Animal husbandry environment monitoring and early warning system based on big data platform

Abstract

The invention discloses an animal husbandry environment monitoring and early warning system based on a big data platform, and relates to the technical field of intelligent animal husbandry environment monitoring and early warning systems, and the system comprises an acquisition module which acquires farm environment data and animal behavior image data and uploads the data to an edge computing node, a preprocessing module, a data processing module and a data processing module, the edge computing node is used for preprocessing the farm environment data and the animal behavior image data, the fusion module is used for fusing the preprocessed farm environment data and animal behavior image data to form a multi-modal data set, and the prediction module is used for establishing a health and risk prediction model. Obtaining an animal health condition assessment score and a potential risk level score according to the multi-modal data set; according to the method, data from different sources are integrated, so that health and risk assessment is more comprehensive and accurate, the spatial relevance of the data is adjusted by using a geographic weighting function, and the representativeness of the data is further improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent livestock environment monitoring and early warning systems, and particularly to a livestock environment monitoring and early warning system based on a big data platform. Background Art

[0002] With the rapid development of Internet of Things, big data platform and artificial intelligence technologies, the livestock environment monitoring and early warning system has also undergone an important transformation. Traditional livestock farming relies on manual inspections and empirical judgments, making it difficult to achieve real-time monitoring and precise management of animal health conditions and environmental changes.

[0003] Although existing methods have improved the management efficiency of livestock farming to a certain extent, there are still many deficiencies. First, there is a lack of an effective multi-modal data fusion mechanism, resulting in the underutilization of data from different sources. For example, environmental data and animal behavior image data are often processed independently, ignoring the internal connection between the two, thus limiting the accuracy of comprehensive evaluation. Second, when setting health thresholds, fixed standards are usually adopted, which cannot be dynamically adjusted according to the actual situation. This not only easily causes false alarms or missed alarms, but also is difficult to meet the specific needs of different farms. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a livestock environment monitoring and early warning system based on a big data platform to solve the problem of the underutilization of data from different sources due to the lack of an effective multi-modal data fusion mechanism.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: The present invention provides a livestock environment monitoring and early warning system based on a big data platform, which includes: A collection module that collects environmental data and animal behavior image data of a farm and uploads them to an edge computing node; A preprocessing module that preprocesses the environmental data and animal behavior image data of the farm by the edge computing node; A fusion module that fuses the preprocessed environmental data and animal behavior image data of the farm to form a multi-modal data set; A prediction module that establishes a health and risk prediction model and obtains an animal health condition assessment score and a potential risk level score according to the multi-modal data set; A judgment module that sets a health threshold according to the environmental data of the farm, judges the animal health condition assessment score and the potential risk level score, and dynamically adjusts the health threshold; A feedback module that sets a health threshold evaluation link to optimize the health threshold.

[0007] As a preferred solution of the livestock environment monitoring and early warning system based on the big data platform of the present invention, wherein: the farm environment data includes temperature, humidity, ammonia concentration and carbon dioxide concentration; The animal behavior image data refers to the animal behaviors captured by a camera.

[0008] As a preferred solution of the livestock environment monitoring and early warning system based on the big data platform of the present invention, wherein: the edge computing node preprocesses the farm environment data and the animal behavior image data, specifically including the following steps, Preliminarily clean the farm environment data through the edge computing node, remove outliers and noise points, extract the action features of the animal behavior image data, and convert the features into structured data; The edge node uploads the preprocessed farm environment data and structured data to the central server.

[0009] As a preferred solution of the livestock environment monitoring and early warning system based on the big data platform of the present invention, wherein: extracting the features of the animal behavior image data and converting the features into structured data specifically includes the following steps, Apply the YOLOv5 object detection model to perform real-time analysis on the animal behavior images captured by the camera and execute the object detection task; Obtain the confidence of the animal action postures appearing in the image through the classifier of the YOLOv5 object detection model; Set the posture threshold and retain the detection results with a confidence higher than the posture threshold; The detection results with a confidence higher than the posture threshold are the action posture features; Convert the retained action posture features into a structured CSV file format for storage.

