and respiratory rate coefficient of variation are set to 0.2-0.3, while the weights of auxiliary dimensions such as humidity fluctuation range and peak activity level are set to 0.05-0.15. The feature transformation function is selected according to the data type. Continuous data (such as temperature and respiratory rate) uses a linear normalization function, while discrete data (such as activity level) uses a one-heat encoding function to ensure that different types of data are within a uniform numerical range (0-1) after transformation. The interaction coefficient is determined based on the synergistic effect between each feature dimension. The strength of the synergistic effect is calculated using mutual information entropy, with a value ranging from 0.1 to 0.4. The number of interaction data dimensions is consistent with the number of feature extraction dimensions. The interaction index is dynamically adjusted according to the strength of the synergistic effect, with an index of 1.2-1.5 for high-strength dimensions and 0.8-1.0 for low-strength dimensions. During feature extraction, the effectiveness of the extraction results needs to be verified using a confusion matrix method to ensure that the feature extraction accuracy is above 90%. This process can extract key information from the original monitoring data, reduce redundant data interference, and provide accurate feature input for subsequent data fusion and risk assessment.

[0042] Preferably, the S3 comprises the following steps: S31, screening the feature data related to the metabolic behavior of the dairy cow from the unified data feature set generated in S2, including the activity frequency derived feature of the dairy cow, the body surface temperature change feature and the respiratory frequency fluctuation feature; S32, classifying the screened feature data according to the preset dimension classification rule, and dividing into the metabolic basis feature class and the metabolic correlation feature class; S33, inputting the classified feature data into different input layers of the dairy cow metabolic behavior multi-modal fusion model, and performing nonlinear conversion on the feature data through the activation function of each layer in the model; S34, weighting and fusing the converted feature data of each input layer, combining the feature correlation weight determined in the model training process, and generating the real-time metabolic behavior evaluation result of the dairy cow.

[0043] Specifically, step S3 includes four sub-steps. In the S31 stage, when screening the metabolic behavior related feature data from the unified data feature set generated in S2, the feature selection algorithm (such as variance analysis) is used to remove the features with a metabolic correlation degree lower than 0.3, and the remaining features include the activity frequency derived activity change value of the dairy cow per hour, the fluctuation difference value of the body surface temperature every 10 minutes, and the fluctuation amplitude of the respiratory frequency every 5 minutes. The dimension of the screened feature data needs to be controlled within 15-20 dimensions to ensure that the data focuses on the metabolic core information. In the S32 stage, when classifying according to the preset dimension classification rule, the classification rule is determined according to the direct correlation degree of the feature and the metabolism. The metabolic basis feature class includes features such as body surface temperature and respiratory frequency that directly reflect the metabolic state, and there are 8-10 features in total. The metabolic correlation feature class includes activity frequency and metabolic indirect features under the influence of environmental temperature and humidity, and there are 7-10 features in total. After classification, the class needs to be clearly marked through feature labeling. In the S33 stage, when inputting the classified features into different input layers of the model, the number of input layer neurons corresponds to the dimension of the features one by one, and the activation function uses the ReLU function. This function outputs the original value when the feature value is greater than 0, and outputs 0 when the feature value is less than 0. The feature discrimination is enhanced through nonlinear conversion. The gradient change of the feature data needs to be monitored in real time during the conversion process to ensure that the gradient value is within a reasonable range of 0.1-0.5. In the S34 stage, when weighting and fusing the converted features, the fusion weight is determined according to the importance of the features. The feature importance score is calculated through the random forest algorithm. The feature weight is set to 0.2-0.3 when the score is above 0.8, to 0.1-0.2 when the score is between 0.5 and 0.8, and to 0.05-0.1 when the score is below 0.5. The metabolic behavior evaluation result generated after fusion needs to include the metabolic intensity value (range 0-10, the higher the value, the more active the metabolism) and the metabolic stability value (range 0-1, the closer the value to 1, the more stable the metabolism), which provides accurate basis for subsequent heat stress risk analysis.

