A method and system for monitoring and preventing wild animal disease based on artificial intelligence

Abstract

The application relates to the technical field of animal monitoring, and provides a wild animal epidemic disease monitoring and prevention and control method and system based on artificial intelligence, wherein the system collects data such as video streams, body surface temperatures, air pathogen concentrations, sound characteristics and VOCs spectrum graphs of wild animals in real time through multiple modal sensor nodes, carries out real-time inference through a wildness AI model in an edge computing unit, evaluates the health status and epidemic disease risks of the animals, combines a multi-source risk knowledge graph and a graph neural network, dynamically adjusts an epidemic situation occurrence probability threshold value of the system, generates accurate prevention and control instructions such as risk area division, isolation early warning and material allocation schemes, after the epidemic disease risks are confirmed, the system sends early warning information to a prevention and control center through various communication links, and continuously optimizes the prevention and control strategy through reinforcement learning. The method realizes the intelligentization and precision of wild animal epidemic disease prevention and control, can respond to complex ecological environment changes in real time, and significantly improves the prevention and control efficiency and accuracy.

Description

Technical Field

[0001] This invention relates to the field of animal monitoring technology, specifically to a method and system for monitoring and controlling wildlife diseases based on artificial intelligence. Background Technology

[0002] With the continuous expansion of human activities, the ecological environment of wildlife habitats is facing increasingly serious threats. The spread and outbreak of wildlife diseases not only affect the survival of species but may also trigger cross-species disease transmission, thus posing a potential threat to human health. Traditional wildlife disease monitoring methods mostly rely on manual observation or limited monitoring equipment, resulting in monitoring blind spots and a lack of effective real-time data processing and analysis methods, making it difficult to make accurate predictions and timely interventions in the early stages of an outbreak.

[0003] In recent years, with the rapid development of artificial intelligence technology, especially the advancements in image recognition, deep learning, and sensing technologies, AI-based disease monitoring methods have gradually become a new research hotspot. By combining AI algorithms with sensing technologies, it is possible to collect and analyze real-time data on the behavior, vital signs, and environment of wild animals, thereby achieving accurate assessment of their health status and early warning of disease risks.

[0004] Despite the progress made in animal health monitoring, existing technologies still face challenges: disease prediction and risk assessment often rely on single sensors or data sources, lacking fusion analysis of cross-modal data. Therefore, there is an urgent need for an intelligent wildlife disease monitoring and control method that integrates multiple sensors and is based on edge computing and artificial intelligence technologies. This would enable comprehensive and accurate disease monitoring and real-time early warning, effectively improving the response speed and accuracy of wildlife disease prevention and control. Summary of the Invention

[0005] This invention provides a method and system for monitoring and controlling wildlife diseases based on artificial intelligence, aiming to solve the technical problems of single data sources for disease prediction and risk assessment and inaccurate disease monitoring in the prior art.

[0006] In a first aspect, the present invention provides a method for monitoring and controlling wildlife diseases based on artificial intelligence, the method comprising:

[0007] Step S1: Deploy multimodal monitoring nodes consisting of various types of sensors in wildlife habitats;

[0008] Step S2: Acquire wildlife video streams, perform real-time inference on the video streams using a wildness AI model deployed in an edge computing unit, and output the wildness score for each individual wildlife species.

[0009] Step S3: When the wildness score is lower than the preset wildness threshold, the wild animal is determined to be a high-risk individual for disease.

[0010] Step S4: Collect the surface temperature, airborne pathogen concentration, sound characteristics, and VOCs characteristics of the wild animal using the various types of sensors to obtain the probability of the wild animal contracting an epidemic.

[0011] Step S5: Integrate the probability of the animal contracting the disease with the multi-source risk knowledge graph in the regional gateway, and use graph neural network reasoning to obtain the probability of the wild animal contracting the disease in the current region;

[0012] Step S6: When the probability of an outbreak of the wild animal disease exceeds the preset dynamic probability threshold, the regional gateway immediately sends an early warning to the prevention and control center.

[0013] Further, step S1 specifically includes: deploying multimodal monitoring nodes in wildlife habitats, consisting of a visible-infrared dual-spectrum imaging module, a bioaerosol detection module, a photoparticle gas detection module, an environmental parameter sensor array, and an acoustic sensor array;

[0014] A regional gateway is set up to communicate with each multimodal monitoring node, which is used to aggregate the monitoring and sensor datasets detected by the multimodal monitoring nodes and transmit them to the cloud processing unit.

[0015] Furthermore, the acquisition of wildlife video streams in step S2 specifically includes: continuously acquiring wildlife video streams using the visible light-infrared dual-spectrum imaging module among the various types of sensors;

[0016] The process of using a wildness AI model deployed in an edge computing unit to perform real-time inference on the video stream and output the wildness score of individual wild animals specifically includes: identifying and tracking individual wild animals in the video based on a lightweight target detection and tracking model; estimating the key skeleton points of the individual wild animals based on a lightweight pose estimation model; identifying the preset behavior categories and their probabilities of the individual wild animals based on the change sequence of the key skeleton points using a temporal behavior recognition model; and obtaining the wildness score of the individual wild animals based on the probabilities of the identified preset behavior categories and the pre-set weight coefficients of the corresponding behavior categories.

[0017] Furthermore, step S3 specifically includes:

[0018] Based on the species attributes, current time period, and microenvironment of the individual wild animal, the corresponding dynamic wildness threshold is retrieved from a predefined threshold mapping table; wherein, the current time period includes sunrise, sunset, noon, and midnight, and the microenvironment includes water sources, foraging areas, and open areas;

[0019] Within a preset continuous monitoring time window, if the proportion of sampling points with wildness scores below the dynamic wildness threshold in the wild animal individual's wildness score sequence exceeds a preset proportion threshold, then the individual is determined to be a high-risk individual for disease.

[0020] Furthermore, step S4 specifically includes:

[0021] The acoustic sensor array plays simulated foraging or mating sounds of the wild animal's species through its loudspeakers, attracting it to approach the corresponding multimodal monitoring node.

[0022] The surface temperature distribution of the wild animal was acquired using the visible-infrared dual-spectrum imaging module.

