Pig health state early warning system based on multi-mode AI

Abstract

The invention discloses a live pig health state early warning system based on multi-mode AI. The system comprises an edge sensing layer, a network transmission layer, an AI processing layer and an application interaction layer. The edge sensing layer is used for collecting multi-modal data of vision, acoustics, environment, biological signs and vibration; the network transmission layer adopts national secret encryption and a heterogeneous network to realize safe and efficient transmission of data; the AI processing layer generates a quantized health score through multi-modal data cleaning, feature extraction and cross-modal fusion analysis on the basis of a health assessment model of a time sequence convolutional network and comparative learning, and starts graded early warning through an early warning decision engine; and the application interaction layer converts an analysis result into a visual monitoring and early warning function. According to the invention, the problems of single monitoring, lagging early warning and high false alarm rate in the prior art are solved, the early, accurate and self-adaptive early warning of the health state of the live pig is realized, and the intelligent level and epidemic disease prevention and control capability of breeding management are improved.

Description

Technical Field

[0001] This invention relates to the field of livestock breeding technology, specifically to an early warning system for the health status of pigs based on multimodal AI. Background Technology

[0002] Currently, pig farming relies heavily on manual inspections and experience-based judgment, making it difficult to achieve real-time and accurate monitoring of individual pig health status. Existing technologies often use single sensors (such as thermometers and cameras) for localized monitoring, resulting in problems such as isolated data, delayed early warnings, high false alarm rates, and susceptibility to environmental interference. These technologies fail to effectively identify sub-healthy states or early-stage diseases in pigs, leading to passive disease control and increased farming costs. Existing solutions often analyze and simply overlay various data independently, lacking cross-modal collaborative perception and decision-making fusion mechanisms. They fail to comprehensively utilize multi-source information for complementary enhancement, resulting in high rates of missed early anomalies, frequent false alarms, and an inability to quantify health risk levels.

[0003] Most early warning models are based on fixed rules or static thresholds, which cannot continuously learn from data and are difficult to adapt to individual differences, different growth stages and environmental changes. The warning thresholds remain unchanged, which can easily generate a large number of false alarms due to the normal fluctuations of the pig herd, or be insensitive to slowly deteriorating health problems. Summary of the Invention

[0004] To address these issues, this invention provides an early warning system for the health status of pigs based on multimodal AI.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] An early warning system for pig health status based on multimodal AI includes:

[0007] The edge perception layer collects multimodal raw data from pig farms; and deploys several edge computing devices as edge computing nodes in pig houses to preprocess the collected multimodal raw data, perform lightweight AI analysis, and output multimodal edge data.

[0008] The network transport layer uses national cryptographic algorithms to achieve end-to-end data encryption, transmitting edge data processed by edge computing nodes to the AI ​​processing layer.

[0009] The AI ​​processing layer extracts health-related information from multimodal edge data and performs fusion analysis. It adopts a hierarchical and progressive analysis strategy, outputs analysis results, and transmits the analysis results to the application interaction layer.

[0010] The application interaction layer transforms the analysis results from the AI ​​processing layer into intuitive and usable business functions, providing customized interactive experiences for users with different roles.

[0011] By working together through the edge perception layer, network transmission layer, AI processing layer and application interaction layer, early monitoring and warning of the health status of pigs can be achieved.

[0012] Furthermore: the edge-aware layer includes:

[0013] The visual perception module collects data on pig behavior and appearance.

[0014] The acoustic sensing module collects the sound signals of the pigs.

[0015] The environmental sensing module collects environmental parameters of the pigsty.

[0016] The biosignature sensing module collects data on pig body temperature and activity levels.

[0017] The vibration sensing module collects vibration data of the collective activity of the pig herd.

[0018] Furthermore, the network transmission layer adopts a heterogeneous network architecture, including 5G / F5G, LoRa / NB-IoT and Wi-Fi6, and supports adaptive bit rate adjustment and key data priority transmission mechanism.