[0010] As a preferred solution of the livestock environment monitoring and early warning system based on the big data platform of the present invention, wherein: fuse the preprocessed farm environment data and the animal behavior image data to form a multi-modal data set, Combine the preprocessed farm environment data and all the extracted action posture features respectively to form an environmental data matrix and an action posture feature set; Define a geographic weight function, use GIS tools, and measure the spatial correlation between data points at different positions according to the installation location coordinates of the sensors and cameras; Determine the analysis time period and calculate the weighted sum of the farm environment data and the action posture features at each time point; Adjust the weighted sum using a geographical weight function, solve and summarize the adjusted results at all time points to obtain a multi-modal dataset, and its expression is: ; Among them, represents the multi-modal dataset, represents the environmental data matrix at time , represents the set of action pose features at time , represents the geographical weight function, represents the calculation result of the geographical weight function at time , represents the end time of the time period, represents the start time of the time period.

[0011] As a preferred solution of the livestock environment monitoring and early warning system based on the big data platform according to the present invention, wherein: establish a health and risk prediction model, and obtain an animal health status evaluation score and a potential risk level score according to the multi-modal dataset. The specific steps are as follows: Define the multi-modal dataset as the input of the health and risk prediction model; Divide the multi-modal dataset into a training set and a validation set; Use a convolutional neural network to process the animal pose features in the multi-modal dataset, and at the same time use a long short-term memory network to process the farm environment data in the multi-modal dataset; Define the cross entropy as the loss function, and use the training set to train the health and risk prediction model until the loss on the training set no longer decreases, then the establishment of the health and risk prediction model is completed; Use the validation set to test the health and risk prediction model to obtain an animal health status evaluation score and a potential risk level score.

[0012] As a preferred solution of the livestock environment monitoring and early warning system based on the big data platform according to the present invention, wherein: the steps for obtaining the animal health status evaluation score and the potential risk level score specifically include the following steps: Input the test set into the health and risk prediction model; The health and risk prediction model fuses the outputs of the convolutional neural network and the long short-term memory network through a fully connected layer to form an intermediate feature vector; Introduce a multi-modal attention mechanism to calculate the correlation between different modal data, and its expression is: ; Among them, represents the th modal attention weight, Indicates the importance score of the th modality, Indicates the importance score of the th modality, Indicates the number of modalities; According to the calculated attention weights, the intermediate feature vectors are weighted to obtain weighted feature vectors; Based on the weighted feature vectors, the animal health status assessment score and the risk level score are obtained through two independent fully connected layers respectively; The animal health status assessment score and the risk level score are combined to form a multi-dimensional output vector; The multi-dimensional output vector is a comprehensive score that includes the animal health status assessment score and the risk level score.

[0013] As a preferred solution of the livestock environment monitoring and early warning system based on the big data platform of the present invention, wherein: a health threshold is set according to the farm environment data, and the animal health status assessment score and the potential risk level score are judged, specifically including the following steps, A health threshold is established based on the environment data of the farm; According to the value range of the multi-dimensional output vector located at the health threshold, the health status and potential risks of the animals are judged; When the multi-dimensional output vector is in the abnormal range, the early warning mechanism is immediately triggered to notify the farmer to take measures.

[0014] As a preferred solution of the livestock environment monitoring and early warning system based on the big data platform of the present invention, wherein: the dynamic adjustment of the health threshold specifically includes the following steps, Based on multi-modal data and historical farm data, the environmental state is defined; An action space is defined based on the specific amplitude of the health threshold adjustment; A deep Q network is introduced, a value function is defined, and a reward mechanism is set at the same time to obtain a threshold adjustment scheme, and its expression is: ; Wherein, Represents the value function, Represents the environmental state at the current time , Represents the environmental state at the current time , Represents the time The reward points, Represents the environmental state at the current time The next time, Represents the discount factor, Represents the current time Actions to be executed at the next time; Dynamically adjust the health threshold according to the value function.

[0015] As a preferred solution of the livestock environment monitoring and early warning system based on the big data platform of the present invention, wherein: a health threshold evaluation link is set up to optimize the health threshold, which specifically includes the following steps Regularly evaluate the performance indicators under the old and new health thresholds, and decide whether to continue using the new health threshold according to the evaluation results; When the false negatives and false positives increase, roll back to the previous version and readjust the threshold adjustment plan.

