Artificial intelligence-based animal digital management method and system

Abstract

This invention discloses an artificial intelligence-based method and system for digital animal management, relating to the field of intelligent livestock control technology. The method includes collecting multi-source monitoring data from animal video images, animal sounds, individual identification information, and environmental sensors within a designated breeding area. Simultaneous analysis of the animal video image and sound data is performed to identify individual animals and extract behavioral feature sequences and environmental correlation features. Individual identification information is read and decoded to establish a mapping relationship between individual animal identity and behavioral feature sequences. By fusing behavioral feature sequences, environmental correlation features, and environmental sensor data, multi-dimensional time-series status data is constructed. This data is input into a pre-trained evaluation model built using machine learning, outputting real-time animal health status scores and potential anomaly warnings. Individual identification is matched to generate feeding and management instructions, which are then issued and executed. This solution achieves collaborative analysis and intelligent evaluation of multi-source data, enabling precise individualized and digital management of farmed animals.

Description

Technical Field

[0001] This invention belongs to the field of intelligent livestock management technology, specifically an artificial intelligence-based digital animal management method and system. Background Technology

[0002] In large-scale animal farming, the industry generally adopts a decentralized data collection approach, collecting single-category data such as video images, sound signals, individual identifiers, and environmental sensors. These data are stored and analyzed independently, without a unified data processing system. Daily animal behavior observation and health status assessment have long relied on subjective human experience, lacking an intelligent data analysis architecture to support digital management processes.

[0003] Current technologies cannot achieve simultaneous analysis and processing of animal video and audio data; they can only analyze single data sources individually, failing to simultaneously identify individual animals and extract behavioral and environmental features. Animal identification information is only used for basic record-keeping and cannot be read or decoded, thus failing to establish a correlation between individual identity and behavioral feature sequences. Various monitoring data are fragmented and cannot be integrated and reconstructed to form continuous, complete, multi-dimensional time-series status data. Traditional methods lack dedicated assessment models based on machine learning, making it impossible to quantify health status scores and provide early warnings of potential anomalies. The use of a uniform standard in husbandry management prevents the development of differentiated control measures based on the actual condition of each animal.

[0004] The industry needs to break through the limitations of independent processing of single-source data, open up channels for collaborative analysis of multi-source data, improve the correlation mechanism between individual identity and behavioral data, rely on intelligent models to achieve automatic assessment of health status, and complete refined digital breeding management. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; Therefore, this invention proposes an artificial intelligence-based digital management method for animals, including: Acquire multi-source monitoring data within a designated breeding area, including animal video image data, animal sound data, animal individual identification information, and environmental sensor data; The animal video image data and animal sound data are analyzed synchronously to identify individual animals and extract the behavioral feature sequences and environmental features of the individual animals. The individual animal identification information is read and decoded to establish a mapping relationship between the individual animal identity and the behavioral feature sequence; The behavioral feature sequence, the environmental correlation features, and the environmental sensor data are fused to form multi-dimensional time-series state data. The multi-dimensional time-series state data is input into a pre-trained animal health status assessment model, which is built based on machine learning algorithms and outputs the real-time health status score and potential abnormality warning of the individual animal. Based on the real-time health status score and potential abnormality warning, combined with the individual animal identity, targeted individualized feeding and management instructions are generated and issued for execution.

[0006] Furthermore, the animal video image data and animal sound data are analyzed synchronously to identify individual animals and extract behavioral feature sequences and environmental association features of the individual animals, including: The animal video image data is subjected to target detection and tracking processing to identify the outline of each individual animal in the video image frame, and cross-frame correlation tracking is performed on each identified individual animal outline in consecutive video image frames to form the movement trajectory of the individual animal. The animal sound data is subjected to sound source separation and feature extraction to separate sound segments emitted by different animal individuals or groups, and the time-spectral features of the sound segments are extracted. The animal's individual movement trajectory is timestamped with the sound segment, and the movement trajectory segment and sound segment belonging to the same animal individual are associated based on the aligned timestamps. Kinematic features are extracted from the motion trajectory segment, including velocity, acceleration, movement path complexity, and attitude change frequency. The behavior-related acoustic features are extracted from the aligned sound segments, including the frequency distribution of specific calls, the duration of the calls, and the pattern of call intervals. Image features related to the animal's surrounding environment are extracted from video image frames. These image features include the distance to neighboring animals and the location information relative to the feeding trough or watering point. These image features related to the animal's surrounding environment are used as the environmental association features.

[0007] Further, the individual animal identification information is read and decoded to establish a mapping relationship between the individual animal's identity and the behavioral characteristic sequence, including: The individual animal identification information includes the encoded information carried by the animal's ear tag, collar tag, or biometric identifier; The animal individual identification information is read and decoded by an RFID reader, image recognition device, or biometric identification device to obtain a unique animal individual identification code; When the outline of an individual animal is identified through target detection and tracking, the individual animal identification information is searched and read within a preset range of the outline. If the reading is successful, the read animal individual identification code will be bound to the animal individual outline and the movement trajectory formed by subsequent tracking; If the animal cannot be read directly, the animal's individual biometrics are matched with a pre-established biometric database, or the animal's individual movement trajectory is correlated with the behavior patterns of known individuals to indirectly infer and bind the animal's individual identity code.

[0008] Furthermore, the behavioral feature sequence, the environmental correlation features, and the environmental sensor data are fused to form multi-dimensional time-series state data, including: From the behavioral feature sequence, statistics of the kinematic features and the acoustic features related to the behavior are extracted according to a fixed time window to form a behavioral time-series vector; From the environmental association features, the statistics of the interaction between individual animals and key environmental elements within each time window are extracted to form an environmental interaction time series vector; Data on temperature, humidity, light intensity, and ammonia concentration within the same time window are extracted from the environmental sensor data to form a time-series vector of environmental parameters; The behavior time sequence vector, the environmental interaction time sequence vector, and the environmental parameter time sequence vector are aligned and concatenated according to timestamps to form the multi-dimensional time sequence state data; The multi-dimensional time-series state data is imputed for missing values ​​and standardized to form a standardized multi-dimensional time-series sequence.

