Disease early warning method and system fusing deep probability map model and bayesian inference

Abstract

The present application relates to the technical field of artificial intelligence monitoring, and more particularly to a disease early warning method and system fusing a deep probabilistic graph model and Bayesian inference, which inputs multi-source time series monitoring data and static attribute features into a deep generative model, maps to an independent latent space and decouples output pathological, environmental and noise variables, purifies the pathological variables using the latter two, acquires and propagates structure particles based on the purified features, inputs a continuous time evolution model to generate multiple epidemic evolution trajectories, and finally calculates a comprehensive risk value through a risk sensitive evaluation function to trigger an early warning, thereby achieving accurate separation of environmental drift and random noise from mixed signals and significantly reducing the false positive rate. Meanwhile, multiple propagation hypotheses are deduced in parallel, and the weight of high-risk trajectories is amplified through nonlinear aggregation, thereby ensuring that long-tail disaster risks can be effectively captured in the face of uncertainty.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence monitoring technology, and in particular to a disease early warning method and system that integrates deep probabilistic graphical models and Bayesian inference. Background Technology

[0002] In modern large-scale farms or regional disease prevention and monitoring, IoT sensors deployed on-site are typically used to collect real-time data on livestock and poultry biological characteristics such as body temperature and feed intake, as well as environmental parameters such as temperature and humidity in pigsties and ammonia concentration. The health status of livestock and poultry groups is monitored by analyzing the fluctuation trends of these multi-source time-series data.

[0003] However, in real-world applications, the raw data collected by monitoring equipment often contains a variety of mixed components, making it difficult for existing conventional monitoring methods to effectively distinguish the root cause of data fluctuations. Specifically, the system cannot accurately identify whether an abnormal data fluctuation (such as elevated body temperature) stems from pathological changes caused by viral infection in individual livestock or poultry, a collective environmental heat stress response caused by seasonal changes and sudden temperature increases, or simply random noise or outliers caused by aging sensor equipment and unstable signal transmission.

[0004] Due to the lack of ability to decouple and purify the different attributes mentioned above, existing early warning systems have significant limitations when facing complex aquaculture environments. For example, when environmental parameters drift significantly, the system is prone to confusing environmental factors with pathological factors, leading to widespread false alarms and increasing the cost of manual investigation. On the other hand, if simple smoothing and noise reduction are performed, it is easy to filter out weak early pathological signals as equipment noise, resulting in the underreporting of key epidemic information. This problem of not being able to accurately extract true infection characteristics from mixed signals seriously affects the reliability and timeliness of early warning results. Summary of the Invention

[0005] The main objective of this invention is to provide a disease early warning method that integrates deep probabilistic graphical models and Bayesian inference, aiming to solve the problem that existing early warning methods are difficult to integrate data with different attribute features, resulting in poor early warning effects.

[0006] To achieve the above objectives, this invention provides a disease early warning method that integrates deep probabilistic graphical models and Bayesian inference. The early warning method includes the following steps:

[0007] Acquire multi-source time-series monitoring data and static attribute characteristics within the monitoring area, wherein the multi-source time-series monitoring data includes biological characteristic data and environmental sensor data;

[0008] The multi-source time-series monitoring data is input into a deep generation model, the static attribute features are used as conditional constraints, and the multi-source time-series monitoring data is mapped to mutually independent latent spaces to decouple and output pathological latent variables, environmental latent variables, and sensor noise latent variables in the latent spaces. The pathological latent variables are then purified through the environmental latent variables and sensor noise latent variables.

[0009] Based on the purified pathological latent variables, several parallel dynamic propagation structure particles are obtained. These dynamic propagation structure particles are used to characterize the uncertain propagation dependencies between nodes within the monitoring area.

[0010] The pathological latent variables and the dynamic propagation structure particles are input into a continuous-time evolution model to obtain multiple epidemic evolution trajectories within a future preset time period;

[0011] The epidemic evolution trajectory is nonlinearly aggregated and calculated based on the risk sensitivity evaluation function to obtain a comprehensive risk assessment value, and a graded early warning signal is triggered when the comprehensive risk assessment value exceeds a preset threshold.

[0012] To achieve the above objectives, the present invention also provides an epidemic early warning system that integrates deep probabilistic graphical models and Bayesian inference, the early warning system comprising:

[0013] The data acquisition module is used to acquire multi-source time-series monitoring data and static attribute characteristics within the monitoring area. The multi-source time-series monitoring data includes biological sign data and environmental sensor data.

[0014] The decoupling and purification module is used to input the multi-source time-series monitoring data into the deep generation model, use the static attribute features as conditional constraints, and then map the multi-source time-series monitoring data to mutually independent latent spaces, so as to decouple and output pathological latent variables, environmental latent variables and sensor noise latent variables in the latent spaces, and purify the pathological latent variables through the environmental latent variables and sensor noise latent variables.

[0015] The structure inference module is used to obtain several parallel dynamic propagation structure particles based on the purified pathological latent variables. The dynamic propagation structure particles are used to characterize the uncertain propagation dependencies between nodes within the monitoring area.

[0016] An evolutionary projection module is used to input the pathological latent variables and the dynamic propagation structure particles into a continuous-time evolution model to obtain multiple epidemic evolution trajectories within a future preset time period.

[0017] The risk warning module is used to perform nonlinear aggregation calculations on the epidemic evolution trajectory based on a risk sensitivity evaluation function to obtain a comprehensive risk assessment value, and to trigger a graded warning signal when the comprehensive risk assessment value exceeds a preset threshold.

[0018] This invention uses static attribute features as constraints for a deep generative model, mapping multi-source time-series monitoring data, including biological signs and environmental sensors, to an independent latent space. It forces decoupling and outputs three latent variables: pathology, environment, and sensor noise. The pathology variable is then purified using environmental and noise variables. Subsequently, based on the purified features, several parallel dynamic propagation structure particles are generated to simulate uncertain propagation dependencies between nodes. These particles are then input into a continuous-time evolution model to obtain multiple evolution trajectories. Finally, a risk-sensitive evaluation function is used to nonlinearly aggregate the trajectories to trigger an early warning. This effectively alleviates the technical problems of being unable to distinguish between environmental heat stress, equipment failure, and real epidemic signals in complex monitoring environments, as well as the underreporting caused by the static model's inability to adapt to dynamic propagation paths.

[0019] Specifically, the following effects were achieved:

[0020] First, through independent mapping and decoupling purification of the latent space, the system can automatically filter out the collective background drift caused by seasonal temperature changes and the random noise of the sensor itself, retaining only the pathological fluctuations caused by infection, which greatly reduces the false alarm rate caused by environmental or equipment factors.

[0021] Secondly, by utilizing parallel dynamic propagation structure particles combined with a continuous-time evolution model, the system no longer relies on a single fixed propagation pattern, but can simultaneously deduce multiple possible propagation paths and evolution trends, covering the uncertainty of propagation relationships.

[0022] Finally, through the nonlinear aggregation of risk-sensitive evaluation functions, the system can automatically amplify the weight of high-risk evolution trajectories during decision-making. This ensures that even when most predicted trajectories are stable but there are a few catastrophic evolution possibilities, it can still output a high level of comprehensive risk assessment value and trigger an alarm. This provides precise early warning support for epidemic prevention and control work that takes into account both anti-interference capabilities and bottom-line safety thinking. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the method in Embodiment 1 of the present invention;

[0024] Figure 2 This is a structural block diagram of the system in Embodiment 8 of the present invention.

