Pig house robot health monitoring and early warning system for pig farming platforms

Abstract

This invention belongs to the field of intelligent farming and robotics technology, specifically relating to a robotic health monitoring and early warning system for pig farms. This system aims to solve the problem of delayed early warnings caused by the discontinuity of existing monitoring data. The system includes: a static sentinel node subsystem for continuously and frequently collecting data, generating anomaly scores, and sending triage requests; a mobile diagnosis and service unit subsystem for autonomously diagnosing upon receiving requests, acquiring precise individual data, and generating reports; and a central control and prediction subsystem for fusing data, constructing a spatiotemporal graph model, predicting disease transmission, and generating early warnings. Through this system, 24 / 7 uninterrupted monitoring of pig herds is achieved, resolving data gaps, forming a collaborative "general screening-precise diagnosis" approach, and utilizing a spatiotemporal graph neural network to provide early warnings, significantly advancing the disease prevention and control efforts.

Description

Technical Field

[0001] This invention belongs to the field of intelligent farming and robotics technology, specifically, it relates to a pig farm robot health monitoring and early warning system for pig farming platforms. Background Technology

[0002] This invention belongs to the field of automation and intelligent monitoring technology in animal husbandry, specifically relating to an animal health status monitoring and early warning system applied in large-scale farming environments. In modern animal husbandry, especially in intensive and large-scale pig farming, efficient and precise disease prevention and control, as well as individual health management, are core elements for ensuring industry stability and improving economic benefits. Traditional farming methods heavily rely on manual inspections to observe the health status of pigs. This method is not only labor-intensive and inefficient, but the observation results are also easily affected by the experience level and subjective judgment of the farmers, making it difficult to accurately and promptly detect early, sporadic diseases in large groups.

[0003] To address the aforementioned challenges, existing technologies have proposed and applied intelligent monitoring solutions based on automated equipment. Specifically, these solutions typically employ pigsty inspection robots equipped with multimodal sensors as data acquisition platforms. For example, by integrating an infrared thermal imager and a sound acquisition module into the robot, it can autonomously inspect the pigsty according to a preset path and schedule (e.g., twice daily). During the inspection, the infrared thermal imager captures the surface temperature distribution map of each pig in a non-contact manner, providing objective evidence for screening for abnormal body temperatures; simultaneously, the sound acquisition module records sounds in the pigsty environment, particularly abnormal breathing sounds such as coughing and wheezing. The collected multidimensional raw data is transmitted to a backend data processing center, where deep learning models (such as the ResNet model based on convolutional neural networks) intelligently analyze the thermal imaging data and audio signals, automatically identifying individuals with body temperatures exceeding a preset threshold (e.g., 39.5℃) or individuals with abnormal respiratory rates. Once a pig with abnormal health is identified, the system generates an analysis report and combines it with its historical health records for a comprehensive assessment. Finally, it sends an early warning to the farm manager through the data platform. Compared to traditional manual inspections, this technology significantly improves the automation and objectivity of monitoring, and to some extent solves the problem of delayed disease detection, representing an important direction for the intelligent development of livestock farming.