[0009] Furthermore, the multi-dimensional time-series state data is input into a pre-trained animal health status assessment model, which is constructed based on a machine learning algorithm and outputs the real-time health status score and potential abnormality warning for the individual animal, including: The animal health status assessment model includes a feature encoder, a temporal pattern learning network, and a status assessment head. The feature encoder performs dimensionality reduction and deep feature extraction on the input multi-dimensional temporal state data, and outputs a high-dimensional feature sequence. The temporal pattern learning network is constructed based on a recurrent neural network or a self-attention mechanism. It models the high-dimensional feature sequence, captures the dependencies and evolution patterns of features in the time dimension, and outputs a comprehensive feature vector containing temporal context information. The status assessment head includes a health score output layer and an abnormality warning output layer. The health score output layer maps the comprehensive feature vector into a continuous score representing the health level. The abnormality warning output layer analyzes the comprehensive feature vector and outputs a potential abnormality warning describing the type and severity of the abnormality when the detected pattern deviates from the preset normal behavior and physiological pattern library.

[0010] Furthermore, based on the real-time health status score and potential abnormality warning, combined with the animal's individual identity, targeted individualized feeding and management instructions are generated, including: Establish a husbandry and management knowledge base that includes animal breeds, growth stages, historical health records, and management knowledge and rules; Based on the individual animal's identity, the basic management parameters of that individual animal are retrieved from the animal husbandry and management knowledge base; The real-time health status score is compared with a preset score threshold, and the type and content of the potential abnormality warning are analyzed. When the real-time health status score is lower than the preset score threshold or there is a specific type of potential abnormality warning, a refined intervention decision process is triggered. The refined intervention decision-making process is based on the basic management parameters of the individual animal, the current health status score, abnormal warning information, and the knowledge rules in the feeding management knowledge base. Through rule reasoning or machine learning decision-making models, it generates at least one specific instruction, including adjusting feed formula, adjusting feeding amount, adjusting feeding time, suggesting isolation and observation, suggesting vaccination, or suggesting veterinary examination, as the individualized feeding management instruction.

[0011] Furthermore, the temporal pattern learning network is constructed based on a recurrent neural network or a self-attention mechanism to model the high-dimensional feature sequence, including: When using a time-series pattern learning network based on a recurrent neural network, the recurrent neural network is a bidirectional long short-term memory network. This network processes the high-dimensional feature sequence step by step, and at each time step, it combines past and future contextual information to update its internal hidden state. Finally, it aggregates the final hidden states of all time steps to generate the comprehensive feature vector. When using a temporal pattern learning network based on a self-attention mechanism, the self-attention mechanism calculates the correlation weights between the features of each time step in the high-dimensional feature sequence and the features of all other time steps. Based on the correlation weights, the features are weighted and summed and information is fused to directly capture long-distance temporal dependencies, outputting a transformed feature sequence. Then, a pooling operation is performed on the transformed feature sequence to generate the comprehensive feature vector.

[0012] Furthermore, the refined intervention decision-making process, based on the animal's basic management parameters, current health status score, abnormal warning information, and knowledge rules in the feeding management knowledge base, generates at least one specific instruction through rule reasoning or machine learning decision-making models. This instruction includes adjusting feed formulation, adjusting feeding amount, adjusting feeding time, suggesting isolation and observation, suggesting vaccination, or suggesting veterinary examination. Retrieve knowledge rule entries from the animal husbandry and management knowledge base that match the current animal breed, growth stage, and the type of potential abnormal warning. Based on the aforementioned knowledge rule entries, an initial set of intervention options is constructed; The initial set of intervention options is evaluated using a machine learning decision model. The inputs of the machine learning decision model include the basic management parameters, real-time health status scores, abnormal warning information, and current environmental parameters. The output is the expected effect score for each intervention option. Select one or more intervention options with the highest expected effect score, and quantify the execution parameters of the intervention options based on the real-time health status score and the severity of abnormal warnings. The quantified parameters include adjusting the specific ingredient ratio of the feed formula, adjusting the specific weight of the feed, and adjusting the specific time of the feed. The selected intervention options are combined with their quantified execution parameters to form the specific instructions that can be automatically executed by the feeding management system.

[0013] Furthermore, target detection and tracking processing is performed on the animal video image data to identify the outline of each individual animal in the video image frame, including: The target detection model based on a deep convolutional neural network is used to perform frame-by-frame analysis on the input animal video image data. The target detection model outputs the bounding box coordinates and confidence score of each individual animal in the image frame. The non-maximum suppression algorithm is used to filter overlapping bounding boxes, retaining the most likely single bounding box for each animal individual as the initial animal individual outline; In continuous video image frames, a tracking algorithm based on appearance features and motion features is used to match and associate the outline of an animal in the current frame with the track of an animal in the previous frame. The appearance features are in-contour image features extracted by a convolutional neural network, and the motion features are motion vectors predicted based on the bounding box positions of the previous and next frames. For newly emerging animal individual silhouettes that fail to be successfully matched and associated, create new tracking tracks; For animal trajectories that are continuously lost, the tracking will be terminated after the preset number of lost frames is exceeded.

[0014] Furthermore, the present invention also includes an artificial intelligence-based animal digital management system, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of the artificial intelligence-based animal digital management method described above.