[0044] Preferably, the S4 comprises the following steps: S41, retrieving environmental temperature and humidity data, physiological index data and metabolic behavior data of the cow when heat stress occurs in the past three years from the historical database to construct a historical heat stress data set; S42, performing feature screening on the historical heat stress data set to retain feature items consistent with the dimension of the current monitoring data to form a standardized historical data set; S43, inputting the real-time metabolic behavior evaluation result output by S3 and the standardized historical data set into a heat stress risk dynamic threshold calculation model, calculating the similarity of the historical data and the real-time data through the model to determine the initial weight of the risk influencing factor; S44, based on the initial weight, adjusting the heat stress risk threshold multiple times through the threshold iteration calculation module of the model to finally determine the heat stress risk threshold interval under the current environment.

[0045] Specifically, step S4 includes four sub-steps. In the S41 stage of retrieving historical heat stress data, the data needs to be screened from the historical database of the multi-source data intelligent pasture pupil early warning platform. The database storage period is set to 3-5 years. The retrieved data needs to include environmental temperature and humidity (temperature 28-35℃, humidity 65%-90%) when heat stress occurs, physiological indicators (body surface temperature 39-42℃, respiratory rate 40-80 times / minute), metabolic behavior data (metabolic intensity value 6-10, metabolic stability value 0.3-0.7). The amount of data retrieved needs to reach 2000+ groups, and includes data of cows of different breeds and different growth stages to ensure data representativeness. In the S32 stage of screening feature items, the 20-30-dimensional features of the current monitoring data are used as a benchmark to eliminate features in the historical data that do not match the dimension, such as eliminating the historical feature if the current monitoring includes activity frequency derived features but the historical data does not. The dimension of the standardized historical data set formed after screening is consistent with the current monitoring data, and the amount of data is maintained at 1500+ groups. In the S33 stage of determining the initial weight of the risk influencing factor, the similarity between the real-time metabolic behavior evaluation result and the standardized historical data set needs to be calculated. The cosine similarity algorithm is adopted, and the similarity threshold is set to 0.7. The historical data with a similarity higher than the threshold is taken as a reference sample. The initial weight of the environmental temperature, respiratory rate, metabolic intensity, and other risk influencing factors is determined through the analytic hierarchy process, wherein the initial weight of the environmental temperature, respiratory rate, and metabolic intensity is set to 0.25-0.35, and the initial weight of the other factors is set to 0.05-0.15. In the S34 stage of iterative optimization of the threshold, the number of iterations is set to 15-20 times. The threshold is adjusted according to the deviation of the real-time data and the historical reference sample in each iteration. The deviation is calculated using the Euclidean distance method. The final heat stress risk threshold interval needs to meet the temperature fluctuation ±1℃, humidity fluctuation ±5%, and respiratory rate fluctuation ±2 times / minute to ensure that the threshold can adapt to the real-time breeding scene.

[0046] Preferably, S5 comprises the following steps: S51, setting a thermal stress risk level classification standard, classifying the risk level into three levels of low risk, medium risk and high risk, each level corresponding to a different threshold range; S52, comparing the calibrated indicators in the real-time metabolic behavior evaluation result output by S3, including metabolic intensity value and physiological indicator deviation, with the thermal stress risk threshold interval determined by S4 one by one; S53, counting the number of calibrated indicators falling into each risk threshold range, and calculating the matching degree of each risk level; S54, determining the current thermal stress risk level of the dairy cow according to the risk level with the highest matching degree.

[0047] Specifically, step S5 comprises four sub-steps. In S51, the risk level classification standard needs to be formulated in combination with the industry thermal stress prevention and control specifications and historical early warning data. The threshold interval corresponding to the low risk level is temperature lower than 28℃, humidity lower than 65%, respiratory rate lower than 40 times / minute, and metabolic intensity value lower than 6. The threshold interval corresponding to the medium risk is temperature 28-32℃, humidity 65%-75%, respiratory rate 40-60 times / minute, and metabolic intensity value 6-8. The threshold interval corresponding to the high risk is temperature higher than 32℃, humidity higher than 75%, respiratory rate higher than 60 times / minute, and metabolic intensity value higher than 8. The threshold range of each level needs to be verified by 500+ historical early warning cases to ensure the rationality of the classification. In S52, three key indicators, i.e. metabolic intensity value and physiological indicator deviation (difference of body surface temperature deviating from the normal range 38-39℃, difference of respiratory rate deviating from the normal range 25-35 times / minute), are extracted from the metabolic behavior evaluation result for comparison. The threshold interval matching method is used for comparison, and the matching result (compliance / incompliance with the threshold of a certain risk level) is recorded after each indicator comparison. In S53, the matching degree of the risk level is calculated by dividing the number of indicators complying with the threshold of a certain risk level by the total number of key indicators. For example, if two indicators comply with the threshold of the medium risk, the matching degree of the medium risk is 66.7%. The calculation process needs to retain two decimal places to ensure accuracy. In S54, the risk level is determined by selecting the level with the highest matching degree as the current risk level. If the matching degrees of two levels are the same (e.g. 66.7% for the medium risk and the high risk), the absolute values of the physiological indicator deviations are further compared, and the greater the deviation, the higher the risk level. After determining the risk level, a level identifier (low risk marked green, medium risk marked yellow, and high risk marked red) is generated to provide a clear basis for subsequent early warning signal transmission.