[0016] The beneficial effects of the present invention are as follows: integrating data from different sources, such as environmental parameters and animal behavior images, to construct a multimodal dataset that comprehensively reflects the actual situation of the farm. The multimodal dataset not only covers information on environmental changes, but also includes the behavior patterns of animals, making health and risk assessment more comprehensive and accurate. Using the geographical weight function to adjust the spatial correlation of data further improves the representativeness of the data; in addition, the dynamic adjustment mechanism ensures that the threshold is always close to the actual needs, minimizing the occurrence of false positives and false negatives; the instant trigger warning mechanism can quickly notify the farmers when abnormal situations occur, improving the speed and efficiency of emergency response. Brief Description of the Drawings

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for description in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0018] Figure 1 It is a module diagram of the livestock environment monitoring and early warning system of the big data platform in Embodiment 1.

[0019] Figure 2 It is a decision diagram of the health and risk prediction model in Embodiment 1. Detailed Embodiments

[0020] To make the above objects, features and advantages of the present invention more obvious and understandable, the detailed embodiments of the present invention will be described in detail below with reference to the drawings of the specification.

[0021] Many specific details are set forth in the following description in order to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described herein. Those skilled in the art can make similar promotions without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0022] Second, the "one embodiment" or "embodiment" mentioned herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation manner of the present invention. The "in one embodiment" that appears in different places in this specification does not all refer to the same embodiment, nor is it an individual or alternative embodiment that is mutually exclusive with other embodiments.

[0023] Embodiment 1, referring to Figure 1 and Figure 2 , is the first embodiment of the present invention. This embodiment provides a livestock environment monitoring and early warning system based on a big data platform, including the following steps: The acquisition module collects the environment data and animal behavior image data of the farm and uploads them to the edge computing node.

[0024] Deploy various types of sensors inside the farm to collect environment data such as temperature, humidity, ammonia concentration, and carbon dioxide concentration, and install high-definition network cameras for capturing animal behavior videos. According to the specific layout of the farm and the livestock activity areas, reasonably arrange the positions of various types of sensors and high-definition network cameras to ensure that data acquisition fully covers key areas. For large farms, a grid layout is adopted to ensure the comprehensiveness and accuracy of data.

[0025] Furthermore, by collecting multiple environmental parameters simultaneously, a comprehensive environmental overview can be provided to accurately understand the actual situation of the farm. By monitoring animal behavior through video analysis technology, abnormal behaviors such as staying still for a long time, sneezing frequently, and walking unsteadily can be identified, which may be early signs of diseases.

[0026] After collection, upload to the edge computing node.

[0027] The preprocessing module preprocesses the farm environment data and animal behavior image data by the edge computing node; Use the Pandas library to initially clean the farm environment data collected by various types of sensors; Define the outlier range. For example, a temperature exceeding 40°C or below 0°C is regarded as an outlier, a humidity exceeding 90% or below 10% is regarded as an outlier, an ammonia concentration exceeding 20 ppm is regarded as an outlier, and a carbon dioxide concentration exceeding 1000 ppm is regarded as an outlier. Remove these outliers and noise points to ensure the reliability of the data.

[0028] Input the animal image frames captured by the camera into the YOLOv5 model to perform object detection tasks. The YOLOv5 model will generate multiple detection frames, and each detection frame corresponds to a possible action posture (such as eating, drinking, standing, lying down, etc.).

[0029] Calculate the possibility that the action posture appears in the given image through the classifier of the YOLOv5 object detection model, that is, the confidence of each detection box, and its expression is: ; Among them, represents the possibility that the action posture appears in the given image, which is obtained by the classifier.

[0030] Set a confidence threshold (for example, 0.7), and only retain the animal images higher than this threshold to ensure the accuracy of the final output.

[0031] Convert the animal images with confidence higher than the threshold into a structured CSV file format for storage, which is the action posture feature.

[0032] Each detection result contains fields such as timestamp, action posture category, confidence, and coordinate position.

[0033] The fusion module fuses the preprocessed farm environment data and animal behavior image data to form a multi-modal dataset.

[0034] Through time series processing, organize the farm environment data at each time point into a time series format, construct the environmental data at each time point into a multi-dimensional array, and form an environmental data matrix.

[0035] Align the action posture features according to the timestamp to ensure the time consistency with the environmental data, and construct the action posture features at each time point into a feature vector to form an action posture feature set.

