Live pig health intelligent evaluation method and system based on multi-modal behavior perception

Abstract

The invention provides an intelligent live pig health assessment method and system based on multi-modal behavior perception, and relates to the technical field of live pig breeding, and the method comprises the steps: obtaining a basic data set; performing multi-modal behavior feature fusion processing according to the basic data set, and extracting joint behavior features; performing graph structure construction according to the joint behavior characteristics to obtain a directed weighted group behavior graph; performing group behavior dynamic evolution according to the directed weighted group behavior graph to generate dynamic embedding representation; performing group health reasoning according to the dynamic embedding representation to output a reasoning result, wherein the reasoning result comprises an individual health risk probability and a group propagation influence factor; and performing group health decision according to the reasoning result, and generating a joint evaluation result. According to the method, the accuracy and operability of health assessment of live pig groups are improved through space-time collaborative fusion of depth video and audio data, graph topology construction constrained by a breeding functional area, dynamic gating evolution of contact exposure double-path propagation and a decision mapping mechanism driven by virtual blocking.

Description

Technical Field

[0001] The present invention relates to the field of pig farming technology, and in particular to a method and system for intelligent pig health assessment based on multimodal behavior perception. Background Art

[0002] With the rapid development of smart farming technologies, automated assessment of pig health status has become a core requirement for large-scale pig farm management, particularly in intensive farming environments, where the risk of disease transmission in pig herds is significantly increased. Current technologies face a key bottleneck: in closed, high-density farming environments, disease transmission in pigs occurs through a dual, coupled pathway: close contact transmission and environmental vector transmission. Crowding and competition at feed troughs lead to direct contact infection through oral and nasal secretions, while retained feces in excretion areas spreads virulence across regions through airflow. This complex transmission mechanism makes it difficult for traditional health assessment methods to accurately quantify risk. Existing technologies primarily rely on two approaches: First, wearable sensors monitor individual behavioral indicators. While these sensors can capture individual pig activity data, they cannot capture the topology of transmission within the group interaction network. Furthermore, these devices are expensive and can easily induce animal stress. Second, single-modal visual analysis, which identifies abnormal individuals, has some application in detecting overt symptoms such as coughing and lameness, but ignores the impact of key behavioral characteristics such as feeding synchrony and group movement rhythms on transmission dynamics, and cannot model the cumulative virulence of environmental vectors.

[0003] Based on the shortcomings of the above-mentioned existing technologies, there is an urgent need for an intelligent pig health assessment method and system based on multimodal behavior perception. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for intelligently assessing pig health based on multimodal behavioral perception to improve the above-mentioned problems. To achieve the above-mentioned purpose, the technical solutions adopted by the present invention are as follows:

[0005] In a first aspect, the present application provides a method for intelligently assessing pig health based on multimodal behavioral perception, comprising:

[0006] Acquire a basic data set, wherein the basic data set includes deep video data, audio data, and individual motion trajectory data of a pig group;

[0007] Performing multimodal behavioral feature fusion processing based on the basic data set to extract joint behavioral features with group spatial correlation;

[0008] A graph structure is constructed based on the joint behavior characteristics, and a directed weighted group behavior graph is constructed by calculating the contact frequency and behavior synchronization between individuals;

[0009] Dynamically evolving group behavior based on the directed weighted group behavior graph, capturing state dependencies between nodes and quantifying the propagation of group health state deviation signals to generate a dynamic embedding representation;

[0010] Performing group health reasoning based on the dynamic embedding representation, and outputting reasoning results by simulating the diffusion path of health status deviation in the group topology, wherein the reasoning results include individual health risk probability and group transmission impact factor;

[0011] Group health decisions are made based on the reasoning results to generate joint evaluation results.

[0012] In a second aspect, the present application also provides a pig health intelligent assessment system based on multimodal behavior perception, comprising:

[0013] An acquisition module is used to acquire a basic data set, wherein the basic data set includes deep video data, audio data, and individual motion trajectory data of a pig group;

[0014] A fusion module, configured to perform multimodal behavioral feature fusion processing based on the basic data set to extract joint behavioral features with group spatial correlation;

[0015] A construction module is used to construct a graph structure based on the joint behavior characteristics, and to construct a directed weighted group behavior graph by calculating the contact frequency and behavior synchronization between individuals;

[0016] An evolution module is used to dynamically evolve group behavior according to the directed weighted group behavior graph, capture the state dependencies between nodes, and quantify the propagation process of group health state deviation signals to generate a dynamic embedding representation;

[0017] An inference module, configured to perform group health inference based on the dynamic embedding representation, and output inference results by simulating the diffusion path of health status deviation in the group topology, wherein the inference results include individual health risk probabilities and group transmission impact factors;

[0018] The decision-making module is used to make group health decisions based on the reasoning results and generate joint evaluation results.

[0019] The beneficial effects of the present invention are:

[0020] The present invention significantly improves the accuracy and operability of pig group health assessment through the spatiotemporal collaborative fusion of deep video and audio data, the graph topology construction of breeding functional area constraints, the dynamic gating evolution of contact-exposure dual-path transmission, and the decision-making mapping mechanism driven by virtual blocking. It realizes the complementary enhancement of cross-source behavioral characteristics at the multimodal level and supports the fine topological expression of the group interaction network. It finally generates a joint assessment result that combines individual isolation operation instructions with environmental control plans, overcoming the defects of traditional methods with a single evaluation dimension and disconnected prevention and control guidelines, and comprehensively improving the timeliness and targeted prevention and control of farm health management responses. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.