[0025] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0027] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0028] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection or an electrical connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0029] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the meaning of "and / or" throughout the text includes three parallel solutions; for example, "A and / or B" includes solution A, solution B, or a solution where both A and B are satisfied simultaneously. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0030] Example 1:

[0031] As attached Figure 1 As shown in the figure, this embodiment provides a disease early warning method that integrates deep probabilistic graphical models and Bayesian inference. The early warning method includes the following steps:

[0032] Acquire multi-source time-series monitoring data and static attribute characteristics within the monitoring area, wherein the multi-source time-series monitoring data includes biological characteristic data and environmental sensor data;

[0033] The multi-source time-series monitoring data is input into a deep generation model, the static attribute features are used as conditional constraints, and the multi-source time-series monitoring data is mapped to mutually independent latent spaces to decouple and output pathological latent variables, environmental latent variables, and sensor noise latent variables in the latent spaces. The pathological latent variables are then purified through the environmental latent variables and sensor noise latent variables.

[0034] Based on the purified pathological latent variables, several parallel dynamic propagation structure particles are obtained. These dynamic propagation structure particles are used to characterize the uncertain propagation dependencies between nodes within the monitoring area.

[0035] The pathological latent variables and the dynamic propagation structure particles are input into a continuous-time evolution model to obtain multiple epidemic evolution trajectories within a future preset time period;

[0036] The epidemic evolution trajectory is nonlinearly aggregated and calculated based on the risk sensitivity evaluation function to obtain a comprehensive risk assessment value, and a graded early warning signal is triggered when the comprehensive risk assessment value exceeds a preset threshold.

[0037] In traditional livestock and poultry health monitoring systems, the root causes of fluctuations in multi-source time-series data are difficult to distinguish effectively. Specifically, abnormal data fluctuations cannot be accurately identified as being caused by pathological changes due to viral infection, environmental heat stress response due to seasonal changes, or random noise generated by aging sensor equipment. As a result, the reliability of monitoring results is affected, and the early warning system generates false alarms when environmental parameters drift significantly. Furthermore, noise reduction processing may lead to the underreporting of real pathological signals. Consequently, the system cannot extract real infection characteristics from mixed signals, thereby reducing the credibility and timeliness of early warning results.

[0038] To address the aforementioned issues, this embodiment utilizes an Internet of Things (IoT) sensor network deployed within the monitoring area (such as large-scale farms, slaughtering and processing centers, and logistics hubs) to acquire multi-source time-series monitoring data in real time. Simultaneously, it retrieves static attribute features corresponding to each monitoring node from the backend database. It is understood that the multi-source time-series monitoring data in this embodiment is dynamically changing, specifically including biological characteristic data collected via infrared thermal imagers or contact ear tags (e.g., real-time body temperature sequences and daily feed intake fluctuation curves for pigs and cattle), and environmental sensor data collected via environmental sensors (e.g., temperature, relative humidity, ammonia concentration, and carbon dioxide concentration inside pigsties and cattle sheds). The static attribute features include the node's geographic spatial coordinates (latitude and longitude), the farm's designed stocking size, and the upstream and downstream logistics relationships determined based on long-term supply chain contracts. Although these static attribute features are relatively stable in a short period of time, they constitute the background physical constraints for data generation. For example, under the same ventilation conditions, the accumulation rate of ammonia concentration in farms of different sizes is quite different. Therefore, introducing static attribute features as conditional constraints into the model can help the model understand the physical background of the monitoring data more accurately.

[0039] After acquiring the aforementioned data, the system performs a core data decoupling and purification step. This step aims to alleviate the problem in existing technologies that cannot distinguish between environmental heat stress and actual pathological reactions. The system inputs multi-source time-series monitoring data into a deep generative model based on a conditional variational autoencoder architecture, and encodes static attribute features into embedding vectors, which are then concatenated as conditional variables into the input layer of the encoder. The deep generative model maps high-dimensional time-series observation data to three mutually independent latent spaces with clear physical meanings through complex nonlinear transformations, decoupling and outputting pathological latent variables, environmental latent variables, and sensor noise latent variables, respectively. To ensure that these three latent variables are statistically strictly independent, i.e., to force orthogonal decoupling of features, this embodiment introduces a mutual information minimization mechanism in the objective function of model training. The mathematical expression of this processing logic is shown in Expression 1 below: In the formula, This represents the total loss function value that the deep generative model needs to minimize during the training phase.

[0040] The lower bound of evidence loss term is derived from variational inference theory and is used to ensure that the latent variables generated by the model can accurately reconstruct the original multi-source time-series monitoring data, that is, to ensure the integrity of the information.

[0041] This represents the orthogonal decoupling penalty coefficient, which is derived from the setting of the model hyperparameters. It is usually set to a large positive number to adjust the model's emphasis on feature decoupling.

[0042] This represents the mutual information estimation operator, used to quantify the statistical dependence between two random variables. The lower the value, the more independent the two variables are.

[0043] This represents the pathological latent variables mapped from multi-source time-series monitoring data, whose values ​​are derived from the extraction of abnormal fluctuations in biological signs.

[0044] This represents potential environmental variables, the values ​​of which are derived from the extraction of periodic or trend changes in environmental sensor data.

[0045] This represents the potential variable of sensor noise, the value of which is derived from the extraction of high-frequency random disturbances in the data.

[0046] Based on the above expression one, it can be understood that traditional loss functions only focus on reconstruction errors, which can easily lead to data inertia in the model. This can result in the model directly encoding the increase in body temperature of the entire herd caused by summer heat into pathological features, thus causing false alarms. The above expression, by introducing mutual information I as a penalty term, mathematically mandates... The information contained therein must be consistent with and None of them are relevant.

[0047] This means that any fluctuations that can be explained by environmental data (such as temperature) will be forcibly stripped away from [the data source]. In the middle; any random jumps without any pattern will be classified as... Of them, the ones that were ultimately preserved It only includes pure anomalous signals (i.e. real signs of viral infection) that cannot be explained by the environment or are random noise. In this way, environmental latent variables and sensor noise latent variables can be discarded, and only the purified pathological latent variables can be used for subsequent processing, thereby greatly improving the signal-to-noise ratio of the early warning signal.

[0048] After feature extraction, the system enters the inference stage of dynamic propagation structure. Given the high uncertainty of disease transmission paths in real-world scenarios (e.g., contact transmission may or may not occur; logistics routes may be interrupted due to blockades), the system cannot rely on a single static adjacency matrix. This embodiment uses a particle flow variational inference algorithm (specifically, Stein variational gradient descent, SVGD) to obtain several parallel dynamic propagation structure particles based on the extracted latent pathological variables. Each structure particle essentially represents a hypothetical propagation dependency graph between nodes at the current time. To ensure that these particles can cover multiple possible propagation modes, the system updates the particle state by introducing a repulsive force generated by a kernel function. The update formula satisfies the following expression:

[0049] ;

[0050] In the formula, This represents the state matrix of the k-th dynamically propagating structural particle at the t-th iteration step, and its value comes from the parameterized representation of the propagation probability between nodes within the monitoring area;

[0051] The learning rate step size represents the structural inference process and is derived from preset optimization parameters;

[0052] M represents the total number of structural particles sampled in parallel, which is derived from the allocation of system computing resources;

[0053] This represents the radial basis function (RBF kernel), used to calculate the similarity distance between two structural particles;

[0054] The log-likelihood term, representing the posterior probability, is used to guide particles to move toward propagation structures with higher data support (i.e., to make the structure more consistent with the current observation data).