[0019] Furthermore: the AI ​​processing layer includes:

[0020] The multimodal data cleaning module performs quality control and spatiotemporal alignment on the received data;

[0021] Feature extraction engine extracts features from visual, acoustic, and time-series data;

[0022] The cross-modal fusion analysis module enables the fusion of feature-level and decision-level information from multiple modalities.

[0023] The health assessment model outputs a health score based on fused features.

[0024] The early warning decision engine initiates a tiered early warning process based on health scores.

[0025] Furthermore, the cross-modal fusion analysis module employs a cross-attention mechanism to achieve feature-level fusion and a dynamic ensemble learning method to achieve decision-level fusion.

[0026] Furthermore, the health assessment model is constructed based on a temporal convolutional network and contrastive learning regularization. It reconstructs the health feature sequence through an encoder-decoder structure and calculates anomaly scores to output a health score.

[0027] Furthermore, the early warning decision engine implements a multi-level triggering mechanism based on health scores, including three levels: low risk, medium risk, and high risk, and supports dynamic threshold updates and early warning priority calculation.

[0028] Furthermore: the application interaction layer includes:

[0029] A health digital twin platform that builds virtual pig farms and supports visualization and simulation of health status;

[0030] Real-time alert dashboards are used to display alert information from multiple perspectives;

[0031] Large data visualization dashboards are used for interactive data analysis.

[0032] A mobile app is used for on-site operation and receiving early warnings.

[0033] Furthermore, the system's workflow includes: multimodal data acquisition, edge preprocessing and transmission, cloud data cleaning and alignment, multimodal feature extraction, cross-modal fusion analysis, health status judgment and early warning decision-making, early warning production and intervention execution, effect feedback and model optimization.

[0034] Furthermore, the system supports a closed-loop workflow, enabling continuous optimization of the AI ​​model based on feedback from intervention effects, thereby improving the accuracy and adaptability of early warnings.

[0035] This invention offers the following advantages: By employing a cross-attention mechanism and dynamic ensemble learning to fuse feature-level and decision-level information, it achieves deep complementarity and synergistic enhancement of visual, acoustic, and temporal information, significantly improving the completeness and robustness of health status analysis. By combining a health assessment model based on TCN and contrastive learning, it can keenly capture subtle early symptoms such as coughing, posture changes, and deviations from circadian rhythms, greatly enhancing the sensitivity and specificity of early warnings. Through an early warning decision engine implementing a three-tiered triggering mechanism (low, medium, and high risk) based on health scores and dynamically adjusting warning thresholds based on historical performance, it effectively balances the timeliness of warnings with false alarm control, making warnings more scientific and reasonable.

[0036] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. Attached Figure Description

[0037] To more intuitively illustrate the prior art and this application, exemplary drawings are provided below. It should be understood that the specific shapes and structures shown in the drawings should not generally be regarded as limiting conditions for implementing this application; for example, based on the technical concept disclosed in this application and the exemplary drawings, those skilled in the art are able to easily make conventional adjustments or further optimizations to the addition / reduction / classification, specific shapes, positional relationships, connection methods, size ratios, etc. of certain units (components).

[0038] Figure 1 This is a system block diagram of the early warning system for the health status of pigs based on multimodal AI according to the present invention.

[0039] Figure 2 This is a data processing flowchart of the early warning system for pig health status based on multimodal AI of the present invention. Detailed Implementation

[0040] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. It should be understood that these embodiments are merely for further explanation of the present invention and should not be construed as limiting the scope of protection of the present invention. Technical engineers in the field can make some non-essential improvements and adjustments to the present invention based on the above-described content. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] Please see Figures 1-2 The early warning system for the health status of pigs based on multimodal AI includes an edge perception layer, a network transmission layer, an AI processing layer, and an application interaction layer.