[0023] The concentration of airborne pathogens around the wild animal is obtained through the bioaerosol detection module;

[0024] The acoustic characteristics of the wild animal are acquired through the microphones of the acoustic sensor array;

[0025] The VOCs spectrum characteristics of the air surrounding the wild animal are obtained through the optical particle gas detection module.

[0026] The surface temperature distribution, sound characteristics, concentration of airborne pathogens, and VOCs spectrum characteristics of the wild animal are input into a pre-trained multimodal fusion classification model. The model outputs the probability value of the wild animal's possibility of having an epidemic, which is used as the degree of possibility of the wild animal having an epidemic.

[0027] Furthermore, step S5 specifically includes:

[0028] In the regional network management system, a multi-source risk knowledge graph is constructed and maintained. The multi-source risk knowledge graph stores and dynamically updates the following nodes and their associated edges in a graph structure: climate change nodes, geological disaster nodes, human epidemic nodes, and historical wildlife epidemic nodes.

[0029] By accessing Internet data APIs and satellite remote sensing data interfaces, when new risk events are detected, corresponding nodes and edges are automatically created in the multi-source risk knowledge graph, and each edge is assigned a dynamic weight that decays over time.

[0030] The probability of the wild animal contracting the disease is added to the knowledge graph as a wild animal disease individual node;

[0031] Triggered by the individual nodes of the wild animal disease and their probability of contracting the disease, the graph neural network model is launched to reason about the subgraph of the current region. The graph neural network aggregates the edge type relationships between the individual nodes of the wild animal disease and other nodes through a message passing mechanism, and generates an embedding vector representing the probability of the occurrence of the epidemic in the region as the probability of the occurrence of the wild animal epidemic in the current region.

[0032] Furthermore, step S6 specifically includes:

[0033] A dynamic probability threshold is set based on the probability of an outbreak occurring in relevant regional nodes within the multi-source risk knowledge graph;

[0034] The probability of an outbreak of an epidemic in the region is continuously monitored. When the probability of an outbreak of an epidemic of a wild animal in the current region exceeds a preset dynamic probability threshold for the first time, a preset first observation period is initiated. If, during the observation period, the proportion of sampling points in the region whose probability of an outbreak of an epidemic continues to exceed the preset dynamic probability threshold reaches a preset sampling point proportion, then the early warning condition is confirmed to be triggered.

[0035] When the warning conditions are triggered, the regional gateway automatically generates a structured warning information message, which includes at least: the geographical coordinate range of the warning area, the probability value of the warning epidemic and the dynamic probability threshold, the list of associated high-risk wild animal individual IDs and their probability of contracting the disease, the corresponding node in the associated main risk knowledge graph, the timestamp and the warning level.

[0036] The structured early warning information message is sent to the prevention and control center server in parallel through at least two different communication links; the communication links use a satellite communication link as the main channel and a cellular network link as the backup channel.

[0037] Furthermore, after step S6, the process includes: constructing a three-dimensional risk tensor of "species-location-time" based on the probability of wildlife epidemics, inputting the risk tensor into a reinforcement learning policy network, and automatically generating differentiated prevention and control instructions, including: risk area division, isolation early warning push, prevention and control material distribution plan, and blockchain evidence storage of the prevention and control process.

[0038] Furthermore, after step S6, the method further includes: using the risk tensor of each confirmed event, prevention and control instructions, and the on-site verified transmission results as new samples to update the graph neural network and the learning strategy network online.

[0039] Secondly, the present invention provides an artificial intelligence-based wildlife disease monitoring and control system for implementing the aforementioned artificial intelligence-based wildlife disease monitoring and control method, comprising:

[0040] The multimodal monitoring unit, deployed in wildlife habitats, consists of a visible-infrared dual-spectrum imaging module, a bioaerosol detection module, a photoparticle gas detection module, and an acoustic / environmental parameter sensor array. It is used to collect wildlife video streams, body surface temperature, airborne pathogen concentration, sound characteristics, VOCs characteristics, and environmental parameters.

[0041] The edge computing unit is communicatively connected to the multimodal monitoring unit and has a built-in wildness AI model for real-time inference of wildlife video streams and output of individual wildness scores.

[0042] The risk assessment unit is connected to the edge computing unit and pre-stores a dynamic wildness threshold mapping table of associated species attributes, time periods and microenvironments, which is used to determine whether wild animals are individuals with high disease risk.

[0043] A multimodal data acquisition unit, connected to the multimodal monitoring unit and the risk determination unit, is used to control the multimodal monitoring unit to collect data of an individual in a targeted manner when determining an individual with a high risk of disease, and output the degree of probability of the individual having the disease through a multimodal fusion classification model;

[0044] The regional gateway unit is communicatively connected to the multimodal data acquisition unit and has a built-in multi-source risk knowledge graph and graph neural network reasoning model, which is used to fuse the probability of disease occurrence with the knowledge graph and reason the probability of regional wildlife epidemics.

[0045] The early warning triggering unit is connected to the regional gateway unit and pre-stores a dynamic probability threshold. It is used to generate a structured early warning and send it to the prevention and control center when the probability of an epidemic occurrence exceeds the threshold.

[0046] The prevention and control instruction generation unit is connected to the regional gateway unit and has a built-in reinforcement learning policy network for constructing a three-dimensional risk tensor based on the probability of epidemic occurrence and generating differentiated prevention and control instructions.