[0015] Compared with the prior art, the beneficial effects of the present invention are: Animal video and audio data are analyzed and processed synchronously. This process accurately identifies individual animals while simultaneously outputting behavioral characteristic sequences and environmental correlation features. The synchronous analysis mode can collaboratively uncover the inherent correlation between visual and acoustic information, fully reconstructing the correspondence between animal behavior and surrounding environmental conditions. After reading and decoding, the individual animal identification information establishes a correspondence between the animal's identity and behavioral characteristic sequences. This allows discrete behavioral data to be accurately mapped to each individual animal, eliminating the misattribution problem between monitoring data and farmed individuals, and giving animal behavioral data traceable individual attributes.

[0016] Behavioral feature sequences, environmental correlation features, and environmental sensor data are integrated and stitched together to construct multi-dimensional time-series state data with temporal continuity. This time-series integration method can connect monitoring information from different time periods and dimensions, completely reconstructing the state change patterns throughout the entire animal growth and activity process. After the multi-dimensional time-series state data is fed into an evaluation model trained by machine learning algorithms, dynamic health status scores and abnormal risk warnings are generated based on the model's calculations. Combined with individual animal identification, appropriate feeding management and scheduling instructions are generated and issued. Differentiated management methods are matched according to the real-time status of the animals, adapting to the growth and physiological changes of different individuals within the breeding group. Attached Figure Description

[0017] Figure 1 This is a flowchart of the artificial intelligence-based digital animal management method described in this invention; Figure 2 A flowchart for mapping individual animal identities; Figure 3 This is a flowchart for the fusion of multi-dimensional time-series state data. Detailed Implementation

[0018] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] See Figure 1A specific implementation method for an artificial intelligence-based digital animal management approach is as follows: Multi-source monitoring data is acquired within a designated breeding area. This multi-source monitoring data includes animal video image data, animal sound data, animal individual identification information, and environmental sensor data. The animal video image data and animal sound data are simultaneously analyzed to identify individual animals and extract their behavioral feature sequences and environmental association features. The animal individual identification information is read and decoded to establish a mapping relationship between the animal individual's identity and the behavioral feature sequences. The behavioral feature sequences, environmental association features, and environmental sensor data are fused to form multi-dimensional time-series state data. This multi-dimensional time-series state data is input into a pre-trained animal health status assessment model, which is constructed based on a machine learning algorithm and outputs a real-time health status score and potential anomaly warning for each individual animal. Based on the real-time health status score and potential anomaly warning, combined with the individual animal's identity, targeted individualized feeding and management instructions are generated and executed.

[0020] In one embodiment of the present invention, target detection and tracking processing is performed on the animal video image data to identify the contour of each individual animal in the video image frame. Cross-frame correlation tracking is then performed on each identified animal contour in consecutive video image frames to form an animal motion trajectory. A target detection model based on a deep convolutional neural network is used to analyze the input animal video image data frame by frame. The target detection model outputs the bounding box coordinates and confidence score of each individual animal in the image frame. A non-maximum suppression algorithm is used to filter overlapping bounding boxes, retaining the most probable single bounding box for each individual animal as the initial animal contour. In consecutive video image frames, a tracking algorithm based on appearance features and motion features is used to match and correlate the animal contour of the current frame with the tracked animal trajectory of the previous frame. The appearance features are in-contour image features extracted by a convolutional neural network, and the motion features are motion vectors predicted based on the bounding box positions of the preceding and following frames. For newly appearing animal contours that fail to match and correlate successfully, a new tracking trajectory is created. For animal trajectories that are continuously lost, tracking is terminated after exceeding a preset threshold of lost frames.

[0021] The animal sound data undergoes sound source separation and feature extraction to isolate sound segments emitted by different individual animals or groups, and the temporal-spectral features of these sound segments are extracted. The animal's movement trajectory is timestamped with the sound segments, and movement trajectory segments and sound segments belonging to the same animal are associated based on the aligned timestamps. Kinematic features are extracted from the movement trajectory segments, including velocity, acceleration, movement path complexity, and posture change frequency. Behavior-related acoustic features are extracted from the aligned sound segments, including the frequency distribution of specific calls, call duration, and call interval patterns. Image features related to the animal's surrounding environment are extracted from video image frames, including distance to neighboring animals and location information relative to feeding troughs or watering points. These image features related to the animal's surrounding environment are used as the environmental association features.

[0022] In practice, target detection and tracking are performed on the acquired animal video image data. The contours of each individual animal in each video frame are identified, and cross-frame correlation tracking is performed on each identified animal contour across consecutive video frames to form the animal's motion trajectory. A target detection model based on a deep convolutional neural network is used to analyze the input animal video image data frame by frame. The target detection model outputs the bounding box coordinates and confidence score of each individual animal in the video frame. A non-maximum suppression algorithm is used to filter overlapping bounding boxes, retaining each... The most probable single bounding box of an individual animal is used as the initial animal outline. In consecutive video image frames, a tracking algorithm based on appearance and motion features matches and associates the animal outline of the current video image frame with the tracked animal trajectory of the previous video image frame. The appearance features are in-contour image features extracted by a convolutional neural network, and the motion features are motion vectors predicted based on the bounding box positions of the preceding and following video image frames. For newly appearing animal outlines that fail to match and associate, a new tracking trajectory is created. For continuously lost animal trajectories, tracking is terminated after exceeding a preset lost frame threshold. In some embodiments, the target detection model based on a deep convolutional neural network employs a region convolutional neural network or a single-shot multi-frame detector structure, and the bounding box coordinates are represented by the pixel coordinates of the top-left and bottom-right corners of the rectangle in the video image frame. In some embodiments, the tracking algorithm based on appearance and motion features uses the Hungarian algorithm for matching and association, and the motion vectors are predicted by a Kalman filter. Optionally, the preset lost frame threshold is set to a fixed integer value within the range of 30 to 50 frames.

