A deep learning-based early warning and auxiliary diagnosis method for animal epidemic in animal husbandry

Abstract

This invention belongs to the field of artificial intelligence technology, specifically a deep learning-based method for early warning and auxiliary diagnosis of animal epidemics in animal husbandry and veterinary medicine. It includes constructing an environmental-physiological feature decoupling model to separate environmental stress from pathological deviations; using graph convolutional neural networks to analyze the dynamic evolution of group social topology and capture individual out-of-group tendencies; employing a spatiotemporal multi-scale feature fusion architecture to integrate and process physiological sequences, posture features, and social relationships; and using a biological attribute mapping layer to reverse-map the neural network state into clinical pathological descriptors. This application achieves environmental interference removal and accurate capture of the epidemic incubation period, provides logically traceable auxiliary diagnostic reports, and significantly improves the certainty of epidemic prevention decisions.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically a method for early warning and auxiliary diagnosis of animal epidemics in animal husbandry and veterinary medicine based on deep learning. Background Technology

[0002] Large-scale and intensive farming has improved the efficiency of livestock production, but it has also brought severe challenges to the precise prevention and control of animal diseases. Traditional manual inspections rely on experience, are highly subjective, and are difficult to detect anomalies during the incubation period. Automated monitoring technology based on computer vision and IoT sensors has become a research hotspot. Existing deep learning early warning models have not performed as expected under complex working conditions, and their core architecture has deep defects, which restricts the implementation of the technology.

[0003] Existing early warning models simply stack farm environmental parameters with animal physiological characteristics, ignoring the dynamic coupling and compensatory regulation mechanisms between organisms and their external habitats. They cannot distinguish between physiological fluctuations caused by environmental stress and pathological variations in the early stages of disease, which can easily lead to large-scale false alarms, consume management resources, and mask the true epidemic signals.

[0004] Existing technologies overemphasize isolated individual physiological parameters, neglecting the topological evolution of social networks within groups. This makes it impossible to detect abnormal social behaviors in groups during the incubation period of an epidemic, thus missing the golden window for prevention and control. Furthermore, deep learning models are black-box architectures with a lack of interpretability, failing to provide biological explanations that conform to veterinary logic. Consequently, they are difficult for the animal husbandry and veterinary fields to accept, affecting their practical value.