[0042] The edge perception layer collects multimodal raw data from pig farming scenarios; through a heterogeneous sensor collaborative deployment strategy, a comprehensive, blind-spot-free data acquisition network is constructed. The edge perception layer includes a visual perception module, an acoustic perception module, an environmental perception module, a biological characteristic perception module, and a vibration perception module.

[0043] The visual perception module consists of multiple multispectral industrial cameras (including visible and near-infrared channels) and infrared thermal imagers deployed in the farm, which can continuously monitor the behavior of each pig. In addition, wide-angle dynamic range imaging technology is used to effectively overcome challenges such as drastic changes in lighting and dust interference in the pigsty. Through a lightweight target detection algorithm, key body parts (ears, eyes, back, and limbs) of each pig are tracked in real time and their apparent feature values ​​are calculated.

[0044] The acoustic sensing module consists of several high-sensitivity microphone arrays deployed in the farm to collect acoustic signals such as coughing, squealing, and chewing from each pig. The acoustic sensing module uses blind source separation technology to effectively distinguish individual voiceprints from environmental noise and accurately locate the sound source. Through time-frequency domain feature analysis of the audio signals, it can identify typical symptoms of respiratory diseases (such as cough frequency and pitch changes).

[0045] The environmental sensing module integrates environmental sensors with multiple parameters, enabling real-time monitoring of key environmental indicators such as temperature, humidity, ammonia concentration, and carbon dioxide concentration in the pigsty. The data from the environmental sensors are processed through a self-calibration algorithm to eliminate drift errors and ensure the stability of long-term monitoring.

[0046] In this embodiment, the biometric sensing module uses a low-power smart ear tag to collect vital sign data such as body temperature and activity level of each pig every 5 minutes; the ear tag has a built-in three-axis accelerometer, which can accurately identify the basic behavioral patterns of each pig, such as standing, lying down, and eating.

[0047] Several vibration sensing modules are installed on the pigsty building structure to monitor the collective activity patterns of the pigs. The vibration signals are analyzed by wavelet transform to extract behavioral rhythm features related to health status, such as the frequency and intensity of the group getting up / lying down.

[0048] The network transmission layer uses the national cryptographic SM9 algorithm to achieve end-to-end data encryption, ensuring the security of aquaculture data; the edge data processed by the edge computing nodes is transmitted to the AI ​​processing layer; the network transmission layer adopts a hierarchical networking strategy to adapt to the transmission needs of different data types.

[0049] The network transmission layer adopts a heterogeneous network architecture. For high-bandwidth visual data, 5G / F5G high-speed transmission is used; for low-frequency small data packets (such as environmental sensor data), low-power wide-area networks such as LoRa / NB-IoT are used; in addition, Wi-Fi 6 edge gateways are deployed in sensor-dense areas to achieve local data aggregation and preprocessing. This heterogeneous networking method effectively controls communication costs while ensuring data transmission quality.

[0050] The network transport layer deploys edge computing devices locally in the pigsty, using these devices as edge computing nodes to perform tasks such as data cleaning, compression, and lightweight AI analysis. The edge computing nodes continuously optimize model performance through online knowledge distillation technology, reducing the amount of visual data transmitted while retaining most of the effective information.

[0051] In addition, the transport layer is designed with an adaptive bit rate adjustment mechanism, which can dynamically adjust the transmission strategy according to network conditions to ensure the priority transmission of critical health data.

[0052] The AI ​​processing layer extracts health-related information from multimodal edge data and performs fusion analysis; it adopts a hierarchical and progressive analysis strategy, outputs analysis results, and transmits them to the application interaction layer; the AI ​​processing layer includes a multimodal data cleaning module, a feature extraction engine, a cross-modal fusion analysis module, a health assessment model, and an early warning decision engine.

[0053] The multimodal data cleaning module is designed with a multi-level data quality control mechanism to address the unique characteristics of the pig farming environment. This includes anomaly detection based on the isolated forest algorithm, missing data imputation based on generative adversarial networks, and timestamp alignment technology to ensure the spatiotemporal consistency of multi-source data.

