Real-time recognition and early warning method for abnormal behavior of live pigs and edge computing system

Abstract

The application provides a pig abnormal behavior real-time identification and early warning method and an edge computing system, the method collects pigsty video data and pig wearable device activity data, performs target detection and tracking on the video data, and performs filtering and feature extraction on the activity data; video space-time behavior features are extracted based on space-time graph convolution modeling, and are fused with activity time change features to construct a pig individual and group behavior correlation model; then, an abnormal behavior identification model is combined with a preset behavior threshold to determine fighting, coughing and abnormal feeding, and local and remote early warning is triggered. The application fuses pigsty video data and pig activity data, constructs a pig individual and group behavior correlation model based on space-time graph convolution modeling, comprehensively analyzes video space-time behavior features and activity time change features, and can more effectively distinguish normal behavior from abnormal behavior such as fighting, coughing and abnormal feeding.

Description

Technical Field

[0001] This invention relates to the field of monitoring technology for livestock and poultry, and particularly to a method for real-time identification and early warning of abnormal behavior in pigs and an edge computing system. Background Technology

[0002] With the development of large-scale pig farming, the demand for real-time monitoring of pig health status and abnormal behavior is constantly increasing. Behaviors such as fighting, coughing, and abnormal feeding are usually external manifestations of stress, disease, or abnormal feeding in pigs. If they are not detected in time, they can easily affect the health of pigs and the efficiency of farming.

[0003] In existing technologies, abnormal behavior monitoring mostly adopts manual observation, single video recognition, or single activity sensor monitoring methods. Manual observation has problems such as poor continuity and low efficiency. Single video recognition is prone to inaccurate identification, misjudgment, or missed judgment in crowded and mutually obstructed scenarios of pigs. Although single activity monitoring can reflect changes in movement, it is difficult to accurately distinguish different types of behavior such as fighting, coughing, and abnormal feeding.

[0004] In addition, some existing solutions rely on uploading raw video data to the cloud for processing, which has problems such as large data transmission volume, slow response, and insufficient on-site adaptability, making it difficult to meet the application needs for real-time early warning in the pig farming scenario; To address this, a method for real-time identification and early warning of abnormal pig behavior and an edge computing system are proposed. Summary of the Invention

[0005] In view of this, the present invention provides a method for real-time identification and early warning of abnormal behavior in pigs and an edge computing system to solve or alleviate the technical problems existing in the prior art, and at least provide a beneficial option.

[0006] The technical solution of this invention is implemented as follows: a real-time identification and early warning method for abnormal pig behavior, based on spatiotemporal graph convolutional modeling and fusing video data and activity data, to achieve identification and early warning of abnormal pig behavior, including the following steps: S1. Collect video data through cameras installed in pig pens and collect activity data through wearable devices worn by individual pigs; The video data was denoised and normalized, and the location information, movement trajectory and posture characteristics of individual pigs were obtained through target detection and multi-target tracking. The activity data is filtered and normalized, and the peak, mean and fluctuation characteristics of the activity are extracted. S2. Using individual pigs as graph nodes and the spatial relationships between individuals as graph edges, construct a pig group behavior graph structure, and extract spatiotemporal behavior features from video data based on a spatiotemporal graph convolutional network. Meanwhile, based on the time series feature extraction model, the temporal variation characteristics of activity data are obtained, and the two types of features are weighted and fused to construct a behavioral association model between individual pigs and groups, thereby obtaining comprehensive behavioral characteristics; S3. Input the comprehensive behavioral features into the pre-trained abnormal behavior recognition model, output the behavior category and corresponding confidence level, and make a judgment in combination with the preset behavior threshold. The abnormal behavior includes fighting, coughing and abnormal eating. S4. When abnormal behavior is detected, local sound and light warnings and remote push warnings are triggered, and the identification results are stored locally and the aggregated data is uploaded to the cloud at a preset frequency.

[0007] More preferably, the video data acquisition frame rate is 15-30 frames per second; the activity data is acquired by a smart ear tag or smart collar, and the sampling frequency is synchronized with the video data acquisition.

[0008] More preferably, the target detection uses a deep learning target detection model to identify individual pigs, and the multi-target tracking uses a data association algorithm to continuously track individuals in crowded or occluded scenarios.