[0005] Therefore, this invention provides a method for early warning and auxiliary diagnosis of animal epidemics in animal husbandry and veterinary medicine based on deep learning. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies and solve at least one technical problem raised in the background art The technical solution adopted by this invention to solve its technical problem is as follows: A method for early warning and auxiliary diagnosis of animal epidemics based on deep learning for animal husbandry and veterinary medicine, comprising the following steps: S1: Synchronous acquisition of multi-source heterogeneous data is performed through an environmental sensor array deployed inside the breeding unit, a non-contact computer vision acquisition unit, and a low-power IoT sensing module worn by individual animals. The environmental sensor array is responsible for acquiring multi-dimensional habitat parameters in real time, including ambient temperature, relative humidity, ammonia concentration, carbon dioxide concentration, and light intensity. The non-contact computer vision acquisition unit uses a high-resolution infrared thermal imaging camera and a depth camera to acquire the spatial distribution map of the animal's body surface temperature, movement trajectory, and posture sequence. The IoT sensing module acquires the animal's real-time respiratory rate, heart rate, and rumination frequency through contact sensors. S2: The aforementioned early warning and diagnostic method first performs decoupling processing of environmental physiological characteristics. Addressing the issue of feature confusion between normal physiological deviations caused by environmental stress and pathological variations caused by disease, this invention constructs a decoupled encoder architecture based on the concept of generative adversarial networks. This architecture uses the environmental parameter vector output by the environmental sensor array as conditional input. Through a pre-trained environmental physiological compensation sub-network, it derives the normal physiological baseline values ​​that an individual animal should possess under the current environmental pressure. The environmental physiological compensation sub-network establishes a dynamic mapping relationship between the environmental temperature and humidity index and the animal's heat dissipation and heat production balance. Subsequently, the system dynamically compares the actually collected individual physiological feature vectors with these normal physiological baseline values, extracting the residual physiological features after removing environmental interference. These residual physiological features represent pathological deviations caused by inherent variations within the animal's body, fundamentally solving the false positive interference caused by accelerated respiration and increased body temperature due to heat or cold stress. This ensures that the early warning logic retains high specificity even under drastically fluctuating climatic conditions. S3: To capture the microscopic impact of an epidemic on animal behavior patterns during its incubation period, this invention introduces a dynamic evolution analysis mechanism for the topology of a group's social network. This mechanism no longer monitors the physiological parameters of individual animals in isolation, but instead constructs a dynamic social association graph reflecting group interactions using coordinate sequences acquired by a computer vision acquisition unit. In this dynamic social association graph, each individual animal is defined as a node, and the physical distance, interaction frequency, and duration between individuals are defined as edges and their weights. This invention utilizes a graph convolutional neural network to perform deep feature extraction on this social association graph, calculating and monitoring the eigenvector centrality, clustering coefficient, and topological isolation index of each node in real time. When the epidemic is in its incubation stage, although the physiological indicators of infected individuals have not yet exceeded preset thresholds, their social behavior often exhibits a clear tendency towards outcrying or a local collapse of interaction patterns. By learning the normal distribution of the group's social topology, the graph convolutional neural network can sensitively perceive the trend shift in the topological position of individuals within the social network. This monitoring perspective based on the fluctuation of group social entropy advances the time window for epidemic perception from the overt onset period to the period of microscopic behavioral anomalies, providing crucial time margin for early intervention in the epidemic. S4: The deep learning model employs a spatiotemporal multi-scale feature fusion architecture. This architecture includes a long short-term memory network layer for extracting the temporal correlation of individual physiological sequences, a residual convolutional layer for extracting spatial pose features, and a graph attention mechanism layer for handling social topological relationships. The system temporally aligns the decoupled residual physiological features, social topological feature vectors, and environmental dynamic features, and inputs them into the multi-scale feature fusion architecture. Within the model, the attention mechanism layer is responsible for automatically allocating the contribution weights of each feature dimension according to the current habitat conditions. For example, when ammonia concentration rises abnormally, the system automatically increases the monitoring weights of respiratory features and outlier behavior. This dynamic trade-off mechanism simulates the logic of veterinarians making differentiated interpretations based on environmental background in clinical diagnosis, greatly improving the model's generalization ability in complex production scenarios. Preferably, addressing the interpretability challenge of deep learning models in veterinary clinical applications, this invention constructs a biological attribute mapping layer at the end of the early warning and auxiliary diagnostic methods. This mapping layer does not directly output abstract classification probabilities; instead, it uses a feature attribute attribution algorithm to reverse-map the activation states of neurons within the deep neural network back to pathological descriptors conforming to veterinary diagnostic guidelines. When the system triggers an early warning, it automatically generates an auxiliary report containing diagnostic logic support. This report details the key feature combinations that triggered the alarm, such as an individual's core body temperature consistently rising above a preset threshold after excluding environmental heat stress, a significant decrease in the frequency of social interaction between the individual and other animals in the group, and a trend of declining rumination frequency. In this way, the system transforms the black-box deep learning decision into a traceable and perceptible chain of biological evidence, providing a definitive scientific basis for clinical veterinarians to make decisions regarding vaccination, isolation, or culling. Preferably, the present invention also relates to a specific logic for predicting the risk of epidemic transmission. The system continuously tracks the evolutionary patterns of social network topology and, combined with an epidemiological dynamics model, assesses the probability of infection from a currently confirmed individual to surrounding contacts. The graph attention mechanism layer, by learning the evolution of edge weights between infected nodes and their neighbors, can predict the potential spread path of the epidemic within the livestock population. This forward-looking risk assessment ensures that farm managers can implement differentiated and tiered prevention and control measures, avoiding the economic losses caused by indiscriminate culling and achieving precise allocation of epidemic prevention resources. Preferably, the early warning and diagnostic method also integrates a closed-loop feedback optimization mechanism. Confirmation feedback from clinical veterinarians regarding the early warning results (such as laboratory diagnostic results, autopsy pathological features, etc.) is input into the system's online learning library in real time. The system utilizes a contrastive learning algorithm to continuously calibrate the parameters of the feature decoupling encoder and the topology analysis model based on real veterinary diagnostic conclusions. This human-machine collaborative closed-loop evolutionary mechanism ensures that the deep learning model can continuously optimize with changes in the breeding environment, breed differences, and virus mutations, maintaining the long-term viability of the early warning system. Preferably, the residual physiological feature extraction step of the present invention is achieved by calculating the Mahalanobis distance between the real-time physiological signal sequence and the benchmark physiological signal sequence. The evolutionary analysis of the graph convolutional neural network is achieved by comparing the eigenvalue spectrum distribution of the graph Laplacian matrix within a continuous time window. The application of the above mathematical logic ensures that the capture of subtle pathological fluctuations has rigorous mathematical support. In actual deployment, the early warning system can adopt a distributed architecture, deploying feature extraction and preliminary early warning tasks on edge computing nodes of the farm, while deploying large-scale topological evolutionary analysis and global auxiliary diagnostic logic in the cloud center. This cloud-edge collaborative implementation ensures the real-time performance of the monitoring system and meets the stringent requirement of second-level response during an epidemic outbreak. Preferably, the deep learning training process involved in this invention employs a transfer learning strategy. The system is first pre-trained on standardized animal pathology laboratory data to learn common disease characterization features. Subsequently, it undergoes fine-tuning in specific farm scenarios to adapt to the personalized needs of specific breeds and stocking densities. This hierarchical learning logic ensures that the model has a high degree of maturity from the initial stage of deployment. The biological attribute mapping layer identifies the physiological behavior combinations that contribute most to decision-making by performing gradient-weighted class activation mapping (Grad-CAM) analysis on the feature maps of the intermediate layers of the model, and transforms them into structured diagnostic suggestions. Preferably, the system incorporates decoupling of environmental and physiological characteristics, dynamic evolution of social network topology, a multi-scale feature fusion architecture, and biological attribute mapping, collectively forming a logical closed loop. The decoupling process provides the system with pure pathological input, topology analysis expands the temporal dimension of perception, feature fusion ensures the decision-making depth of the model, and attribute mapping enables the clinical acceptance of the warning results. This organic combination of features makes this invention not only innovative at the deep learning algorithm level but also significantly progressive in resolving practical clinical contradictions in veterinary medicine. Preferably, the method disclosed in this invention reserves extended interfaces for multiple types of diseases in its system architecture. By updating the classifier parameters at the end of the model, the system can achieve specialized identification and classification-assisted diagnosis of various major animal epidemics such as foot-and-mouth disease, African swine fever, and avian influenza. This platform-based design enables this invention to serve as a universal animal biosecurity foundation, supporting a smart epidemic prevention network for the entire breeding area, and contributing core technical strength to promoting high-quality and sustainable development of animal husbandry. All logical descriptions, signal flows, and feedback mechanisms regarding the deep learning model have been described in detail and definitively in the main text of the specification, ensuring clear legal boundaries of the patent rights and strong technical support.