[0048] The dairy cow metabolic behavior multi-modal fusion model in the application is an algorithmic model for integrating multi-dimensional physiological and behavioral data of dairy cows and accurately evaluating the real-time metabolic state of dairy cows. Specifically, the model takes multi-type data related to metabolism such as activity frequency, body surface temperature, and respiratory rate as input, performs weighted fusion and correlation analysis on multi-modal data through a multi-layer neural network structure, and finally outputs evaluation results including metabolic intensity level and metabolic stability index. In the implementation process, first, 8-12 metabolic evaluation dimensions are determined, based on 5000+ historical metabolic data, the gradient descent algorithm is used to iteratively optimize the weight coefficients of each dimension (core dimension weight 0.2-0.3, auxiliary dimension 0.05-0.15), and a correlation adjustment coefficient of 0.1-0.5 is set according to the correlation between dimensions, and an error correction term of -0.05-0.05 is introduced to reduce the prediction bias, so that the model error is controlled within 5%. The role of the model is to break through the limitations of single data evaluation, comprehensively capture the metabolic changes of dairy cows, provide reliable metabolic state basis for heat stress risk judgment, solve the one-sided and inaccurate problem of traditional monitoring, enable breeders to real-time grasp the physiological basic state of dairy cows, provide accurate data support for subsequent heat stress prevention and control, and help the fine management of dairy cow health in large-scale breeding.

[0049] The federated learning heterogeneous data fusion model in the application is a distributed algorithmic model for feature alignment and fusion of multi-source heterogeneous dairy cow monitoring data under the premise of protecting data privacy. Specifically, the model relies on multiple distributed training nodes to process different types and formats of data such as environmental temperature and humidity, dairy cow activity, and physiological indicators, and generates a unified data feature set. In implementation, first, 5-12 nodes are set according to the breeding scale (corresponding to different breeding units), and the node weight of 0.1-0.3 is allocated according to the number of dairy cows, and the local data processing coefficient of 0.6-0.9 is set according to the data quality; the feature data is shared between nodes through 2048-bit homomorphic encryption technology to avoid leakage of original data, and a data interaction weight matrix is generated through 300-500 iterations of training, and finally a unified feature set with 20-30 dimensions and more than 90% completeness is obtained. The model breaks through the barriers of multi-source heterogeneous data, fully excavates the associated value of different types of data, provides a high-quality data basis for subsequent metabolic evaluation, solves the problems of privacy leakage and low data utilization in traditional data fusion, improves the data integration efficiency while ensuring the safety of breeding data, and provides a feasible solution for collaborative analysis of multi-source data in large-scale dairy farming.