[0009] More preferably, the behavioral association model is used to help distinguish between normal and abnormal behaviors by analyzing the synergy and differences between individual and group behaviors.

[0010] More preferably, the preset behavior threshold includes: Fighting behavior corresponds to an activity level exceeding the normal range by a preset multiple and lasting for a preset time; Coughing behavior corresponds to a periodic fluctuation in activity level that lasts for a preset time. Abnormal feeding corresponds to a time spent in the feeding area or an activity level that is below the normal range and continues for a preset time.

[0011] This invention also provides an edge computing system for real-time identification and early warning of abnormal pig behavior, used to implement the aforementioned method for real-time identification and early warning of abnormal pig behavior. The system is deployed locally in the pigpen and includes: The multi-source data acquisition module is used to collect video data and pig activity data; The data preprocessing module is used to denoise, normalize, and extract features from the collected data. The feature fusion and recognition module is used to fuse multi-source data based on spatiotemporal graph convolution modeling and time series analysis and output abnormal behavior recognition results. Edge computing terminal module is used to perform data processing, result storage and communication control; The early warning module is used to output early warning information for abnormal behavior; The edge computing terminal module is used to process data locally to reduce the amount of data uploaded.

[0012] More preferably, the multi-source data acquisition module includes a high-definition camera and a wearable device, wherein the wearable device is a smart ear tag or a smart collar, has a built-in accelerometer sensor, and communicates with an edge computing terminal via low-power wireless communication.

[0013] More preferably, the feature fusion and recognition module includes a spatiotemporal graph convolution processing unit, a time series feature extraction unit, and a feature fusion unit.

[0014] More preferably, the edge computing terminal module includes a processor, a local storage unit, and a wireless communication unit, used to realize local data processing and low-latency response, with processing latency not exceeding a preset threshold, local storage capacity being a preset capacity, and supporting local data retention for a preset duration.

[0015] More preferably, the early warning module includes a local audible and visual early warning unit and a remote push unit, wherein the remote push unit is used to send abnormal behavior information to the mobile terminal.

[0016] The embodiments of the present invention have the following advantages due to the adoption of the above technical solutions: I. This invention integrates video data from pigpen cameras with activity data from wearable devices for pigs, and introduces a spatiotemporal graph convolutional modeling approach to construct a behavioral correlation model between individual pigs and the group. This allows for comprehensive analysis of pig behavior from multiple dimensions, including spatial relationships, temporal dynamics, and individual activity status, effectively overcoming the limitations of single video monitoring or single sensor monitoring in terms of information dimensionality. Especially in scenarios where pigs are densely packed and mutually obstructing each other, it can still effectively distinguish between normal behavior and abnormal behaviors such as fighting, coughing, and abnormal feeding, improving the accuracy and stability of abnormal behavior identification. Second, this invention deploys data preprocessing, feature extraction, behavior recognition, and early warning determination processes locally in the pigsty, and only uploads the summary information of abnormal events to the cloud at a preset frequency, thereby effectively reducing the long-term and large-scale transmission of raw video data, reducing network bandwidth consumption and cloud computing pressure; at the same time, the localized processing method can shorten the abnormal behavior recognition and early warning response time, meeting the application needs of real-time monitoring, rapid early warning, and low-cost deployment in large-scale pig farming scenarios.

[0017] The above overview is for illustrative purposes only and is not intended to be limiting in any way. In addition to the illustrative aspects, embodiments, and features described above, further aspects, embodiments, and features of the invention will become readily apparent from the accompanying drawings and the following detailed description. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the early warning method steps of the present invention. Detailed Implementation

[0020] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0021] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0022] like Figure 1 As shown, this embodiment of the invention provides a real-time identification and early warning method for abnormal pig behavior and an edge computing system, mainly targeting large-scale pig farming scenarios. By integrating pen video data and pig activity data, it can identify behaviors such as fighting, coughing, and abnormal feeding in pigs in real time. Based on the edge computing architecture, it realizes local analysis, local early warning, and cloud aggregation, thereby reducing the pressure of remote transmission of raw data and improving the real-time performance and reliability of abnormal behavior identification.