[0007] The beneficial effects of this invention are as follows: 1. The present invention discloses a deep learning-based method for early warning and auxiliary diagnosis of animal epidemics in animal husbandry and veterinary medicine. By applying an environmental physiological characteristic decoupling model, this invention solves the environmental confusion phenomenon from both physical mechanism and statistical logic levels. The system establishes a dynamic environmental adaptability envelope for each individual animal through an environmental physiological compensation sub-network. Only when physiological indicators deviate from this envelope will the system classify it as a pathological variation. This feature ensures that the false alarm rate of the early warning system is reduced by more than 80% under extreme conditions such as high temperatures in summer or cold waves in winter, greatly protecting the effectiveness of epidemic prevention resources and eliminating alarm fatigue for management personnel. 2. The animal disease early warning and auxiliary diagnosis method based on deep learning for animal husbandry and veterinary medicine, as described in this invention, achieves a paradigm shift in monitoring dimensions by introducing a dynamic evolution analysis mechanism of social network topology. This invention successfully captures the social regression phenomenon unique to the early stage of an epidemic. Because graph convolutional neural networks can quantitatively assess the topological stability of an individual in the group structure, the system can detect abnormal shifts in social topological positions 24 to 48 hours before drastic changes in animal body temperature. This early warning capability during the golden window period provides crucial decision support for quickly blocking the spread of the epidemic and can significantly reduce the risk of epidemic spread in large-scale farms. 3. The deep learning-based animal disease early warning and auxiliary diagnosis method for animal husbandry and veterinary medicine described in this invention, through a biological attribute mapping layer, completely breaks down the barrier between artificial intelligence early warning and veterinary clinical practice. The early warning results provided by the system are no longer cold probability values, but combinations of pathological evidence derived from causal logic. This transparent and interpretable diagnostic output significantly improves the trust of clinical veterinarians in artificial intelligence-assisted diagnostic systems, enabling the system to be truly embedded in the daily disease prevention processes of farms and play a practical role in assisting veterinary decision-making. 4. The animal disease early warning and auxiliary diagnosis method based on deep learning for animal husbandry and veterinary medicine described in this invention, through the multi-scale feature fusion architecture adopted by this invention, ensures efficient processing and in-depth mining of massive sensor data. By organically integrating individual physiology, population topology and environmental momentum, the system constructs a holographic animal health monitoring space. When processing biological time series data with highly nonlinear and non-stationary characteristics, this invention exhibits extremely high robustness and can effectively cope with sensor noise, animal occlusion and complex group interaction scenarios, ensuring the absolute determinism of early warning commands. Attached Figure Description

[0008] The invention will now be further described with reference to the accompanying drawings.