[0055] This represents the repulsive gradient term, which is the derivative of the kernel function.

[0056] The core logic of expression two above lies in the repulsive force term. Specifically, in traditional Monte Carlo sampling, particles tend to get trapped in the same local optimum, meaning all particles predict the same propagation path. The expression above uses the gradient of the kernel function to generate a repulsive force term, forcing each structural particle to... They are far apart in the probability space. This has excellent adaptability in technical solutions: it allows the system to maintain multiple reasonable propagation assumptions simultaneously. For example, particle 1 might simulate the scenario of "the virus spreading through a feed transport network," while particle 2 might simulate the scenario of "the virus spreading through short-range aerosols." This mechanism of multiple assumptions in parallel is a key means to solve the problem of uncertainty in propagation paths.

[0057] Subsequently, the system will include pathological latent variables. With the aforementioned group of dynamically propagating structural particles Simultaneously, the data is input into a continuous-time evolution model. To simulate the temporal continuity of the biological pathological process (rather than discrete time steps), this embodiment employs neural stochastic differential equations to construct the evolution model, thereby obtaining multiple epidemic evolution trajectories within a preset future time period. The mathematical description of its evolutionary process is shown in Expression Three:

[0058] ;

[0059] In the formula, This represents the hidden state vector of the system at time t, whose initial values ​​are derived from the input pathological latent variables. ;

[0060] The drift function, fitted by a neural network, is used to describe the deterministic trend of the epidemic's development (such as the exponential growth of the number of infections), and its parameter θ is derived from model training.

[0061] The kth dynamically propagating structural particle is injected as a parameter into the drift function, affecting the evolving topology.

[0062] The diffusion function, fitted by a neural network, describes the intensity of the inherent random fluctuations in the system, and its parameters... φ Derived from model training;

[0063] The increment representing Brownian motion is derived from random sampling of a standard normal distribution and is used to introduce random noise.

[0064] It represents the smallest infinitesimal component in continuous time.

[0065] Understandably, traditional RNN models are discrete, making it difficult to handle data with uneven sampling intervals and quantify the uncertainty of predictions. Neural SDEs, on the other hand, introduce Brownian motion terms, making each generated trajectory a sample of a stochastic process. Combined with the input M structural particles, the system eventually integrates to solve for M distinct evolutionary trajectories. These trajectories form a deep probability map, which intuitively demonstrates the multiple possible paths the epidemic may take under different propagation assumptions and random fluctuations.

[0066] Finally, the system executes a decision-making step based on a risk-sensitive evaluation function. To avoid underreporting due to a few highly destructive high-risk trajectories being averaged out by the majority of stable trajectories among multiple predicted trajectories, the system performs nonlinear aggregation calculations on the output M epidemic evolution trajectories to obtain a comprehensive risk assessment value. This calculation process uses the risk-sensitive functional of the following expression:

[0067] ;

[0068] In the formula, This represents the final comprehensive risk assessment value, used for comparison with the early warning threshold;

[0069] γ represents the risk aversion coefficient, which is an adjustable hyperparameter greater than 0. It is derived from the setting of the safety level of the monitoring area (the higher the safety level requirement, the larger this value).

[0070] This represents the natural exponential function, used for nonlinear amplification of high-loss trajectories;

[0071] This represents the k-th trajectory of the epidemic's evolution. The corresponding predicted loss value (such as the predicted scale of infection or economic loss) is derived from the assessment of the state at the end of the trajectory.

[0072] This represents the natural logarithm function, used to restore amplified values ​​to a linear scale.

[0073] Understandably, when γ > 0, the above expression four is actually a soft maximum operator. If 99 out of M trajectories predict no epidemic (loss close to 0), but only one trajectory predicts a major epidemic (extremely large loss), this trajectory would be submerged in the conventional average calculation. However, in this embodiment, the extremely large loss value will grow exponentially after the exp operation, thus dominating the result of the entire summation term, making the final... It remains at a very high level. This design gives the system a bottom-line thinking ability similar to that of human experts: when faced with scientific uncertainty, as long as there is a reasonable catastrophic hypothesis, the system should trigger a graded early warning signal, thereby completely solving the problem of major epidemic underreporting caused by the majority rule in traditional models.

[0074] In summary, this embodiment constructs a disease early warning system that deeply integrates deep probabilistic graphical models and Bayesian inference theory. At the data representation and modeling level, firstly, based on the generative framework of the deep probabilistic graphical model, a conditional variational autoencoder is used to model the generation process of multi-source time-series monitoring data as a nonlinear transformation process driven by mutually independent latent variables such as pathology, environment, and noise. By introducing mutual information minimization constraints into the objective function of variational inference, orthogonal decoupling of different causal factors is enforced in the latent probability space. This allows for the accurate extraction of latent pathological variables affected only by viral infection from mixed observation data, eliminating interference from environmental drift and equipment noise at the source of the probability distribution.

[0075] At the structural inference and evolution level, the core idea of ​​Bayesian inference is introduced to handle the uncertainty of the propagation network. That is, the propagation dependencies between monitoring nodes are no longer treated as a deterministic static graph structure, but rather as random variables to be inferred. By applying the particle flow variational inference algorithm (Stein variational gradient descent), a set of parallel dynamic propagation structure particles is generated to approximate the posterior distribution of the actual propagation structure. Each particle represents a potential propagation network hypothesis. Subsequently, these structure particles representing different Bayesian hypotheses, along with purified pathological variables, are input into a neural stochastic differential equation to deduce multiple epidemic evolution trajectories containing randomness and structural uncertainty in the continuous time domain. Finally, these probability trajectories are nonlinearly aggregated based on a risk-sensitive evaluation function to ensure that, in the face of scientific uncertainty, the system can capture high-risk scenarios in the long-tail distribution and trigger early warnings.

[0076] Example 2:

[0077] In this embodiment, the step of inputting the multi-source time-series monitoring data into a deep generative model, using the static attribute features as conditional constraints, and then mapping the multi-source time-series monitoring data to mutually independent latent spaces includes:

[0078] A conditional variational autoencoder is constructed, and an embedding vector is obtained based on the static attribute features. The embedding vector is then concatenated into the encoding process of the multi-source time-series monitoring data to apply distribution constraints.

[0079] During model training, the mutual information minimization mechanism is used to force the covariance between each pair of the pathological latent variables, environmental latent variables and sensor noise latent variables to approach zero.

[0080] It should be noted that in the engineering application of deep learning, especially in the actual construction of variational autoencoders (VAE / CVAE), calculating the precise mutual information of high-dimensional continuous variables usually requires training an additional discriminator network, such as MINE, which is computationally costly. Therefore, while the theoretical goal of this invention is to minimize mutual information, in the specific operation of this embodiment, it is achieved by penalizing the Frobenius norm of the covariance matrix, that is, satisfying the following variation of expression one:

[0081] ;

[0082] In the formula, λ represents the decoupling penalty coefficient, which is a hyperparameter greater than zero. It is a preset value obtained before model training and is used to adjust the model's constraint strength on feature independence.