[0023] Example 1 This embodiment discloses a method for real-time identification and early warning of abnormal pig behavior, which can be deployed in pig farming environments such as fattening pens, nursery pens, or breeding pig pens; the following description uses a large-scale fattening pen as an example. This fattening pen houses 300 pigs, with a pen area of ​​approximately 150 square meters. The pen is equipped with feeding troughs, drinking water areas, and passageways to facilitate subsequent identification and judgment of areas related to feeding behavior.

[0024] The method in this embodiment includes the following steps: S1. Complete the installation of video acquisition devices and activity tracking devices in the enclosure. High-definition infrared cameras are preferred for video acquisition devices in order to meet the monitoring needs during both day and night. In this embodiment, four high-definition infrared cameras are evenly installed on the top of the pen, with each camera spaced apart to ensure that its field of view covers the main activity area of ​​the pen, thereby reducing blind spots in video acquisition. The activity acquisition device preferably uses wearable devices for pigs, such as smart ear tags or smart collars. In this embodiment, each pig wears one smart ear tag, which has a built-in three-axis accelerometer sensor for collecting individual pig activity information. During the data acquisition phase, the camera acquires video data of the enclosure at a frame rate of 15 to 30 frames per second, with 25 frames per second being preferred in this embodiment; the smart ear tag acquires activity data at a frequency that matches the video acquisition, thereby keeping the video data and activity data synchronized in the time dimension, so as to facilitate subsequent multi-source feature alignment and fusion. The collected video data first enters the preprocessing process. Preferably, the video frames are denoised to reduce interference from factors such as dust, light fluctuations, and equipment vibrations in the pen. Then, normalization is performed to reduce the impact of differences in imaging brightness between different time periods and different cameras on the subsequent analysis results. Afterward, based on the target detection model, individual pigs in the video frames are identified to obtain basic information such as the bounding box and center position of each pig target. Furthermore, combined with a multi-target tracking algorithm, individual pigs in consecutive video frames are correlated across frames to obtain the continuous trajectory information of the corresponding individuals. Through the above processing, the position information, movement trajectory, and posture characteristics of each pig in the pen over time can be obtained. For scenarios where pigs are concentrated and there is significant mutual occlusion, the above tracking process can reduce the problem of short-term loss of individual targets and improve the continuity of subsequent abnormal behavior identification.

[0025] The collected activity data also undergoes preprocessing. Preferably, the raw activity signal is first filtered to remove invalid fluctuations introduced by equipment noise, short-term collision jitter, or occasional abnormal sampling; then, normalization is performed to reduce the impact of calibration differences between different devices; after completing the basic preprocessing, peak features, mean features, and fluctuation features are extracted from the activity sequence; peak features are used to characterize the instantaneous intensity of activity per unit time, mean features are used to characterize the overall activity level over a period of time, and fluctuation features are used to reflect the amplitude and rhythmic characteristics of activity state changes.

[0026] In some implementations, the feeding area can be pre-defined based on the location of the feeding trough in the enclosure, or the feeding area can be automatically identified through image segmentation, so as to make regional correlation judgments on abnormal feeding behavior in the future.

[0027] S2. After completing the preprocessing of video data and activity data, behavioral features are extracted and fused to form a model for the two types of data.

[0028] A group behavior graph structure is constructed based on individual pigs identified in consecutive video frames; Specifically, each pig is used as a node in the graph, and the spatial relationship between any two pigs is used as the basis for constructing the graph edges. The weights of the graph edges can be set according to the distance between nodes, their relative movement relationship, or their proximity. The graph structure constructed in this way can dynamically reflect the group distribution status of the pigs in the pen at a certain moment and the spatial relationship between individuals.

[0029] Building upon this, feature extraction is performed on the video data using a spatiotemporal graph convolutional network. The graph convolution part of the network is used to extract spatial relationship features between individual pigs, such as proximity relationships, local clustering states, and relative positional changes. The temporal convolution part is used to extract dynamic features of behavior evolution over time, such as action duration, changes in movement trends, and posture evolution. After this processing, spatiotemporal behavioral features that characterize the spatial behavior patterns and individual action changes of pig herds can be obtained.

[0030] Simultaneously, feature extraction is performed on the time series of activity data. The preferred method is to use a time series feature extraction model to analyze the activity series and extract time-varying features that reflect sudden changes, sustained stability, and periodic fluctuations in activity. For abnormal pig behavior, a sudden increase in activity is usually associated with violent behaviors such as fighting, while periodic fluctuations in activity can provide auxiliary judgment for rhythmic actions such as coughing. A continuous decrease in activity may be related to insufficient feed intake, depression, and other states.