[0047] This invention offers the following advantages: First, it uses collected video stream data to infer the wildness of individual wild animals and predict disease risks in real time. Then, it performs multimodal fusion using data from different sensors, such as sound characteristics, temperature, and pathogen concentration. Utilizing a pre-trained multimodal fusion classification model and graph neural network technology, it comprehensively infers the probability of wild animal disease occurrence, obtaining the likelihood of the wild animal contracting the disease. Finally, it combines this likelihood with a multi-source risk knowledge graph in a regional gateway and uses graph neural network inference to obtain the probability of disease occurrence in the current region. Compared to traditional single-data-source monitoring methods, this invention provides more comprehensive and accurate analysis results, effectively improving the accuracy of wild animal disease monitoring. Furthermore, through the deep integration of cutting-edge technologies such as multimodal perception, edge intelligence, knowledge graphs, graph neural networks, and reinforcement learning, this invention constructs an integrated intelligent system of "sensing-knowledge-judgment-alarm-control," ultimately achieving early detection, accurate diagnosis, risk prediction, rapid response, and precise prevention and control of wild animal diseases, which is beneficial for maintaining biosecurity, ecological security, and public health security. Attached Figure Description

[0048] Figure 1 This is a flowchart of an artificial intelligence-based method for monitoring and controlling wildlife diseases according to the present invention;

[0049] Figure 2 This is a schematic block diagram of an artificial intelligence-based wildlife disease monitoring and control system according to the present invention. Detailed Implementation

[0050] To further understand the content of this invention, a detailed description of the invention is provided in conjunction with the accompanying drawings and embodiments. The specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0051] like Figure 1 This embodiment provides a method for monitoring and controlling wildlife diseases based on artificial intelligence, including:

[0052] Step S1: Deploy multimodal monitoring nodes consisting of various types of sensors in wildlife habitats.

[0053] Specifically, step S1 includes: deploying multimodal monitoring nodes in wildlife habitats, consisting of a visible-infrared dual-spectrum imaging module, a bioaerosol detection module, a photoparticle gas detection module, an environmental parameter sensor array, and an acoustic sensor array; and setting up a regional gateway that communicates with each multimodal monitoring node to aggregate the monitoring and sensing data sets detected by the multimodal monitoring nodes and transmit them to the cloud processing unit.

[0054] More specifically, firstly, multimodal monitoring nodes are set up in wildlife habitats, consisting of visible-infrared dual-spectrum imaging modules, bioaerosol detection modules, photoparticle gas detection modules, environmental parameter sensor arrays, and acoustic sensor arrays. These sensor modules can collect monitoring data on various aspects, including animal behavior, animal temperature, environmental changes, and whether animals are diseased.

[0055] The system utilizes a visible-infrared dual-spectrum imaging module to continuously monitor animal activity, capture video streams, and analyze wildlife behavior. A bioaerosol detection module monitors airborne pathogen concentrations, providing crucial information for assessing animal health. Furthermore, a photoparticle-based gas detection module monitors VOCs (volatile organic compounds) in the surrounding environment to determine if wildlife is experiencing diarrhea. An acoustic sensor array captures animal sound characteristics, not only identifying coughs but also providing additional information for behavioral analysis. This data helps identify animal population structure and activity patterns. An environmental parameter sensor array collects environmental information such as temperature, humidity, air pressure, and wind speed, supporting the assessment of external environmental conditions for wildlife activity. All these data sources are collected through multimodal monitoring nodes, forming a comprehensive and accurate dataset.

[0056] Then, the data collected by these sensor modules is transmitted via wireless communication to the area gateway, which is connected to the multimodal monitoring nodes. The area gateway's function is to aggregate, process, and transmit the various types of data collected by the multimodal monitoring nodes to the cloud processing unit. Through the area gateway, all raw data can undergo simple preprocessing, such as noise reduction and data format conversion, before being transmitted to the cloud, to facilitate subsequent more complex data analysis and processing. During data transmission, efficient, low-latency communication protocols ensure the real-time nature and integrity of the data, guaranteeing timely transmission to the cloud processing system in any emergency to support further analysis and decision-making.

[0057] The cloud processing unit is responsible for more in-depth analysis and processing of the data transmitted through the regional gateway. At this point, the data, after being processed by algorithmic models, enables real-time monitoring of wildlife health status, disease prediction, and identification of environmental changes. The cloud processing unit combines the collected sensor data with other environmental data, utilizes artificial intelligence and big data technologies to intelligently analyze the data, and feeds the results back to the control center or relevant personnel to achieve comprehensive monitoring and precise management of wildlife habitats.

[0058] This technology integrates and coordinates multiple data sources to form a comprehensive, real-time monitoring system. This system not only monitors the behavior and health status of wild animals in real time but also adjusts and responds accordingly to environmental changes. The sensor data, aggregated and transmitted through regional gateways and deeply analyzed by cloud processing units, provides strong technical support for subsequent disease prevention and control and wildlife conservation efforts, significantly improving the accuracy and speed of disease early warning. This integrated multimodal data acquisition and processing approach solves the problems of traditional monitoring methods, such as the inability to comprehensively collect various types of data, poor real-time performance, and low prediction accuracy.

[0059] Step S2: Acquire wildlife video streams, perform real-time inference on the video streams using a wildness AI model deployed in an edge computing unit, and output the wildness score for each individual wildlife species.

[0060] Specifically, step S2 involves acquiring wildlife video streams by continuously acquiring wildlife video streams using the visible-infrared dual-spectrum imaging module among the various types of sensors. The step of performing real-time inference on the video stream using a wildness AI model deployed in the edge computing unit to output the wildness score of an individual wildlife species specifically includes: identifying and tracking individual wildlife species in the video based on a lightweight target detection and tracking model; estimating the key skeletal points of the individual wildlife species based on a lightweight pose estimation model; identifying the preset behavior categories and their probabilities of the individual wildlife species based on the change sequence of the key skeletal points using a temporal behavior recognition model; and obtaining the wildness score of the individual wildlife species based on the probabilities of the identified preset behavior categories and the pre-set weight coefficients of the corresponding behavior categories.

[0061] More specifically, this step involves using a wildness score AI model within an edge computing unit to perform real-time inference on wildlife video streams to output wildness scores for each individual. The implementation of this technical feature in the overall solution relies primarily on multiple processing steps and algorithm models. It combines a visible-infrared dual-spectrum imaging module, a target detection and tracking model, a pose estimation model, and a temporal behavior recognition model to process the acquired video stream and obtain accurate wildness scores. The video stream is acquired by continuously capturing wildlife behavior through a visible-infrared dual-spectrum imaging module. This dual-mode module provides clear video images under different lighting conditions, providing foundational data for subsequent image processing and analysis.