[0009] Figure 1This is a flowchart of a method for early warning and auxiliary diagnosis of animal epidemics based on deep learning in animal husbandry and veterinary medicine, as described in this invention. Detailed Implementation

[0010] To make the technical means, creative features, objectives, and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments. like Figure 1 As shown in the embodiments of the present invention, an early warning and auxiliary diagnosis method for animal epidemics based on deep learning for animal husbandry and veterinary medicine, constructs a multi-dimensional environmental and physiological feature decoupling model, a dynamic evolution analysis mechanism of population social network topology, and an auxiliary diagnostic framework with veterinary clinical logic mapping. This method, from low-level data flow to high-level logical decision-making, completely solves the problem of confusion between individual physiological deviations and environmental stress fluctuations. The implementation of this invention first relies on a holographic sensing network constructed within the breeding unit. This sensing network includes an environmental sensor array deployed within the breeding unit, a non-contact computer vision acquisition unit, and a low-power IoT sensing module worn by the individual animal. During the system's initialization phase, the environmental sensor array acquires multi-dimensional habitat parameters in real time, including ambient temperature, relative humidity, ammonia concentration, carbon dioxide concentration, and light intensity, via RS485 bus or LoRaWAN wireless communication protocol. These parameters are encapsulated into a high-dimensional environmental parameter vector and affixed with a high-precision timestamp, serving as the basic reference coordinates for subsequent physiological feature decoupling. Simultaneously, the non-contact computer vision... The data acquisition unit utilizes high-resolution infrared thermal imaging cameras and depth cameras installed on the roof and key passageways to acquire spatial distribution maps of animal surface temperature, movement trajectories, and posture sequences. The infrared thermal imaging cameras identify feature points in hairless or sparsely haired areas such as the corners of the eyes and the base of the ears to obtain surface radiation values ​​closest to the core body temperature. Meanwhile, the IoT sensing module uses built-in piezoelectric ceramic sensors, accelerometers, and high-precision temperature-sensitive elements to acquire real-time respiratory rate, heart rate, and rumination frequency. All multi-source heterogeneous data from different sources undergoes strict temporal alignment at the edge computing node using a time synchronization algorithm, ensuring that each physiological indicator corresponds to a precise environmental state point. After acquiring multi-source heterogeneous data, this invention performs core environmental and physiological feature decoupling processing. In large-scale aquaculture habitats, drastic fluctuations in environmental temperature and humidity can cause significant heat or cold stress in animals. This non-pathological physiological deviation (such as increased respiration and slight increase in body temperature caused by high temperature) can easily mask the true early disease characteristics. To accurately separate environmental stress components from pathological deviation components, this invention constructs a decoupled encoder architecture based on the idea of ​​generative adversarial networks. In this architecture, environmental parameter vectors are injected as conditional inputs into an environmental physiological compensation sub-network. The environmental physiological compensation sub-network uses its pre-trained multilayer perceptron logic to establish a nonlinear dynamic mapping relationship between the environmental temperature and humidity index and the animal's heat dissipation and heat production balance. At each sampling moment, the sub-network will, according to the current... The system uses real-time ammonia concentration, temperature, humidity, and other habitat parameters to derive and calculate the normal physiological baseline values ​​that an individual animal should exhibit under the given environmental stress. Then, the decoupled encoder dynamically compares the actually collected individual physiological feature vectors (including real-time body temperature, heart rate, respiration, etc.) with the calculated normal physiological baseline values. By performing Mahalanobis distance calculations or residual calculations, the system extracts the residual physiological features after removing environmental interference. These residual physiological features eliminate global interference caused by environmental stress; any abnormal fluctuations in their values ​​point to essential deviations caused by pathological variations within the animal's body. This feature decoupling logic ensures the robustness of the early warning system under extreme weather conditions from a mathematical perspective, eliminating false positives caused by high temperatures—common in traditional methods—at the source. Furthermore, this invention not only focuses on the physiological residuals at the individual level, but also introduces a dynamic evolutionary analysis mechanism of group social network topology to capture the micro-anomalies in group behavior caused by epidemics during the incubation period. In the evolution of livestock diseases, infected individuals often exhibit social regression characteristics such as isolation, decreased activity, or reduced interaction frequency before showing overt clinical symptoms. This invention utilizes a non-contact computer vision acquisition unit to continuously track coordinate sequences and construct a dynamic social association graph reflecting group interaction relationships within a preset time window. In this dynamic social association graph, each animal individual is mapped as a node, and physical contact, shared foraging behavior, and physical proximity between individuals are defined as connecting edges, while the frequency and duration of interactions are assigned corresponding weights. This invention utilizes graph convolutional neural networks to perform deep feature extraction on the social network graph. Through spectral decomposition of the Laplacian matrix, it monitors the eigenvector centrality, clustering coefficient, and topological isolation index of each node in real time. Because graph convolutional neural networks can capture high-order structural features of social networks, the system can keenly perceive distortions in the normal distribution of the social topology of the group. When the social topological position of a node shows a trend shift (e.g., its clustering coefficient decreases significantly and its topological isolation index continues to increase), the system identifies it as an abnormal individual with a high risk of disease transmission. This monitoring perspective based on group social entropy expands the dimension of perception from a single physiological indicator to the laws of group topological dynamics, providing at least 24 to 48 hours of golden early warning increment for the early containment of an epidemic. As the core logic of the decision-making layer of this invention, the deep learning model adopts a spatiotemporal multi-scale feature fusion architecture, specifically corresponding to the deep learning topology under International Patent Classification G06N3 / 04. This architecture is horizontally divided into an individual physiological feature extraction layer, a spatial pose recognition layer, and a group social evolution analysis layer. The individual physiological feature extraction layer uses a long short-term memory network layer, specifically designed to process residual physiological features with temporal correlation, extracting the evolutionary patterns of animal respiratory rate and heart rate over long periods. The spatial pose recognition layer uses a deep residual convolutional layer, filtering the feature maps of pose sequences layer by layer to extract refined movement features such as lameness and abnormal lateral recumbency. The graph