[0083] This indicates a summation operation over all different types of combinations of latent variables, covering pairwise combinations between pathology and environment, pathology and noise, and environment and noise.

[0084] This represents the covariance operator, used to calculate the covariance matrix of two random variable vectors on a batch of data. This value is derived from the statistical calculation of the latent variable sample values ​​in the current training batch.

[0085] , This indicates different types of latent variables, specifically referring to pathological latent variables, environmental latent variables, or sensor noise latent variables in the decoupling output;

[0086] This represents the square of the Frobenius norm, used to quantify the overall size of the covariance matrix, i.e., to measure the overall strength of the correlation between variables.

[0087] In this variation, the static attribute features of the monitoring area are first preprocessed. Because static data such as geographical coordinates, livestock size, and supply chain relationships are discrete or high-dimensional and sparse, directly inputting them into a neural network can easily lead to low gradient update efficiency. Therefore, an embedding layer is first used to map these static attribute features into low-dimensional, dense embedding vectors. This embedding vector is considered a background constraint for data generation and is concatenated and fused with the feature vectors of multi-source time-series monitoring data during the encoding stage. This informs the model of the current observed body temperature or environmental data under what farming scale and geographical environment it was generated. This is equivalent to providing the model with a physical prior, enabling it to distinguish data benchmarks under different backgrounds.

[0088] Subsequently, the time-series data, incorporating static constraints, is input into the encoder network and mapped to three predefined, independent latent spaces, generating pathological latent variables, environmental latent variables, and sensor noise latent variables, respectively. To mathematically ensure that these three latent variables do not interfere with each other—that is, to prevent information from environmental factors (such as rising temperatures) from leaking into the pathological variables—a decoupling mechanism based on minimizing mutual information is introduced during the model training phase. This mechanism, by adding a specific penalty term to the loss function, forces the statistical dependence between the distributions of different latent variables to be reduced to a minimum.

[0089] It's also worth noting that in conventional autoencoders, models often tend to reconstruct data using the most salient features (such as population fluctuations caused by ambient temperature), while ignoring weak pathological features or conflating the two. By introducing a covariance penalty term, an orthogonal filter is mathematically constructed. When the model attempts to encode environmental information into pathological variables, the covariance between the two is considered. This will significantly increase, leading to a greater total loss. Increase.

[0090] To minimize loss, the gradient descent algorithm forces the encoder to adjust the weights and find an encoding method that makes the covariance approach zero. This forced constraint ensures that the potential spaces are statistically independent (i.e. orthogonal), thus achieving feature decoupling based solely on statistical data characteristics rather than artificial rules. In terms of technical effectiveness, this means that the system can automatically identify and remove background noise that changes synchronously with the environment, retaining only independent fluctuations unrelated to the environment as pathological features, thereby providing a clean data foundation for subsequent accurate early warning.

[0091] Example 3:

[0092] In this embodiment, obtaining several parallel, dynamically propagating structural particles based on the purified pathological latent variables includes:

[0093] A set of dynamic propagation structure particles is initialized based on the particle rheometry variational inference algorithm. The dynamic propagation structure particles are used to characterize the propagation adjacency matrix hypothesis between nodes.

[0094] Define a kernel function to calculate the repulsive force between particles in a dynamically propagating structure;

[0095] Obtain the gradient of the target posterior distribution with respect to the current dynamic propagation structure particles, and iteratively update the state of each dynamic propagation structure particle according to the repulsive force until all dynamic propagation structure particles approach the true propagation structure posterior distribution.

[0096] In this embodiment, the propagation dependencies represented by the dynamic propagation structure particles include contact propagation relationships based on geographic proximity and cross-regional propagation relationships based on supply chain logistics and transportation networks; different dynamic propagation structure particles correspond to different combinations of propagation path weights.

[0097] It should be noted that although Expression 2 in Example 1 is concise in form, it obscures the specific mathematical origin of the repulsive force term. In order to enable those skilled in the art to accurately reproduce the specific algorithmic logic of generating repulsive force using the kernel function gradient, this embodiment specifically adopts a complete mathematical expression including the kernel function gradient expansion term, namely, a variation of Expression 2:

[0098] ;

[0099] ;

[0100] In the formula, Indicates the first l In the nth iteration step, the 1st iThe state matrix of each dynamically propagating structure particle is derived from the parameterized representation of the propagation probability weights between nodes within the monitoring area.

[0101] The learning rate step size represents the learning rate step size in the structure inference process. It is derived from the preset optimization hyperparameter and is used to control the magnitude of each iteration update.

[0102] The optimal transformation direction function, i.e., the Stein variational gradient, is used to guide the direction of particle movement in probability space.

[0103] The kernel function is used to calculate the similarity between two structural particles, and determines the range of interaction between particles. In this embodiment, the radial basis kernel function shown in Example 1 is preferred.

[0104] It represents the gradient of the logarithmic posterior probability calculated based on the purified pathological latent variables, i.e., the driving force term. Its value comes from the data likelihood assessment and guides the particles to move towards the propagation structure that can better explain the current pathological data.

[0105] This represents the gradient of the kernel function, i.e., the repulsive force term. Its value comes from the calculation of the derivative of the kernel function. When two particles are too close, this term will generate a vector repulsive force that forces them to separate.

[0106] Understandably, given the highly concealed and dynamic nature of disease transmission paths in actual farming and disease prevention scenarios, a single static adjacency matrix cannot cover all possible risk transmission paths. For example, viruses may spread short-range through airborne aerosols or human contact between geographically adjacent farmhouses, or they may spread long-range through supply chain logistics networks such as cross-regional feed transport vehicles and breeding pig transport vehicles. To accurately capture these uncertain transmission dependencies, this embodiment does not seek a single "optimal" transmission structure, but instead uses a particle flow variational inference algorithm, specifically the Stein variational gradient descent algorithm, to initialize and update a set of parallel dynamic transmission structure particles. Each dynamic transmission structure particle is essentially an independent instance of the adjacency matrix hypothesis, representing the transmission network with different weight combinations that may exist between nodes within the monitoring area at the current moment. For example, one particle might focus on characterizing contact transmission relationships based on geographic proximity, assuming the virus mainly spreads between adjacent farms; while another particle might focus on characterizing cross-regional transmission relationships based on supply chain logistics networks, assuming the virus is spreading along specific slaughter or feed transport routes.

[0107] To mathematically maintain this set of particles representing different hypotheses simultaneously and ensure that the system can approximate the true propagation structure posterior distribution, the system defines an iterative update strategy that includes a repulsive force mechanism. Specifically, the system first defines a kernel function (usually a radial basis function RBF) to measure the similarity of particles with different structures in the probability space and calculates the repulsive force between particles accordingly to prevent all particles from collapsing to the same local optimum. Subsequently, the system calculates the gradient of the target posterior distribution with respect to the current particle (i.e., the direction of data support for the hypothesis) and updates the state of each particle in combination with the repulsive force.