[0062] The acquired video stream is then fed into an AI model deployed in an edge computing unit. The AI ​​model performs preliminary processing on the video stream using a target detection and tracking module, identifying and tracking individual wild animals. This module uses computer vision technology to identify individual animals in the video and track their movement trajectories. The target detection and tracking model, after training, can accurately identify the species and individuals of wild animals in complex environments, ensuring the correct tracking of each individual in the dynamic video stream.

[0063] Subsequently, the movements and postures of each individual animal in the video stream are further processed, and key skeletal points of the animal are estimated using a posture estimation model. This model, based on deep learning technology, can accurately identify and locate key points in the body posture of wild animals, forming a sequence of skeletal points. By observing changes in the position of these skeletal points, the model can effectively capture changes in the animal's posture and activity patterns, providing accurate information for subsequent behavior recognition. Based on the captured skeletal point sequence, a temporal behavior recognition model analyzes the animal's behavior. This model identifies preset animal behavior categories and their corresponding probability values ​​by performing temporal analysis on the sequence of changes in skeletal points. The preset behavior categories can include common behaviors such as foraging, running, resting, and alertness. The temporal behavior recognition model can determine the animal's activity state within a specific time period based on the temporal changes in skeletal points and assign a probability value to each behavior category. Based on these probability values ​​and the weight coefficients of the corresponding behavior categories, the model finally calculates the wildness score of the individual wild animal.

[0064] Wildness score assesses an animal's condition based on its behavioral patterns, activity intensity, and behavioral characteristics, reflecting its wildness. A high wildness score reflects an animal's adaptability to its habitat, alertness, and the naturalness of its behavior; a higher score indicates better health, while a lower score suggests the animal may be diseased. By deploying AI models in edge computing units, real-time inference of video streams can be performed simultaneously with data acquisition, significantly reducing latency and bandwidth pressure on data transmission to the cloud, and improving the overall system's response speed and processing capacity. This step, by integrating multiple data processing technologies and algorithms, comprehensively and accurately reflects animal behavioral patterns, thereby improving the accuracy and response speed of wildlife disease monitoring and control. It effectively addresses the shortcomings of traditional methods that rely on manual observation or single data collection methods, resulting in low processing efficiency and susceptibility to environmental interference.

[0065] Step S3: When the wildness score is lower than the preset wildness threshold, the wild animal is determined to be an individual with a high risk of disease.

[0066] Specifically, step S3 includes: based on the species attributes, current time period, and microenvironment of the wild animal individual, calling the corresponding dynamic wildness threshold from a predefined threshold mapping table; wherein the current time period includes sunrise, sunset, noon, and midnight, and the microenvironment includes water sources, foraging areas, and open areas; within a preset continuous monitoring time window, if the proportion of sampling points with wildness scores below the dynamic wildness threshold in the wild animal individual's wildness score sequence exceeds a preset proportion threshold, then the individual is determined to be a high-risk individual for disease.

[0067] More specifically, the key to this step lies in determining whether a wild animal is at high risk of disease by comparing its wildness score with a dynamic threshold. This process first uses a predefined threshold mapping table to dynamically adjust and apply appropriate wildness thresholds based on the individual wild animal's species attributes, the current time period, and its microenvironment. The time periods include sunrise, sunset, noon, and midnight, which significantly impact animal behavior. For example, some animals may be more active or aggressive in the early morning or at dusk, so their dynamic wildness thresholds are correspondingly higher than those at other times. Furthermore, microenvironments such as water sources, foraging areas, or open areas also influence animal behavior, resulting in higher dynamic wildness thresholds in these areas compared to others. These different environmental conditions are reflected in the mapping table to accurately set the appropriate wildness thresholds.

[0068] After setting a dynamic wildness threshold, the system assesses whether an individual wild animal is at high risk of disease by analyzing its wildness score sequence. Specifically, this process is conducted within a pre-defined continuous monitoring window. If, within a certain time period, the proportion of sampling points in an individual wild animal's wildness score sequence that are below the set dynamic wildness threshold exceeds a pre-defined percentage threshold, that individual is identified as high-risk for disease. Through this mechanism, the system can dynamically adjust its risk assessment criteria based on changes in each individual's behavior and their environment. This dynamic assessment mechanism utilizes real-time data collected from sensors, processed by algorithms, to accurately determine disease risk. By combining time period and microenvironmental information and dynamically adjusting the threshold, the impact of different times and environments on animal behavior is systematically considered, thereby improving the system's accuracy in identifying high-risk individuals for disease.

[0069] Step S4: Collect the surface temperature of the wild animal, the concentration of airborne pathogens, sound characteristics, and VOCs characteristics using the various types of sensors to obtain the probability of the wild animal contracting an epidemic.

[0070] Specifically, step S4 includes: playing the foraging or courtship sounds of the wild animal's species through the loudspeaker of the acoustic sensor array to attract it to the multimodal monitoring node; acquiring the surface temperature distribution of the wild animal through the visible-infrared dual-spectrum imaging module; obtaining the concentration of airborne pathogens around the wild animal through the bioaerosol detection module; acquiring the sound characteristics of the wild animal through the microphone of the acoustic sensor array; acquiring the VOCs spectral characteristics of the air around the wild animal through the photoparticle gas detection module; inputting the surface temperature distribution, sound characteristics, airborne pathogen concentration, and VOCs spectral characteristics of the wild animal into a pre-trained multimodal fusion classification model, and having the model output the probability value of the wild animal's disease as the degree of possibility of the wild animal having a disease.

[0071] More specifically, step S4 uses various types of sensors to collect data on the surface temperature of wild animals, concentration of airborne pathogens, sound characteristics, and VOCs characteristics, thereby determining the likelihood of wild animals contracting diseases. This technique integrates multimodal sensor data and combines it with a pre-trained multimodal fusion classification model to analyze the health status and disease risk of wild animals.

