CN-121726096-B - Epidemic early warning method and system integrating depth probability map model and Bayesian inference

CN121726096BCN 121726096 BCN121726096 BCN 121726096BCN-121726096-B

Abstract

The invention relates to the technical field of artificial intelligent monitoring, in particular to a epidemic early warning method and a epidemic early warning system integrating a depth probability map model and Bayesian inference, wherein multisource time sequence monitoring data and static attribute characteristics are input into a depth generation model, mapped into independent potential space, and are decoupled to output pathological, environmental and noise variables, and the latter two are utilized to purify the pathological variables; the method comprises the steps of obtaining parallel dynamic propagation structure particles based on purification characteristics, inputting a continuous time evolution model to generate a plurality of epidemic situation evolution tracks, finally calculating comprehensive risk values through a risk sensitive evaluation function to trigger early warning, realizing accurate stripping of environmental drift and random noise from clutter, remarkably reducing false alarm rate, simultaneously deducing multiple propagation assumptions in parallel, amplifying high risk track weights through nonlinear aggregation, and guaranteeing that long-tail disaster risks can be effectively captured when the risks face uncertainty.

Inventors

WANG WEI
CHEN YING
YANG XIAO
FANG DONGHUI
YI JUN
DENG XIAODONG
CHEN XIAOYUN
MA XIAOQIN
SHI YI
Agoyoda

Assignees

四川省畜牧科学研究院
四川云牧智眸科技有限责任公司

Dates

Publication Date: 20260512
Application Date: 20260225

Claims (7)