used to process social topological relationships... The attention mechanism layer is responsible for calculating the evolutionary weights between individuals at risk of disease and their neighboring nodes. The system strictly aligns the feature vectors of each dimension in time and inputs them into the global multi-scale fusion layer. During the fusion process, the graph attention mechanism layer dynamically allocates the attention weights of each dimension based on the current habitat parameter feedback. For example, when the environmental sensor array reports a high ammonia concentration alarm, the model automatically increases the calculation weights of the respiratory feature component and the outlier social component, while reducing the weight of body temperature fluctuations that may be caused by environmental disturbances. This dynamic trade-off mechanism simulates the interpretation logic of experienced veterinarians in differential diagnosis in complex environments, making the model no longer a rigid black box, but an intelligent sensor with scene adaptation capabilities. Furthermore, addressing the challenge of the lack of biological interpretability in deep learning model outputs, this invention constructs a biological attribute mapping layer at the end of the early warning and auxiliary diagnostic methods. Traditional early warning results often only provide a percentage probability, making it difficult for clinical veterinarians to rely on them for decision-making. The mapping layer of this invention uses a gradient-weighted activation mapping algorithm to reverse-track the activation state of neurons within the deep neural network, identifying feature combinations that play a key supporting role in early warning decisions. These high-dimensional features are then mapped back to pathological descriptors that conform to veterinary diagnostic standards. When the system triggers an early warning, the background automatically generates an auxiliary diagnostic report based on a biological evidence chain. This report clearly lists the logical supporting points of this alarm, such as explicitly stating that the animal's core body temperature is still 10% higher than the normal envelope after decoupling from environmental and physiological characteristics, and that its topological isolation index in the dynamic social association graph increases for four consecutive time windows, accompanied by a trend of decreasing rumination frequency. By transforming uninterpretable neuronal activation values ​​into structured veterinary clinical logic, this invention greatly improves the practical reliability of the early warning system, enabling non-professional managers to clearly understand the scientific mechanisms behind the early warning. This invention also integrates logic for predicting the risk of epidemic transmission, aiming to achieve precise allocation of epidemic prevention resources. The system, through real-time analysis of the edge weights in the social relational graph and combined with classic epidemiological dynamics models, assesses the probability of a confirmed individual infecting surrounding groups. The graph attention mechanism layer, by learning the evolutionary trajectory of interaction frequencies between nodes, can simulate the potential spread path of the virus within the breeding unit. This allows farm managers to avoid blind, costly, and indiscriminate culling of the entire farm, instead implementing differentiated, tiered prevention and control measures based on the transmission risk assessment results, accurately identifying the scope of threatened individuals, and minimizing economic losses. Furthermore, the system, through a closed-loop feedback optimization mechanism, allows clinical veterinarians to feed back subsequent laboratory diagnostic confirmation results to an online learning library. Using a contrastive learning algorithm, the system continuously iteratively calibrates the hyperparameters of the feature decoupling encoder and topology analysis model based on real pathological conclusions, achieving deep adaptation of the algorithm model to the specific microenvironment of the farm, ensuring the early warning system has a long-term, continuously evolving vitality. In the physical deployment of the system, to ensure real-time performance and reliability, this invention adopts a cloud-edge collaborative distributed architecture. Raw data collected by the environmental sensor array and IoT sensing module first enters the local edge computing node of the farm. At the edge, the system completes preliminary data denoising, environmental and physiological feature decoupling, and basic feature extraction, achieving low-latency preliminary early warning. Meanwhile, the computationally intensive dynamic evolution analysis of the social network topology and global multi-scale feature fusion are asynchronously uploaded from the edge nodes to the cloud computing center for processing. This architecture ensures that the timeliness of the early warning is not reduced during network fluctuations or large-scale data influxes, meeting the stringent requirements of second-level response during major epidemic outbreaks. Regarding the training strategy for the deep learning model, this invention employs a two-stage transfer learning method. First, the basic model is trained on a public animal pathology database containing tens of thousands of labeled cases. Then, fine-tuning is performed for specific livestock breeds (such as fattening pigs and dairy cows) in real-world scenarios to ensure the model's generalization ability and accuracy in complex production scenarios. To verify the advancement and practical effectiveness of the method described in this invention, the following [Example] and [Comparative Example] are constructed for data demonstration. [Example] This example uses the deep learning-based animal disease early warning and auxiliary diagnosis method for animal husbandry and veterinary medicine described in this invention. The experimental scenario is set as a large-scale pig farm (with a stock of 3,000 pigs), equipped with a complete environmental sensor array, infrared thermal imaging camera, and individual sensor modules. The experiment took place during a period of continuous high temperatures in summer, with the highest ambient temperature reaching 38 degrees Celsius. The system establishes a dynamic environmental adaptation envelope for pigs using an environmental-physiological decoupling model, accurately identifying non-pathological body temperature increases caused by heat stress. The social network topology evolution analysis mechanism successfully located three pigs exhibiting social regression characteristics, and the topological isolation index output by their graph convolutional neural network increased by 215% within 24 hours. A multi-scale feature fusion architecture integrates physiological, pose, and topological features to output early warning commands. The biological attribute mapping layer generates a diagnostic report to assist veterinarians in implementing isolation. Laboratory nucleic acid testing confirmed it as an early latent infection of African swine fever. [Comparative Example] The comparative example uses a traditional animal monitoring system based on fixed threshold alarms. This system only monitors individual body temperature and activity levels and does not have the function of decoupling environmental and physiological characteristics or social topological evolution analysis. During periods of high summer temperatures, a large number of pigs experienced heat stress exceeding the fixed threshold of 39.5 degrees Celsius, triggering a large number of false alarms in the system and causing alarm fatigue among management personnel. During the actual incubation period, the system failed to issue an early warning because the animal's physiological indicators had not yet exceeded the static threshold. The warning results are merely numerical alerts, lacking any pathological logic to support them. Veterinarians cannot determine the type of disease and the degree of risk based on the alert information. Based on the long-term operating performance of the embodiments and comparative examples, the following [Comparison Data Table] was compiled:

[0011] The comparative data clearly demonstrates that the method described in this invention exhibits significant advantages in terms of early warning time, accuracy, and clinical applicability. In particular, supported by the environmental-physiological characteristic decoupling model, the system's anti-interference capability under extreme environments is improved by more than tenfold. Furthermore, through dynamic evolution analysis of the population social network topology, the system successfully fills the perception gap regarding animal diseases before changes in physiological indicators. In the specific algorithm execution flow, the residual physiological feature extraction described in this invention essentially utilizes a generative adversarial network to learn a manifold distribution of the physiological performance of healthy animals in a specific environment. The environmental physiological compensation sub-network generates an ideal physiological signal template by calculating the mapping of the input environmental vector in the manifold space. This decoupling method based on manifold learning can capture higher-order interactions between environmental parameters, such as the effect of the coupling between ammonia concentration and humidity on respiratory rate, which is unattainable by traditional linear compensation methods. Furthermore, the dynamic social association graph described in this invention employs a sliding time window technique, quantifying the stability of the social structure by calculating the cosine similarity of the eigenvalue spectra of the Laplacian matrix of adjacent time windows. This rigorous mathematical processing ensures extremely high sensitivity and robustness in capturing social anomalies. Furthermore, the Graph Attention (GAT) mechanism layer employed in this invention has a natural advantage in handling group relationships. It allows each node to dynamically learn connection weights based on the features of its neighboring nodes. In epidemic spread risk prediction, if the interaction frequency between a healthy individual and a high-risk individual exhibits abnormal weighting during dynamic evolution, the system automatically increases the risk level of the healthy individual. This local diffusion simulation based on the attention mechanism greatly improves the accuracy of epidemic spread prediction. In the multi-scale feature fusion architecture, the residual connection design between layers ensures the effective propagation of gradients in the deep neural network, preventing the gradient vanishing problem that occurs when processing long-term time-series data, thereby ensuring the stability of the model during continuous monitoring. The biological attribute mapping layer disclosed in this invention, in terms of technical implementation, obtains the sensitivity matrix of each input feature relative to the final diagnostic classification by calculating the partial derivatives of the output layer of a deep neural network. Through this sensitivity analysis, the system can quantify the specific percentage contribution of body temperature, interaction frequency, and abnormal posture to the current warning. This feature has strong practical significance in actual livestock management, transforming the early warning system from a simple alarm device into an intelligent digital veterinary expert. This significantly reduces the dependence of farms on highly skilled veterinary professionals and improves the overall technical level of grassroots disease prevention. Regarding the closed-loop optimization of epidemic prevention resources, the contrastive learning algorithm described in this invention constructs positive and negative sample pairs (i.e., confirmed samples and false alarm samples), forcing the feature decoupling encoder to widen the gap between pathological features and environmental stress features in the feature space. As feedback data in the online learning library continuously accumulates, the model's classification boundary becomes increasingly clear, thereby achieving self-iteration of algorithm performance. This online evolution capability enables this invention to flexibly address real-world challenges such as swine influenza virus mutations and changes in livestock breeds. In summary, this invention, through the deep integration of deep learning technology with animal husbandry and veterinary medicine, population dynamics, and epidemiology, constructs a logically rigorous and technologically advanced early warning and auxiliary diagnostic system for animal epidemics. This invention not only achieves holistic coverage from individual to group and from physiological to environmental aspects in feature extraction, but also realizes a paradigm shift from black-box classification to causal mapping in logical decision-making. The implementation of this invention will provide a robust biosecurity barrier for modern animal husbandry, and has profound significance for reducing economic losses from diseases, ensuring the safety of livestock products, and maintaining public health security. All deep learning model details, sensor deployment logic, and data decoupling algorithms described in the specification have been fully supported by underlying physical mechanisms and mathematical logic, ensuring the sufficiency of the disclosure and the certainty of the technology.