[0031] After obtaining the spatiotemporal behavioral features and activity level variation features from the video, the two types of features are fused. In this embodiment, a weighted fusion method can be used for feature integration, allowing both video features and activity level features to participate in the abnormal behavior determination. The fusion weights can be optimized based on training samples to give features with higher correlation to the target abnormal behavior a greater discriminative effect. Through the above fusion process, comprehensive behavioral characteristics that take into account both "video appearance and motion information" and "individual activity status information" can be formed.

[0032] Furthermore, a behavioral correlation model between individuals and groups is constructed based on comprehensive behavioral characteristics. This behavioral correlation model is used to analyze the degree of consistency or deviation between individual behavior and group behavior. For example, during normal feeding, the activity rhythm of most pigs in the same area usually shows a certain consistency; when fighting occurs, the intensity of activity of local individuals is often significantly higher than that of surrounding individuals; when coughing occurs, related individuals may exhibit local periodic movements or fluctuation patterns that differ from the group's regular activity rhythm. By introducing individual-group correlation analysis, the system can not only make judgments based on the local characteristics of a single pig, but also make corrections based on group background information, thereby reducing the probability of misidentification under conditions such as crowding, obstruction, and poor local perspective.

[0033] S3. After completing the comprehensive behavioral feature extraction, input the comprehensive behavioral features into the pre-trained abnormal behavior recognition model to classify and identify the pig behavior.

[0034] The abnormal behavior recognition model can be trained based on pre-collected and labeled sample data. The sample data includes normal behavior samples and abnormal behavior samples. Normal behavior samples can include normal feeding, lying still, slow walking, and group movement, while abnormal behavior samples include at least fighting behavior, coughing behavior, and abnormal feeding behavior. By manually labeling the sample data, corresponding training and validation datasets are established. The abnormal behavior recognition model is then trained based on comprehensive behavioral features, enabling the model to learn the feature representation patterns corresponding to different types of behavior. After training, the model can output the corresponding behavior category and behavior confidence score for the input comprehensive behavioral features.

[0035] During real-time operation, the system inputs the comprehensive behavioral features extracted at the current moment or within the current time window into the recognition model. The model outputs the current behavior category and its confidence level, and further determines the outcome by combining this with a preset behavior threshold. The preset behavior threshold is used to combine the model output with business rules in the aquaculture scenario to enhance the interpretability and executability of the recognition results.

[0036] Specifically, for fighting behavior, it can be set that when the target pig's activity level is significantly higher than the normal range and continues for a certain period of time, it can be judged as abnormal fighting based on characteristics such as rapid displacement, local collisions, and increased body swaying in the video. For coughing behavior, it can be set that when the target pig's activity level shows periodic fluctuations and is accompanied by regular movements of the head, chest, or body, it can be judged as abnormal coughing. For abnormal feeding behavior, it can be set that when the target pig stays in the feeding area for less time than the corresponding range of normal feeding status, or when it is in the feeding period but its activity level is consistently lower than the normal feeding level for a preset duration, it can be judged as abnormal feeding.

[0037] In a preferred embodiment, the determination of abnormal behavior can be divided into two levels: valid abnormality and suspected abnormality. When the confidence level of the behavior output by the model reaches a first preset threshold and simultaneously meets the corresponding behavior threshold condition, it is determined to be a valid abnormality and an early warning is triggered directly. When the confidence level of the behavior output by the model is below the first preset threshold but above the second preset threshold, it is determined to be a suspected abnormality, and the monitoring time can be extended or a secondary verification can be initiated. When the confidence level is below the second preset threshold, the current behavior is considered to be within the normal range or does not meet the triggering conditions.

[0038] By combining model recognition with threshold verification, the stability of abnormal behavior recognition is effectively improved, and false alarms caused by transient noise, short-term congestion, or local posture changes are reduced.

[0039] S4. When the system determines that there is valid abnormal behavior, it will output an early warning message. The early warning method can include both local and remote early warning.