[0054] The feature extraction engine can design dedicated feature extractors for different modal data, including visual feature data, acoustic feature data, and temporal feature data.

[0055] For visual feature data, an improved ResNet-50 network was used to extract the appearance features of each pig. At the same time, an improved pose estimation algorithm for each pig using the OpenPose architecture was used to extract the spatial coordinates of multiple key body parts.

[0056] Acoustic feature data is extracted by combining Mel spectrograms with ConvNeXt networks to identify abnormal vocal patterns such as coughing and sneezing.

[0057] Temporal feature data uses a multi-head self-attention mechanism to capture long-term dependencies in vital sign data and identify subtle changes in physiological rhythms.

[0058] The cross-modal fusion analysis module employs a multi-level fusion strategy to achieve information complementarity and enhancement. In the feature-level fusion stage, an adaptive weighting of features from different modalities is achieved through a cross-attention mechanism. In the decision-level fusion stage, a dynamic ensemble learning method is used to integrate the analysis results of each modality and generate a consistent health assessment.

[0059] In this embodiment, the calculation process for feature-level fusion (cross-attention) is as follows:

[0060] (1) Input visual features Acoustic characteristics Time series characteristics ;

[0061] (2) Design cross attention, with acoustic features as the query and visual / temporal features as the key-value pair; cross attention enables acoustic features to guide visual / temporal features to focus on relevant regions, thus alleviating the modality gap;

[0062] (3) The query transformation formula is: ;

[0063] The key-value transformation formula is: ;

[0064] (4) The formula for calculating attention is:

[0065] ;

[0066] ;

[0067] (5) Calculation of fusion features:

[0068] ;

[0069] The cross-attention output formula is:

[0070] ;

[0071] b. The computational process for decision-level fusion (dynamic ensemble learning) is as follows:

[0072] (1) Input each modality independently and output y v y a y t (Probability of health / abnormality);

[0073] (2) Calculate the dynamic weights;

[0074] Based on modal history performance, the weight of mode m at time t is calculated. The calculation formula is as follows:

[0075] ;

[0076] Where Z is the normalization factor; This represents the loss of mode m at the previous time step;

[0077] (3) The formula for calculating the weighted sum is:

[0078] ;

[0079] (4) Consistency check, when This will trigger a recalibration.

[0080] In summary, the formula for calculating the dynamic integration weights is: ;

[0081] ;

[0082] The final decision formula is:

[0083] ;

[0084] Where Q is the query matrix; K is the key matrix; and V is the value matrix; For attention head dimension; For visual features, acoustic characteristics It is a time-series feature; The attenuation coefficient; The loss is the value of mode m in the previous time step. Let m be the predicted probability of mode m. For the post-integration health assessment; y is the consistency threshold; vFor visual abnormalities (such as redness of the ears); y a For acoustic anomalies (such as cough frequency); y t This represents the probability of time-series anomalies (such as body temperature fluctuations).

[0085] A health assessment model is constructed based on deep temporal anomaly detection, which captures subtle deviations in health status through contrastive learning methods.

[0086] That is, an encoder-decoder based on Temporal Convolutional Network (TCN) combined with contrastive learning regularization; the training strategy is: positive samples are data augmentation in a healthy state (temporal distortion, adding small noise); negative samples are artificially injected anomalies (such as simulated cough, sudden rise in body temperature); the output is a health score h, h∈[0,1], the lower the h, the higher the risk.