[0072] In practice, the loudspeakers of the acoustic sensor array play simulated foraging or courtship sounds of the wild animals' species to attract them to the multimodal monitoring nodes. Attracting wild animals into the monitoring range ensures the validity and reliability of the sensor data and enhances the accuracy of data acquisition. The visible-infrared dual-spectrum imaging module collects real-time data on the animals' body surface temperature distribution. Body surface temperature is a crucial indicator of animal health, especially during disease outbreaks, where temperature changes are often an early symptom. Animal body surface temperature is primarily used to determine if an animal has a fever, reflecting signs of infection or physiological abnormalities. The bioaerosol detection module acquires the concentration of airborne pathogens around the wild animals, determining whether they are carrying or in an environment with pathogens. Timely monitoring of these changes helps in the early detection of potential disease transmission sources.

[0073] Meanwhile, the microphones of the acoustic sensor array collect the sound characteristics of wild animals, using coughing features in their vocalizations to determine if they have respiratory infections. The photoparticle-based gas detection module acquires VOCs spectral characteristics from the surrounding air; these VOCs spectra are used to determine if animals are experiencing diarrhea caused by intestinal infectious diseases, further providing data support for assessing environmental pollution or disease transmission.

[0074] The collected data is fed into a pre-trained multimodal fusion classification model. The model combines various feature data and, through algorithmic processing, outputs a probability value indicating the animal's likely disease. This multimodal fusion classification model integrates data from different sensors, including body surface temperature, pathogen concentration, sound characteristics, and VOC spectra, to comprehensively assess animal health risks from multiple dimensions. Through learning and optimization from historical data, the model can more accurately identify disease risks and calculate a probability value for each individual animal, representing its risk level. This multi-dimensional data analysis allows for a more precise assessment of animal health and disease risk, enabling timely detection of potential disease-carrying individuals and improving the accuracy and speed of disease early warning. By comprehensively processing data from different sources, the system fully utilizes the advantages of different sensors, compensating for the shortcomings of single-sensor data, ultimately achieving precise monitoring and effective control of wildlife diseases. This method can provide accurate disease risk assessments in complex and ever-changing environments, offering strong technical support for ecological protection and public health.

[0075] Step S5: Integrate the probability of the animal contracting the disease with the multi-source risk knowledge graph in the regional gateway, and use graph neural network reasoning to obtain the probability of the wild animal contracting the disease in the current region.

[0076] Specifically, step S5 includes: constructing and maintaining a multi-source risk knowledge graph in the regional network management system. The multi-source risk knowledge graph stores and dynamically updates the following nodes and their associated edges in a graph structure: climate change nodes, geological disaster nodes, human epidemic nodes, and historical wildlife epidemic nodes. By accessing internet data APIs and satellite remote sensing data interfaces, when a new risk event is detected, corresponding nodes and edges are automatically created in the multi-source risk knowledge graph, and each edge is assigned a dynamic weight that decays over time. The probability of the wild animal contracting an epidemic is added to the knowledge graph as a wild animal epidemic individual node. Triggered by the wild animal epidemic individual node and its probability of contracting an epidemic, a graph neural network model is started to infer the sub-graph of the current region. The graph neural network, through a message passing mechanism, aggregates the edge type relationships between the wild animal epidemic individual node and other nodes, generating an embedding vector representing the probability of an epidemic occurring in the region as the probability of the wild animal epidemic occurring in the current region.

[0077] More specifically, step S5 achieves intelligent assessment of wildlife disease risk by fusing the probability of animals contracting diseases with a multi-source risk knowledge graph in the regional gateway and using graph neural network inference to derive the probability of wildlife disease outbreaks in the current region. The implementation of this technology relies on several key steps, including constructing and maintaining a multi-source risk knowledge graph, dynamic data updates, graph neural network inference, and data fusion.

[0078] First, a multi-source risk knowledge graph is constructed, maintained, and dynamically updated within the regional gateway. Node types include: individual wildlife nodes, wildlife population nodes, climate change event nodes, geological disaster event nodes, human community epidemic nodes, historical wildlife epidemic nodes, and geographic region nodes. Edge types represent relationships between nodes, including: location, occurrence, belonging, adjacency, temporal sequence, and causal relationships. Dynamic updates involve automatically creating corresponding nodes and edges in the knowledge graph when new risk events or data are detected, through access to internet data APIs, satellite remote sensing data interfaces, and regional sensor networks. Each edge is assigned a dynamic weight that decays over time. This weight represents the change and decay of risk over time, ensuring the system can dynamically reflect the impact of various environmental and external conditions on wildlife health. Through these steps, the multi-source risk knowledge graph is continuously updated and optimized, providing accurate data support reflecting the real-world situation.

[0079] Corresponding to the potentially diseased wild animals monitored in step S4, the occurrence of historical wildlife epidemic nodes in the time dimension, or the occurrence of climate change events, geological disaster events, or human community epidemic nodes in the surrounding human settlements at the current time point, will greatly increase the actual probability of the wild animal contracting the disease. Therefore, these nodes need to be considered as factors in determining the accurate probability of the wild animal contracting the disease, based on the degree of likelihood of it contracting the disease.

[0080] Then, the high-risk individuals and their probability values ​​of contracting the disease obtained in step S4 are added to the knowledge graph as new "wildlife individual nodes". Relationship edges of "located in" and "belonging to" are established between these individuals and their respective "geographical region nodes" and "wildlife population nodes". The newly added individual node and its probability value trigger the inference of the subgraph of the current region by the graph neural network model. The graph neural network aggregates the features and relational weights of the multi-hop neighbor nodes (such as climate, geology, and historical epidemic nodes) of the individual node through a message passing mechanism, ultimately generating an embedding vector representing the probability of an epidemic occurring in the corresponding "geographical region node". The embedding vector of the geographic region node is then mapped to a regional epidemic probability value between 0 and 100% through a fully connected layer classifier, serving as the probability of an epidemic occurring in the current region for that wild animal.

[0081] In this way, compared with traditional single-data monitoring systems, this system not only considers the health data of wild animals, but also incorporates various external factors such as the environment, climate change, and historical epidemics. Through multi-source data fusion and intelligent reasoning using graph neural networks, it can more comprehensively and dynamically reflect epidemic risks, avoiding misjudgments caused by isolated or outdated data in traditional methods. The system's flexibility and adaptability improve the accuracy of risk warnings and provide more efficient technical support for the prevention and control of wildlife diseases.