[0050] The heat stress risk dynamic threshold calculation model in the application is an algorithmic model that dynamically generates heat stress risk judgment criteria by combining real-time state and historical data of dairy cows. Specifically, the model takes the real-time metabolic assessment results and historical heat stress data of dairy cows as input, determines the heat stress risk threshold interval that adapts to the current scene through weight distribution and iterative calculation of environmental and physiological impact factors. In the implementation process, first, based on 3000+ groups of historical non-stress data, the baseline values of temperature 25-27℃, humidity 55%-65% and other baseline values are determined, combined with the season to set the environmental factor fluctuation adjustment coefficient of 0.8-1.5, and according to the physiological state of dairy cows to set the physiological factor deviation adjustment coefficient of 0.9-1.3; then 6-8 environmental factors (weight 0.05-0.35) and 4-6 physiological factors (weight 0.1-0.4) are selected, and the threshold is optimized through 10-15 times of sliding window iteration, so that the temperature accuracy is ±1℃, the humidity is ±5%, and the respiratory rate is ±2 times / minute. The model replaces the fixed threshold to generate risk judgment criteria that match the real-time scene, improves the accuracy of heat stress early warning, solves the problem that the traditional fixed threshold cannot adapt to environmental fluctuations and individual differences of dairy cows, avoids early warning lag or misjudgment, provides scientific and dynamic decision-making basis for heat stress prevention and control, and reduces breeding losses caused by heat stress.

[0051] The multi-source data intelligent pasture pupil early warning platform algorithm in the application is a comprehensive algorithm system supporting multi-source data acquisition, processing, analysis and early warning signal transmission. Specifically, the algorithm includes data acquisition scheduling, transmission efficiency monitoring, risk level display and early warning instruction generation functions, and runs through the whole process of dairy cow heat stress monitoring and early warning. In the implementation, the data acquisition link schedules the sensors at a frequency of 5 minutes (temperature and humidity), 1 minute (activity), and 3 minutes (physiological indicators), the transmission link transmits data through 5G / LoRa technology (delay 100-300 milliseconds), uses CRC32 check to ensure data integrity (coefficient mean value above 0.95), and monitors the transmission state through the efficiency formula (92 transmissions per hour, 0.8MB / s threshold); the early warning link generates text, sound and light, and device control instructions according to the risk level, and synchronously records the early warning log (saved for 1-3 years). The algorithm integrates the technologies of each link, realizes the real-time circulation of monitoring data and efficient transmission of early warning information, ensures the stable operation of the whole system, solves the problems of data dispersion, low transmission efficiency and slow response in traditional monitoring, and builds a closed loop from data acquisition to early warning execution, providing an integrated heat stress management tool for large-scale dairy farming, and promoting the upgrading of intelligent and automated farming management.

[0052] As Figure 2As shown, a dairy cow heat stress online monitoring and early warning system based on multi-source data, the system is applied to a dairy cow heat stress online monitoring and early warning method based on multi-source data, comprising: a multi-source data acquisition unit, the unit is connected with multi-source data intelligent pasture pupil early warning platform, is used for through humidity sensor, activity amount collector, physiological index monitoring module respectively collect dairy cow living environment temperature and humidity data, dairy cow individual activity frequency data, dairy cow body surface temperature and respiratory rate data, and transmit the collected data to data fusion unit; federated learning heterogeneous data fusion unit, the unit is connected with multi-source data acquisition unit and metabolic behavior evaluation unit respectively, is used for receiving the data transmitted by multi-source data acquisition unit, through federated learning heterogeneous data fusion model, the feature alignment and fusion processing of heterogeneous data are carried out, the unified data feature set is generated, and the unified data feature set is transmitted to metabolic behavior evaluation unit; dairy cow metabolic behavior multi-modal evaluation unit, the unit is connected with federated learning heterogeneous data fusion unit and risk threshold calculation unit respectively, is used for receiving unified data feature set, through dairy cow metabolic behavior multi-modal fusion model, the weighted fusion and correlation analysis of dairy cow metabolism related multi-dimensional features are carried out, the real-time metabolic behavior evaluation result of dairy cow is output, and the evaluation result is transmitted to risk threshold calculation unit; heat stress risk threshold calculation unit, the unit is connected with dairy cow metabolic behavior multi-modal evaluation unit and risk grade judgment unit respectively, is used for receiving real-time metabolic behavior evaluation result, combining historical heat stress data, through heat stress risk dynamic threshold calculation model, the dynamic weight distribution and threshold iteration calculation of risk influence factor are carried out, the heat stress risk threshold interval is determined, and the threshold interval is transmitted to risk grade judgment unit; heat stress risk grade judgment unit, the unit is connected with heat stress risk threshold calculation unit and early warning signal transmission unit respectively, is used for receiving real-time metabolic behavior evaluation result and heat stress risk threshold interval, through comparison and analysis, the current heat stress risk grade of dairy cow is judged, and the risk grade is transmitted to early warning signal transmission unit; early warning signal processing and transmission unit, the unit is connected with risk grade judgment unit and display terminal, management terminal of multi-source data intelligent pasture pupil early warning platform respectively, is used for transmitting risk grade to display terminal in the form of data signal, and according to risk grade, corresponding early warning signal is generated and transmitted to management terminal, the display and transmission of early warning information are carried out.