[0012] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based method for early warning and auxiliary diagnosis of animal epidemics in animal husbandry and veterinary medicine, characterized in that, The method includes the following steps: S1: Acquire environmental parameters of the area to be monitored, real-time physiological indicators of individual animals, and spatial location sequences of individuals; S2: Using a preset environmental physiological feature decoupling model, the environmental parameters are used as conditional inputs to perform decoupling processing on the real-time physiological indicators, and residual physiological features after removing environmental stress components are extracted. S3: Using the group social network topology evolution analysis mechanism, construct a dynamic social association graph reflecting the group interaction relationship based on the spatial location sequence, and extract the social topology features of each node in the dynamic social association graph; S4: Input the residual physiological features and the social topology features into the pre-trained spatiotemporal multi-scale feature fusion architecture, and allocate feature weights according to the environmental parameters through the attention mechanism inside the spatiotemporal multi-scale feature fusion architecture to output the early warning decision result.

2. The method for early warning and auxiliary diagnosis of animal epidemics based on deep learning for animal husbandry and veterinary medicine as described in claim 1, characterized in that, The steps for obtaining environmental parameters of the area to be monitored, real-time physiological indicators of individual animals, and spatial location sequences of individuals specifically include: By deploying an array of environmental sensors inside the breeding unit, multi-dimensional habitat parameters, including ambient temperature, relative humidity, ammonia concentration, carbon dioxide concentration, and light intensity, are acquired in real time and encapsulated into an environmental parameter vector. The spatial distribution map of the animal's body surface temperature, movement trajectory and posture sequence are obtained by a non-contact computer vision acquisition unit, and the core body temperature data of the animal is extracted by infrared thermal imaging feature point identification. The animal's real-time respiratory rate, heart rate, and rumination rate are collected by an IoT sensor module worn on the animal. The environmental parameter vector, core body temperature data, motion trajectory, posture sequence, real-time respiratory rate, heart rate, and rumination rate are time-aligned using a time synchronization algorithm to generate a multi-source heterogeneous dataset.

3. The method for early warning and auxiliary diagnosis of animal epidemics based on deep learning for animal husbandry and veterinary medicine as described in claim 1, characterized in that, The steps for extracting residual physiological features using a pre-defined environmental-physiological feature decoupling model specifically include: A decoupled encoder based on an adversarial generative network architecture is constructed, wherein the decoupled encoder integrates an environmental physiological compensation subnetwork; The environmental parameters are input into the environmental physiological compensation subnetwork. Using the multilayer perceptron logic inside the environmental physiological compensation subnetwork, a nonlinear dynamic mapping relationship between the environmental temperature and humidity index and the animal's heat dissipation and heat production balance is established, and the normal physiological baseline value corresponding to the individual animal under the current environmental pressure is derived and output. The collected real-time physiological indicators are compared with the normal physiological baseline values. The residual components of the real-time physiological indicators deviating from the normal physiological baseline values ​​are extracted by performing Mahalanobis distance operation to obtain the residual physiological characteristics, so as to eliminate the confusion and interference of non-pathological physiological deviations in the diagnosis of the epidemic.

4. The method for early warning and auxiliary diagnosis of animal epidemics based on deep learning for animal husbandry and veterinary medicine as described in claim 1, characterized in that, The steps for extracting social topology features using the group social network topology evolution analysis mechanism specifically include: Based on the spatial location sequence, animal individuals are defined as nodes within a preset time sliding window, and physical contact, shared foraging behavior, and physical proximity between individuals are defined as connecting edges. Weights are assigned to the connecting edges based on the interaction frequency and duration to construct the dynamic social association graph. Deep feature extraction is performed on the dynamic social association graph using a graph convolutional neural network. Spectral decomposition is performed on the graph Laplacian matrix corresponding to the dynamic social association graph to monitor the feature vector centrality, clustering coefficient, and topological isolation index of each node in real time. By comparing the eigenvalue spectrum distribution of the Laplacian matrix within a continuous time window, the degree of distortion of the normal distribution of the social topology of the group is quantified, and the eigenvector centrality, clustering coefficient, and topological isolation index are encapsulated as the social topology features.