[0036] Two independent but related datasets are formed, the environmental data matrix and the action posture feature set.

[0037] Define the geographic weight function. Using GIS tools, according to the installation location coordinates of sensors and cameras, measure the spatial correlation between data points at different positions, and its expression is: ; Among them, represents the calculation result of the geographic weight function at time , represents the distance between two positions measured at time , represents the standard deviation of the geographic location distance, which is used to control the attenuation speed of the geographic weight function.

[0038] Furthermore, by introducing the geographic weight function, the spatial relationship between data points at different positions can be more accurately reflected, avoiding misjudgment caused by distance differences.

[0039] Determine the time period for analysis (usually the time period you want to analyze, such as one day, one week, or a longer time). At each time point, obtain the corresponding environmental data matrix and the set of action posture features , and calculate the weighted sum of the environmental data and the action posture features; Use the geographical weight function to adjust the weighted sum, and solve and summarize the adjustment results at all time points to obtain a multi-modal data set, and its expression is: ; Among them, represents the multi-modal data set, represents the environmental data matrix at time , represents the set of action posture features at time , represents the geographical weight function, represents the calculation result of the geographical weight function at time .

[0040] Furthermore, the multi-modal data set integrates environmental data and animal behavior characteristics, and at the same time considers time and space factors, providing more comprehensive and accurate evaluation results.

[0041] Prediction module, establish a health and risk prediction model, and obtain the animal health status evaluation score and potential risk level score according to the multi-modal data set.

[0042] Define the multi-modal data set as the input of the health and risk prediction model; Divide the multi-modal data set into a training set and a validation set; Use a convolutional neural network to process the animal behavior characteristics in the multi-modal data set, and its expression is: ; Among them, represents the output of the convolutional neural network, represents the animal behavior characteristic image, represents the activation function, represents the convolutional weight matrix, represents the convolutional bias term, represents the convolutional operator.

[0043] At the same time, use a long short-term memory network to process the farm environment data in the multi-modal data set, and its expression is: ; Among them, represents the hidden state vector of the LSTM at time , Indicates time The hidden state vector, which is the hidden state vector of the previous time Represents the recursive function of the long short-term memory network

[0044] The hidden state contains information passed down from previous time steps and new information at the current time step, which is the key for LSTM to capture long-term dependencies in time series data

[0045] Furthermore, the convolutional neural network can extract meaningful action and pose features from complex image data, improving the ability to understand animal behavior. The long short-term memory network is good at processing time series data with long-term dependencies and is suitable for capturing the trends and patterns of environmental parameters changing over time. By combining the convolutional neural network and the long short-term memory network, the health and risk prediction model can more comprehensively understand the complex relationships between multimodal data and improve the overall prediction accuracy

[0046] Define the cross-entropy as the loss function and use the training set to train the health and risk prediction model until the loss on the training set no longer decreases, then the establishment of the health and risk prediction model is completed. Its expression is ; Among them Represents the loss function Represents the true score Represents the predicted score Represents the number of categories Represents the th true score Represents the th predicted score

[0047] Use the validation set to test the health and risk prediction model to obtain the animal health status assessment score and the potential risk level score

[0048] Input the test set into the health and risk prediction model The health and risk prediction model fuses the outputs of the convolutional neural network and the long short-term memory network through a fully connected layer to form an intermediate feature vector. Its expression is ; Among them Represents the intermediate feature vector Represents the weight matrix of the fully connected layer Represents the output of the convolutional neural network for the animal behavior feature image Represents the input animal behavior feature image Represents the output obtained by the long short-term memory network in processing the time series of environmental parameters, that is, the final hidden state Denotes the feature concatenation operation; Introduce a multi-modal attention mechanism to calculate the correlation between different modal data, and its expression is: ; Among them, Denotes the attention weight of the th modality, Denotes the importance score of the th modality, which is used to calculate the attention weight of the current modality, Denotes the importance score of the th modality, which is used to traverse all modalities and calculate the sum of their importance scores, Denotes the number of modalities; According to the calculated attention weights, weight the intermediate feature vectors to obtain weighted feature vectors, and its expression is: ; Among them, Denotes the weighted feature vector, Denotes the part of the feature vector corresponding to each modality; Furthermore, the weighted feature vector incorporates the importance and relevance between modalities, making the content more abundant.