[0053] The method and system for online monitoring and early warning of heat stress of dairy cows based on multi-source data have strong data integration and processing capabilities, can synchronously collect multi-dimensional data such as dairy cow living environment, individual activity and physiological indexes through a multi-source data intelligent pasture pupil early warning platform, perform feature alignment and fusion on various heterogeneous data through a federated learning heterogeneous data fusion model, avoid one-sidedness of single data information, provide a comprehensive data basis for subsequent analysis, have high evaluation and early warning accuracy, can deeply analyze dairy cow metabolism related features by means of a dairy cow metabolism behavior multi-modal fusion model, accurately output metabolism behavior evaluation results, generate a risk threshold adaptive to a real-time scene in combination with a heat stress risk dynamic threshold calculation model, and then accurately judge a heat stress risk grade, thereby providing accurate reference for breeding management, and have strong practicability, can display the risk grade in real time through a platform terminal and transmit an early warning signal, and meet real-time management needs of large-scale breeding.

[0054] The method and system can break the barriers of different types of data through the federated learning heterogeneous data fusion model, perform systematic fusion processing on data such as environmental temperature and humidity, dairy cow activity frequency, body surface temperature and respiratory frequency, fully mine the correlation value of multi-source data, comprehensively reflect the heat stress state of dairy cows, and solve the limitations of single data monitoring; the method and system do not use a unified fixed threshold, but combine dairy cow real-time metabolism behavior evaluation results and historical heat stress data, perform dynamic weight distribution on various factors affecting heat stress through a dynamic threshold calculation model, and continuously iterate and optimize the threshold, so that the threshold can adapt to individual differences of dairy cows, different growth stages and environmental changes, effectively avoid early warning lag or misjudgment, and improve the scientificity and effectiveness of early warning.

[0055] In the description of the present application, it should be noted that, unless otherwise explicitly specified and limited, the terms "arrangement", "installation", "connection", "link", "fixing" should be understood in a broad sense, for example, can be fixed connection, can also be detachable connection, or integrally connected; can be mechanically connected, or can be electrically connected; can be directly connected, or can be indirectly connected through an intermediate medium, or can be internal communication of two elements. For ordinary skilled in the art, the specific meaning of the above terms in the present application can be understood through specific circumstances.

[0056] Although the embodiments of the present application have been shown and described, it can be understood by those skilled in the art that various equivalent changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalent scope.

Claims

1. A method for online monitoring and early warning of heat stress in dairy cows based on multi-source data, characterized in that, Includes the following steps: S1. The temperature and humidity sensor, activity level collector, and physiological indicator monitoring module of the multi-source data intelligent shepherd eye early warning platform collect data on the temperature and humidity of the dairy cow's living environment, the frequency of individual dairy cow activity, and the body surface temperature and respiratory rate of the dairy cow, respectively. S2. Input the multi-type data collected in S1 into the federated learning heterogeneous data fusion model, and perform feature alignment and fusion processing on the heterogeneous data through the distributed training mechanism between nodes in the model to generate a unified data feature set. S3. Input the unified data feature set generated in S2 into the multimodal fusion model of dairy cow metabolic behavior. Through the weighted fusion and correlation analysis of the multi-dimensional features related to dairy cow metabolism, the model outputs the real-time metabolic behavior assessment results of dairy cows. S4. Based on the metabolic behavior assessment results output by S3, combined with the environmental and physiological data of historical heat stress occurrence, input the dynamic threshold calculation model for heat stress risk. Through the dynamic weight allocation of risk influencing factors and threshold iteration calculation by the model, determine the heat stress risk threshold range of dairy cows under the current environment. S5. Compare and analyze the real-time metabolic behavior assessment results output by S3 with the heat stress risk threshold range determined by S4 to determine the current heat stress risk level of the dairy cow. S6. The heat stress risk level determined by S5 is transmitted to the platform display terminal in the form of a data signal through the multi-source data intelligent Mutong early warning platform, and the corresponding risk level early warning signal is transmitted to the designated management terminal.