[0040] Local early warning systems preferably employ audible and visual alerts. For example, warning lights and alarms are installed in the designated areas of the pens. When fighting, coughing, or abnormal feeding behavior meets the warning criteria, the warning lights flash and the alarm sounds to alert on-site management personnel to take timely action. Remote early warning systems transmit abnormal behavior information to the mobile terminals of farm management personnel via communication networks, such as mobile applications, SMS platforms, or other information receiving terminals. The information pushed remotely may include the type of abnormal behavior, the time of occurrence, the location of occurrence, the corresponding pig number or pen number, and the confidence level of the behavior.

[0041] In terms of data processing, the system prioritizes storing the identification results, early warning records, and feature data related to anomaly determination in the local storage unit for subsequent querying, tracing, and manual review. In this embodiment, it is preferable to upload summary data of abnormal events to the cloud platform only at a preset frequency, rather than continuously uploading all the original video data, thereby reducing network bandwidth consumption and cloud data processing pressure. The uploaded content may include the type of abnormal behavior, the number of occurrences, the time period of occurrence, the area of ​​occurrence, the target pig identification, and the equipment operating status, etc. Through this method of local processing at the edge and summary management in the cloud, it can not only meet the needs of real-time early warning, but also facilitate the subsequent generation of abnormal behavior statistical reports and health management analysis results of the farm.

[0042] Example 2 This embodiment discloses an edge computing system for real-time identification and early warning of abnormal pig behavior, used to implement the real-time identification and early warning method for abnormal pig behavior in Embodiment 1; the system is deployed locally in the pig pen and mainly includes a multi-source data acquisition module, a data preprocessing module, a feature fusion and recognition module, an edge computing terminal module, and an early warning module; The multi-source data acquisition module is used to collect video data and pig activity data. The multi-source data acquisition module includes a high-definition camera and a wearable device. The high-definition camera is installed on the top or side of the pen to collect real-time video of pig activity in the pen. The wearable device is preferably a smart ear tag or smart collar, which is worn on the individual pig to collect the pig's activity data and transmit the data to the edge computing terminal through low-power wireless communication. The data preprocessing module is connected to the multi-source data acquisition module and is used to preprocess the acquired video data and activity data. Specifically, the video data is processed by denoising, normalization, target detection and target tracking to obtain the location information, movement trajectory and posture characteristics of individual pigs. The activity data is processed by filtering, normalization and feature extraction to obtain the peak value, mean value and fluctuation characteristics of activity. The feature fusion and recognition module is used to perform fusion analysis on preprocessed multi-source data and output abnormal behavior recognition results. The feature fusion and recognition module includes a spatiotemporal graph convolution processing unit, a time series feature extraction unit, and a feature fusion unit. The spatiotemporal graph convolution processing unit is used to extract spatiotemporal behavioral features from video data. The time series feature extraction unit is used to extract time variation features from activity data. The feature fusion unit is used to fuse the two types of features and, combined with a pre-trained abnormal behavior recognition model, output recognition results for fighting, coughing, and abnormal feeding. The edge computing terminal module is used to perform local data processing, result storage, and communication control. The edge computing terminal module includes a processor, a local storage unit, and a communication unit. It is used to process data locally, reduce the amount of raw data uploaded, and store recognition results and early warning records. When necessary, the edge computing terminal module can also monitor the device's operating status and output fault prompts when the camera is offline, the wearable device has low power, or communication is abnormal. The early warning module is used to output early warning information for abnormal behavior; the early warning module includes a local audible and visual early warning unit and a remote push unit; when abnormal behavior is detected, the local audible and visual early warning unit issues an on-site alarm, and the remote push unit sends abnormal behavior information to the mobile terminal of the farmer to remind the farmer to deal with it in time. In this embodiment, during system operation, the multi-source data acquisition module collects video data and activity data in real time. After the data preprocessing module completes preprocessing, it sends the processing results to the feature fusion and recognition module. After the feature fusion and recognition module completes the abnormal behavior recognition, it sends the recognition results to the edge computing terminal module for local storage. When the warning conditions are met, the warning module is triggered to output warning information. At the same time, the edge computing terminal module can also upload the summary information of abnormal events to the cloud platform at a preset frequency for subsequent statistics and management.