[0022] Figure 1 This is a flow chart of a method for intelligently assessing pig health based on multimodal behavior perception according to an embodiment of the present invention;

[0023] Figure 2 This is a schematic diagram of the structure of a pig health intelligent assessment system based on multimodal behavior perception according to an embodiment of the present invention;

[0024] Figure 3 This is a schematic structural diagram of a pig health intelligent assessment device based on multimodal behavior perception described in an embodiment of the present invention.

[0025] Markings in the figure: 800, an intelligent pig health assessment device based on multimodal behavior perception; 801, processor; 802, memory; 803, multimedia component; 804, I / O interface; 805, communication component; 901, acquisition module; 902, fusion module; 903, construction module; 904, evolution module; 905, reasoning module; 906, decision module. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0027] It should be noted that similar reference numerals and letters represent similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings. At the same time, in the description of the present invention, the terms "first", "second", etc. are used only to distinguish the description and should not be understood as indicating or implying relative importance.

[0028] Example 1:

[0029] This embodiment provides a method for intelligently assessing pig health based on multimodal behavior perception.

[0030] See also Figure 1 , the figure shows that the method includes steps S100 to S600.

[0031] Step S100: obtaining a basic data set, wherein the basic data set includes depth video data, audio data, and individual motion trajectory data of a pig group;

[0032] It can be understood that this step uses a depth camera system deployed in the pig house to continuously capture three-dimensional dynamic images containing spatial dimension information, and simultaneously uses a microphone array to record environmental sound signals; at the same time, a positioning system (such as tracking technology based on UWB or computer vision) is used to automatically obtain and record the position coordinates of each pig and its motion trajectory data, so as to comprehensively obtain the pig's behavioral performance, sound characteristics and spatial position information.

[0033] Step S200: performing multimodal behavior feature fusion processing based on the basic data set to extract joint behavior features with group spatial correlation;

[0034] It should be noted that this step performs strict time synchronization processing on the collected visual images, sound signals, and position motion data to achieve temporal alignment; based on a unified timestamp, the intrinsic temporal and spatial correlations of different data modalities (visual, auditory, and position) are analyzed, and the sound characteristics or group movement patterns corresponding to specific behavioral patterns (such as fighting and feeding) are identified, thereby inferring the changes in the individual physiological state of pigs and the overall group behavior patterns.

[0035] Step S300: constructing a graph structure based on the joint behavior characteristics, and calculating the contact frequency and behavior synchronization between individuals to construct a directed weighted group behavior graph;

[0036] It can be understood that this step abstracts each pig as a node in the graph model; by quantitatively analyzing the associated data processed by step S200, the direct interaction behavior between individual pigs (such as approach distance, following speed, contact frequency and direction) is identified; according to the interaction type, frequency and dominant direction, directed weighted edge connections with directionality and intensity weights are established between nodes, and a social relationship network model is constructed that can quantitatively characterize the social hierarchy structure within the pig herd and the distribution of influence between individuals.

[0037] Step S400: performing dynamic evolution of group behavior according to the directed weighted group behavior graph, capturing state dependencies between nodes, and quantifying the propagation process of group health state deviation signals to generate a dynamic embedding representation;

[0038] It should be noted that this step conducts continuous dynamic monitoring of the social relationship network model, tracking the changes in node attributes (such as pig activity, visual / auditory indicators representing health) and edge connection weights (interaction strength and direction) in the network topology over time; quantitatively analyzes how individual status abnormalities (such as a sudden drop in the activity of a node) are transmitted through the network's connection relationships and affect the behavioral performance of its associated nodes (other pigs), thereby evaluating the potential diffusion path and transmission trend of health risk factors within the group.

[0039] Step S500: Perform group health reasoning based on the dynamic embedding representation, and output reasoning results by simulating the diffusion path of health status deviation in the group topology. The reasoning results include individual health risk probabilities and group transmission impact factors.

[0040] Based on the dynamically evolving network obtained in step S400, combined with individual key health characteristic indicators extracted from multimodal data (such as abnormal gait, cough frequency, and local body temperature changes), a probability model is applied to calculate the individual disease risk value of each pig; at the same time, based on the topological position of each pig in the network (such as centrality), the strength and directionality of the associated edges, the factors affecting the possible disease transmission of the pig on neighboring pigs and the entire group are calculated, thereby realizing a comprehensive risk assessment that integrates the degree of individual health abnormality and its potential for transmission in the group.

[0041] Step S600: Make group health decisions based on the inference results and generate joint evaluation results.

[0042] It can be understood that this step divides all pigs into risk levels based on the calculated individual disease risk values ​​and transmission influence factors; based on this, a specific set of executable intervention instructions is generated, including marking high-risk individuals who need to be immediately isolated and treated separately, identifying medium-risk individuals and their directly related network members who need to be strengthened observation and monitoring, and proposing management suggestions such as adjusting pig house environmental parameters (such as ventilation volume, temperature and humidity) and strengthening epidemic prevention and disinfection measures in specific areas based on the overall exposure level of the group, which directly guide the daily operations of the farm.

[0043] Furthermore, step S200 includes steps S210 to S230.