[0087] Maximizing the similarity between positive sample pairs forces the encoder to learn robust and healthy representations; therefore, the contrastive learning regularization process is as follows:

[0088] Assuming health data x, generate two enhanced views x. + x - The encoder Output embedding ;

[0089] The formula for calculating the loss function (NT-Xent) is as follows:

[0090] ;

[0091] Among them, cosine similarity ; N represents the temperature parameter; N represents the batch size. Embedding for negative samples;

[0092] The process of Temporal Anomaly Detection (TCN) is as follows:

[0093] Input fusion feature sequence The output formula of the TCN encoder is:

[0094] ;

[0095] The formula for reconstructing the decoder from the encoder content is as follows:

[0096] ;

[0097] Then abnormal scores for: ;

[0098] Health score: ;in, for function, This is the scaling factor;

[0099] In summary, the contrastive loss function is:

[0100] ;

[0101] The formula for calculating TCN reconstruction loss is:

[0102] ;

[0103] The formula for calculating the health score is:

[0104] ;

[0105] in, To compare the losses; Embedding of positive sample pairs; For reconstruction losses; True characteristics; is the reconstructed feature; h is the health score; T is the time series length.

[0106] The early warning decision engine, based on the output of the health assessment model, adopts a multi-level triggering mechanism to initiate corresponding early warning processes according to the risk level (low, medium, high), ensuring timely and non-intrusive early warning responses.

[0107] The early warning decision engine implements a multi-level triggering mechanism based on the health score h to balance the timeliness of early warnings with false alarm control.

[0108] Risk levels are categorized as follows: low risk is h>0.85, in which case no intervention is required; medium risk is 0.6≤h≤0.85, in which case enhanced monitoring is needed; and high risk is h<0.6, in which case immediate action is required.

[0109] Thresholds are dynamically adjusted based on historical events to prevent fixed thresholds from becoming invalid; data is pushed to the corresponding channels (SMS / APP / digital twin) according to risk level.

[0110] The calculation process for risk level determination is as follows:

[0111] Define the threshold function:

[0112] ;

[0113] Initial threshold In this embodiment, =0.85; =0.6;

[0114] Dynamic threshold updates: if k consecutive low-risk cases are misclassified as medium-risk, then... Therefore, update the rules: ; where FP is a false alarm indication; The learning rate; To update the step size; The current low / medium threshold;

[0115] Calculate the warning priority, where the priority score is:

[0116] ;

[0117] in , This indicates the rate of change in the health score; a negative value indicates deterioration. and As the weight, in this embodiment, =0.7; =0.3; Contribute to static risk;

[0118] Only when >p min An alert is triggered at any time.

[0119] The formula for determining the risk level is:

[0120] The dynamic threshold update formula is:

[0121] ;

[0122] The formula for warning priority is:

[0123] ;

[0124] ;

[0125] Where h represents the health score; , Risk threshold; For indicator functions; p represents the rate of change in health scores. min The lowest priority threshold; For time intervals.

[0126] The application interaction layer transforms AI analysis results into intuitive and usable business functions, providing customized interactive experiences for users with different roles; the application interaction layer includes a health digital twin platform, real-time early warning dashboards, visual data dashboards, and mobile terminal apps.

[0127] The health digital twin platform constructs a virtual pig farm environment that is completely mapped to the physical pig farm, enabling visualized monitoring and simulation of the health status of pigs. Users can use the twin system to trace back historical health events or simulate the potential impact of different intervention measures on the health of the pig herd, providing a scientific basis for management decisions.

[0128] The real-time early warning dashboard provides multi-dimensional early warning displays, including individual abnormality warnings, group health trend warnings, and environmental risk warnings. Early warning information is pushed in real time through multiple channels (PC, mobile APP, SMS) to ensure that relevant personnel receive key information as soon as possible.

[0129] The data visualization dashboard, through visualization libraries such as ECharts, transforms complex health data into intuitive charts, supporting multi-dimensional data drill-down and interactive analysis; aquaculture managers can customize the analysis perspective through drag-and-drop, deeply mining the value of the data.

[0130] The mobile app provides a user-friendly interface for on-site staff, supporting functions such as receiving early warning information, viewing health reports, and reporting response measures. The app is designed for offline use, ensuring core functions remain operational even in aquaculture areas with poor network signal.