[0108] It's also understandable that the above variation consists of two parts (driving force term and repulsive force term): the driving force term ensures accuracy, forcing all structural particles to evolve in the direction that best matches the current observation data (i.e., the latent pathological variables). However, if there is only this term, the algorithm will degenerate into a conventional optimization that searches for the maximum a posteriori probability (MAP), causing all particles to eventually converge to the same most probable structure, thus losing diversity. However, in actual epidemics, there are often hidden logistics transmissions, and a single optimal solution is very likely to lead to missed cases.

[0109] Therefore, this embodiment introduces a second repulsive force term. This term ensures diversity by utilizing the gradient properties of the kernel function to generate a mutually repelling vector field between particles. When multiple particles attempt to converge on the same solution, the repulsive force increases, forcibly pushing them towards other high-probability regions in the probability space. This mechanism allows the system to simultaneously retain multiple reasonable propagation hypotheses (e.g., retaining hypothetical particles for both close-range contact propagation and long-range material propagation), thus mathematically constructing a multi-prevention model covering multiple possibilities. This significantly improves the system's robustness and adaptability in complex propagation scenarios.

[0110] Example 4:

[0111] In this embodiment, the step of inputting the pathological latent variables and the dynamic propagation structural particles into a continuous-time evolution model to obtain multiple epidemic evolution trajectories within a future preset time period includes:

[0112] Construct a neural stochastic differential equation model that includes drift and diffusion terms;

[0113] For each dynamic propagation structural particle, the pathological latent variables are solved by integral over the continuous time domain according to the neural stochastic differential equation model to obtain the continuous state evolution curve corresponding to the dynamic propagation structural particle, and the set of all dynamic propagation structural particle evolution curves is taken as the epidemic evolution trajectory.

[0114] It should be noted that the spread and diffusion of biological viruses is essentially a stochastic process that changes continuously over time, rather than a discrete-time step-jumping process assumed by traditional deep learning models (such as recurrent neural networks, RNNs). This embodiment uses neural stochastic differential equations to construct an evolutionary model. This model no longer directly predicts the state value at the next moment, but models the time derivative of the system state. By introducing Brownian motion, it simulates the random fluctuations inside the system, thereby describing the dynamic mechanism of the epidemic evolution in the continuous time domain.

[0115] Combining with Expression 3, by explicitly embedding the dynamically propagating structural particles into the drift function, the system mandates that the evolutionary process must follow the propagation topology assumed by the particle. This means that if particle k assumes "a strong connection exists between node A and node B," then the drift function will drive the virus to numerically spread from A to B. Since the system maintains M structural particles in parallel, this mechanism ensures that the structural uncertainties inferred from previous steps are fully preserved and propagated into future evolutionary trajectories, rather than being prematurely averaged.

[0116] It should also be noted that the diffusion term Responsible for introducing random uncertainty.

[0117] In actual epidemic transmission, even if the environment and contact relationship are determined, the spread of the virus is still affected by micro-random factors (such as individual immune differences and randomness of contact duration). The diffusion term adaptively learns the noise level of the current state through a neural network and injects random perturbations using Brownian motion, so that the generated trajectory is no longer a rigid curve, but a tubular region with a probability distribution.

[0118] This neural SDE-based design exhibits excellent adaptability in technical solutions: it allows the system to make predictions at any consecutive point in time, without being limited by a fixed data acquisition frequency (e.g., even if the sensor transmits data only once every 3 hours, the model can still predict the state at the 1.5-hour mark). Finally, the system uses a numerical integrator to solve the above equations, generating a unique continuous state evolution curve for each input dynamic propagation structure particle. The collection of all these curves together constitutes a bundle of epidemic evolution trajectories containing multiple propagation hypotheses and random possibilities, intuitively and completely demonstrating the entire probability space of epidemic development within a predetermined time period. This provides a quantitative basis rich in causal logic and uncertainty information for subsequent risk-sensitive decision-making.

[0119] Example 5:

[0120] In this embodiment, the step of performing nonlinear aggregation calculation on the epidemic evolution trajectory based on the risk sensitivity evaluation function to obtain a comprehensive risk assessment value includes:

[0121] Obtain the predicted loss value corresponding to each trajectory of the epidemic's evolution;

[0122] By introducing a risk aversion coefficient, the predicted loss values ​​for all epidemic evolution trajectories are summed exponentially.

[0123] The weighted summation result is logarithmically transformed to obtain the comprehensive risk assessment value.

[0124] The introduction of the risk aversion coefficient specifically includes:

[0125] The distribution dispersion among the multiple epidemic evolution trajectories is obtained, and the distribution dispersion is used as an indicator of cognitive uncertainty.

[0126] When the cognitive uncertainty index increases, the risk aversion coefficient is increased to improve the sensitivity of the comprehensive risk assessment value.

[0127] It should be noted that after the system obtains multiple epidemic evolution trajectories within a preset future time period through a continuous-time evolution model, in order to transform these highly random and structurally uncertain projection results into a single quantitative indicator with decision-making reference value, the system enters the core risk-sensitive decision-making stage. Given the typical asymmetric loss characteristics in the field of epidemic prevention and control—that is, the socio-economic losses caused by underreporting of major epidemics far outweigh the costs of false reporting—a simple arithmetic average aggregation method would lead to missed detections due to smoothing out a few highly destructive high-risk trajectories. Therefore, this embodiment employs a risk-sensitive evaluation function based on exponential utility theory to perform nonlinear aggregation calculations on all generated epidemic evolution trajectories to obtain a comprehensive risk assessment value, as shown in Expression Four.

[0128] It's also worth noting that traditional Bayesian model averaging typically uses an equal-weighted average (i.e., the extreme case of γ→0), which implicitly assumes risk neutrality. However, in an epidemic early warning scenario, if 99 out of M trajectories show no epidemic, and only one trajectory (based on a specific assumption of covert transmission) indicates a major epidemic, the arithmetic average will dilute the risk value, thus ignoring the danger. Expression four utilizes the properties of the exponential function exp, ensuring that trajectories with high loss values ​​dominate the summation term. When γ is large, the overall calculation result will closely approximate the maximum loss value among all trajectories, rather than the average. This mathematical treatment gives the system a bottom-line thinking capability, ensuring that as long as there is a reasonable possibility of catastrophic evolution, the system's overall risk assessment value will remain high, thereby avoiding underreporting.

[0129] To address the problem that fixed risk coefficients in traditional methods are ill-suited to complex and ever-changing environments, this embodiment proposes a dynamic parameter adjustment mechanism based on cognitive uncertainty. Instead of using a fixed γ value, the system adjusts in real-time according to the current prediction status. Specifically, the system first calculates the dispersion of multiple epidemic evolution trajectories, using it as an indicator of the model's cognitive uncertainty. When the trajectory bundles are highly divergent, it indicates significant disagreement in the model's judgments about the future. In this case, based on the early warning principle, the system should exhibit a higher degree of risk aversion. This dynamic adjustment process follows the following linear mapping logic:

[0130] ;

[0131] In the formula, This represents the basic risk aversion coefficient, which is derived from the system's preset safety level parameters under low uncertainty (baseline state).