[0023] In specific implementation, the acquired animal sound data undergoes sound source separation and feature extraction to separate sound segments emitted by different animal individuals or groups. The time-spectrum features of these sound segments are then extracted. These time-spectrum features are represented by the amplitude values ​​of the complex spectrum matrix obtained by performing a short-time Fourier transform on each sound segment. In specific implementation, the formed animal individual motion trajectories are timestamped with the separated sound segments. Based on the aligned timestamps, motion trajectory segments and sound segments belonging to the same animal individual are associated. The timestamps represent the acquisition time of each frame of the animal video image data and the acquisition time of each sampling point of the animal sound data, both synchronized using the same system clock. It can be understood that kinematic features are extracted from the associated motion trajectory segments. These kinematic features include velocity, acceleration, path complexity, and attitude change frequency. The path complexity is calculated based on the variance of the turning angles of the lines connecting the trajectory points in the motion trajectory segment. The formula for calculating path complexity is: Where: C represents the movement path complexity. This represents the total number of trajectory points in a motion trajectory segment. Indicates the first The angle between the line segments before and after each trajectory point Indicates all The arithmetic mean. It can be understood that behavior-related acoustic features are extracted from the aligned sound segments. These behavior-related acoustic features include the frequency distribution of a specific call, the duration of the call, and the call interval pattern. The frequency distribution of the specific call is obtained by peak detection on the frequency axis of the time-spectrum characteristics of the sound segment. The duration of the call is the duration during which the continuous energy in the sound segment exceeds a preset energy threshold. The call interval pattern is the difference sequence between the start time points of two adjacent call segments.

[0024] Optionally, image features related to the animal's surrounding environment are extracted from video image frames. These image features include the distance to neighboring animals and the location information relative to the feeding trough or watering point. The image features related to the animal's surrounding environment are used as the environmental association features. The distance to neighboring animals is obtained by calculating the Euclidean distance between the center of the current animal's outline and the centers of the bounding boxes of other animal outlines. The location information relative to the feeding trough or watering point is obtained by pre-marking the coordinate range of the feeding trough area and the watering point area in the video image frame and comparing it with the center of the bounding box of the current animal outline.

[0025] In one embodiment of the present invention, the animal individual identification information includes coded information carried by an animal ear tag, collar tag, or biometric identifier, see below. Figure 2The animal's individual identification information is read and decoded using a radio frequency identification reader, image recognition device, or biometric identification device to obtain a unique animal identification code. When an animal's outline is identified through target detection and tracking processing, the animal's individual identification information is searched and read within a preset range of the outline. If the reading is successful, the read animal identification code is bound to the animal's outline and its subsequent tracking trajectory. If the reading fails directly, the animal's individual identification code is indirectly inferred and bound by matching the biometric features of the outline with a pre-established biometric database, or by analyzing the correlation between the animal's trajectory and the behavioral patterns of known individuals.

[0026] In specific implementation, the animal individual identification information includes the encoded information carried by animal ear tags, collar tags, or biometric tags. The animal individual identification information is read and decoded by an RFID reader, image recognition device, or biometric identification device to obtain a unique animal individual identification code. When the outline of an animal individual is identified through target detection and tracking processing, the animal individual identification information is searched and read within a preset range of the animal individual outline. If the reading is successful, the read animal individual identification code is bound to the animal individual outline and the movement trajectory formed by its subsequent tracking. If the reading cannot be performed directly, the animal individual identification code is indirectly inferred and bound by matching the biometric features of the animal individual outline with a pre-established biometric feature database, or by analyzing the correlation between the animal individual's movement trajectory and the behavior patterns of known individuals.

[0027] In some embodiments, the biometric features are animal facial features or body spot distribution features, and the pre-established biometric database stores biometric templates for each known animal individual. In some embodiments, the behavioral pattern includes a sequence of animal individual movement trajectory features and a sequence of vocal features, and the behavioral pattern of the known individual is pre-extracted and stored from historical monitoring data. Optionally, when matching the biometric features of the animal individual's outline with the pre-established biometric database, the similarity between the biometric features of the animal individual's outline and each biometric template in the biometric database is calculated, and the animal individual identification code corresponding to the biometric template with the highest similarity exceeding a preset similarity threshold is selected as the binding result. The similarity calculation formula is: in: Indicates similarity. This represents the biological feature vector extracted from the individual animal's silhouette. This represents the feature vector of a biometric template in a biometric database. This represents the vector dot product operation. The Euclidean norm of the biological feature vector is represented by the following: This represents the Euclidean norm of the feature vector of the biometric template. It can be understood that when performing correlation analysis between the animal's individual movement trajectory and the behavioral patterns of known individuals, the time series of the animal's individual movement trajectory is dynamically time-normalized and matched with the movement trajectory feature sequences in the behavioral patterns of each known individual. The normalized distance is calculated, and the animal's individual identification code corresponding to the known individual with the smallest normalized distance that is less than a preset distance threshold is selected as the binding result.

[0028] It is understandable that if both biometric matching and behavioral pattern association analysis are used simultaneously, candidate identity codes obtained by the two methods are calculated separately. When the candidate identity codes obtained by the two methods are consistent, the animal individual's identity code is confirmed to be bound. When the candidate identity codes obtained by the two methods are inconsistent, the biometric matching result is selected as the binding basis. Optionally, the preset range is a circular area with the center of the bounding box of the animal individual's outline as the center and the length of the diagonal of the bounding box as the radius.

[0029] In one embodiment of the present invention, see [reference] Figure 3 From the behavioral feature sequence, statistics of the kinematic features and the behavior-related acoustic features are extracted according to fixed time windows to form a behavioral time-series vector. From the environmental association features, statistics of the interaction between the animal individual and key environmental elements within each time window are extracted to form an environmental interaction time-series vector. Data on temperature, humidity, light intensity, and ammonia concentration within the same time window are extracted from the environmental sensor data to form an environmental parameter time-series vector. The behavioral time-series vector, the environmental interaction time-series vector, and the environmental parameter time-series vector are aligned and concatenated according to timestamps to form the multi-dimensional time-series state data. Missing value imputation and standardization are performed on the multi-dimensional time-series state data to form a standardized multi-dimensional time-series sequence.