[0082] Step S6: When the probability of an outbreak of the wild animal disease exceeds the preset dynamic probability threshold, the regional gateway immediately sends an early warning to the prevention and control center.

[0083] Specifically, step S6 includes: setting a dynamic probability threshold based on the probability of an outbreak in relevant regional nodes in the multi-source risk knowledge graph; continuously monitoring the probability of an outbreak in the region, and initiating a preset first-duration observation period when the probability of an outbreak of this wild animal in the current region exceeds the preset dynamic probability threshold for the first time; if, during the observation period, the proportion of sampling points whose probability of an outbreak in the region continuously exceeds the preset dynamic probability threshold reaches a preset sampling point proportion, then a warning condition is confirmed to be triggered; when the warning condition is triggered, the regional gateway automatically generates a structured warning information message, which includes at least: the geographical coordinate range of the warning area, the warning outbreak probability value and dynamic probability threshold, the list of associated high-risk wild animal individual IDs and their disease probability, the corresponding node in the associated main risk knowledge graph, the timestamp, and the warning level; and sending the structured warning information message in parallel to the prevention and control center server through at least two different communication links; the communication links use a satellite communication link as the main channel and a cellular network link as the backup channel.

[0084] More specifically, a dynamic probability threshold is first set based on the probability of outbreaks in relevant regional nodes within a multi-source risk knowledge graph. This dynamic threshold is adjusted according to changes in different risk factors within the region, such as climate change, environmental pollution, and historical outbreaks, ensuring the system can respond in real time and adapt to different environmental and health risks. When a regional gateway begins continuously monitoring the probability of outbreaks in its region, the system determines whether the probability of outbreaks in wild animals in the current region exceeds the preset dynamic probability threshold. If the probability exceeds the threshold for the first time, a preset observation period is initiated. This time window is used to further confirm whether the outbreak is persistent. During the observation period, if the probability of outbreaks in the region continues to exceed the set threshold, and the proportion of sampling points below the threshold exceeds a preset percentage threshold, then an early warning condition is triggered. This process, by combining historical outbreak data with real-time monitoring data, ensures a timely response in the early stages of an outbreak, avoiding the possibility of false alarms and missed reports.

[0085] When the warning conditions are met, the system automatically generates a structured warning message. This message includes the geographic coordinates of the warning area, the probability value and dynamic probability threshold of the epidemic, a list of associated high-risk wild animal individual IDs and their disease probability, key risk knowledge graph nodes, timestamps, and the warning level. This information is organized into a standardized message format for easy reception and subsequent processing by the control center. By combining epidemic risk data with geographic information and other important indicators, the warning message provides the control center with comprehensive and clear epidemic information.

[0086] To ensure the reliability and real-time nature of information transmission, the structured early warning messages are sent to the control center server in parallel through at least two different communication links. The primary communication link uses a satellite communication link, serving as the primary channel for information transmission. This link has strong anti-interference capabilities and long-distance transmission capacity, making it suitable for remote areas and locations with unstable signals. The backup link uses a cellular network link to ensure the stability of data transmission in the event of a failure of the primary link. Through this multi-link transmission design, the system improves the reliability and real-time nature of information transmission, ensuring that the control center can obtain early warning information and respond to emergencies in the first instance.

[0087] Further, after step S6, the process includes: constructing a three-dimensional risk tensor of "species-location-time" based on the probability of wildlife epidemic occurrence; inputting the risk tensor into a reinforcement learning strategy network to automatically generate differentiated prevention and control instructions, including: risk area division, isolation early warning push, prevention and control material distribution plan, and blockchain-based evidence storage of the prevention and control process. The risk tensor, prevention and control instructions, and on-site verified transmission results of each confirmed case are used as new samples to update the graph neural network and the learning strategy network online.

[0088] More specifically, after step S6, a three-dimensional risk tensor based on the probability of wildlife epidemic occurrence is constructed as "species-location-time." This risk tensor is then input into a reinforcement learning policy network to automatically generate differentiated prevention and control instructions. This process involves combining the probability of epidemic occurrence with the changing relationships between different species, locations, and times to form a three-dimensional risk tensor, providing more comprehensive and accurate epidemic predictions. In this process, the risk tensor integrates epidemic data from various regions, considers the risk assessment of different species at different locations and times, and then analyzes and makes decisions within the reinforcement learning policy network.

[0089] The three-dimensional risk tensor quantifies the risk of an outbreak by combining species attributes, location, and time. Changes in each dimension are crucial for outbreak prediction. For example, the activity patterns of a specific species at different times may affect its probability of exposure to pathogens, while different geographical locations may have different environmental conditions, affecting the speed and scope of disease transmission. Through this three-dimensional structure, the system can capture the complex relationships in the spread of an outbreak, thus providing rich information for subsequent prevention and control decisions.

[0090] After inputting the risk tensor into the reinforcement learning policy network, the system automatically generates differentiated prevention and control instructions based on the current epidemic data and risk assessment within the region. These instructions include the delineation of risk areas, the dissemination of isolation and early warning systems, the allocation of prevention and control materials, and blockchain-based evidence storage of the prevention and control process. Risk area delineation classifies different regions based on the probability of epidemic occurrence, ensuring that high-risk areas receive priority attention; isolation and early warning systems notify relevant personnel in high-risk areas to take measures; the allocation of prevention and control materials ensures that materials reach severely affected areas as quickly as possible, reducing the speed of epidemic spread; and blockchain-based evidence storage is used for traceability of the prevention and control process, ensuring that all operations are recorded and that the implementation of prevention and control measures is transparent and auditable.

[0091] After prevention and control instructions are generated, the risk tensor of each confirmed event, the prevention and control instructions, and the on-site verified transmission results are used as new samples and input into the graph neural network and learning strategy network for online updates. Through continuous online learning, the system can continuously optimize and adjust prevention and control strategies based on data feedback from real-world scenarios. By processing these new samples, the graph neural network can continuously adjust the relationships between nodes, update the risk assessment model in real time, and enhance the system's adaptability to complex epidemic transmission patterns. The reinforcement learning strategy network, based on continuously updated risk data, adjusts the strategy generation method to further improve the accuracy and timeliness of epidemic prevention and control decisions.