[0004] However, with the continuous expansion of farming scale and increasingly stringent requirements for disease transmission control, a deep-seated contradiction inherent in the principle of the aforementioned monitoring system based on timed inspections has gradually emerged. The reason lies in the fact that the core of this technological paradigm is built upon a "discrete snapshot" data acquisition model, while the occurrence and spread of swine diseases is essentially a "continuous dynamic evolution." This fundamental mismatch between the monitoring model and the time dimension of the monitored objects leads to a bottleneck in the timeliness of early warning. Further analysis reveals that there is an incubation period and a subclinical period between pig infection with pathogens and the appearance of obvious clinical symptoms (such as a significant increase in body temperature or severe coughing). Within the fixed interval of only twice a day for inspections, one or a few pigs may have already entered the early stage of infection, and their body temperature or behavior may have begun to show weak but indicative fluctuations. However, due to the discontinuous nature of the inspections, these valuable early signals are very likely to be missed. When the robot collects data at the next inspection point on schedule, the results may already be those with clearly defined symptoms and a delayed effect. By this time, the virus may have already completed its initial spread within the pen, and the system's "warning" becomes, in effect, a "confirmation" of the existing situation, thus missing the optimal window for precise intervention in the early stages of an outbreak. Furthermore, this discrete data collection method means that while the backend intelligent algorithm is powerful, its analysis is based on sparse and isolated data points, making it impossible to construct a continuous curve of the health status changes of individual pigs or even the entire herd. The algorithm can determine whether a pig has a fever at a certain moment, but it struggles to dynamically model and predict the upward trend of body temperature and the frequency of abnormal behavior based on a continuous data stream, thus failing to achieve the qualitative leap from "diagnosis" to "prediction."

[0005] Therefore, how to design a method that can break through the limitations of timed and discrete data acquisition modes, and achieve continuous, high-frequency capture and accurate prediction of the health status of pig herds in the breeding environment and the dynamic evolution trend, so as to issue early warnings at the earliest stage of disease outbreaks, has become a key challenge and a technical problem that needs to be solved by those skilled in the art. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to overcome the non-continuous data acquisition defect of the existing pig house health monitoring system based on timed and discrete inspection mode. This defect causes the system to be unable to capture the continuous dynamic evolution process of the pig's health status, resulting in a serious lag in disease early warning and missing the best opportunity to intervene in the early stage of disease transmission.

[0007] To address the technical problem of delayed early warning caused by the discontinuous monitoring of pig herd health status in the aforementioned background technology, this invention provides a pig farm robot health monitoring and early warning system for pig farming platforms.

[0008] The technical solution adopted in this invention is: a pig farm robot health monitoring and early warning system for pig farming platforms, characterized in that the system includes:

[0009] A static sentinel node subsystem consists of at least one static sentinel node deployed in each independent pig pen. Each static sentinel node is used to continuously and frequently collect data on the environment and pig activities in its pen, and to perform real-time analysis on the collected raw data based on a preset edge computing model to generate a pen abnormality score that represents the overall abnormal state of the pen. When the pen abnormality score exceeds a dynamically adjusted trigger threshold for a continuous period of time within a preset time window, the static sentinel node generates and sends a triage request containing the unique identifier of the pen.

[0010] A mobile diagnostic and service unit subsystem is included, which is a pigsty inspection robot with autonomous navigation and refined operation capabilities. The mobile diagnostic and service unit subsystem is configured to perform two interchangeable working modes: baseline inspection mode and on-demand triage mode. In baseline inspection mode, the mobile diagnostic and service unit subsystem conducts inspections throughout the pigsty according to a preset low-frequency scheduling strategy to obtain baseline health data of individual pigs and the environment. In on-demand triage mode, upon receiving a triage request from the static sentinel node subsystem, the mobile diagnostic and service unit subsystem immediately interrupts the current baseline inspection task, autonomously navigates to the pen specified in the triage request, and initiates a targeted, high-precision individual diagnostic protocol for the pigs in that pen to obtain accurate physiological and pathological data of the target pigs and generate a structured individual diagnostic report.

[0011] It also includes a central control and prediction subsystem, which communicates bidirectionally with the static sentinel node subsystem and the mobile diagnosis and service unit subsystem. The central control and prediction subsystem receives and integrates the data stream of continuous pen anomaly scores sent by the static sentinel node subsystem and the discrete high-precision individual diagnostic reports sent by the mobile diagnosis and service unit subsystem. The central control and prediction subsystem is further used to construct and update a spatiotemporal map model of pig herd health status in real time based on the received data, and to predict the potential disease transmission paths and future health status evolution trends in the pig herd through the model, thereby generating disease early warning information.