[0131] See Figure 2 The complete implementation process of the early warning system of the present invention is as follows:

[0132] (1) Multimodal data acquisition

[0133] The system first comprehensively collects data related to pigs through multimodal sensing devices deployed within the farm; the visual perception module continuously monitors pig behavior and body surface temperature using multispectral industrial cameras and infrared thermal imagers; the acoustic perception module collects individual sound signals using a high-sensitivity microphone array; the environmental perception module monitors environmental indicators such as temperature, humidity, and ammonia levels in real time; the biometrics perception module periodically collects body temperature and activity levels through smart ear tags; and the vibration perception module monitors the collective activity rhythm of the pig herd. All sensors are deployed collaboratively to form a comprehensive, high-precision data acquisition network.

[0134] (2) Edge preprocessing and transmission

[0135] The raw data undergoes preliminary processing at the local edge computing node in the pigsty, including data cleaning, lightweight AI analysis (such as object detection and sound source separation), and data compression. The edge model is continuously optimized through online knowledge distillation. The processed data is sent to the cloud via the network transmission layer. This layer uses the national cryptographic SM9 encryption algorithm to ensure security and achieves hierarchical and efficient transmission based on a heterogeneous network. It also supports adaptive bitrate adjustment and a mechanism for prioritizing the transmission of critical data.

[0136] (3) Cloud data cleaning and alignment

[0137] The cloud-based multimodal data cleaning module performs in-depth quality control on the received data, including anomaly detection based on isolated forests, missing data imputation based on generative adversarial networks, and timestamp alignment of multi-source data, ensuring the spatiotemporal consistency and reliability of cross-modal data.

[0138] (4) Multimodal feature extraction

[0139] The feature extraction engine calls dedicated models for deep feature extraction for different modal data: visual data uses an improved ResNet-50 and pose estimation algorithm to extract appearance and pose features; acoustic data uses Mel spectrograms and ConvNeXt network to identify abnormal audio patterns; and time series data uses a multi-head self-attention mechanism to capture long-term dependencies of physiological rhythms.

[0140] (5) Cross-modal fusion analysis

[0141] The cross-modal fusion analysis module adopts a feature-level and decision-level fusion strategy. It achieves adaptive weighting of multimodal features through a cross-attention mechanism and combines dynamic ensemble learning methods to integrate the analysis results of each modality to form a consistent health status assessment.

[0142] (6) Health status assessment and branch processing

[0143] The system determines whether the health status of pigs is normal based on the results of the fusion analysis; when the status is normal, the system updates the health records of individuals and groups, conducts health risk assessments based on historical data, and continuously monitors trend changes.

[0144] When an anomaly is detected, the system triggers the early warning decision engine, initiates a graded early warning process based on the risk level (low, medium, high), generates corresponding intervention instructions, and pushes them to the application interaction layer for execution.

[0145] (7) Early warning production and intervention implementation

[0146] In response to abnormal situations, early warning information is pushed through multiple channels such as real-time early warning dashboards and mobile apps. The health digital twin platform supports simulation and decision support. On-site personnel receive instructions and execute interventions through mobile terminals, and the system records the execution process in real time.

[0147] (8) Effect feedback and model optimization

[0148] After the intervention is implemented, the system collects feedback data on the effect and combines it with a new round of multimodal data collection to continuously optimize and iterate the AI ​​model, forming a closed-loop workflow of "monitoring-analysis-early warning-intervention-optimization" to continuously improve the system's early warning accuracy and adaptability.

[0149] This invention employs a cross-attention mechanism and dynamic ensemble learning to fuse feature-level and decision-level information, achieving deep complementarity and synergistic enhancement of visual, acoustic, and temporal information, significantly improving the completeness and robustness of health status analysis. By combining a health assessment model based on TCN and contrastive learning, it can keenly capture subtle early symptoms such as coughing, posture changes, and deviations from physiological rhythms, greatly improving the sensitivity and specificity of early warnings. The early warning decision engine implements a three-tiered triggering mechanism (low, medium, and high risk) based on health scores, and dynamically adjusts the warning threshold based on historical performance, effectively balancing the timeliness of warnings with false alarm control, making warnings more scientific and reasonable.