[0049] Based on the weighted feature vector, obtain the animal health status assessment score and the risk level score respectively through two independent fully connected layers, and use the activation function to map the results to between [0,1], and its expression is: ; ; Among them, Denotes the animal health status score, Denotes the risk level score, Denotes the activation function, Denotes the weight matrix of the health status score, Denotes the bias term of the health status score, Denotes the weight matrix of the risk level score, Denotes the bias term of the risk level score, Denotes the intermediate feature vector.

[0050] Combine the animal health status assessment score and the risk level score to form a multi-dimensional output vector, and its expression is: ; Among them, Denotes the multi-dimensional output vector, which is the comprehensive prediction score output by the health and risk prediction model.

[0051] Further explanation, the multi-dimensional output vector represents the comprehensive score of the animal's health status assessment result and potential risk reminder.

[0052] The judgment module sets a health threshold based on the farm environment data, judges the animal health status assessment score and the potential risk level score, and dynamically adjusts the health threshold; Establish a health threshold based on the environment data of the farm; Obtain the multi-dimensional output vector through the health and risk prediction model; Judge the animal's health status and potential risks according to the value range of the multi-dimensional output vector located at the health threshold; The health threshold is divided into the animal health status assessment threshold and the potential risk level assessment threshold.

[0053] Since the value ranges of both the animal health status assessment score and the potential risk level assessment score are [0, 1], the low-risk area will be uniformly set as [0.8, 1], the medium-risk area as [0.5, 0.8], and the high-risk area as [0, 0.5].

[0054] When there is a contradictory situation where the animal health status assessment score indicates low risk, but the potential risk level assessment score indicates medium risk or high risk, the solution is as follows. If both the animal health status assessment score and the potential risk level assessment score are in the low-risk range, the overall assessment is good health; if one of the scores is in the medium-risk range and the other is in the low-risk range, a reminder will be issued, suggesting that the farmer further observe or take preventive measures; if any one of the scores is in the high-risk range, an alarm will be immediately triggered to notify the farmer to take emergency measures.

[0055] Define the environmental state based on the multi-modal data set and historical farm data; Synchronize the timestamps of the multi-modal data set and historical farm data. After synchronization, identify long-term trends and periodic changes through time series analysis, and integrate them into the environmental state ; The environmental state is used to describe the current environmental conditions of the farm, including all factors affecting animal health and growth, such as environmental parameters like temperature, humidity, ammonia concentration, carbon dioxide concentration, etc., as well as the behavioral characteristics of animals; Further explanation, by integrating multi-modal data and historical farm data and defining it as the environmental state, long-term trends and periodic changes can be identified, enhancing the understanding of the environmental state.

[0056] Define the action space based on the specific amplitude of threshold adjustment; The action space contains a series of possible operations, such as increasing the health threshold, decreasing the health threshold, and keeping the health threshold unchanged; it also includes the amplitude of the operation, for example, increasing the health threshold by 0.1 and decreasing the health threshold by 0.2.

[0057] Introduce a deep Q-network, define the value function, and at the same time set up a reward mechanism to obtain a threshold adjustment scheme, and its expression is: ; Among them, represents the value function, represents the current environmental state, represents the current action space, represents the immediate reward score, represents the next environmental state, represents the discount factor, represents the next action to be executed; The reward mechanism is specifically that if an abnormal situation is correctly detected, a positive reward is given, otherwise a negative reward is given.

[0058] Furthermore, through the deep Q-network, the threshold adjustment can be automatically learned without manual intervention, and the reward mechanism selects those actions that can maximize the long-term benefits when adjusting the threshold, improving the adaptive ability.

[0059] Dynamically adjust the health threshold according to the value function. For example, the low-risk area is adjusted to [0.9, 1].

[0060] The evaluation module sets up a health threshold evaluation session to optimize the health threshold.

[0061] Regularly evaluate the performance of the currently used health threshold (new threshold) and the health threshold of the previous version (old threshold) in actual applications.

[0062] Specifically, extract the multi-dimensional output vectors and the actual results of the corresponding animal health status and risk levels over a period of time. Evaluate the accuracy rate, compare the performance indicators under the new and old health thresholds, and evaluate whether the new health threshold is significantly better than the old health threshold.