[0092] Secondly, please refer to Figure 2 This application provides an artificial intelligence-based wildlife disease monitoring and control system. This system is used to implement an artificial intelligence-based wildlife disease monitoring and control method as described in this embodiment, including:

[0093] The multimodal monitoring unit 101, deployed in wildlife habitats, consists of a visible-infrared dual-spectrum imaging module, a bioaerosol detection module, a photoparticle gas detection module, and an acoustic / environmental parameter sensor array. It is used to collect wildlife video streams, body surface temperature, airborne pathogen concentration, sound characteristics, VOCs characteristics, and environmental parameters.

[0094] Edge computing unit 102 is communicatively connected to the multimodal monitoring unit and has a built-in wildness AI model for real-time inference of wildlife video streams and output of individual wildness scores;

[0095] The risk determination unit 103 is connected to the edge computing unit and pre-stores a dynamic wildness threshold mapping table of associated species attributes, time periods and microenvironments, which is used to determine whether wild animals are individuals with high disease risk.

[0096] The multimodal data acquisition unit 104 is connected to the multimodal monitoring unit and the risk determination unit. When determining an individual with a high risk of disease, it controls the multimodal monitoring unit to collect data of that individual in a targeted manner and outputs the probability of that individual having the disease through a multimodal fusion classification model.

[0097] The regional gateway unit 105 is communicatively connected to the multimodal data acquisition unit and has a built-in multi-source risk knowledge graph and graph neural network reasoning model, which is used to fuse the probability of disease occurrence with the knowledge graph and reason the probability of regional wildlife epidemics.

[0098] The early warning triggering unit 106 is connected to the regional gateway unit and pre-stores a dynamic probability threshold. It is used to generate a structured early warning and send it to the prevention and control center when the probability of an epidemic occurrence exceeds the threshold.

[0099] The prevention and control instruction generation unit 107 is connected to the regional gateway unit and has a built-in reinforcement learning policy network for constructing a three-dimensional risk tensor based on the probability of an epidemic and generating differentiated prevention and control instructions.

[0100] This embodiment integrates advanced technologies such as multimodal sensors, edge computing, graph neural networks, and reinforcement learning to provide an efficient and accurate wildlife disease monitoring and control system. This system can comprehensively monitor wildlife behavior, vital signs, and environmental data, and assess the risk of disease outbreaks in real time based on this data. Through the integration of multi-source data and dynamically adjusted risk thresholds, the system can flexibly respond to complex ecological environments and epidemic risks, achieving accurate monitoring of wildlife health. The Wildness Degree AI model, combined with target detection, posture estimation, and behavior recognition technologies, ensures accurate tracking and analysis of wildlife behavior, thus providing a reliable basis for subsequent disease risk assessment.

[0101] Furthermore, by constructing and maintaining a multi-source risk knowledge graph, the system incorporates environmental factors such as climate change and historical epidemics, and combines this with graph neural networks for intelligent reasoning, further enhancing the accuracy of epidemic probability prediction. Based on this, the system can automatically generate differentiated prevention and control instructions according to dynamic risk changes, including risk area delineation, isolation warnings, material distribution plans, and blockchain-based evidence storage of the prevention and control process, ensuring the real-time nature and transparency of prevention and control measures. Compared with traditional methods, the technical solution of this invention significantly improves the accuracy, real-time nature, and response speed of epidemic prevention and control, avoiding the problem of untimely prevention and control caused by isolated or delayed data processing. By continuously updating the learning strategy network and graph neural network online, the system can continuously optimize prevention and control strategies to cope with complex epidemic transmission patterns. This system provides an innovative technical means for the prevention and control of wildlife diseases, with strong adaptability and scalability, effectively controlling epidemics and providing strong support for ecological protection and public health security.

[0102] The above description is merely a specific implementation of this specification. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the scope of protection of this specification is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this specification, and these modifications or substitutions should all be covered within the scope of protection of this specification.