[0044] Step S210: extracting group spatial topological features based on the depth video data, extracting skeleton posture features of each pig individual through a spatiotemporal aligned three-dimensional convolutional neural network, and calculating the relative spatial position matrix between individuals;

[0045] Step S220: performing group acoustic behavior correlation analysis based on the audio data and individual motion trajectory data, identifying the spatial source of a specific sound event based on a preset sound source localization-trajectory matching algorithm, and constructing an acoustic-motion joint distribution feature;

[0046] Step S230: performing cross-modal group behavior fusion based on the relative spatial position matrix, the skeleton posture features, and the acoustic-motion joint distribution features, and splicing them to generate a joint behavior feature tensor.

[0047] Specifically, in step S210, a three-dimensional convolutional neural network is used to process the depth video data: the network adopts a spatiotemporal filter design, which simultaneously captures the posture changes and spatial continuity of pig movements by applying three-dimensional convolution kernels (spatial dimensions X / Y and temporal dimension T) on consecutive frames of the video; the network architecture includes multiple layers of convolution and pooling operations, and finally outputs the accurate positions of 17 key points of the pig skeleton; based on these skeleton points, the relative spatial position matrix is ​​further calculated, which quantifies the topological structure of the pig herd by comprehensively evaluating the Euclidean distance and movement direction differences between individuals.

[0048] In step S220, the sound source localization and trajectory matching algorithm is implemented: the sound source localization adopts the improved arrival time difference principle, and the time difference of sound arrival is recorded by multiple microphones deployed in the pig house, and the position of the sound source is accurately calculated in combination with the spatial coordinates of the microphone array; at the same time, a trajectory matching mechanism is designed to dynamically correlate the located sound source position with the position trajectory data of individual pigs. This process uses a dynamic time warping algorithm to achieve temporal alignment of audio events and motion trajectories, and finally generates a multi-dimensional joint feature that integrates acoustic characteristics (such as fundamental frequency changes) and motion patterns (such as changes in moving speed).

[0049] In step S230, cross-modal feature fusion is performed: the spatial position matrix obtained in step S210 is first flattened into a feature vector, and the skeleton posture features are encoded into a high-dimensional representation through a fully connected layer; then these visual modal features are tensor-concatenated with the acoustic-motion joint features generated in step S220; a feature normalization mechanism is introduced in the entire fusion process to ensure the scale consistency of data of different modalities, and finally a unified composite feature tensor is constructed.

[0050] The multimodal behavior feature synthesis formula is:

[0051]

[0052] in, represents the joint behavior feature tensor; N represents the number of individuals in the pig herd; i represents the number of the individual pig; D i represents the depth video data cube of the i-th pig; Γ(·) represents the skeleton key point extraction network; M pos Represents the spatial relative position relationship matrix; represents the feature space relationship weighting operator; σ(·) represents the spatial topology activation function; Λ(·) represents the sound source localization function; A represents the multi-channel audio signal matrix; Υ(·) represents the motion trajectory dynamic regularization function; T v represents the visual motion trajectory tensor; represents the optimal path dynamic time warping difference; δ(·) represents the acoustic-motion feature encoder; Ψ(·) represents the cross-modal feature fusion operator; Represents a multimodal feature concatenation operator.

[0053] Furthermore, step S300 includes steps S310 to S330.

[0054] Step S310: quantify physical contact under spatial constraints of the farm based on the joint behavioral characteristics, calculate the co-occurrence frequency of individuals in the feed trough area and the lying area, and generate a physical contact intensity matrix using trajectory collision detection;

[0055] Step S320: performing synchronization analysis of the breeding behavior cycle based on the joint behavior characteristics, matching the fixed event group of eating or drinking through the time window, calculating the phase difference of individual actions in the group, and generating a scenario-based behavior synchronization matrix;

[0056] Step S330: construct a hierarchical directed graph based on the physical contact strength matrix and the scenario-based behavior synchronization matrix, determine the leadership node through resource competition dominance analysis and perform adversarial contact detection, and establish a directed weighted group behavior graph including leadership follower edges, competitive confrontation edges, and neutral coexistence edges.

[0057] Specifically, the process first realizes spatial behavior analysis by quantifying the physical contact intensity: using the joint behavioral characteristic data of live pigs, the co-occurrence frequency between individuals is calculated in fixed areas of the farm (the feeding trough area and the lying area), and at the same time, a physical contact intensity matrix is ​​generated based on the trajectory intersection detection algorithm (when the distance function value of the motion trajectory of two bodies is lower than the set threshold, it is determined to be a contact event); then, group behavior synchronization analysis is performed: for periodic behaviors such as eating and drinking, a time window matching algorithm is used to align similar event groups, and the phase angle difference of different individual action sequences is calculated (preferably, cosine similarity is used to align the phase angle difference of different individual action sequences). The method uses a metric (e.g., a metric of physical contact intensity and synchronicity) to quantify the degree of behavioral coordination and output a scenario-based behavioral synchronization matrix. Finally, a hierarchical relationship network is constructed: integrating the results of physical contact intensity and synchronization analysis, first performing resource competition dominance analysis (identifying the leading node that occupies a priority position in the competition for feed), then conducting adversarial contact detection (identifying individual combinations with high contact intensity but low synchronization), and finally establishing a directed weighted group behavior graph with three types of edges (leader-follower edges are determined by dominance analysis to determine direction and weight, competitive adversarial edges are constructed based on adversarial detection results, and neutral coexistence edges are connections between individuals without obvious interactive relationships). The behavioral phase difference measurement formula is:

[0058]

[0059] in, is the behavioral phase angle difference between individuals i and j; T is the total number of time points in the behavioral cycle; t represents the sequence number of the time point; A i(t) represents the action intensity value of individual i at time point t (such as eating speed, drinking time); A j (t) represents the action intensity value of individual j at time point t.