[0063] Furthermore, based on the extracted multi-dimensional output vectors and the actual data and performance indicators of the corresponding animal health status and risk levels over a period of time, ensure that the health threshold adjustment is based on facts and actual data-driven, rather than by experience or guesswork, and through regular evaluation, the threshold setting can be continuously optimized to improve its adaptability and accuracy.

[0064] Define clear evaluation criteria. For example, the new health threshold must be better than the old health threshold by a certain percentage (such as 5%) in the main performance indicators to be considered valid.

[0065] If the new health threshold meets the standard, it continues to be used; if not, other adjustment schemes are considered or rolled back to the old health threshold.

[0066] Monitor the operation in real time. Once it is found that the number of missed reports and false reports increases, immediately activate the rollback mechanism, restore to the health threshold setting of the previous version, analyze the cause of the problem, and readjust the threshold scheme. Furthermore, the timely rollback mechanism can take measures quickly when problems are discovered, avoiding the expansion of losses.

[0067] In summary, the present invention integrates data from different sources, such as environmental parameters and animal behavior images, to construct a multimodal dataset that comprehensively reflects the actual situation of the farm. The multimodal dataset not only covers information on environmental changes but also includes the behavior patterns of animals, making health and risk assessment more comprehensive and accurate. By using a geographical weight function to adjust the spatial correlation of the data, the representativeness of the data is further improved. In addition, the dynamic adjustment mechanism ensures that the threshold is always close to the actual needs, minimizing the occurrence of false reports and missed reports. The instant trigger warning mechanism can quickly notify the farmers when abnormal situations occur, improving the speed and efficiency of emergency response.

[0068] Example 2, referring to Table 1, is the second example of the present invention. To further verify the technical solution of the present invention, experimental simulation data of the livestock environment monitoring and warning system based on the big data platform are given.

[0069] To verify the effectiveness of the livestock environment monitoring and warning system based on the big data platform, a large dairy farm was selected as the test object. The farm has 500 dairy cows and is equipped with multiple cameras and sensors for real-time collection of environmental data and animal behavior image data. The test period was from January 1, 2024, to March 31, 2024, for a total of three months.

[0070] Test equipment and settings, sensor network: Temperature, humidity, ammonia concentration, and carbon dioxide concentration sensors were installed in the farm to ensure that each area is covered.

[0071] Camera arrangement: High-definition cameras were installed at key locations (such as drinking areas, feeding areas, and resting areas) to capture the behavior images of dairy cows.

[0072] Edge computing nodes: Multiple edge computing nodes were deployed to be responsible for preliminary processing of the collected data and uploading it to the central server.

[0073] Preprocessing module: The edge computing node preliminarily cleans the collected environmental data, removing outliers and noise points. For animal behavior image data, the YOLOv5 object detection model is applied for real-time analysis to extract the action posture features of livestock. The specific steps are as follows: Data cleaning: Through statistical analysis and threshold filtering, outliers in the environmental data are removed to ensure data quality.

[0074] Feature extraction: Use the YOLOv5 model to perform real-time analysis on the images to obtain the confidence level of each action posture. Set the posture threshold to 0.7, retain the detection results above this threshold, and convert these features into a structured CSV file format for storage.

[0075] Fusion module: The preprocessed environmental data and the set of action posture features are fused through a geographical weight function to form a multi-modal dataset. The specific steps include: Data combination: Combine the preprocessed environmental data and action posture features respectively to form an environmental data matrix and a set of action posture features.

[0076] Geographical weight function: Use GIS tools to measure the spatial correlation between data points at different locations based on the installation location coordinates of sensors and cameras. Define the geographical weight function.

[0077] Weighted sum calculation: Determine the analysis time period, calculate the weighted sum of the environmental data and action posture features at each time point, and adjust it using the geographical weight function. Finally, summarize the adjusted results at all time points to obtain the multi-modal dataset.

[0078] Prediction module: Based on the multi-modal dataset, a health and risk prediction model is established. The specific steps include: Data division: Divide the multi-modal dataset into a training set and a validation set with a ratio of 8:2.