[0043] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in the present invention, and these should all be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for real-time identification and early warning of abnormal behavior in pigs, characterized in that, Based on spatiotemporal graph convolutional modeling and fusion of video data and activity data, abnormal behavior identification and early warning in pigs are achieved, including the following steps: S1. Collect video data through cameras installed in pig pens and collect activity data through wearable devices worn by individual pigs; The video data was denoised and normalized, and the location information, movement trajectory and posture characteristics of individual pigs were obtained through target detection and multi-target tracking. The activity data is filtered and normalized, and the peak, mean and fluctuation characteristics of the activity are extracted. S2. Using individual pigs as graph nodes and the spatial relationships between individuals as graph edges, construct a pig group behavior graph structure, and extract spatiotemporal behavior features from video data based on a spatiotemporal graph convolutional network. Meanwhile, based on the time series feature extraction model, the temporal variation characteristics of activity data are obtained, and the two types of features are weighted and fused to construct a behavioral association model between individual pigs and groups, thereby obtaining comprehensive behavioral characteristics; S3. Input the comprehensive behavioral features into the pre-trained abnormal behavior recognition model, output the behavior category and corresponding confidence level, and make a judgment in combination with the preset behavior threshold. The abnormal behavior includes fighting, coughing and abnormal eating. S4. When abnormal behavior is detected, local sound and light warnings and remote push warnings are triggered, and the identification results are stored locally and the aggregated data is uploaded to the cloud at a preset frequency.

2. The method for real-time identification and early warning of abnormal pig behavior according to claim 1, characterized in that: The video data acquisition frame rate is 15-30 frames per second; the activity data is collected by smart ear tags or smart collars, and the sampling frequency is synchronized with the video data acquisition.

3. The method for real-time identification and early warning of abnormal pig behavior according to claim 1, characterized in that: The target detection uses a deep learning target detection model to identify individual pigs, and the multi-target tracking uses a data association algorithm to continuously track individuals in crowded or occluded scenarios.

4. The method for real-time identification and early warning of abnormal pig behavior according to claim 1, characterized in that: The behavioral association model analyzes the synergy and differences between individual and group behaviors to help distinguish between normal and abnormal behaviors.

5. The method for real-time identification and early warning of abnormal pig behavior according to claim 1, characterized in that: The preset behavior thresholds include: Fighting behavior corresponds to an activity level exceeding the normal range by a preset multiple and lasting for a preset time; Coughing behavior corresponds to a periodic fluctuation in activity level that lasts for a preset time. Abnormal feeding corresponds to a time spent in the feeding area or an activity level that is below the normal range and continues for a preset time.

6. A real-time edge computing system for identifying and warning of abnormal pig behavior, used to implement the real-time identification and warning method for abnormal pig behavior as described in any one of claims 1 to 5, characterized in that, The system is deployed locally in the pigsty and includes: The multi-source data acquisition module is used to collect video data and pig activity data; The data preprocessing module is used to denoise, normalize, and extract features from the collected data. The feature fusion and recognition module is used to fuse multi-source data based on spatiotemporal graph convolution modeling and time series analysis and output abnormal behavior recognition results. Edge computing terminal module is used to perform data processing, result storage and communication control; The early warning module is used to output early warning information for abnormal behavior; The edge computing terminal module is used to process data locally to reduce the amount of data uploaded.

7. The edge computing system for real-time identification and early warning of abnormal pig behavior according to claim 6, characterized in that: The multi-source data acquisition module includes a high-definition camera and a wearable device. The wearable device is a smart ear tag or a smart collar, with a built-in accelerometer sensor, and communicates with an edge computing terminal via low-power wireless communication.

8. The edge computing system for real-time identification and early warning of abnormal pig behavior according to claim 6, characterized in that: The feature fusion and recognition module includes a spatiotemporal graph convolution processing unit, a time series feature extraction unit, and a feature fusion unit.

9. The edge computing system for real-time identification and early warning of abnormal pig behavior according to claim 6, characterized in that: The edge computing terminal module includes a processor, a local storage unit, and a wireless communication unit, which are used to realize local data processing and low-latency response. The processing latency is no higher than a preset threshold, the local storage capacity is a preset capacity, and it supports local data retention for a preset duration.

10. The edge computing system for real-time identification and early warning of abnormal pig behavior according to claim 6, characterized in that: The early warning module includes a local audible and visual early warning unit and a remote push unit, wherein the remote push unit is used to send abnormal behavior information to the mobile terminal.

Images