Claims

1. A method for monitoring and controlling wildlife diseases based on artificial intelligence, characterized in that, include: Step S1: Deploy multimodal monitoring nodes consisting of various types of sensors in wildlife habitats; Step S2: Acquire wildlife video streams, perform real-time inference on the video streams using a wildness AI model deployed in an edge computing unit, and output the wildness score for each individual wildlife species. Step S3: When the wildness score is lower than the preset wildness threshold, the wild animal is determined to be an individual with a high risk of disease. Step S4: Collect the surface temperature, airborne pathogen concentration, sound characteristics, and VOCs characteristics of the wild animal using the various types of sensors to obtain the probability of the wild animal contracting an epidemic. Step S5: Integrate the probability of the animal contracting the disease with the multi-source risk knowledge graph in the regional gateway, and use graph neural network reasoning to obtain the probability of the wild animal contracting the disease in the current region; Step S6: When the probability of an outbreak of the wild animal disease exceeds the preset dynamic probability threshold, the regional gateway immediately sends an early warning to the prevention and control center; Specifically, step S1 includes: deploying multimodal monitoring nodes in wildlife habitats, consisting of a visible-infrared dual-spectrum imaging module, a bioaerosol detection module, a photoparticle gas detection module, an environmental parameter sensor array, and an acoustic sensor array; and setting up a regional gateway that communicates with each multimodal monitoring node to aggregate the monitoring and sensing data sets detected by the multimodal monitoring nodes and transmit them to the cloud processing unit. The acquisition of wildlife video streams in step S2 specifically includes: continuously acquiring wildlife video streams using the visible light-infrared dual-spectrum imaging module among the various types of sensors; the real-time inference of the video streams by the wildness AI model deployed in the edge computing unit to output the wildness score of individual wildlife specifically includes: identifying and tracking individual wildlife in the video based on a lightweight target detection and tracking model, estimating the key skeleton points of the individual wildlife based on a lightweight pose estimation model, identifying the preset behavior categories and their probabilities of the individual wildlife based on the change sequence of the key skeleton points using a temporal behavior recognition model, and obtaining the wildness score of the individual wildlife based on the probabilities of the identified preset behavior categories and the corresponding pre-set weight coefficients of the behavior categories; Step S3 specifically includes: based on the species attributes, current time period, and microenvironment of the wild animal individual, calling the corresponding dynamic wildness threshold from a predefined threshold mapping table; wherein, the current time period includes sunrise, sunset, noon, and midnight, and the microenvironment includes water sources, foraging areas, and open areas; within a preset continuous monitoring time window, if the proportion of sampling points with wildness scores below the dynamic wildness threshold in the wild animal individual's wildness score sequence exceeds a preset proportion threshold, then the individual is determined to be a high-risk individual for disease. Step S5 specifically includes: constructing and maintaining a multi-source risk knowledge graph in the regional network management system, wherein the multi-source risk knowledge graph stores and dynamically updates the following nodes and their associated edges in a graph structure: climate change nodes, geological disaster nodes, human epidemic nodes, and historical wildlife epidemic nodes; by accessing Internet data APIs and satellite remote sensing data interfaces, when a new risk event is detected, corresponding nodes and edges are automatically created in the multi-source risk knowledge graph, and each edge is assigned a dynamic weight that decays over time; the probability of the wildlife contracting an epidemic is added to the knowledge graph as a wildlife epidemic individual node; triggered by the wildlife epidemic individual node and its probability of contracting an epidemic, a graph neural network model is started to reason about the subgraph of the current region. The graph neural network aggregates the edge type relationships between the wildlife epidemic individual node and other nodes through a message passing mechanism, and generates an embedding vector representing the probability of an epidemic occurring in the region as the probability of the wildlife epidemic occurring in the current region; Step S6 specifically includes: setting a dynamic probability threshold based on the probability of an outbreak in relevant regional nodes in the multi-source risk knowledge graph; continuously monitoring the probability of an outbreak in the region, and initiating a preset first-duration observation period when the probability of an outbreak of this wild animal in the current region exceeds the preset dynamic probability threshold for the first time; if, during the observation period, the proportion of sampling points whose probability of an outbreak in the region continuously exceeds the preset dynamic probability threshold reaches a preset sampling point proportion, then a warning condition is confirmed to be triggered; when the warning condition is triggered, the regional gateway automatically generates a structured warning information message, which includes at least: the geographical coordinate range of the warning area, the warning outbreak probability value and dynamic probability threshold, the list of associated high-risk wild animal individual IDs and their disease probability, the corresponding node in the associated main risk knowledge graph, the timestamp and the warning level; and sending the structured warning information message to the prevention and control center server in parallel through at least two different communication links; the communication links use a satellite communication link as the main channel and a cellular network link as the backup channel.

2. The method for monitoring and controlling wildlife diseases based on artificial intelligence as described in claim 1, characterized in that, Step S4 specifically includes: The acoustic sensor array plays simulated foraging or mating sounds of the wild animal's species through its loudspeakers, attracting it to approach the corresponding multimodal monitoring node. The surface temperature distribution of the wild animal was acquired using the visible-infrared dual-spectrum imaging module. The concentration of airborne pathogens around the wild animal is obtained through the bioaerosol detection module; The acoustic characteristics of the wild animal are acquired through the microphones of the acoustic sensor array; The VOCs spectrum characteristics of the air surrounding the wild animal are obtained through the optical particle gas detection module. The surface temperature distribution, sound characteristics, concentration of airborne pathogens, and VOCs spectrum characteristics of the wild animal are input into a pre-trained multimodal fusion classification model. The model outputs the probability value of the wild animal's possibility of having an epidemic, which is used as the degree of possibility of the wild animal having an epidemic.

3. The method for monitoring and controlling wildlife diseases based on artificial intelligence as described in claim 1, characterized in that, Step S6 and beyond includes: constructing a three-dimensional risk tensor of "species-location-time" based on the probability of wildlife epidemics, inputting the risk tensor into a reinforcement learning policy network, and automatically generating differentiated prevention and control instructions, including: risk area division, isolation early warning push, prevention and control material distribution plan, and blockchain evidence storage of the prevention and control process.

4. The method for monitoring and controlling wildlife diseases based on artificial intelligence as described in claim 3, characterized in that, Step S6 is followed by: using the risk tensor of each confirmed event, prevention and control instructions, and the on-site verified transmission results as new samples to update the graph neural network and the learning strategy network online.

5. A wildlife disease monitoring and control system based on artificial intelligence, characterized in that, A method for monitoring and controlling wildlife diseases based on artificial intelligence, as described in any one of claims 1 to 4, comprises: The multimodal monitoring unit, deployed in wildlife habitats, consists of a visible-infrared dual-spectrum imaging module, a bioaerosol detection module, a photoparticle gas detection module, and an acoustic / environmental parameter sensor array. It is used to collect wildlife video streams, body surface temperature, airborne pathogen concentration, sound characteristics, VOCs characteristics, and environmental parameters. The edge computing unit is communicatively connected to the multimodal monitoring unit and has a built-in wildness AI model for real-time inference of wildlife video streams and output of individual wildness scores. The risk assessment unit is connected to the edge computing unit and pre-stores a dynamic wildness threshold mapping table of associated species attributes, time periods and microenvironments, which is used to determine whether wild animals are individuals with high disease risk. A multimodal data acquisition unit, connected to the multimodal monitoring unit and the risk determination unit, is used to control the multimodal monitoring unit to collect data of an individual in a targeted manner when determining an individual with a high risk of disease, and output the degree of probability of the individual having the disease through a multimodal fusion classification model; The regional gateway unit is communicatively connected to the multimodal data acquisition unit and has a built-in multi-source risk knowledge graph and graph neural network reasoning model, which is used to fuse the probability of disease occurrence with the knowledge graph and reason the probability of regional wildlife epidemics. The early warning triggering unit is connected to the regional gateway unit and pre-stores a dynamic probability threshold. It is used to generate a structured early warning and send it to the prevention and control center when the probability of an epidemic occurrence exceeds the threshold. The prevention and control instruction generation unit is connected to the regional gateway unit and has a built-in reinforcement learning policy network for constructing a three-dimensional risk tensor based on the probability of epidemic occurrence and generating differentiated prevention and control instructions.

Images