[0079] Model training: Use a convolutional neural network (CNN) to process the animal behavior features in the multi-modal dataset, and use a long short-term memory network (LSTM) to process the environmental parameters. Define the cross-entropy as the loss function, and use the training set to train the model until the loss on the training set no longer decreases.

[0080] Model testing: Use the validation set to test the health and risk prediction model to obtain the animal health status assessment score and the potential risk level score.

[0081] Judgment module: Set a health threshold based on the farm environmental data, and judge the animal health status assessment score and the potential risk level score. The specific steps include: Threshold setting: Dynamically set the health threshold based on historical data and current environmental conditions.

[0082] Status judgment: Determine the health status and potential risks of the animal according to the value range of the multi-dimensional output vector within the healthy threshold.

[0083] Early warning mechanism: When the multi-dimensional output vector is within the abnormal range, immediately trigger the early warning mechanism to notify the farmer to take measures.

[0084] The evaluation module regularly evaluates the performance indicators under the new and old thresholds, and decides whether to continue using the new threshold according to the evaluation results. If the false negative and false positive rates increase, roll back to the previous version and readjust the threshold adjustment plan.

[0085] Specifically as shown in Table 1 below: Table 1 Invention comparison table Through the analysis of the above table data, it can be clearly seen that the livestock environment monitoring and early warning system based on the big data platform has significant advantages compared with the prior art.

[0086] In terms of the false negative rate, the false negative rate of the present invention is significantly lower than that of the comparative invention. For example, when the temperature is 20.3 °C, the humidity is 72%, the ammonia concentration is 11 ppm, and the carbon dioxide concentration is 458 ppm, the false negative rate of the present invention is 1.8%, while that of the comparative invention is 10.5%. This shows that the present invention can detect potential problems more timely and reduce the risks brought by false negatives.

[0087] In terms of the false positive rate, the false positive rate of the present invention is also much lower than that of the comparative invention. For example, under the same environment, the false positive rate of the present invention is 2.6%, while that of the comparative invention is 7.2%. This shows that the present invention can judge the actual risks more accurately and reduce unnecessary early warning interference.

[0088] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A livestock environment monitoring and early warning system based on a big data platform, characterized in that: Including, A collection module that collects the environmental data of the farm and the animal behavior image data, and uploads them to the edge computing node; A preprocessing module that preprocesses the environmental data of the farm and the animal behavior image data by the edge computing node; A fusion module that fuses the preprocessed environmental data of the farm and the animal behavior image data to form a multi-modal data set; A prediction module that establishes a health and risk prediction model, and obtains the animal health status assessment score and the potential risk level score according to the multi-modal data set; A judgment module that sets a health threshold according to the environmental data of the farm, judges the animal health status assessment score and the potential risk level score, and dynamically adjusts the health threshold; An evaluation module that sets a health threshold evaluation link to optimize the health threshold.

2. The livestock environment monitoring and early warning system based on a big data platform according to claim 1, characterized in that: The environmental data of the farm includes temperature, humidity, ammonia concentration and carbon dioxide concentration; The animal behavior image data refers to the animal behaviors captured by the camera.

3. The livestock environment monitoring and early warning system based on the big data platform according to claim 2, characterized in that: The edge computing node preprocesses the environmental data of the farm and the animal behavior image data, which specifically includes the following steps. Through the edge computing node, the environmental data of the farm is preliminarily cleaned, outliers and noise points are removed, the action features of the animal behavior image data are extracted, and the features are converted into structured data; The edge node uploads the preprocessed environmental data of the farm and the structured data to the central server.

4. The livestock environment monitoring and early warning system based on the big data platform according to claim 3, wherein: The features of the animal behavior image data are extracted and the features are converted into structured data, which specifically includes the following steps. Apply the YOLOv5 object detection model to perform real-time analysis on the animal behavior images captured by the camera and execute the object detection task; Obtain the confidence of the animal action posture appearing in the image through the classifier of the YOLOv5 object detection model; Set the posture threshold to retain the detection results with a confidence higher than the posture threshold; The detection results with a confidence higher than the posture threshold are the action posture features; Convert the retained action posture features into a structured CSV file format for storage.