[0060] Furthermore, step S400 includes steps S410 to S430.

[0061] Step S410: performing propagation path topology analysis based on the directed weighted group behavior graph, extracting high-propagation-risk edges through edge weight gradient threshold segmentation, and assigning the high-propagation-risk edges state inheritance attributes or bidirectional contamination attributes to obtain high-risk edge information;

[0062] Step S420: quantifying environmental exposure hotspots based on the depth video data and the individual motion trajectory data, and generating an environmental exposure vector by calculating the individual's cumulative exposure time in a preset functional area and combining it with a preset regional toxin load attenuation model;

[0063] Step S430: Perform state evolution based on the high-risk edge information and the environmental exposure vector, fuse contact propagation and environmental exposure paths through a gated spatiotemporal graph neural network, introduce a latency correction function to update the node health offset, and generate a dynamic embedding representation carrying the spatiotemporal propagation imprint.

[0064] In a specific implementation, the topology of the transmission path is first analyzed: based on the edge weight distribution characteristics of the directed weighted group behavior graph, the gradient threshold segmentation algorithm is applied to identify high-transmission risk edges whose weight change rate exceeds the critical threshold, and these edges are assigned two types of infectious attributes - state inheritance attributes (disease state propagation along the edge direction) or bidirectional pollution attributes (mutual cross-infection risk); then the environmental exposure hotspot is quantified: based on the individual movement trajectory data, the cumulative residence time in the preset functional areas such as the trough area and the drinking water area is calculated, and the regional viral load attenuation model is combined to generate an environmental exposure vector that represents the exposure risk of each area; finally, the health state evolution is performed: the high-risk edge transmission path and the environmental exposure vector are integrated through the gated spatiotemporal graph neural network. The network architecture design includes a time gating unit (memory latent period temporal characteristics) and a space gating unit (encoding contact transmission path). At the same time, the latent period correction function (defined as a nonlinear transformation of the latent period coefficient and the infection pressure) is introduced to dynamically update the node health offset, and finally the dynamic embedding representation carrying the spatiotemporal transmission characteristics is output. Among them, the health state evolution formula is:

[0065]

[0066] Where Δh v Indicates the updated health offset value of node v; f latency (τ v ) represents the latent period correction function; τv is the individual incubation period coefficient; α represents the influence weight of the contact transmission path; GST-GNN(·) represents the gated spatiotemporal graph neural network operator; represents the set of high-risk neighbors of node v; w uv represents the edge weight from node u to v; h u represents the current health status vector of neighbor node u; β represents the environmental exposure impact weight; γ(·) represents the environmental exposure feature encoder; represents the environmental exposure vector of node v.

[0067] Furthermore, step S500 includes steps S510 to S530.

[0068] Step S510: Construct an individual symptom evidence reasoning chain based on the dynamic embedding representation, infer the causal relationship between cough fundamental frequency shift and respiratory tract infection, the reverse reasoning relationship between the rate of decrease in eating speed and digestive tract disease, and the conditional probability distribution between hot zone deviation and systemic disease, thereby generating an initial reasoning chain for individual health risk probability.

[0069] Step S520: Calculate the contact transmission intensity of the trough and the air diffusion attenuation effect based on the dynamic embedding representation, integrate the spatial distance attenuation by the directed edge weights, and combine the ventilation direction correction factor to generate an initial inference matrix of the group transmission impact;

[0070] Step S530: Based on the initial reasoning chain and the initial reasoning matrix, simulate the changes in the transmission chain after removing high-risk nodes, and calculate the individual health risk probability and the group transmission impact factor.

[0071] As you can understand, the process begins by constructing a causal inference network for individual health risks: based on features extracted from the dynamic embedding representation, probabilistic modeling is performed for three typical health issues. Specifically, for respiratory infections, a Bayesian causal inference model is used to establish a conditional relationship between cough fundamental frequency shift and infection probability. For digestive tract diseases, a reverse inference mechanism is employed to fit the inverse proportional relationship between the rate of decline in eating speed and disease occurrence. For systemic diseases, a conditional probability distribution of hotspot deviation is established. These three chains of evidence constitute the initial health risk inference network.

[0072] Furthermore, the respiratory infection probability model is:

[0073] P(resp|Δf)=exp(-k1·(Δf-μ f ) 2 );

[0074] Where P(resp|Δf) represents the probability of respiratory infection under a given fundamental frequency offset; Δf represents the fundamental frequency offset of cough sound; μ frepresents the baseline value of coughing frequency in healthy pig herds; k1 represents the infection risk sensitivity coefficient.

[0075] The formula for calculating the risk coefficient of digestive tract diseases is:

[0076]

[0077] Among them, R dig represents the risk factor of digestive tract diseases; v e Indicates the current eating speed; v c Indicates the standard eating speed for the same weight; v h represents the highest historical eating speed; δ represents the basic risk factor for gastrointestinal diseases.

[0078] The probability distribution formula for systemic diseases is:

[0079]

[0080] Among them, P(sys|d) represents the probability of occurrence of systemic diseases; d represents the thermal zone deviation, which is the spatial distance between individual pigs and the average thermal zone of the group; w represents the deviation impact weight; and b represents the basic disease offset.