[0030] In specific implementation, statistics of the kinematic features and the acoustic features related to the behavior are extracted from the behavioral feature sequence according to a fixed time window to form a behavioral time-series vector. The length of the fixed time window is a preset time length value, and the statistics include the arithmetic mean, standard deviation, maximum value and minimum value of each feature within the fixed time window. In some embodiments, the fixed time window is set as an overlapping time window, and the step size between two adjacent fixed time windows is smaller than the length of the fixed time window.

[0031] In specific implementation, statistical quantities of animal interactions with key environmental elements within each fixed time window are extracted from the environmental association features to form an environmental interaction time-series vector. The key environmental elements include feeding troughs, watering points, and neighboring individuals. The statistical quantities of interactions with key environmental elements include the average distance between an animal and the feeding trough, the average distance between an animal and the watering point, the average distance between an animal and its neighboring individuals, the number of times an animal enters a preset neighboring area of ​​the feeding trough, and the number of times an animal enters a preset neighboring area of ​​the watering point within each fixed time window. It can be understood that data on temperature, humidity, light intensity, and ammonia concentration within the same fixed time window are extracted from the environmental sensor data to form an environmental parameter time-series vector, wherein the data on temperature, humidity, light intensity, and ammonia concentration are the arithmetic mean of the corresponding sensor data collected within each fixed time window. In specific implementation, the behavior time-series vector, the environmental interaction time-series vector, and the environmental parameter time-series vector are aligned and concatenated according to timestamps to form the multi-dimensional time-series state data. The timestamps are the start or end times of each fixed time window. The concatenation operation is performed in the order of behavior time-series vector, environmental interaction time-series vector, and environmental parameter time-series vector along the feature dimensions. Optionally, the multi-dimensional time-series state data undergoes missing value imputation and standardization processing to form a standardized multi-dimensional time-series sequence. The missing value imputation uses a linear interpolation method; for multiple consecutive missing time points, linear filling is performed based on the known values ​​at the beginning and end of the missing sequence. Optionally, the standardization processing uses the Z-score standardization method, standardizing each feature dimension in the multi-dimensional time-series state data separately. The standardization calculation formula is: in: Represents the standardized time point First The numerical values ​​of each feature dimension, Indicates the time point before standardization First The numerical values ​​of each feature dimension, Indicates the first The arithmetic mean of each feature dimension at all time points Indicates the first The standard deviation of a feature dimension over all time points. This can be understood as the standard deviation of a certain feature dimension... When the value is zero, the standardization process sets all values ​​of that feature dimension to zero. In some embodiments, the missing value imputation can also employ a forward padding method, using the previous non-missing value to fill the current missing value. In some embodiments, the standardized multidimensional time series sequence formed after the standardization process is stored in a tensor data structure, the shape of which is the number of time steps multiplied by the total number of feature dimensions. It can be understood that the total number of feature dimensions is the sum of the feature dimensions of the behavioral time series vector, the feature dimensions of the environmental interaction time series vector, and the feature dimensions of the environmental parameter time series vector.

[0032] In one embodiment of the present invention, the animal health status assessment model includes a feature encoder, a temporal pattern learning network, and a state assessment head. The feature encoder performs dimensionality reduction and deep feature extraction on the input multi-dimensional temporal state data, outputting a high-dimensional feature sequence. The temporal pattern learning network is constructed based on a recurrent neural network or a self-attention mechanism. It models the high-dimensional feature sequence, captures the dependencies and evolution patterns of features in the time dimension, and outputs a comprehensive feature vector containing temporal context information. When the temporal pattern learning network is based on a recurrent neural network, the recurrent neural network is a bidirectional long short-term memory network. This network processes the high-dimensional feature sequence step by step, and at each time step, it combines past and future context information to update its internal hidden state. The final hidden states of all time steps are aggregated to generate the comprehensive feature vector. When the temporal pattern learning network is based on a self-attention mechanism, the self-attention mechanism calculates the correlation weight between the features of each time step in the high-dimensional feature sequence and the features of all other time steps. Based on the correlation weight, the features are weighted and summed and information is fused to directly capture long-distance temporal dependencies, output the transformed feature sequence, and then the transformed feature sequence is pooled to generate the comprehensive feature vector.

[0033] The status assessment head includes a health score output layer and an abnormality warning output layer. The health score output layer maps the comprehensive feature vector into a continuous score representing the health level. The abnormality warning output layer analyzes the comprehensive feature vector and outputs a potential abnormality warning describing the type and severity of the abnormality when the detected pattern deviates from the preset normal behavior and physiological pattern library.

[0034] In its implementation, the animal health status assessment model comprises a feature encoder, a temporal pattern learning network, and a status assessment head. The feature encoder performs dimensionality reduction and deep feature extraction on the input multi-dimensional temporal status data, outputting a high-dimensional feature sequence. The temporal pattern learning network, constructed based on a recurrent neural network or a self-attention mechanism, models the high-dimensional feature sequence, captures the dependencies and evolutionary patterns of features in the temporal dimension, and outputs a comprehensive feature vector containing temporal context information. The status assessment head comprises a health score output layer and an anomaly warning output layer. The health score output layer maps the comprehensive feature vector into a continuous score representing the degree of health. The anomaly warning output layer analyzes the comprehensive feature vector, and when the detected pattern deviates from a preset library of normal behavior and physiological patterns, it outputs a potential anomaly warning describing the type and severity of the anomaly.