1. The epidemic early warning method integrating the depth probability map model and the Bayesian inference is characterized by comprising the following steps of: acquiring multi-source time sequence monitoring data and static attribute characteristics in a monitoring area, wherein the multi-source time sequence monitoring data comprises biological sign data and environmental sensing data; Inputting the multi-source time sequence monitoring data into a depth generation model, taking the static attribute characteristics as condition constraints, mapping the multi-source time sequence monitoring data into mutually independent potential spaces so as to decouple and output a pathological potential variable, an environment potential variable and a sensor noise potential variable in the potential spaces, and purifying the pathological potential variable through the environment potential variable and the sensor noise potential variable; Obtaining a plurality of parallel dynamic propagation structure particles according to the purified pathological potential variables, wherein the dynamic propagation structure particles are used for representing uncertain propagation dependency relations among nodes in a monitoring area, the propagation dependency relations represented by the dynamic propagation structure particles comprise a contact propagation relation of geographic space adjacency and a trans-regional propagation relation based on a supply chain logistics transportation network, different dynamic propagation structure particles correspond to combination assumptions of different propagation path weights, and concretely comprise initializing a group of dynamic propagation structure particles according to a particle flow variation inference algorithm, wherein the dynamic propagation structure particles are used for representing propagation adjacency matrix assumptions among the nodes, defining a kernel function, utilizing gradients of the kernel function to generate repulsive force items so as to force each dynamic propagation structure particle to be far away from each other in a probability space, simultaneously keeping a plurality of reasonable propagation assumptions, calculating the repulsive force gradient of target posterior distribution on current dynamic propagation structure particles, updating the state of each dynamic propagation structure particle according to iteration until all the dynamic propagation structure particles approach real propagation structure posterior distribution; inputting the pathological potential variables and the dynamic propagation structure particles into a continuous time evolution model to acquire a plurality of epidemic situation evolution tracks in a preset time period in the future, wherein the method comprises the steps of constructing a neural random differential equation model comprising a drift term and a diffusion term, explicitly embedding the dynamic propagation structure particles into a drift function for each dynamic propagation structure particle, carrying out integral solution on the pathological potential variables in a continuous time domain according to the neural random differential equation model to acquire a continuous state evolution curve corresponding to the dynamic propagation structure particles, and taking a set of all dynamic propagation structure particle evolution curves as the epidemic situation evolution tracks; and carrying out nonlinear aggregation calculation on the epidemic evolution track according to a risk sensitivity evaluation function to obtain a comprehensive risk evaluation value, and triggering a grading early warning signal when the comprehensive risk evaluation value exceeds a preset threshold.
2. The epidemic early warning method of claim 1, wherein the inputting the multi-source time series monitoring data into the depth generation model, using the static attribute feature as a condition constraint, and mapping the multi-source time series monitoring data to mutually independent potential spaces comprises: Constructing a conditional variation self-encoder, acquiring an embedded vector according to the static attribute characteristics, and splicing the embedded vector into the encoding process of the multi-source time sequence monitoring data to apply distribution constraint; The covariance between the pathological latent variable, the environment latent variable and the sensor noise latent variable is constrained to be close to zero by mutual information minimization mechanism in the model training process.
3. The epidemic early warning method of claim 1, wherein the performing nonlinear aggregate calculation on the epidemic evolution track according to a risk sensitivity evaluation function to obtain a comprehensive risk evaluation value comprises: obtaining a predicted loss value corresponding to each epidemic situation evolution track; introducing a risk aversion coefficient, and carrying out exponential weighting summation on the predicted loss values of all epidemic situation evolution tracks; and carrying out logarithmic transformation on the weighted and summed result to obtain the comprehensive risk assessment value.
4. The epidemic early warning method based on the depth probability map model and Bayesian inference, as set forth in claim 1, wherein the multi-source time sequence monitoring data comprises livestock and poultry body temperature data, feed intake data, breeding environment temperature and humidity data and ammonia concentration data obtained through an internet of things sensor; The static attribute features include geographical location coordinates of the monitoring nodes, stock scale, and upstream and downstream supply chain relationship data.
5. The epidemic early warning method by combining a depth probability map model and Bayesian inference according to claim 1, wherein, The pathological latent variables are used for characterizing abnormal fluctuation characteristics of organism signs caused by virus infection; The environment latent variable is used for representing the background drift characteristics of the group physiological indexes caused by season replacement or air temperature change; the sensor noise latent variable is used to characterize random outliers caused by equipment failure or signal transmission disturbances.
6. The epidemic early warning method of combining a depth probability map model and bayesian inference as set forth in claim 3, wherein the introducing risk aversion coefficient specifically comprises: Acquiring distribution dispersion among the epidemic situation evolution tracks, and taking the distribution dispersion as a cognitive uncertainty index; When the cognitive uncertainty index increases, the risk aversion coefficient is increased to increase the sensitivity of the integrated risk assessment value.
7. The epidemic early warning system integrating the depth probability map model and the Bayesian inference is characterized by comprising: the data acquisition module is used for acquiring multi-source time sequence monitoring data and static attribute characteristics in a monitoring area, wherein the multi-source time sequence monitoring data comprises biological sign data and environment sensing data; The decoupling and purifying module is used for inputting the multi-source time sequence monitoring data into a depth generation model, taking the static attribute characteristics as conditional constraints, mapping the multi-source time sequence monitoring data into mutually independent potential spaces so as to decouple and output a pathological potential variable, an environment potential variable and a sensor noise potential variable in the potential spaces, and purifying the pathological potential variable through the environment potential variable and the sensor noise potential variable; The structure inference module is used for obtaining a plurality of parallel dynamic propagation structure particles according to the purified pathological potential variables, wherein the dynamic propagation structure particles are used for representing uncertain propagation dependency relations among nodes in a monitoring area, the propagation dependency relations represented by the dynamic propagation structure particles comprise contact propagation relations of geographic space adjacency and cross-region propagation relations based on a supply chain logistics network, different dynamic propagation structure particles correspond to combination assumptions of different propagation path weights, and the structure inference module specifically comprises initializing a group of dynamic propagation structure particles according to a particle flow variation inference algorithm, wherein the dynamic propagation structure particles are used for representing propagation adjacency matrix assumptions among the nodes, defining a kernel function, utilizing gradients of the kernel function to generate repulsive force items, forcing all the dynamic propagation structure particles to be far away from each other in a probability space so as to simultaneously maintain a plurality of reasonable propagation assumptions, calculating repulsive force among the dynamic propagation structure particles, obtaining gradients of target posterior distribution versus current dynamic propagation structure particles, and iteratively updating the state of each dynamic propagation structure particle until all the dynamic propagation structure particles approach real propagation structure posterior distribution; The evolution deduction module is used for inputting the pathological potential variables and the dynamic propagation structure particles into a continuous time evolution model to acquire a plurality of epidemic situation evolution tracks in a preset time period in the future, and comprises the steps of constructing a neural random differential equation model comprising drift items and diffusion items, explicitly embedding the dynamic propagation structure particles into a drift function for each dynamic propagation structure particle, carrying out integral solution on the pathological potential variables on a continuous time domain according to the neural random differential equation model to acquire continuous state evolution curves corresponding to the dynamic propagation structure particles, and taking a set of all dynamic propagation structure particle evolution curves as the epidemic situation evolution tracks; The risk early warning module is used for carrying out nonlinear aggregation calculation on the epidemic situation evolution track according to a risk sensitivity evaluation function so as to obtain a comprehensive risk evaluation value, and triggering a grading early warning signal when the comprehensive risk evaluation value exceeds a preset threshold value.