[0081] Secondly, group transmission impact modeling is carried out: the contact transmission intensity in the trough area is calculated, and the initial transmission impact matrix is ​​generated by combining the air diffusion model.

[0082] The contact transmission intensity is calculated as:

[0083]

[0084] Among them, C ij represents the contact transmission intensity of pig individual i to j; w ij represents the edge weight from individual i to j in the behavior graph; t ij represents the time individuals i and j spend together in the trough area; d ij represents the average spatial distance between individuals i and j; ρ represents the virus retention coefficient on the surface of the trough.

[0085] The air diffusion propagation model is:

[0086] A ij =C ij ·exp(-λ·d′ ij )·cosθ;

[0087] Among them, A ij represents the airborne transmission influence factor of pig individual i on pig j; d′ ij represents the straight-line distance from individual i to j; λ represents the air virus attenuation coefficient; cosθ represents the ventilation direction correction factor.

[0088] Finally, the transmission chain simulation is implemented: the Monte Carlo method is used to simulate the removal of high-risk nodes, calculate the changes in network connectivity and the changes in the length of the transmission path, and accordingly modify the individual risk probability and the group transmission impact factor. The quantitative formula for the change in the transmission chain is:

[0089]

[0090] Where ζ represents the node removal impact factor; M pre represents the initial propagation influence matrix; M post represents the propagation impact matrix after node removal; det(·) represents the matrix determinant calculation, which is used to describe the connectivity of the propagation network; represents the initial propagation path set; represents the set of propagation paths after the node is removed; L p represents the effective length of the propagation path p; γ represents the network connectivity influence weight; μ represents the propagation path influence weight.

[0091] Furthermore, step S600 includes steps S610 to S630.

[0092] Step S610: Classify the risk levels according to the individual health risk probability, and generate risk level labels through probability density clustering;

[0093] Step S620: Perform simulation processing based on the group transmission influencing factors, simulate the blocking effect of key transmission paths, implement virtual isolation for high-transmission nodes in the trough area and excretion area, calculate the group infection rate decline gradient, and generate a dynamic early warning baseline;

[0094] Step S630: Based on the risk level label and the dynamic warning baseline, a dual-channel decision matrix is ​​constructed. By mapping the individual risk level to the isolation and disposal plan and synchronously associating the group warning baseline to the environmental control strategy, a joint evaluation result integrating the individual disposal instructions and the group prevention and control plan is output.

[0095] Preferably, the process first performs risk grading based on individual health risk probability: a probability density clustering algorithm is used to perform multimodal analysis on the health risk probability distribution of individual pigs, and an adaptive bandwidth optimization algorithm is used to identify risk aggregation intervals. The risk values ​​are divided into three levels: high, medium, and low, and corresponding labels are generated (high risk: probability > 0.7; medium risk: 0.4≤probability≤0.7; low risk: probability < 0.4).

[0096] Secondly, dynamic early warning of group transmission is implemented: key prevention and control scenarios are simulated through the transmission chain simulation system, that is, virtual isolation is implemented for high transmission impact nodes (transmission impact factor > 0.8) in the trough area and excretion area (temporarily removing nodes from the transmission network), and the gradient of the group infection rate decline in different time windows is calculated to generate a dynamic early warning baseline that changes over time.

[0097] In the final phase of building a dual-channel decision-making mechanism, precise prevention and control measures were achieved by establishing a two-dimensional decision-making matrix: at the horizontal channel level, individual risk levels were directly mapped to specific treatment measures: high-risk individuals were immediately isolated and treated, medium-risk individuals were transferred to separate observation areas for continuous monitoring, and low-risk individuals maintained routine feeding and management. At the vertical channel level, the group warning baseline was dynamically linked to the environmental control strategy. When the system issued a yellow warning, it automatically triggered enhanced ventilation and regional disinfection procedures. If it escalated to a red warning, a thorough disinfection of the entire house was initiated and a feed additive intervention plan was simultaneously implemented. Ultimately, a comprehensive joint assessment report was generated, which included both individual handling instructions, such as the operational instructions for moving pig No. 23 to an isolation house, and group-oriented prevention and control plans, such as the implementation specifications for spray disinfection of the No. 1 feed trough area three times a day.

[0098] Example 2:

[0099] like Figure 2 As shown, this embodiment provides a pig health intelligent assessment system based on multimodal behavior perception, the system comprising:

[0100] An acquisition module 901 is used to acquire a basic data set, wherein the basic data set includes depth video data, audio data, and individual motion trajectory data of a pig group;

[0101] A fusion module 902 is configured to perform multimodal behavior feature fusion processing based on the basic data set to extract joint behavior features with group spatial correlation;

[0102] A construction module 903 is configured to construct a graph structure based on the joint behavior characteristics, and to construct a directed weighted group behavior graph by calculating the contact frequency and behavior synchronization between individuals;

[0103] An evolution module 904 is configured to dynamically evolve group behavior based on the directed weighted group behavior graph, capture state dependencies between nodes, and quantify the propagation of group health state deviation signals to generate a dynamic embedding representation;

[0104] An inference module 905 is configured to perform group health inference based on the dynamic embedding representation, and output an inference result by simulating the diffusion path of health status deviation in the group topology. The inference result includes individual health risk probability and group transmission impact factor;

[0105] The decision module 906 is used to make group health decisions based on the reasoning results and generate joint evaluation results.