[0150] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An early warning system for the health status of pigs based on multimodal AI, characterized in that, include: The edge sensing layer collects multimodal raw data from pig farms; Several edge computing devices were deployed in the pigsty as edge computing nodes to preprocess the collected multimodal raw data, perform lightweight AI analysis, and output multimodal edge data. The network transport layer uses national cryptographic algorithms to achieve end-to-end data encryption, transmitting edge data processed by edge computing nodes to the AI ​​processing layer. The AI ​​processing layer extracts health-related information from multimodal edge data and performs fusion analysis. It adopts a hierarchical and progressive analysis strategy, outputs analysis results, and transmits the analysis results to the application interaction layer. The application interaction layer transforms the analysis results from the AI ​​processing layer into intuitive and usable business functions, providing customized interactive experiences for users with different roles. By working together through the edge perception layer, network transmission layer, AI processing layer and application interaction layer, early monitoring and warning of the health status of pigs can be achieved.

2. The early warning system for pig health status based on multimodal AI according to claim 1, characterized in that, The edge-aware layer includes: The visual perception module collects data on pig behavior and appearance. The acoustic sensing module collects the sound signals of the pigs. The environmental sensing module collects environmental parameters of the pigsty. The biosignature sensing module collects data on pig body temperature and activity levels. The vibration sensing module collects vibration data of the collective activity of the pig herd.

3. The early warning system for pig health status based on multimodal AI according to claim 1, characterized in that, The network transport layer adopts a heterogeneous network architecture, including 5G / F5G, LoRa / NB-IoT and Wi-Fi 6. The network transport layer supports adaptive bit rate adjustment and key data priority transmission mechanism.

4. The early warning system for pig health status based on multimodal AI according to claim 1, characterized in that, The AI ​​processing layer includes: The multimodal data cleaning module performs quality control and spatiotemporal alignment on the received data; Feature extraction engine extracts features from visual, acoustic, and time-series data; The cross-modal fusion analysis module enables the fusion of feature-level and decision-level information from multiple modalities. The health assessment model outputs a health score based on fused features. The early warning decision engine initiates a tiered early warning process based on health scores.

5. The early warning system for pig health status based on multimodal AI according to claim 4, characterized in that, The cross-modal fusion analysis module uses a cross-attention mechanism to achieve feature-level fusion and a dynamic ensemble learning method to achieve decision-level fusion.

6. The early warning system for pig health status based on multimodal AI according to claim 4, characterized in that, The health assessment model is built based on a temporal convolutional network and contrastive learning regularization. It reconstructs the health feature sequence through an encoder-decoder structure and calculates anomaly scores to output a health score.

7. The early warning system for pig health status based on multimodal AI according to claim 4, characterized in that, The early warning decision engine implements a multi-level triggering mechanism based on health scores, including three levels: low risk, medium risk, and high risk, and supports dynamic threshold updates and early warning priority calculation.

8. The early warning system for pig health status based on multimodal AI according to claim 1, characterized in that, The application interaction layer includes: A health digital twin platform that builds virtual pig farms and supports visualization and simulation of health status; Real-time alert dashboards are used to display alert information from multiple perspectives; Large data visualization dashboards are used for interactive data analysis. A mobile app is used for on-site operation and receiving early warnings.

9. The early warning system for pig health status based on multimodal AI according to claim 1, characterized in that, The system's workflow includes: multimodal data acquisition, edge preprocessing and transmission, cloud data cleaning and alignment, multimodal feature extraction, cross-modal fusion analysis, health status assessment and early warning decision-making, early warning production and intervention execution, and effect feedback and model optimization.

10. The early warning system for pig health status based on multimodal AI according to any one of claims 1-9, characterized in that, The system supports a closed-loop workflow and can continuously optimize the AI ​​model based on feedback on the intervention effect, thereby improving the accuracy and adaptability of early warnings.

Images