[0035] In some embodiments, the feature encoder consists of two fully connected layers and a batch normalization layer stacked together, with each fully connected layer followed by a linear correction unit activation function. The multi-dimensional temporal state data is sequentially input into the feature encoder according to time steps, and a corresponding high-dimensional feature vector is output at each time step. The high-dimensional feature vectors of all time steps are arranged in chronological order to form the high-dimensional feature sequence. Optionally, the dimension of the high-dimensional feature sequence is higher than the original dimension of the multi-dimensional temporal state data.

[0036] It is understood that when the temporal pattern learning network is based on a recurrent neural network, the recurrent neural network is a bidirectional long short-term memory network. The bidirectional long short-term memory network includes a forward long short-term memory layer and a backward long short-term memory layer. The forward long short-term memory layer processes the high-dimensional feature sequence in ascending order of time, and the backward long short-term memory layer processes the high-dimensional feature sequence in descending order of time. At each time step, the bidirectional long short-term memory network concatenates the hidden states of the forward long short-term memory layer and the backward long short-term memory layer as the bidirectional output of that time step. The bidirectional long short-term memory network performs average pooling on the bidirectional outputs of all time steps to generate the comprehensive feature vector.

[0037] It is understood that when the temporal pattern learning network is based on a self-attention mechanism, the self-attention mechanism performs a linear transformation on the high-dimensional feature vector at each time step in the high-dimensional feature sequence to obtain a query vector, a key vector, and a value vector. Then, it calculates the dot product between the query vector at each time step and the key vectors at all time steps and divides it by a scaling factor to obtain an attention weight matrix. The value vector is then weighted and summed using the attention weight matrix to obtain the self-attention output at each time step. The self-attention outputs at all time steps constitute the transformed feature sequence. Global max pooling is then performed on the transformed feature sequence to generate the comprehensive feature vector.

[0038] In a specific implementation, the health score output layer uses a fully connected layer with a linear activation function to map the comprehensive feature vector into a scalar value as the real-time health status score. In another implementation, the anomaly warning output layer uses a fully connected layer followed by a multi-classifier structure. The fully connected layer maps the comprehensive feature vector into the log odds of multiple anomaly categories. The multi-classifier converts the log odds into a probability value for each anomaly category. When the probability value of an anomaly category exceeds a preset anomaly probability threshold, the anomaly category is output as the anomaly type for the potential anomaly warning, and the probability value is used as the severity of the potential anomaly warning. Optionally, the preset normal behavior and physiological pattern library is established by collecting multi-dimensional time-series state data of healthy animal individuals and extracting the distribution range of the corresponding comprehensive feature vectors. In some embodiments, the real-time health status score output is a continuous value between 0 and 1. The health score output layer uses a sigmoid function as the output layer activation function, and the expression of the sigmoid function is: in: This represents the real-time health status score. This represents the unactivated raw score output by the fully connected portion of the health score output layer. Represented by natural constant base Exponential function.

[0039] In one embodiment of the present invention, a husbandry management knowledge base is established, comprising animal breeds, growth stages, historical health records, and management knowledge rules. Based on the individual animal's identity, basic management parameters for that animal are retrieved from the husbandry management knowledge base. The real-time health status score is compared with a preset scoring threshold, and the type and content of potential abnormality warnings are analyzed.

[0040] When the real-time health status score falls below a preset threshold or a specific type of potential abnormal warning is detected, a refined intervention decision-making process is triggered. The process retrieves knowledge rule entries from the feeding management knowledge base that match the current animal breed, growth stage, and the type of potential abnormal warning. Based on these knowledge rule entries, an initial set of intervention options is constructed. A machine learning decision model is then used to evaluate this initial set of intervention options. The inputs to the machine learning decision model include the basic management parameters, the real-time health status score, the abnormal warning information, and the current environmental parameters. The output is the expected effect score for each intervention option. One or more intervention options with the highest expected effect scores are selected. Based on the real-time health status score and the severity of the abnormal warning, the execution parameters of the intervention options are quantified. These quantified parameters include adjusting the specific component ratios of the feed formula, adjusting the specific weight of the feed, and adjusting the specific time of feeding. The selected intervention option is combined with its quantified execution parameters to form at least one specific instruction that can be automatically executed by the feeding management system. This instruction includes adjusting the feed formula, adjusting the feeding amount, adjusting the feeding time, suggesting isolation and observation, suggesting vaccination, or suggesting a veterinary examination. This instruction serves as the individualized feeding management instruction.

[0041] In specific implementation, a husbandry management knowledge base is established, containing animal breed, growth stage, historical health records, and management knowledge rules. Based on the individual animal's identity, the basic management parameters of the individual animal are queried from the husbandry management knowledge base. The real-time health status score is compared with a preset score threshold, and the type and content of potential abnormal warnings are analyzed. When the real-time health status score is lower than the preset score threshold or a specific type of potential abnormal warning exists, a refined intervention decision-making process is triggered. This refined intervention decision-making process, based on the individual animal's basic management parameters, current health status score, abnormal warning information, and knowledge rules in the husbandry management knowledge base, generates at least one specific instruction—including adjusting feed formulation, adjusting feeding amount, adjusting feeding time, suggesting isolation and observation, suggesting vaccination, or suggesting veterinary examination—as the individualized husbandry management instruction through rule reasoning or machine learning decision-making models. In some embodiments, the husbandry management knowledge base is stored in the form of a relational database, where the animal breed field, growth stage field, historical health record field, and management knowledge rule field are linked through foreign keys. In some embodiments, the basic management parameters include the animal's standard daily feed intake, standard water intake, suitable ambient temperature range, and vaccination schedule.