[0106] In a specific embodiment disclosed in this application, the fusion module 902 includes:

[0107] a first fusion unit, configured to extract group spatial topological features based on the depth video data, extract skeletal posture features of each individual pig through a spatiotemporally aligned three-dimensional convolutional neural network, and calculate a relative spatial position matrix between the individuals;

[0108] A second fusion unit is configured to perform group acoustic behavior correlation analysis based on the audio data and individual motion trajectory data, identify the spatial source of a specific sound event based on a preset sound source localization-trajectory matching algorithm, and construct an acoustic-motion joint distribution feature;

[0109] The third fusion unit is used to perform cross-modal group behavior fusion according to the relative spatial position matrix, the skeleton posture feature and the acoustic-motion joint distribution feature, and splice them to generate a joint behavior feature tensor.

[0110] In a specific embodiment disclosed in this application, the building module 903 includes:

[0111] The first construction unit is used to quantify physical contact under the spatial constraints of the farm based on the joint behavioral characteristics, by calculating the co-occurrence frequency of individuals in the feeding trough area and the lying area, and using trajectory collision detection to generate a physical contact intensity matrix;

[0112] The second construction unit is used to perform synchronization analysis of the breeding behavior cycle based on the joint behavior characteristics, match the fixed event group of eating or drinking through the time window, calculate the phase difference of individual actions in the group, and generate a scenario-based behavior synchronization matrix;

[0113] The third construction unit is used to construct a hierarchical directed graph based on the physical contact intensity matrix and the scenario-based behavior synchronization matrix, determine the leadership node through resource competition dominance analysis and perform adversarial contact detection, and establish a directed weighted group behavior graph including leadership follower edges, competitive confrontation edges and neutral coexistence edges.

[0114] In a specific embodiment disclosed in this application, the evolution module 904 includes:

[0115] A first evolution unit is configured to perform propagation path topology analysis based on the directed weighted group behavior graph, extract high-propagation-risk edges through edge weight gradient threshold segmentation, and assign a state inheritance attribute or a bidirectional contamination attribute to the high-propagation-risk edges to obtain high-risk edge information;

[0116] The second evolution unit is used to quantify environmental exposure hotspots based on the depth video data and the individual motion trajectory data, and generate an environmental exposure vector by calculating the individual's cumulative exposure time in a preset functional area and combining it with a preset regional toxin load attenuation model;

[0117] The third evolution unit is used to perform state evolution based on the high-risk edge information and the environmental exposure vector, fuse contact propagation and environmental exposure paths through a gated spatiotemporal graph neural network, introduce a latent period correction function to update the node health offset, and generate a dynamic embedding representation carrying the spatiotemporal propagation imprint.

[0118] In a specific embodiment disclosed in this application, the reasoning module 905 includes:

[0119] A first reasoning unit is configured to construct an individual symptom evidence reasoning chain based on the dynamic embedding representation, infer the causal relationship between cough fundamental frequency shift and respiratory tract infection, the reverse reasoning relationship between the rate of decrease in eating speed and digestive tract disease, and the conditional probability distribution between hot zone deviation and systemic disease, thereby generating an initial reasoning chain for individual health risk probability;

[0120] The second inference unit is used to calculate the contact transmission intensity of the trough and the air diffusion attenuation effect based on the dynamic embedding representation, and generate an initial inference matrix of the group transmission impact by integrating the spatial distance attenuation amount with the directed edge weights and combining it with the ventilation direction correction factor;

[0121] The third reasoning unit is used to simulate the changes in the transmission chain after removing high-risk nodes based on the initial reasoning chain and the initial reasoning matrix, and calculate the individual health risk probability and the group transmission impact factor.

[0122] In a specific embodiment disclosed in this application, the decision module 906 includes:

[0123] A first decision-making unit is configured to classify the risk level according to the individual health risk probability and generate a risk level label through probability density clustering;

[0124] The second decision-making unit is used to perform simulation processing based on the group transmission influencing factors, simulate the blocking effect of key transmission paths, implement virtual isolation for high-transmission nodes in the trough area and excretion area, and calculate the group infection rate decline gradient to generate a dynamic early warning baseline;

[0125] The third decision-making unit is used to construct a dual-channel decision matrix based on the risk level label and the dynamic warning baseline, and output a joint evaluation result that integrates individual disposal instructions and group prevention and control plans by mapping individual risk levels to isolation and disposal plans and synchronously associating group warning baselines to environmental control strategies.

[0126] Example 3:

[0127] Corresponding to the above method embodiment, this embodiment also provides a pig health intelligent assessment device based on multimodal behavior perception. The pig health intelligent assessment device based on multimodal behavior perception described below and the pig health intelligent assessment method based on multimodal behavior perception described above can be referenced to each other.

[0128] Figure 3 FIG is a block diagram of a pig health intelligent assessment device 800 based on multimodal behavior perception according to an exemplary embodiment. Figure 3 As shown, the pig health intelligent assessment device 800 based on multimodal behavior perception may include: a processor 801, a memory 802. The pig health intelligent assessment device 800 based on multimodal behavior perception may also include one or more of a multimedia component 803, an I / O interface 804, and a communication component 805.