[0132] α represents the sensitivity adjustment factor, which is a preset positive number used to control the weight of the impact of uncertainty on the risk coefficient;

[0133] The term and its internal terms represent the sample standard deviation of the predicted loss value, i.e. the dispersion of the distribution between trajectories, which is used to quantify cognitive uncertainty;

[0134] This represents the arithmetic mean of all trajectory prediction loss values, derived from statistical calculations of the current M loss values.

[0135] The standard deviation is used as a quantitative indicator of cognitive uncertainty because it intuitively reflects the degree of difference in outcomes caused by different propagation hypotheses (structural particles). When all particles point to similar outcomes, the standard deviation approaches 0, and γ reverts to its baseline value. The system provides routine early warnings; however, when different particles give drastically different results (e.g., some predict no problem, while others predict an outbreak), the standard deviation increases significantly, causing γ to rise rapidly. This mechanism mathematically forms an adaptive negative feedback protection loop: the higher the model uncertainty, the more inclined the decision-making process is to believe in the worst-case scenario. This design significantly improves the system's robustness in the face of novel variant viruses or complex transmission scenarios, ensuring that when scientific evidence is insufficient (i.e., the model cannot determine the outcome), alarms can be triggered first based on legal early warning principles, thus buying valuable response time for epidemic prevention.

[0136] Example 6:

[0137] In this embodiment, the multi-source time-series monitoring data includes livestock and poultry body temperature data, feed intake data, breeding environment temperature and humidity data, and ammonia concentration data acquired through IoT sensors;

[0138] The static attribute features include the geographical coordinates of the monitoring node, the scale of the livestock, and data on upstream and downstream supply chain relationships.

[0139] It should be noted that the biological characteristics data mainly include livestock and poultry body temperature data and feed intake data. Body temperature data is usually acquired by non-contact infrared thermal imaging arrays or implanted RFID temperature-sensing ear tags at high frequency (e.g., once every 10 minutes) to form a time series that can capture micro-thermal states. Feed intake data is recorded in real time by weighing sensors in smart feeding troughs to reflect changes in livestock and poultry appetite, which is a typical early symptom of many diseases (such as African swine fever). At the same time, the temperature and humidity data of the breeding environment and ammonia concentration data constitute environmental sensing data. Ammonia concentration is introduced because high concentrations of ammonia can damage the respiratory mucosa of livestock and poultry, leading to non-viral respiratory symptoms. If the model lacks this input, it is very easy to misjudge such environmental stress as infectious respiratory diseases.

[0140] For static attribute features, this part of the data acts as a conditional constraint variable in the conditional variational autoencoder, used to define the physical background and potential propagation topology of the monitoring nodes. The geographical coordinates establish the Euclidean distance of the nodes in space, which is the basis for calculating the probability of short-range aerosol propagation. The scale of stock determines the theoretical rate of disease spread within the nodes. Meanwhile, the upstream and downstream supply chain relationship data (such as the source of breeding pigs, feed distribution routes, and slaughter and transportation flows) constitute the topological skeleton of the cross-regional propagation network, providing prior knowledge for the initialization of dynamic propagation structure particles in the previous embodiment.

[0141] To transform the aforementioned multi-source heterogeneous data into a unified format that can be processed by deep neural networks, and to eliminate numerical biases caused by different units (such as temperature in degrees Celsius and ammonia in PPM), this embodiment uses the following data fusion and vectorization formulas to construct the input tensor and conditional vector:

[0142] ;

[0143] ;

[0144] In the formula, This represents the normalized input vector of the multi-source time-series monitoring data at time t, which will be directly input into the inference network (encoder) of the deep generative model.

[0145] This represents a vector concatenation operation, used to merge data features from different sources into a single tensor.

[0146] This indicates the raw livestock and poultry body temperature data collected at time t, which comes from a contact or non-contact body temperature sensor;

[0147] This indicates that the raw feed intake data collected at time t comes from the weighing records of the intelligent feeding equipment;

[0148] This represents the set of raw environmental sensor data (including temperature, humidity, and ammonia concentration) collected at time t, which originated from an environmental monitoring station.

[0149] , , This represents the historical statistical mean of each corresponding data type, derived from a sliding window calculation of the historical monitoring database;

[0150] , , This represents the historical standard deviation of each corresponding data type, used to perform Z-Score standardization and eliminate the impact of dimensional differences on gradient descent.

[0151] c represents the encoded static attribute feature vector, which is used as a conditional constraint input into the decoder and prior network of CVAE;

[0152] This represents the embedding mapping function, typically implemented by a fully connected neural network layer, used to map discrete or high-dimensional sparse static features into low-dimensional dense continuous vectors.

[0153] The geographical coordinates (latitude and longitude) of the monitoring nodes are represented by the GIS geographic information system.

[0154] The numerical value representing the stock size of the monitoring node is derived from the aquaculture record database;

[0155] The adjacency vector or ID code representing the upstream and downstream supply chain relationship is derived from the order records of the logistics and transportation management system.

[0156] It should also be noted that, for In actual farming scenarios, the baseline body temperature and environmental indicators vary greatly across different seasons (for example, pigsty temperatures are generally low in winter). By subtracting the mean and dividing by the standard deviation, the formula transforms absolute values ​​into "relative fluctuation values." This allows deep generative models to focus on capturing anomalous changes relative to the current context, rather than being misled by the magnitude of absolute values. This is crucial for decoupling environmental and pathological latent variables in the preceding examples. Only after eliminating the influence of dimensions can the model distinguish between environmental and pathological causes based on the statistical correlation of fluctuations (rather than numerical magnitude).

[0157] Secondly, for the construction of c, the introduction of the embedding mapping function is to address the sparsity problem of supply chain data. Supply chain relationships are usually discrete graph data (such as "node A is transported by vehicle B"), and directly using them as numerical inputs can lead to difficulties in model convergence. By mapping them to points in a continuous vector space through the embedding layer, the model can learn the potential similarities between different supply chain nodes (for example, two farms served by the same fleet are close to each other in the embedding space). This processing provides rich topological priors for the dynamic propagation structure particles in Example 3, enabling the particle flow inference algorithm to quickly identify high-risk propagation paths using supply chain information during the initialization phase, rather than blindly searching from scratch, thereby significantly improving the convergence speed and accuracy of structure inference.

[0158] Example 7:

[0159] In this embodiment, the pathological latent variables are used to characterize the abnormal fluctuations in biological signs caused by viral infection;

[0160] The environmental latent variables are used to characterize the background drift characteristics of population physiological indicators caused by seasonal changes or temperature variations;

[0161] The sensor noise latent variable is used to characterize random outlier features caused by equipment failure or signal transmission interference.

[0162] Understandably, in actual farming scenarios, viral infections (such as African swine fever or porcine reproductive and respiratory syndrome (PRRS) often lead to physiological abnormalities in livestock and poultry individuals independent of the environmental background. For example, sudden, persistent high fever or a sharp decrease in feed intake may occur even under suitable environmental conditions. Pathological latent variables aim to capture these anomalous signal components, while environmental latent variables characterize the background drift characteristics of group physiological indicators caused by seasonal changes or temperature variations. Because organisms exchange heat with their environment, an increase in environmental temperature inevitably leads to an overall upward shift in the baseline of livestock and poultry body temperature; this change is group-wide, periodic, and conforms to physical laws. Environmental latent variables, combined with static attribute characteristics (such as the latitude and longitude of the farm, and its enclosed structure), are responsible for explaining this part of the "reasonable" data fluctuations. Sensor noise latent variables are used to characterize random outliers caused by equipment failure or signal transmission interference. Because IoT devices are prone to aging drift, electromagnetic interference, or data loss in harsh farming environments, generating high-frequency, irregular spike noise, this variable is responsible for absorbing these biologically meaningless data residuals.