[0012] In a preferred embodiment of the present invention, the static sentinel node is an integrated, low-power electronic device. Its hardware includes a central microcontroller unit, a wireless communication transceiver module, a power supply module, and a sensor array. Specifically, the sensor array includes: a triaxial microelectromechanical system (MEMS) accelerometer with a sampling frequency set to no less than 4096Hz and a measurement range of ±16g, used to continuously capture weak vibration signals in the pen structure caused by pigs' coughing, panting, and abnormal limb movements; a temperature and humidity composite sensor with a temperature measurement accuracy of ±0.2℃ and a humidity measurement accuracy of ±2%RH, used to monitor changes in the microenvironment within the pen; and an 8x8 pixel array infrared thermopile sensor with a field of view of 60 degrees and a temperature measurement range covering 20℃ to 100℃, used to acquire a low-resolution thermal map of the overall body surface temperature distribution of the pigs within the pen. The static sentry node is encapsulated in a corrosion-resistant and bite-resistant housing with an IP68 protection rating, made of polycarbonate injection molding and covered with 304 stainless steel mesh.

[0013] Furthermore, the end-side computational model fixed within the static sentinel node is specifically a pen anomaly scoring calculation module. The execution flow of the pen anomaly scoring calculation module is as follows: It performs a Fast Fourier Transform (FFT) on the raw vibration data collected by the MEMS accelerometer within a continuous 1-second time window, converting the time-domain signal into a frequency-domain signal; in the frequency domain, the module integrates the energy in the preset characteristic frequency bands (50Hz to 200Hz) related to pig coughing and the characteristic frequency bands (400Hz to 800Hz) related to panting to obtain vibration anomaly indicators; it calculates the mean and standard deviation of the temperature of 64 pixels output by the infrared thermopile sensor and compares it with a dynamic baseline temperature issued by a central control and prediction subsystem to obtain a heat map anomaly indicator; it performs a weighted summation of the vibration anomaly indicator and the heat map anomaly indicator, with the weighting coefficients dynamically configured by the central control and prediction subsystem based on historical data and the daily activity rhythm of the pig herd, and the summation result is the pen anomaly score at the current moment. The enclosure anomaly score is a dimensionless normalized value ranging from 0 to 1.

[0014] In another embodiment of the present invention, the wireless communication between the static sentinel node subsystem and the central control and prediction subsystem adopts the LoRaWAN protocol. The static sentinel node acts as the terminal device, and a LoRaWAN gateway is deployed on the central control and prediction subsystem side. When the moving average of a static sentinel node's fence anomaly score is continuously higher than the dynamically adjusted trigger threshold set by the central control and prediction subsystem for that fence over a continuous 300 seconds, the static sentinel node is triggered and constructs an uplink data packet. The uplink data packet is a structured data frame, whose payload always includes: a 16-bit unique fence identifier, a 32-bit timestamp, an 8-bit trigger reason code (indicating that the anomaly mainly originates from vibration or heatmap), and a 16-bit floating-point number representing the current anomaly score value. This data packet constitutes a triage request and is sent to the central control and prediction subsystem via the LoRaWAN network.

[0015] Specifically, the hardware platform of the mobile diagnostics and service unit subsystem, in addition to including a four-wheel differential drive chassis, an inertial measurement unit (IMU), and a battery management system, includes the following core payloads: a 3D lidar (LiDAR) for autonomous navigation and three-dimensional environmental perception, with a scanning frequency of 15Hz and a ranging accuracy of ±2cm; and a high-resolution infrared thermal imager with an uncooled vanadium oxide (VOx) microbolometer detector, a resolution of 640x512 pixels, and a thermal sensitivity (NETD) of less than 40mK for temperature measurement. The accuracy is ±0.3℃ after calibration; a planar circular microphone array consisting of eight MEMS microphones spaced 10cm apart, mounted on a two-dimensional gimbal, is used for sound source localization; a long-range high-frequency RFID reader conforming to ISO11784 / 11785 standards and operating at a frequency of 134.2kHz, along with a directional antenna; and a service end effector mounted on the end of a six-axis collaborative robotic arm, integrating a wireless charging coil and a high-pressure cleaning nozzle, for energy replenishment and lens cleaning of static sentry nodes.