[0129] The processor 801 is used to control the overall operation of the multimodal behavior-based intelligent pig health assessment device 800 to complete all or part of the steps of the multimodal behavior-based intelligent pig health assessment method. The memory 802 is used to store various types of data to support the operation of the multimodal behavior-based intelligent pig health assessment device 800. This data may include, for example, instructions for any application or method operating on the multimodal behavior-based intelligent pig health assessment device 800, as well as application-related data such as contact information, sent and received messages, pictures, audio, video, etc. The memory 802 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk. The multimedia component 803 may include a screen and an audio component. The screen may be, for example, a touch screen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signal may be further stored in the memory 802 or transmitted via the communication component 805. The audio component also includes at least one speaker for outputting audio signals. The I / O interface 804 provides an interface between the processor 801 and other interface modules, and the above-mentioned other interface modules can be a keyboard, a mouse, buttons, etc. These buttons can be virtual buttons or physical buttons. The communication component 805 is used for wired or wireless communication between the pig health intelligent assessment device 800 based on multimodal behavior perception and other devices. Wireless communication, such as Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G or 4G, or a combination of one or more of them, so the corresponding communication component 805 may include: Wi-Fi module, Bluetooth module, NFC module.

[0130] In an exemplary embodiment, a pig health intelligent assessment device 800 based on multimodal behavioral perception can be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components to execute the above-mentioned pig health intelligent assessment method based on multimodal behavioral perception.

[0131] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided. When executed by a processor, the program instructions implement the steps of the aforementioned method for intelligent swine health assessment based on multimodal behavior perception. For example, the computer-readable storage medium may be the aforementioned memory 802 including the program instructions. The program instructions may be executed by the processor 801 of the device 800 for intelligent swine health assessment based on multimodal behavior perception to implement the aforementioned method for intelligent swine health assessment based on multimodal behavior perception.

[0132] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered by the scope of protection of the present invention.

Claims

1. A pig health intelligent assessment method based on multimodal behavior perception, characterized in that: include: Acquire a basic data set, wherein the basic data set includes deep video data, audio data, and individual motion trajectory data of a pig group; Performing multimodal behavioral feature fusion processing based on the basic data set to extract joint behavioral features with group spatial correlation; A graph structure is constructed based on the joint behavior characteristics, and a directed weighted group behavior graph is constructed by calculating the contact frequency and behavior synchronization between individuals; Dynamically evolving group behavior based on the directed weighted group behavior graph, capturing state dependencies between nodes and quantifying the propagation of group health state deviation signals to generate a dynamic embedding representation; Performing group health reasoning based on the dynamic embedding representation, and outputting reasoning results by simulating the diffusion path of health status deviation in the group topology, wherein the reasoning results include individual health risk probability and group transmission impact factor; Group health decisions are made based on the reasoning results to generate joint evaluation results.

2. The method for intelligent pig health assessment based on multimodal behavior perception according to claim 1 is characterized in that: Performing multimodal behavior feature fusion processing based on the basic data set includes: Extracting group spatial topological features based on the depth video data, extracting skeletal posture features of each pig individual through a spatiotemporal aligned three-dimensional convolutional neural network, and calculating the relative spatial position matrix between individuals; Performing group acoustic behavior correlation analysis based on the audio data and individual motion trajectory data, identifying the spatial source of a specific sound event based on a preset sound source localization-trajectory matching algorithm, and constructing an acoustic-motion joint distribution feature; Cross-modal group behavior fusion is performed based on the relative spatial position matrix, the skeleton posture features and the acoustic-motion joint distribution features, and a joint behavior feature tensor is generated by splicing.

3. The method for intelligent pig health assessment based on multimodal behavior perception according to claim 1 is characterized in that: A graph structure is constructed according to the joint behavior characteristics, including: Based on the joint behavioral characteristics, physical contact under the spatial constraints of the farm is quantified by calculating the co-occurrence frequency of individuals in the feeding trough area and the lying area, and using trajectory collision detection to generate a physical contact intensity matrix; Perform synchronization analysis of the breeding behavior cycle based on the joint behavior characteristics, match the fixed event group of eating or drinking through the time window, calculate the phase difference of individual actions in the group, and generate a scenario-based behavior synchronization matrix; Based on the physical contact intensity matrix and the scenario-based behavior synchronization matrix, a hierarchical directed graph is constructed. The leadership node is determined through resource competition dominance analysis and adversarial contact detection is performed to establish a directed weighted group behavior graph containing leadership follower edges, competitive confrontation edges, and neutral coexistence edges.

4. The method for intelligent pig health assessment based on multimodal behavior perception according to claim 1 is characterized in that: Performing dynamic evolution of group behavior according to the directed weighted group behavior graph includes: Performing propagation path topology analysis based on the directed weighted group behavior graph, extracting high-propagation risk edges through edge weight gradient threshold segmentation, and assigning state inheritance attributes or bidirectional pollution attributes to the high-propagation risk edges to obtain high-risk edge information; Quantify environmental exposure hotspots based on the depth video data and the individual motion trajectory data, and generate an environmental exposure vector by calculating the individual's cumulative exposure time in a preset functional area and combining it with a preset regional toxin load attenuation model; State evolution is performed based on the high-risk edge information and the environmental exposure vector. Contact propagation and environmental exposure paths are fused through a gated spatiotemporal graph neural network. A latent period correction function is introduced to update the node health offset, generating a dynamic embedding representation carrying the spatiotemporal propagation imprint.