[0163] Example 8:

[0164] As attached Figure 2 As shown, this embodiment provides a disease early warning system that integrates deep probabilistic graphical models and Bayesian inference. The early warning system includes:

[0165] The data acquisition module is used to acquire multi-source time-series monitoring data and static attribute characteristics within the monitoring area. The multi-source time-series monitoring data includes biological sign data and environmental sensor data.

[0166] The decoupling and purification module is used to input the multi-source time-series monitoring data into the deep generation model, use the static attribute features as conditional constraints, and then map the multi-source time-series monitoring data to mutually independent latent spaces, so as to decouple and output pathological latent variables, environmental latent variables and sensor noise latent variables in the latent spaces, and purify the pathological latent variables through the environmental latent variables and sensor noise latent variables.

[0167] The structure inference module is used to obtain several parallel dynamic propagation structure particles based on the purified pathological latent variables. The dynamic propagation structure particles are used to characterize the uncertain propagation dependencies between nodes within the monitoring area.

[0168] An evolutionary projection module is used to input the pathological latent variables and the dynamic propagation structure particles into a continuous-time evolution model to obtain multiple epidemic evolution trajectories within a future preset time period.

[0169] The risk warning module is used to perform nonlinear aggregation calculations on the epidemic evolution trajectory based on a risk sensitivity evaluation function to obtain a comprehensive risk assessment value, and to trigger a graded warning signal when the comprehensive risk assessment value exceeds a preset threshold.

[0170] It should be noted that the system in this embodiment, as a highly integrated intelligent monitoring platform, achieves a closed-loop process from multi-source heterogeneous data perception to risk-sensitive scenario decision-making through the collaborative work between modules.

[0171] First, the data acquisition module, as the perception front end of the entire early warning system, is responsible for constructing a high-dimensional and complete input data space. Through an IoT sensor network deployed within the monitoring area, it continuously and in real-time collects multi-source time-series monitoring data, covering biological characteristics that directly reflect the health status of individual livestock (such as high-frequency sampled body temperature sequences and feed intake fluctuations) and environmental sensor data that reflects external environmental pressures (such as temperature and humidity in pigsties, ammonia and carbon dioxide concentrations). Simultaneously, this module also retrieves static attribute features corresponding to each monitoring node from the backend database, including geographical coordinates, livestock size, and supply chain logistics relationships. These static features not only provide physical background constraints for subsequent models but also establish the topological framework of the potential propagation network, ensuring that the system can understand the environmental context in which the data is generated.

[0172] Secondly, the decoupling and purification module takes over the original data stream and uses a deep generative model (specifically constructed as a conditional variational autoencoder architecture) to perform the core feature separation task. Static attribute features are concatenated as conditional constraints into the encoding process, guiding the model to map high-dimensional, mixed multi-source time-series monitoring data into three independent latent spaces with clear physical semantics. To overcome the difficulty in distinguishing between environmental heat stress and actual pathological reactions in traditional methods, this module introduces an orthogonal constraint mechanism based on minimizing mutual information during the mapping process, forcing that pathological latent variables, environmental latent variables, and sensor noise latent variables are statistically uncorrelated. Through this mechanism, the system can automatically remove the collective background drift (environmental latent variables) caused by seasonal temperature changes and the random outlier interference (sensor noise latent variables) caused by equipment failure, thereby accurately purifying only the pathological latent variables affected by viral infection, laying a clean data foundation for subsequent accurate inference.

[0173] Next, the structure inference module addresses the high uncertainty of transmission paths based on purified pathological latent variables. Given that viruses in real-world epidemic prevention scenarios may spread through various covert paths, such as short-range aerosols or long-range logistics chains, this module abandons the rigid setting of a static adjacency matrix and instead applies a particle flow variational inference algorithm (Stein variational gradient descent). This module initializes and updates a set of parallel, dynamically evolving propagation structure particles, each representing a potential hypothesis about the propagation dependencies between current nodes. By introducing the repulsive force generated by the kernel function, this module forces these particles to maintain diversity in the probability space, thereby simultaneously covering multiple possible propagation topologies and ensuring that the system does not miss any low-probability but high-risk covert propagation paths.

[0174] Subsequently, the evolutionary deduction module uses the aforementioned latent pathological variables as the initial state and the set of dynamic propagation structure particles representing different propagation hypotheses as topological parameters, inputting them into the continuous-time evolution model. This module constructs neural stochastic differential equations, using drift terms to describe the deterministic dynamic trend of epidemic development and diffusion terms to introduce Brownian motion to simulate random fluctuations. Through integral solving in the continuous-time domain, the module generates a bundle of possible epidemic evolution trajectories. These trajectories visually demonstrate the entire probability space of epidemic development within a predetermined time period under the combined influence of different propagation structure hypotheses and random factors, achieving dynamic deduction of future states.

[0175] Finally, the risk warning module executes a decision-making logic based on a bottom-line mentality. To avoid the few predicted trajectories pointing to major outbreaks being averaged out by the majority of stable trajectories, this module uses a risk-sensitive evaluation function to perform nonlinear aggregation calculations on the output epidemic evolution trajectories. By introducing a dynamically adjustable risk aversion coefficient, the module uses an exponential weighting mechanism to amplify high-loss trajectories and calculate a comprehensive risk assessment value. When this assessment value exceeds a preset threshold, the module triggers a tiered warning signal. This design ensures that when faced with model cognitive uncertainty, the system can prioritize worst-case scenarios based on warning principles, thereby effectively avoiding underreporting of major outbreaks and providing quantitative support for epidemic prevention decisions that balances scientific rigor and safety.

[0176] Furthermore, in one embodiment, this application also provides a computer storage medium storing a computer program, which, when executed by a processor, implements the steps of the methods described in the foregoing embodiments.

[0177] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or it may be a device including one or any combination of the above-mentioned memories. The computer may be a variety of computing devices, including smart terminals and servers.

[0178] In some embodiments, executable instructions may take the form of a program, software, software module, script, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a standalone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0179] As an example, executable instructions may, but do not necessarily, correspond to files in a file system. They may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a Hyper Text Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple collaborating files (e.g., a file that stores one or more modules, subroutines, or code sections).

[0180] As an example, executable instructions can be deployed to execute on a single computing device, or on multiple computing devices located in one location, or on multiple computing devices distributed across multiple locations and interconnected via a communication network.