[0016] Furthermore, the individual diagnostic protocol executed by the mobile diagnostic and service unit subsystem is a deterministic procedure consisting of multiple steps. Upon receiving a scheduling instruction containing the target pen identifier forwarded by the central control and prediction subsystem, the protocol is activated: First, path planning and arrival: the unit uses 3D LiDAR data and a pre-stored high-precision map to plan and execute a collision-free optimal path to reach the entrance of the target pen. Second, individual segmentation and locking: after entering the pen, the unit uses 3D LiDAR to scan the space inside the pen, and combined with its onboard RGB-D depth camera, performs 3D point cloud instance segmentation of all pigs in the pen. Simultaneously, it activates the microphone array to locate the source of abnormal acoustic signals (such as coughing) indicated in the triage request. By fusing the sound source localization vector with the 3D point cloud segmentation results, the individual pig emitting the abnormal sound is locked as the primary diagnostic target. Third, non-contact precise measurement of vital signs: the robotic arm adjusts its posture and precisely positions the high-resolution infrared thermal imager... The first step involves capturing a 5-second video stream of the head and core torso of the targeted individual. Then, an RFID reader is used to read the unique identification code from the individual's ear tag. The second step involves data analysis and report generation. The unit's onboard computing unit (an embedded computer with an integrated graphics processing unit) processes the acquired infrared video stream, extracting multiple quantitative indicators such as maximum body surface temperature, eye temperature, and uneven body surface temperature distribution. Combined with the read individual identification code, a structured individual diagnostic report is generated, containing multi-dimensional information including individual ID, diagnosis time, maximum body surface temperature, thermal snapshot, and acoustic characteristics. The third step involves data reporting. The unit sends this report to the central control and prediction subsystem via a 5G or Wi-Fi 6 wireless network, subsequently continuing the simplified diagnostic process for other individuals in the enclosure or deciding on the next action based on new instructions.

[0017] As a core technical feature of this invention, the spatiotemporal graph model of pig herd health status in the central control and prediction subsystem is specifically a spatiotemporal graph neural network (ST-GNN). The construction and operation mechanism of this model is as follows: In the model construction phase, the system abstracts the entire pig farm into a dynamic graph structure G=(V,E,X). Here, V is the set of nodes in the graph, where each node v_i∈V uniquely corresponds to a pig wearing an RFID ear tag. E is the set of edges in the graph, where edge e_ij∈E represents a potential disease transmission association between nodes v_i and v_j. The initial weight of this association is set based on the physical pen adjacency relationship (such as shared partitions or shared passageways) and a fixed simulated air circulation path. X is the feature matrix of the nodes, where X_i(t) is the feature vector of node v_i at time t. The feature vector contains the pig's historical health data, such as age, weight, vaccination records, and other static information. In the real-time model update phase, the data fusion module of the central control and prediction subsystem performs the following operations: it continuously receives data streams of pen anomaly scores from all static sentinel nodes. When the anomaly score S_k(t) of pen k increases, the system strengthens the weights of the edges e_ij connecting all nodes v_i∈V_k within pen k, and simultaneously appends S_k(t) as a time-varying pen-level environmental feature to the feature vector X_i(t) of all nodes within that pen. Upon receiving an individual diagnostic report for pig v_i generated by the mobile diagnostic and service unit subsystem, the system directly updates the node's latest feature vector X_i(t) with the precise physiological and pathological data from the report (such as precise body temperature T_i and cough frequency C_i), using it as the node's true state label at that moment. Finally, during the model prediction phase, the ST-GNN model performs a forward propagation at a fixed time step (e.g., every 10 minutes). The model propagates node information spatially through graph convolution operations, aggregating the health status of neighboring nodes. Simultaneously, it models the feature sequences of each node temporally using a gated recurrent unit (GRU) or a long short-term memory network (LSTM) to capture the evolutionary trend of its health status. The model's output is a prediction matrix P(t+Δt), where P_i(t+Δt) is a vector containing a series of predictive indicators, such as the probability that pig v_i will have a body temperature exceeding the clinical threshold and the probability of exhibiting severe coughing symptoms within the next Δt time (e.g., the next 12 hours).