5. The method for intelligent pig health assessment based on multimodal behavior perception according to claim 1 is characterized in that: Performing group health reasoning based on the dynamic embedding representation includes: Based on the dynamic embedding representation, an individual disease evidence reasoning chain is constructed to infer the causal relationship between cough fundamental frequency shift and respiratory tract infection, the reverse reasoning relationship between the rate of decline in eating speed and digestive tract disease, and the conditional probability distribution between hot zone deviation and systemic disease, thereby generating an initial reasoning chain for individual health risk probability. Based on the dynamic embedding representation, the contact transmission intensity of the trough and the air diffusion attenuation effect are calculated, and the spatial distance attenuation is integrated by the directed edge weights and combined with the ventilation direction correction factor to generate the initial inference matrix of the group transmission impact; According to the initial reasoning chain and the initial reasoning matrix, the changes in the transmission chain after removing high-risk nodes are simulated, and the individual health risk probability and the group transmission impact factor are calculated.

6. An intelligent pig health assessment system based on multimodal behavior perception, characterized in that: include: An acquisition module is used to acquire a basic data set, wherein the basic data set includes deep video data, audio data, and individual motion trajectory data of a pig group; A fusion module, configured to perform multimodal behavioral feature fusion processing based on the basic data set to extract joint behavioral features with group spatial correlation; A construction module is used to construct a graph structure based on the joint behavior characteristics, and to construct a directed weighted group behavior graph by calculating the contact frequency and behavior synchronization between individuals; An evolution module is used to dynamically evolve group behavior according to the directed weighted group behavior graph, generate a dynamic embedding representation by capturing state dependencies between nodes and quantifying the propagation process of group health state deviation signals; An inference module, configured to perform group health inference based on the dynamic embedding representation, and output inference results by simulating the diffusion path of health status deviation in the group topology, wherein the inference results include individual health risk probabilities and group transmission impact factors; The decision-making module is used to make group health decisions based on the reasoning results and generate joint evaluation results.

7. The pig health intelligent assessment system based on multimodal behavior perception according to claim 6 is characterized in that: The fusion module includes: a first fusion unit, configured to extract group spatial topological features based on the depth video data, extract skeletal posture features of each individual pig through a spatiotemporally aligned three-dimensional convolutional neural network, and calculate a relative spatial position matrix between the individuals; A second fusion unit is configured to perform group acoustic behavior correlation analysis based on the audio data and individual motion trajectory data, identify the spatial source of a specific sound event based on a preset sound source localization-trajectory matching algorithm, and construct an acoustic-motion joint distribution feature; The third fusion unit is used to perform cross-modal group behavior fusion according to the relative spatial position matrix, the skeleton posture feature and the acoustic-motion joint distribution feature, and splice them to generate a joint behavior feature tensor.

8. The pig health intelligent assessment system based on multimodal behavior perception according to claim 6 is characterized in that: The building blocks include: The first construction unit is used to quantify physical contact under the spatial constraints of the farm based on the joint behavioral characteristics, by calculating the co-occurrence frequency of individuals in the feeding trough area and the lying area, and using trajectory collision detection to generate a physical contact intensity matrix; The second construction unit is used to perform synchronization analysis of the breeding behavior cycle based on the joint behavior characteristics, match the fixed event group of eating or drinking through the time window, calculate the phase difference of individual actions in the group, and generate a scenario-based behavior synchronization matrix; The third construction unit is used to construct a hierarchical directed graph based on the physical contact intensity matrix and the scenario-based behavior synchronization matrix, determine the leadership node through resource competition dominance analysis and perform adversarial contact detection, and establish a directed weighted group behavior graph including leadership follower edges, competitive confrontation edges and neutral coexistence edges.

9. The pig health intelligent assessment system based on multimodal behavior perception according to claim 6 is characterized in that: The evolution module includes: A first evolution unit is configured to perform propagation path topology analysis based on the directed weighted group behavior graph, extract high-propagation-risk edges through edge weight gradient threshold segmentation, and assign a state inheritance attribute or a bidirectional contamination attribute to the high-propagation-risk edges to obtain high-risk edge information; The second evolution unit is used to quantify environmental exposure hotspots based on the depth video data and the individual motion trajectory data, and generate an environmental exposure vector by calculating the individual's cumulative exposure time in a preset functional area and combining it with a preset regional toxin load attenuation model; The third evolution unit is used to perform state evolution based on the high-risk edge information and the environmental exposure vector, fuse contact propagation and environmental exposure paths through a gated spatiotemporal graph neural network, introduce a latent period correction function to update the node health offset, and generate a dynamic embedding representation carrying the spatiotemporal propagation imprint.

10. The pig health intelligent assessment system based on multimodal behavior perception according to claim 6 is characterized in that: The reasoning module includes: A first reasoning unit is configured to construct an individual symptom evidence reasoning chain based on the dynamic embedding representation, infer the causal relationship between cough fundamental frequency shift and respiratory tract infection, the reverse reasoning relationship between the rate of decrease in eating speed and digestive tract disease, and the conditional probability distribution between hot zone deviation and systemic disease, thereby generating an initial reasoning chain for individual health risk probability; The second inference unit is used to calculate the contact transmission intensity of the trough and the air diffusion attenuation effect based on the dynamic embedding representation, and generate an initial inference matrix of the group transmission impact by integrating the spatial distance attenuation amount with the directed edge weights and combining it with the ventilation direction correction factor; The third reasoning unit is used to simulate the changes in the transmission chain after removing high-risk nodes based on the initial reasoning chain and the initial reasoning matrix, and calculate the individual health risk probability and the group transmission impact factor.

Images