2. The method for online monitoring and early warning of heat stress in dairy cows based on multi-source data according to claim 1, characterized in that, The expression for the multimodal fusion model of dairy cow metabolic behavior is as follows: ,in, This indicates the results of real-time metabolic behavior assessment in dairy cows. This indicates the number of dimensions for assessing metabolic behavior. Indicates the first The weighting coefficients of each evaluation dimension, Indicates the first Feature mapping function for each evaluation dimension Indicates the first The original data corresponding to each evaluation dimension Indicates the first The correlation moderating coefficients for each assessment dimension, Indicates the number of dimensions of the associated data. Indicates the first The assessment dimension and the first The correlation coefficient of each related data dimension. Indicates the first The original data corresponding to each related data dimension This represents the model error correction term.

3. The method for online monitoring and early warning of heat stress in dairy cows based on multi-source data according to claim 1, characterized in that, The expression for the federated learning heterogeneous data fusion model is: ,in, Represents a unified data feature set. Indicates the number of federated learning nodes. Indicates the first The weight coefficients of each node, Indicates the first Local data processing coefficients for each node Indicates the first Heterogeneous data matrix collected by each node Indicates the first Local data feature weight matrix of each node Indicates the first The data interaction coefficient between each node and other nodes Indicates the first The node and the first The data interaction weight matrix of each node.

4. The method for online monitoring and early warning of heat stress in dairy cows based on multi-source data according to claim 1, characterized in that, The expression for the dynamic threshold calculation model for heat stress risk is: ,in, This represents the critical value within the heat stress risk threshold range. Indicates the historical threshold baseline value. This represents the adjustment coefficient for fluctuations in environmental factors. Indicates the number of environmental impact factors. Indicates the first The weights of each environmental impact factor, Indicates the first Real-time values ​​of several environmental impact factors. Indicates the first Historical average values ​​of environmental impact factors This represents the physiological factor deviation adjustment coefficient. Indicates the number of physiological influencing factors. Indicates the first The weights of each physiological influencing factor, Indicates the first Real-time values ​​of several physiological influencing factors Indicates the first The historical average value of each physiological influencing factor.

5. The method for online monitoring and early warning of heat stress in dairy cows based on multi-source data according to claim 1, characterized in that, The expression for calculating the data transmission efficiency of the multi-source data intelligent early warning platform is as follows: ,in, This indicates the platform's data transmission efficiency. Indicates the number of data transmissions. Indicates the first The amount of data transmitted in this step Indicates the first Integrity coefficient of the transmitted data Indicates the first The start time of this data transmission Indicates the first End time of this data transmission This represents the stability coefficient of the platform's transmission channel.

6. The method for online monitoring and early warning of heat stress in dairy cows based on multi-source data according to claim 1, characterized in that, The feature extraction expression for the online monitoring data of heat stress in dairy cows is: ,in, This indicates the feature extraction results of the monitoring data. Indicates the number of dimensions for feature extraction. Indicates the first Weights for each feature extraction dimension Indicates the first Feature transformation function for each feature extraction dimension Indicates the first The original monitoring data corresponding to each feature extraction dimension Indicates the first Interaction coefficients of each feature extraction dimension Indicates the number of dimensions in the interactive data. Indicates the first The feature extraction dimension and the first Interaction index of each interaction data dimension Indicates the first The original monitoring data corresponding to each interactive data dimension.

7. The method for online monitoring and early warning of heat stress in dairy cows based on multi-source data according to claim 1, characterized in that, S3 includes the following steps: S31, filtering out feature data related to dairy cow metabolic behavior from the unified data feature set generated in S2, including dairy cow activity frequency-derived features, body surface temperature change features, and respiratory rate fluctuation features; S32, classifying the filtered feature data according to preset dimensional classification rules, dividing it into basic metabolic feature class and metabolic related feature class; S33, inputting the classified feature data into different input layers of the dairy cow metabolic behavior multimodal fusion model, and performing nonlinear transformation on the feature data through the activation functions of each layer in the model; S34, performing weighted fusion on the transformed feature data of each input layer, and combining the feature association weights determined during model training to generate real-time dairy cow metabolic behavior assessment results.