[0018] Furthermore, the central control and prediction subsystem also includes a dynamic threshold and scheduling decision module. The scheduling decision module calculates and distributes a private, time-varying trigger threshold for each static sentinel node. The calculation of the dynamic threshold is based on: the historical anomaly score distribution of the pen, the overall activity rhythm of the pig herd (e.g., determining whether it is currently in a peak feeding, sleeping, or activity period based on time), and global environmental parameters (e.g., the overall temperature of the pig pen). By dynamically adjusting the threshold, the system can appropriately increase the threshold during periods of high pig activity to reduce false alarm rates and decrease the threshold during periods of rest to increase sensitivity. In addition, when multiple triage requests arrive simultaneously, the scheduling decision module is responsible for prioritizing the tasks of the mobile diagnosis and service unit subsystem based on the anomaly score of the request, the value of the pigs in the pen (e.g., replacement gilts or fattening pigs), and the disease spread risk level of the pen predicted by the ST-GNN model, generating an optimal scheduling sequence.

[0019] The beneficial effects of this invention are as follows:

[0020] First, by deploying a static sentinel node subsystem, this invention achieves 24 / 7 uninterrupted and high-frequency monitoring of pig herd health-related indicators, fundamentally solving the data gap problem caused by the traditional discrete inspection mode, and ensuring that any weak and early abnormal signals of the group can be captured in a timely manner.

[0021] Secondly, this invention establishes a two-tiered diagnostic and treatment model of "general screening-precision diagnosis" in which static sentinel points and mobile robots work together. Static nodes are responsible for low-cost, wide-coverage anomaly detection, while mobile robots are scheduled on demand and precisely as high-precision diagnostic resources, which greatly improves the utilization efficiency of scarce diagnostic resources and achieves the best balance between overall system performance and cost.

[0022] Third, by introducing a prediction model based on spatiotemporal graph neural networks, this invention transforms monitoring data from isolated "snapshots" into continuous, intrinsically correlated "spatiotemporal data streams." The system no longer merely identifies and alerts to existing symptoms; instead, it learns the dynamics of disease transmission within pig herds to quantitatively predict and warn of future health risks. This provides unprecedented decision support for pig farmers to implement ultra-early intervention measures and break the disease transmission chain, significantly advancing the disease prevention and control efforts.

[0023] Fourth, the system architecture of this invention has high scalability and adaptability. The dynamic threshold adjustment mechanism enables the system to adapt to the activity patterns of pig herds at different growth stages, reducing interference from environmental changes. The modular design allows for easy integration of new sensors (such as gas sensors for monitoring ammonia concentration) or more advanced prediction algorithms in the future. Attached Figure Description

[0024] Figure 1 This is a block diagram of the overall architecture of the system of the present invention.

[0025] Figure 2 This is a schematic diagram of the structure of the static sentinel node in this invention.

[0026] Figure 3 This is a schematic diagram of the structure of the mobile diagnostics and service unit subsystem in this invention.

[0027] Figure 4 This is a schematic diagram of the system's workflow in this invention.

[0028] Figure 5 This is a schematic diagram of the spatiotemporal map model of pig herd health status in this invention.