[0181] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0182] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0183] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory / random access memory, magnetic disk, optical disk) and includes several instructions to cause a multimedia terminal device (which may be a mobile phone, computer, television receiver, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0184] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A disease early warning method integrating deep probabilistic graphical models and Bayesian inference, characterized in that, The early warning method includes the following steps: Acquire multi-source time-series monitoring data and static attribute characteristics within the monitoring area, wherein the multi-source time-series monitoring data includes biological characteristic data and environmental sensor data; The multi-source time-series monitoring data is input into a deep generation model, the static attribute features are used as conditional constraints, and the multi-source time-series monitoring data is mapped to mutually independent latent spaces to decouple and output pathological latent variables, environmental latent variables, and sensor noise latent variables in the latent spaces. The pathological latent variables are then purified through the environmental latent variables and sensor noise latent variables. Based on the purified pathological latent variables, several parallel dynamic propagation structure particles are obtained. These dynamic propagation structure particles are used to characterize the uncertain propagation dependencies between nodes within the monitoring area. The propagation dependencies characterized by these dynamic propagation structure particles include contact propagation relationships based on geographic proximity and cross-regional propagation relationships based on the supply chain logistics network. Different dynamic propagation structure particles correspond to different combinations of propagation path weights. Specifically, this includes: initializing a set of dynamic propagation structure particles according to a particle rheology variational inference algorithm, whereby these dynamic propagation structure particles characterize the propagation adjacency matrix hypothesis between nodes; defining a kernel function and using its gradient to generate a repulsive force term, forcing each dynamic propagation structure particle to move away from each other in the probability space, while simultaneously maintaining multiple reasonable propagation hypotheses; calculating the repulsive force between the dynamic propagation structure particles; obtaining the gradient of the target posterior distribution with respect to the current dynamic propagation structure particle; and iteratively updating the state of each dynamic propagation structure particle according to the repulsive force until all dynamic propagation structure particles approach the true propagation structure posterior distribution. The pathological latent variables and the dynamic propagation structure particles are input into a continuous-time evolution model to obtain multiple epidemic evolution trajectories within a future preset time period. This includes: constructing a neural stochastic differential equation model containing drift and diffusion terms; explicitly embedding each dynamic propagation structure particle into a drift function; integrating the pathological latent variables in the continuous-time domain according to the neural stochastic differential equation model to obtain the continuous state evolution curve corresponding to the dynamic propagation structure particle; and using the set of all dynamic propagation structure particle evolution curves as the epidemic evolution trajectory. The epidemic evolution trajectory is nonlinearly aggregated and calculated based on the risk sensitivity evaluation function to obtain a comprehensive risk assessment value, and a graded early warning signal is triggered when the comprehensive risk assessment value exceeds a preset threshold.

2. The disease early warning method integrating deep probabilistic graphical models and Bayesian inference as described in claim 1, characterized in that, The step of inputting the multi-source time-series monitoring data into a deep generative model, using the static attribute features as conditional constraints, and then mapping the multi-source time-series monitoring data to mutually independent latent spaces includes: A conditional variational autoencoder is constructed, and an embedding vector is obtained based on the static attribute features. The embedding vector is then concatenated into the encoding process of the multi-source time-series monitoring data to apply distribution constraints. During model training, the mutual information minimization mechanism is used to force the covariance between each pair of the pathological latent variables, environmental latent variables and sensor noise latent variables to approach zero.

3. The disease early warning method integrating deep probabilistic graphical models and Bayesian inference as described in claim 1, characterized in that, The step of performing nonlinear aggregation calculations on the epidemic evolution trajectory based on a risk sensitivity evaluation function to obtain a comprehensive risk assessment value includes: Obtain the predicted loss value corresponding to each trajectory of the epidemic's evolution; By introducing a risk aversion coefficient, the predicted loss values ​​for all epidemic evolution trajectories are summed exponentially. The weighted summation result is logarithmically transformed to obtain the comprehensive risk assessment value.

4. The disease early warning method integrating deep probabilistic graphical models and Bayesian inference as described in claim 1, characterized in that, The multi-source time-series monitoring data includes livestock and poultry body temperature data, feed intake data, breeding environment temperature and humidity data, and ammonia concentration data acquired through IoT sensors. The static attribute features include the geographical coordinates of the monitoring node, the scale of the livestock, and data on upstream and downstream supply chain relationships.

5. The disease early warning method integrating deep probabilistic graphical models and Bayesian inference as described in claim 1, characterized in that, The pathological latent variables are used to characterize the abnormal fluctuations in biological signs caused by viral infection; The environmental latent variables are used to characterize the background drift characteristics of population physiological indicators caused by seasonal changes or temperature variations; The sensor noise latent variable is used to characterize random outlier features caused by equipment failure or signal transmission interference.

6. The disease early warning method integrating deep probabilistic graphical models and Bayesian inference as described in claim 3, characterized in that, The introduction of the risk aversion coefficient specifically includes: The distribution dispersion among the multiple epidemic evolution trajectories is obtained, and the distribution dispersion is used as an indicator of cognitive uncertainty. When the cognitive uncertainty index increases, the risk aversion coefficient is increased to improve the sensitivity of the comprehensive risk assessment value.

7. A disease early warning system integrating deep probabilistic graphical models and Bayesian inference, characterized in that, The early warning system includes: The data acquisition module is used to acquire multi-source time-series monitoring data and static attribute characteristics within the monitoring area. The multi-source time-series monitoring data includes biological characteristic data and environmental sensor data. The decoupling and purification module is used to input the multi-source time-series monitoring data into the deep generation model, use the static attribute features as conditional constraints, and then map the multi-source time-series monitoring data to mutually independent latent spaces, so as to decouple and output pathological latent variables, environmental latent variables and sensor noise latent variables in the latent spaces, and purify the pathological latent variables through the environmental latent variables and sensor noise latent variables. The structural inference module is used to obtain several parallel dynamic propagation structural particles based on the purified pathological latent variables. These dynamic propagation structural particles characterize uncertain propagation dependencies between nodes within the monitoring area. The propagation dependencies characterized by these dynamic propagation structural particles include contact propagation relationships based on geographic proximity and cross-regional propagation relationships based on the supply chain logistics network. Different dynamic propagation structural particles correspond to different combinations of propagation path weights. Specifically, the module includes: initializing a set of dynamic propagation structural particles according to a particle rheological variational inference algorithm, whereby these particles characterize the propagation adjacency matrix assumption between nodes; defining a kernel function and using its gradient to generate a repulsive force term, forcing each dynamic propagation structural particle to move away from each other in the probability space while maintaining multiple reasonable propagation assumptions; calculating the repulsive force between the dynamic propagation structural particles; obtaining the gradient of the target posterior distribution with respect to the current dynamic propagation structural particles; and iteratively updating the state of each dynamic propagation structural particle based on the repulsive force until all dynamic propagation structural particles approach the true propagation structural posterior distribution. An evolutionary deduction module is used to input the pathological latent variables and the dynamic propagation structure particles into a continuous-time evolution model to obtain multiple epidemic evolution trajectories within a preset future time period. This includes: constructing a neural stochastic differential equation model containing drift and diffusion terms; explicitly embedding each dynamic propagation structure particle into a drift function; integrating the pathological latent variables over the continuous-time domain according to the neural stochastic differential equation model to obtain the continuous-state evolution curve corresponding to that dynamic propagation structure particle; and using the set of all dynamic propagation structure particle evolution curves as the epidemic evolution trajectory. The risk warning module is used to perform nonlinear aggregation calculations on the epidemic evolution trajectory based on a risk sensitivity evaluation function to obtain a comprehensive risk assessment value, and to trigger a graded warning signal when the comprehensive risk assessment value exceeds a preset threshold.

Images