8. The method for online monitoring and early warning of heat stress in dairy cows based on multi-source data according to claim 1, characterized in that, S4 includes the following steps: S41, retrieve environmental temperature and humidity data, physiological index data, and metabolic behavior data of dairy cows during heat stress from the historical database over the past three years to construct a historical heat stress dataset; S42, perform feature filtering on the historical heat stress dataset, retaining feature items consistent with the dimensions of the current monitoring data to form a standardized historical dataset; S43, input the real-time metabolic behavior assessment results output from S3 and the standardized historical dataset into the dynamic threshold calculation model for heat stress risk, and determine the initial weights of risk influencing factors by calculating the similarity between historical data and real-time data through the model; S44. Based on the initial weights, the heat stress risk threshold is adjusted multiple times through the threshold iteration calculation module of the model to finally determine the heat stress risk threshold range under the current environment.

9. The method for online monitoring and early warning of heat stress in dairy cows based on multi-source data according to claim 1, characterized in that, S5 includes the following steps: S51, setting a standard for classifying heat stress risk levels, dividing the risk levels into three levels: low risk, medium risk, and high risk, with each level corresponding to a different threshold range; S52, comparing the calibrated indicators in the real-time metabolic behavior assessment results output in S3, including metabolic intensity values ​​and physiological indicator deviations, with the heat stress risk threshold ranges determined in S4 one by one; S53, counting the number of calibrated indicators falling into each risk threshold range, and calculating the matching degree of each risk level; S54, determining the current heat stress risk level of the dairy cow based on the risk level with the highest matching degree.

10. An online monitoring and early warning system for heat stress in dairy cows based on multi-source data, characterized in that, This system is applied to the online monitoring and early warning method for heat stress in dairy cows based on multi-source data as described in claim 1, comprising: The multi-source data acquisition unit is connected to the multi-source data intelligent herdsman early warning platform. It is used to collect data on the temperature and humidity of the dairy cow's living environment, the frequency of individual dairy cow activities, and the body surface temperature and respiratory rate of dairy cows through temperature and humidity sensors, activity data collectors, and physiological index monitoring modules, and transmits the collected data to the data fusion unit. The federated learning heterogeneous data fusion unit is connected to the multi-source data acquisition unit and the metabolic behavior assessment unit respectively. It is used to receive data transmitted by the multi-source data acquisition unit, perform feature alignment and fusion processing on the heterogeneous data through the federated learning heterogeneous data fusion model, generate a unified data feature set, and transmit the unified data feature set to the metabolic behavior assessment unit. The dairy cow metabolic behavior multimodal assessment unit is connected to the federated learning heterogeneous data fusion unit and the risk threshold calculation unit, respectively. It is used to receive a unified data feature set, perform weighted fusion and correlation analysis on the multi-dimensional features related to dairy cow metabolism through the dairy cow metabolic behavior multimodal fusion model, output the real-time metabolic behavior assessment results of dairy cows, and transmit the assessment results to the risk threshold calculation unit. The heat stress risk threshold calculation unit is connected to the dairy cow metabolic behavior multimodal assessment unit and the risk level judgment unit. It is used to receive real-time metabolic behavior assessment results, combine them with historical heat stress data, and use the heat stress risk dynamic threshold calculation model to dynamically assign weights to risk influencing factors and perform threshold iteration calculations to determine the heat stress risk threshold range. The threshold range is then transmitted to the risk level judgment unit. The heat stress risk level judgment unit is connected to the heat stress risk threshold calculation unit and the early warning signal transmission unit respectively. It is used to receive the real-time metabolic behavior assessment results and the heat stress risk threshold range, judge the current heat stress risk level of the dairy cow through comparison and analysis, and transmit the risk level to the early warning signal transmission unit. The early warning signal processing and transmission unit is connected to the risk level judgment unit and the display terminal and management terminal of the multi-source data intelligent early warning platform. It is used to transmit the risk level to the display terminal in the form of data signals, and generate corresponding early warning signals according to the risk level and transmit them to the management terminal for display and transmission of early warning information.

Images