[0029] The attached diagrams are labeled as follows: 10. Static sentry node subsystem; 11. Central microcontroller unit; 12. Wireless communication transceiver module; 13. Power supply module; 14. Sensor array; 15. Corrosion-resistant and bite-resistant housing; 20. Mobile diagnostic and service unit subsystem; 21. Four-wheel differential drive chassis; 22. 3D LiDAR; 23. High-resolution infrared thermal imager; 24. Planar circular microphone array; 25. Long-range high-frequency RFID reader / writer; 26. Six-axis collaborative robotic arm; 30. Central control and prediction subsystem. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0031] Please see Figure 1 This invention illustrates the overall architecture of a pig farm robot health monitoring and early warning system for pig farming platforms. The system aims to overcome the data discontinuity problem caused by traditional discrete inspection modes, achieving continuous monitoring, accurate diagnosis, and prediction of pig herd health status through an innovative multi-level collaborative working mechanism. Logically, the system is divided into three mutually collaborative core subsystems with bidirectional data flow: a static sentinel node subsystem 10, a mobile diagnostic and service unit subsystem 20, and a central control and prediction subsystem 30.

[0032] Specifically, the static sentinel node subsystem 10 is the cornerstone of the entire monitoring network. Its physical entity consists of a large number of static sentinel nodes strategically deployed in each independent pig pen. Each pen is equipped with at least one such node, whose core task is to continuously collect data on the microenvironment and pig activity status within its jurisdiction at a high sampling frequency. These nodes are not simple data forwarding units, but intelligent sensing terminals with end-side computing capabilities. Based on internally fixed computing models, they perform real-time preliminary analysis and feature extraction on the massive amounts of raw data collected. Their output is not the raw data stream, but a highly condensed quantitative indicator that characterizes the overall health risk of the pen, namely, the pen anomaly score. When this score is continuously higher than a dynamically set trigger threshold within a preset time window, the static sentinel node will actively generate and send a triage request to the upper layer of the system. The triage request is a structured data packet, the payload of which must contain the unique identifier of the pen, thus accurately indicating the geographical location of the abnormal event.

[0033] The mobile diagnostic and service unit subsystem 20 physically manifests as a pigsty inspection robot with highly autonomous navigation and refined operational capabilities. The design philosophy of the service unit subsystem 20 is to centralize and mobilize high-value, high-precision diagnostic resources, and to schedule them on demand based on the "intelligence" from the static sentinel node subsystem 10. Therefore, the mobile diagnostic and service unit subsystem 20 is configured to dynamically switch between two distinct operating modes: baseline inspection mode and on-demand triage mode. In baseline inspection mode, the robot follows a preset, low-frequency global scheduling strategy, conducting routine inspections in the common passageways of the entire pigsty. Its main purpose is to collect data for constructing and calibrating a global health baseline, such as updating high-precision 3D maps, recording environmental background noise in different areas, and sampling the normal activity patterns of the pigs in non-abnormal states. In on-demand triage mode, once a triage request is received from the static sentinel node subsystem 10, the service unit subsystem 20 immediately suspends its current baseline task and treats the triage request as the highest priority instruction. It will utilize its autonomous navigation capabilities to quickly and accurately move to the target pen indicated by the triage request. Upon arrival, it will no longer perform a broad scan, but instead initiate a highly targeted and structured high-precision individual diagnostic protocol. This protocol aims to collect in-depth physiological and pathological data on specific pigs within the pen, ultimately generating a structured individual diagnostic report containing multi-dimensional information, which will then be uploaded.

[0034] The central control and prediction subsystem 30 is the "brain" of the entire system. It maintains continuous bidirectional data exchange with the aforementioned static sentinel node subsystem 10 and mobile diagnostic and service unit subsystem 20 via a wireless communication network. The core function of the static sentinel node subsystem 10 lies in data fusion and in-depth analysis. On one hand, it continuously receives and aggregates data streams of pen anomaly scores from all static sentinel nodes; on the other hand, it receives discrete but highly valuable individual diagnostic reports sent by the mobile diagnostic and service unit subsystem 20 after completing its tasks. The key innovation of the central control and prediction subsystem 30 is that it does not process these data in isolation, but rather utilizes these heterogeneous data sources to construct and maintain a spatiotemporal map model of pig herd health status in real time. The spatiotemporal graph model of swine herd health status abstracts the health status of the entire swine herd into a dynamically evolving network. By analyzing the changes in the state of nodes in the network and the mutual influence between nodes, the system can identify potential disease transmission paths and make probabilistic predictions on the evolution trend of swine herd health status within a specific time window in the future, and finally generate disease early warning information, thereby providing decision support for managers.