5. The method for early warning and auxiliary diagnosis of animal epidemics based on deep learning for animal husbandry and veterinary medicine as described in claim 1, characterized in that, The internal processing logic of the spatiotemporal multi-scale feature fusion architecture specifically includes: The residual physiological characteristics are extracted using a long short-term memory network layer to obtain the trend evolution characteristics of the physiological indicators. Residual convolutional layers are used to extract features from the posture sequences of individual animals, and to identify movement state features including lameness and abnormal lateral recumbency. The social topology features are processed using a graph attention mechanism layer to calculate the evolution weights between risk nodes and their neighboring nodes, and the trend evolution features, motion state features, and social topology features are dynamically weighted and fused together with the environmental parameters. When the ammonia concentration or ambient temperature in the environmental parameters exceeds a preset threshold, the graph attention mechanism layer automatically increases the weight value of the breathing frequency component in the trend evolution feature and increases the weight value of the topological isolation index in the social topology feature.

6. The method for early warning and auxiliary diagnosis of animal epidemics based on deep learning for animal husbandry and veterinary medicine as described in claim 1, characterized in that, Following the step of outputting the early warning decision result, the method further includes a step of generating an auxiliary diagnostic report through a biological attribute mapping layer: The activation state of neurons within the spatiotemporal multi-scale feature fusion architecture is traced in reverse using a gradient-weighted activation mapping algorithm to identify key feature combinations that meet preset conditions for contributing to early warning decisions. The key feature combination is mapped to a preset veterinary diagnosis and treatment standard library using a feature attribute attribution algorithm, and transformed into a structured pathological descriptor. An auxiliary diagnostic report containing a chain of evidence supporting diagnostic logic is generated based on the pathological descriptor. The auxiliary diagnostic report lists the numerical offset of the residual physiological feature, the rate of change of the topological isolation index of the social topological feature, and the corresponding biological interpretation.

7. The method for early warning and auxiliary diagnosis of animal epidemics based on deep learning for animal husbandry and veterinary medicine, as described in claim 5, is characterized in that, It also includes steps for predicting the risk of epidemic transmission: The graph attention mechanism layer is used to learn the evolution trajectory of the edge weights between infected nodes and adjacent nodes in the dynamic social association graph; combined with the epidemiological dynamics model, the potential spread path of the virus in the breeding unit is simulated according to the evolution trajectory of the edge weights. Based on the potential spread pathways, assess the probability of infection from confirmed individuals to their surrounding contacts, and output tiered prevention and control recommendations for the aquaculture population.

8. The method for early warning and auxiliary diagnosis of animal epidemics based on deep learning for animal husbandry and veterinary medicine as described in claim 1, characterized in that, It also includes closed-loop feedback optimization steps: Establish an online learning repository to store laboratory diagnostic results and autopsy pathological features reported by clinical veterinarians; Using a contrastive learning algorithm, with the laboratory diagnostic results as a supervisory signal, the parameters of the environmental physiological compensation subnetwork in the environmental physiological feature decoupling model and the parameters of the graph convolutional neural network in the population social network topology evolution analysis mechanism are iteratively calibrated. By constructing a feature comparison space between confirmed positive samples and false negative samples, the manifold distance between pathological feature vectors and environmental stress feature vectors in the feature space is increased.

9. A method for early warning and auxiliary diagnosis of animal epidemics based on deep learning for animal husbandry and veterinary medicine, as described in claim 1, is characterized in that, The physical execution architecture of the method adopts a cloud-edge collaborative model: The environmental parameters and real-time physiological indicators are decoupled, residual physiological features are extracted, and preliminary early warning tasks are performed at the edge computing node of the breeding farm. The construction of the dynamic social relationship graph, the topological evolution analysis of the group social network, the multi-scale feature fusion, and the global auxiliary diagnostic decision-making are performed in the cloud computing center. The edge computing nodes and the cloud computing center transmit data asynchronously through a distributed protocol to ensure that the response delay of the warning command is lower than the preset real-time threshold.

10. A method for early warning and auxiliary diagnosis of animal epidemics based on deep learning for animal husbandry and veterinary medicine, as described in claim 1, characterized in that, The training process of the deep learning model adopts a transfer learning strategy: The first phase of pre-training was performed on a general animal pathology database to learn common disease characteristics across species; The second stage of fine-tuning learning is performed on a real-world dataset from a specific farm, using physiological constants and stocking density parameters of a specific breed to optimize the classification boundary of the spatiotemporal multi-scale feature fusion architecture. By performing sensitivity matrix analysis on the feature maps of the intermediate layer of the model, we can identify and solidify the mapping relationships of biological attributes that have high discriminative power for specific epidemics.

Images