5. The livestock environment monitoring and early warning system based on the big data platform according to claim 4, characterized in that: Fusing the preprocessed environmental data of the farm and the animal behavior image data to form a multi-modal data set, which specifically includes the following steps. Combine the preprocessed environmental data of the farm and all the extracted action posture features respectively to form an environmental data matrix and an action posture feature set; Define a geographical weight function, use GIS tools, and measure the spatial correlation between data points at different positions according to the installation position coordinates of the sensors and cameras; Determine the analysis time period, and calculate the weighted sum of the environmental data of the farm and the action posture features at each time point; Use the geographical weight function to adjust the weighted sum, and solve and summarize the adjustment results at all time points to obtain a multi-modal data set, and its expression is: ； Among them, represents a multi-modal data set, represents the environmental data matrix at time ; represents the set of action posture features at time ; represents the geographic weight function, represents the calculation result of the geographic weight function at time ; represents the end time of the time period, represents the start time of the time period. It should be noted that there seems to be a semicolon missing in the original Chinese text for better logical connection in the English translation. I added semicolons in the appropriate places in the translation for better readability. If this is not allowed according to your strict requirements, you can adjust it back according to the original text's punctuation usage.

6. The livestock environment monitoring and early warning system based on the big data platform according to claim 5, characterized in that: Establish a health and risk prediction model, and obtain the animal health status assessment score and the potential risk level score according to the multi-modal data set, which specifically includes the following steps. Define the multi-modal data set as the input of the health and risk prediction model; Divide the multi-modal data set into a training set and a validation set; Use a convolutional neural network to process the animal posture features in the multi-modal data set, and use a long short-term memory network to process the environmental data of the farm in the multi-modal data set; Define cross-entropy as the loss function and use the training set to train the health and risk prediction model until the loss on the training set no longer decreases, then the establishment of the health and risk prediction model is completed; Use the validation set to test the health and risk prediction model to obtain the animal health status assessment score and the potential risk level score.

7. The livestock environment monitoring and early warning system based on the big data platform according to claim 6, characterized in that: The obtaining of the animal health status assessment score and the potential risk level score specifically includes the following steps: Input the test set into the health and risk prediction model; The health and risk prediction model fuses the outputs of the convolutional neural network and the long short-term memory network through a fully connected layer to form an intermediate feature vector; Introduce a multi-modal attention mechanism to calculate the correlation between different modal data, and its expression is: ； Among them, represents the attention weight of the th modality, represents the importance score of the th modality, represents the importance score of the th modality, represents the number of modalities; According to the calculated attention weights, weight the intermediate feature vector to obtain a weighted feature vector; Based on the weighted feature vector, obtain the animal health status assessment score and the risk level score respectively through two independent fully connected layers; Combine the animal health status assessment score and the risk level score to form a multi-dimensional output vector; The multi-dimensional output vector is a comprehensive score that includes the animal health status assessment score and the risk level score.

8. The livestock environment monitoring and early warning system based on the big data platform according to claim 7, characterized in that: Set a health threshold according to the farm environment data and judge the animal health status assessment score and the potential risk level score, specifically including the following steps: Establish a health threshold based on the farm environment data; Judge the animal health status and potential risks according to the value range of the multi-dimensional output vector located at the health threshold; When the multi-dimensional output vector is in the abnormal range, immediately trigger the warning mechanism to notify the farmer to take measures.

9. The livestock environment monitoring and early warning system based on the big data platform according to claim 8, characterized in that: The dynamic adjustment of the health threshold specifically includes the following steps: Define the environmental state based on the multi-modal data set and the historical farm data; Define the action space based on the specific amplitude of the health threshold adjustment; Introduce a deep Q-network, define the value function, and at the same time set the reward mechanism to obtain the threshold adjustment scheme, and its expression is: ； Among them, represents the value function, represents the environmental state at the current time , represents the action space at the current time , represents the reward points at the current time , represents the environmental state at the next time when the current time represents the discount factor, represents the current time and the action to be executed at the next time; Dynamically adjust the health threshold according to the value function.

10. The livestock environment monitoring and early warning system based on the big data platform according to claim 9, characterized in that: Set up a health threshold evaluation link to optimize the health threshold, specifically including the following steps: Regularly evaluate the performance indicators under the new and old health thresholds and decide whether to continue using the new health threshold according to the evaluation results; When the missed reports and false alarms increase, roll back to the previous version and readjust the threshold adjustment scheme.

Images