[0035] As a preferred embodiment of the present invention, please refer to Figure 2 This section provides a more detailed view of the internal structure of the static sentinel node 10. The static sentinel node 10 is designed as a highly integrated electronic device optimized for low-power operation. Its hardware core is a central microcontroller unit 11, such as an ARM Cortex-M4 core-based microcontroller, which is responsible for executing all data processing, logic control, and communication tasks within the node. Connected to the microcontroller unit 11 is a wireless communication transceiver module 12, used to establish a data link with the upper-layer system. Simultaneously, a power supply module 13 provides energy to the entire node. The power supply module 13 can be designed to be powered by a long-life lithium thionyl chloride battery, or integrate energy harvesting circuitry to support the wireless charging function of the mobile diagnostic unit.

[0036] The core function of the static sentinel node 10 lies in its integrated sensor array 14. The sensor array 14 is carefully selected and configured to capture multimodal signals most relevant to respiratory diseases and stress states in pigs in a non-contact, non-invasive manner. Specifically, the sensor array 14 includes: a triaxial microelectromechanical system (MEMS) accelerometer with strictly defined specifications and a sampling frequency set at no less than 4096Hz to ensure the capture of complete waveforms of weak vibrations in structures such as pen fences, floors, or feed troughs caused by high-frequency events such as coughing or panting in pigs; its measurement range is set to ±16g to cope with large-amplitude vibrations such as pig impacts without saturation. Furthermore, the array includes a temperature and humidity composite sensor with a temperature measurement accuracy better than ±0.2℃ and a relative humidity measurement accuracy better than ±2%RH, used to accurately monitor changes in the microenvironmental temperature and humidity within the pen, changes closely related to the comfort of the pig herd and the occurrence of certain diseases. In addition, the sensor array 14 also integrates an 8x8 pixel array infrared thermopile sensor. This sensor offers low-resolution thermal imaging capabilities with a 60-degree field of view, covering key areas of most standard pig pens. Its temperature measurement range spans from 20°C to 100°C, sufficient to monitor both normal and abnormal body temperatures in pigs after fever. The sensor does not prioritize high resolution; its purpose is to capture the overall surface temperature distribution trend of the pig herd within the pen, serving as an early indicator of group fever.

[0037] To ensure long-term stable operation in the harsh environment of pigsties, characterized by high humidity, high corrosiveness, and the risk of animal gnawing, the entire static sentry node is enclosed in a robust housing 15 with an IP68 protection rating. The main structure of the housing 15 is made of polycarbonate (PC) material through precision injection molding, ensuring excellent waterproof sealing performance and impact resistance. Externally, especially in areas potentially accessible to pigs, it is additionally covered with a layer of woven 304 stainless steel wire mesh. This mesh structure effectively resists gnawing and friction from pigs, protecting the internal sensor windows and the housing itself from physical damage.