Description

Epidemic early warning method and system integrating depth probability map model and Bayesian inference Technical Field The invention relates to the technical field of artificial intelligence monitoring, in particular to a epidemic early warning method and a epidemic early warning system integrating a depth probability map model and Bayesian inference. Background In modern large-scale farms or regional epidemic prevention monitoring, the on-site Internet of things sensors are generally utilized to collect vital sign data such as body temperature, feed intake and the like of livestock and poultry in real time, and environmental parameters such as temperature, humidity, ammonia concentration and the like of a pig house, and the health state of livestock and poultry groups is monitored by analyzing fluctuation trend of the multi-source time sequence data. However, in practical application scenarios, the raw data collected by the monitoring device often contains multiple mixed components, and it is difficult for the existing conventional monitoring means to effectively distinguish the root sources of the data fluctuation. Specifically, the system cannot accurately distinguish whether an abnormal data fluctuation (such as temperature rise) is caused by pathological changes caused by virus infection of livestock and poultry individuals, or a group environmental heat stress response caused by season replacement and temperature rising, or a random noise or outlier generated by ageing of sensor equipment and unstable signal transmission. Due to the lack of decoupling and purifying capabilities of the different attribute characteristics, the conventional early warning system has remarkable limitations when facing a complex culture environment, for example, when the environment parameters are greatly drifted, the system easily confuses the environment factors and pathological factors to cause large-area false alarm, so that the manual investigation cost is increased, and if the smooth noise reduction treatment is simply carried out, the early weak real pathological signals are extremely easy to filter as equipment noise, so that the critical epidemic information is not reported, and the problem that the real infection characteristics cannot be accurately extracted from the mixed signals seriously affects the reliability and timeliness of early warning results. Disclosure of Invention The invention mainly aims to provide a epidemic early warning method integrating a depth probability map model and Bayesian inference, and aims to solve the problem that the early warning method in the prior art is difficult to integrate characteristic data with different attributes, so that the early warning effect is poor. In order to achieve the above purpose, the present invention provides a epidemic early warning method integrating a depth probability map model and bayesian inference, the early warning method comprising the following steps: acquiring multi-source time sequence monitoring data and static attribute characteristics in a monitoring area, wherein the multi-source time sequence monitoring data comprises biological sign data and environmental sensing data; Inputting the multi-source time sequence monitoring data into a depth generation model, taking the static attribute characteristics as condition constraints, mapping the multi-source time sequence monitoring data into mutually independent potential spaces so as to decouple and output a pathological potential variable, an environment potential variable and a sensor noise potential variable in the potential spaces, and purifying the pathological potential variable through the environment potential variable and the sensor noise potential variable; Acquiring a plurality of parallel dynamic propagation structure particles according to the purified pathological potential variables, wherein the dynamic propagation structure particles are used for representing uncertain propagation dependency relations among nodes in a monitoring area; inputting the pathological potential variables and the dynamic propagation structure particles into a continuous time evolution model to obtain a plurality of epidemic situation evolution tracks in a future preset time period; and carrying out nonlinear aggregation calculation on the epidemic evolution track according to a risk sensitivity evaluation function to obtain a comprehensive risk evaluation value, and triggering a grading early warning signal when the comprehensive risk evaluation value exceeds a preset threshold. In order to achieve the above object, the present invention further provides a epidemic early warning system that fuses a depth probability map model and bayesian inference, the early warning system comprising: the data acquisition module is used for acquiring multi-source time sequence monitoring data and static attribute characteristics in a monitoring area, wherein the multi-source time sequence monitoring data comprises biol