[0042] In practice, the following steps are taken: First, knowledge rule entries matching the current animal breed, growth stage, and potential abnormal warning type are retrieved from the feeding management knowledge base. Second, an initial set of intervention options is constructed based on these knowledge rule entries. Third, a machine learning decision model is used to evaluate the initial set of intervention options. The inputs to the machine learning decision model include the basic management parameters, real-time health status score, abnormal warning information, and current environmental parameters. The output is an expected effect score for each intervention option. Fourth, one or more intervention options with the highest expected effect score are selected. Based on the real-time health status score and the severity of the abnormal warning, the execution parameters of the intervention options are quantified. These quantified parameters include adjusting the specific component ratios of the feed formula, adjusting the specific weight of the feed, and adjusting the specific time of feeding. Fifth, the selected intervention options are combined with the quantified execution parameters to form specific instructions that can be automatically executed by the feeding management system.

[0043] It is understood that the knowledge rule entries are represented using production rules. Each knowledge rule entry's antecedent includes animal breed conditions, growth stage conditions, and abnormal warning type conditions, while the consequent includes one or more candidate intervention options. It is understood that the expected outcome score is calculated using a machine learning decision model, specifically a gradient boosting decision tree model. The formula for calculating the expected outcome score is: in: Indicates the first Expected effect scores for each intervention option This represents the total number of decision trees in the gradient boosting decision tree model. Indicates the first The weights of the leaf nodes of a decision tree. For the indicator function, when the input vector Falling into In the decision tree and the first Leaf node area corresponding to each intervention option The input vector takes a value of 1 when the condition is met and 0 otherwise. The system is composed of the basic management parameters, real-time health status score, anomaly type code value in the anomaly warning information, and current environmental parameters. Optionally, the severity of the anomaly warning is represented by a value between 0 and 1. When quantifying the execution parameters of the intervention option, the severity value is multiplied by a preset maximum adjustment range to obtain the actual adjustment range. Optionally, the specific component ratio of the adjusted feed formula is represented as a percentage vector of each feed ingredient in the total feeding amount. In some embodiments, the machine learning decision model uses a random forest model instead of a gradient boosting decision tree model. It can be understood that the feeding management system automatically executes the specific instructions by sending the specific instructions in structured text format to the feed feeding control terminal, the isolation fence control terminal, or the veterinary workstation information system.

[0044] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. An artificial intelligence-based digital management method for animals, characterized in that, The method includes: Acquire multi-source monitoring data within a designated breeding area, including animal video image data, animal sound data, animal individual identification information, and environmental sensor data; The animal video image data and animal sound data are analyzed synchronously to identify individual animals and extract the behavioral feature sequences and environmental features of the individual animals. The individual animal identification information is read and decoded to establish a mapping relationship between the individual animal identity and the behavioral feature sequence; The behavioral feature sequence, the environmental correlation features, and the environmental sensor data are fused to form multi-dimensional time-series state data. The multi-dimensional time-series state data is input into a pre-trained animal health status assessment model, which is built based on machine learning algorithms and outputs the real-time health status score and potential abnormality warning of the individual animal. Based on the real-time health status score and potential abnormality warning, combined with the individual animal identity, targeted individualized feeding and management instructions are generated and issued for execution.

2. The artificial intelligence-based digital management method for animals according to claim 1, characterized in that, The animal video image data and animal sound data are analyzed synchronously to identify individual animals and extract behavioral feature sequences and environmental association features of the individual animals, including: The animal video image data is subjected to target detection and tracking processing to identify the outline of each individual animal in the video image frame, and cross-frame correlation tracking is performed on each identified individual animal outline in consecutive video image frames to form the movement trajectory of the individual animal. The animal sound data is subjected to sound source separation and feature extraction to separate sound segments emitted by different animal individuals or groups, and the time-spectral features of the sound segments are extracted. The animal's individual movement trajectory is timestamped with the sound segment, and the movement trajectory segment and sound segment belonging to the same animal individual are associated based on the aligned timestamps. Kinematic features are extracted from the motion trajectory segment, including velocity, acceleration, movement path complexity, and attitude change frequency. The behavior-related acoustic features are extracted from the aligned sound segments, including the frequency distribution of specific calls, the duration of the calls, and the pattern of call intervals. Image features related to the animal's surrounding environment are extracted from video image frames. These image features include the distance to neighboring animals and the location information relative to the feeding trough or watering point. These image features related to the animal's surrounding environment are used as the environmental association features.

3. The artificial intelligence-based digital management method for animals according to claim 2, characterized in that, Reading and decoding the individual animal identification information to establish a mapping relationship between the individual animal's identity and the behavioral feature sequence includes: The individual animal identification information includes the encoded information carried by the animal's ear tag, collar tag, or biometric identifier; The animal individual identification information is read and decoded by an RFID reader, image recognition device, or biometric identification device to obtain a unique animal individual identification code; When the outline of an individual animal is identified through target detection and tracking, the individual animal identification information is searched and read within a preset range of the outline. If the reading is successful, the read animal individual identification code will be bound to the animal individual outline and the movement trajectory formed by subsequent tracking; If the animal cannot be read directly, the animal's individual biometrics are matched with a pre-established biometric database, or the animal's individual movement trajectory is correlated with the behavior patterns of known individuals to indirectly infer and bind the animal's individual identity code.

4. The artificial intelligence-based digital animal management method according to claim 3, characterized in that, The behavioral feature sequence, the environmental correlation features, and the environmental sensor data are fused to form multi-dimensional time-series state data, including: From the behavioral feature sequence, statistics of the kinematic features and the acoustic features related to the behavior are extracted according to a fixed time window to form a behavioral time-series vector; From the environmental association features, the statistics of the interaction between individual animals and key environmental elements within each time window are extracted to form an environmental interaction time series vector; Data on temperature, humidity, light intensity, and ammonia concentration within the same time window are extracted from the environmental sensor data to form a time-series vector of environmental parameters; The behavior time sequence vector, the environmental interaction time sequence vector, and the environmental parameter time sequence vector are aligned and concatenated according to timestamps to form the multi-dimensional time sequence state data; The multi-dimensional time-series state data is imputed for missing values ​​and standardized to form a standardized multi-dimensional time-series sequence.

5. The artificial intelligence-based digital management method for animals according to claim 1, characterized in that, The multi-dimensional time-series state data is input into a pre-trained animal health status assessment model, which is built based on machine learning algorithms and outputs a real-time health status score and potential abnormality warning for the individual animal, including: The animal health status assessment model includes a feature encoder, a temporal pattern learning network, and a status assessment head. The feature encoder performs dimensionality reduction and deep feature extraction on the input multi-dimensional temporal state data, and outputs a high-dimensional feature sequence. The temporal pattern learning network is constructed based on a recurrent neural network or a self-attention mechanism. It models the high-dimensional feature sequence, captures the dependencies and evolution patterns of features in the time dimension, and outputs a comprehensive feature vector containing temporal context information. The status assessment head includes a health score output layer and an abnormality warning output layer. The health score output layer maps the comprehensive feature vector into a continuous score representing the health level. The abnormality warning output layer analyzes the comprehensive feature vector and outputs a potential abnormality warning describing the type and severity of the abnormality when the detected pattern deviates from the preset normal behavior and physiological pattern library.

6. The artificial intelligence-based digital management method for animals according to claim 5, characterized in that, Based on the real-time health status score and potential abnormality warning, combined with the animal's individual identity, targeted individualized feeding and management instructions are generated, including: Establish a husbandry and management knowledge base that includes animal breeds, growth stages, historical health records, and management knowledge and rules; Based on the individual animal's identity, the basic management parameters of that individual animal are retrieved from the animal husbandry and management knowledge base; The real-time health status score is compared with a preset score threshold, and the type and content of the potential abnormality warning are analyzed. When the real-time health status score is lower than the preset score threshold or there is a specific type of potential abnormality warning, a refined intervention decision process is triggered. The refined intervention decision-making process is based on the basic management parameters of the individual animal, the current health status score, abnormal warning information, and the knowledge rules in the feeding management knowledge base. Through rule reasoning or machine learning decision-making models, it generates at least one specific instruction, including adjusting feed formula, adjusting feeding amount, adjusting feeding time, suggesting isolation and observation, suggesting vaccination, or suggesting veterinary examination, as the individualized feeding management instruction.

7. The artificial intelligence-based digital management method for animals according to claim 5, characterized in that, The temporal pattern learning network is constructed based on a recurrent neural network or a self-attention mechanism to model the high-dimensional feature sequence, including: When using a time-series pattern learning network based on a recurrent neural network, the recurrent neural network is a bidirectional long short-term memory network. This network processes the high-dimensional feature sequence step by step, and at each time step, it combines past and future contextual information to update its internal hidden state. Finally, it aggregates the final hidden states of all time steps to generate the comprehensive feature vector. When using a temporal pattern learning network based on a self-attention mechanism, the self-attention mechanism calculates the correlation weights between the features of each time step in the high-dimensional feature sequence and the features of all other time steps. Based on the correlation weights, the features are weighted and summed and information is fused to directly capture long-distance temporal dependencies, outputting a transformed feature sequence. Then, a pooling operation is performed on the transformed feature sequence to generate the comprehensive feature vector.

8. The artificial intelligence-based digital management method for animals according to claim 6, characterized in that, The refined intervention decision-making process is based on the animal's basic management parameters, current health status score, abnormal warning information, and knowledge rules in the feeding management knowledge base. Through rule-based reasoning or machine learning decision models, it generates at least one specific instruction, including adjusting feed formulation, adjusting feeding amount, adjusting feeding time, recommending isolation and observation, recommending vaccination, or recommending veterinary examination. Retrieve knowledge rule entries from the animal husbandry and management knowledge base that match the current animal breed, growth stage, and the type of potential abnormal warning. Based on the aforementioned knowledge rule entries, an initial set of intervention options is constructed; The initial set of intervention options is evaluated using a machine learning decision model. The inputs of the machine learning decision model include the basic management parameters, real-time health status scores, abnormal warning information, and current environmental parameters. The output is the expected effect score for each intervention option. Select one or more intervention options with the highest expected effect score, and quantify the execution parameters of the intervention options based on the real-time health status score and the severity of abnormal warnings. The quantified parameters include adjusting the specific ingredient ratio of the feed formula, adjusting the specific weight of the feed, and adjusting the specific time of the feed. The selected intervention options are combined with their quantified execution parameters to form the specific instructions that can be automatically executed by the feeding management system.

9. The artificial intelligence-based digital management method for animals according to claim 2, characterized in that, Target detection and tracking processing is performed on the animal video image data to identify the outline of each individual animal in the video image frame, including: The target detection model based on a deep convolutional neural network is used to perform frame-by-frame analysis on the input animal video image data. The target detection model outputs the bounding box coordinates and confidence score of each individual animal in the image frame. The non-maximum suppression algorithm is used to filter overlapping bounding boxes, retaining the most likely single bounding box for each animal individual as the initial animal individual outline; In continuous video image frames, a tracking algorithm based on appearance features and motion features is used to match and associate the outline of an animal in the current frame with the track of an animal in the previous frame. The appearance features are in-contour image features extracted by a convolutional neural network, and the motion features are motion vectors predicted based on the bounding box positions of the previous and next frames. For newly emerging animal individual silhouettes that fail to be successfully matched and associated, create new tracking tracks; For animal trajectories that are continuously lost, the tracking will be terminated after the preset number of lost frames is exceeded.

10. An artificial intelligence-based digital animal management system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the artificial intelligence-based digital management method for animals as described in any one of claims 1 to 9.

Images