[0038] Furthermore, the edge computing model embedded in the central microcontroller unit 11 of the static sentinel node 10 is specifically implemented as a pen anomaly scoring calculation module. The introduction of this module aims to transform raw, high-bandwidth sensor data into low-bandwidth, high-information-density feature indicators at the data source, thereby significantly reducing the demand for wireless communication bandwidth and the overall power consumption of the node. The internal execution flow of the pen anomaly scoring calculation module is a deterministic algorithm sequence. The module performs a Fast Fourier Transform (FFT) on 4096 raw vibration data points collected by the MEMS accelerometer within each consecutive 1-second time window, converting the complex time-domain vibration signal into a more easily analyzed frequency-domain signal. In the frequency domain spectrum, the module focuses on two preset feature bands highly correlated with common respiratory symptoms in pigs: one is a band from 50Hz to 200Hz, whose energy changes mainly correspond to coughing events in pigs; the other is a band from 400Hz to 800Hz, whose energy changes are related to wheezing symptoms. The module numerically integrates the energy within these two frequency bands and normalizes the results to obtain a quantified vibration anomaly index. The module processes temperature readings from 64 pixels of an 8x8 infrared thermopile sensor, calculating the arithmetic mean and standard deviation of these 64 data points. The mean reflects the average body surface temperature within the field of view, while the standard deviation characterizes the dispersion or non-uniformity of the temperature distribution. The module compares these two calculated values ​​with a dynamic baseline temperature for the pen periodically issued by the central control and prediction subsystem 30 (e.g., the expected average temperature of the pen's pigs in a healthy state at the current time point) to obtain a thermal anomaly index. Crucially, the module performs a weighted sum of the aforementioned vibration anomaly index and the thermal anomaly index. The weighting coefficients here are not fixed but dynamically calculated and remotely configured by the central control and prediction subsystem 30 based on the pen's historical data, the pigs' age, and their current daily activity rhythms (e.g., feeding, sleeping, activity peaks). For example, during the pigs' sleep period, the weight of the vibration index is increased accordingly to enhance sensitivity to nighttime coughing. The final summation result is normalized to form a dimensionless floating-point number between 0 and 1, which is the pen anomaly score at the current moment.

[0039] In another embodiment of the present invention, the wireless data link between the static sentinel node subsystem 10 and the central control and prediction subsystem 30 preferably adopts the LoRaWAN (Long-Range Wide Area Network) protocol. This choice is based on comprehensive considerations of the complex pig house environment, wide coverage, and extremely stringent requirements for node power consumption. In this architecture, each static sentinel node acts as an end device in the LoRaWAN network, while the central control and prediction subsystem 30 deploys one or more LoRaWAN gateways to receive data sent by all nodes in the vast area. The mechanism for triggering triage requests is designed to be both robust and sensitive. Specifically, a static sentinel node is only considered to be in a triggered state when the pen anomaly score calculated by its internal algorithm is continuously higher than a specific trigger threshold within a continuous 300-second (i.e., 5-minute) time window. This trigger threshold is not a globally static value, but a private dynamic threshold independently calculated and issued by the central control and prediction subsystem 30 for each pen, thereby achieving adaptation to different pig herd characteristics. Once the triggering conditions are met, the static sentinel node immediately constructs a LoRaWAN uplink data packet. This packet is a highly compact, structured data frame with a carefully designed payload portion that carries the most critical information in the smallest number of bytes. Its fixed format includes: a 16-bit unique identifier for precise location; a 32-bit Unix timestamp recording the time of the event; an 8-bit trigger reason code indicating whether the increase in the anomaly score is mainly contributed by vibration indicators or heatmap indicators, providing initial clues for subsequent detailed diagnosis; and a 16-bit compressed floating-point value of the current anomaly score. This uplink data packet is sent out through the LoRaWAN network, ultimately reaching the central control and prediction subsystem 30 via the gateway, and it itself constitutes the aforementioned triage request.

[0040] Specifically, please refer to Figure 3This demonstrates the hardware platform configuration of the mobile diagnostics and service unit subsystem 20. Besides conventional mobile robot platform components, such as a four-wheel differential drive chassis 21 for flexible movement and in-situ turning, an inertial measurement unit (IMU) for attitude estimation, and a battery management system (BMS) for energy distribution and monitoring, its core value lies in its highly specialized payload. This payload includes a 3D LiDAR 22 for high-precision autonomous navigation and 3D environmental perception, such as a 16-line LiDAR with a scanning frequency set to 15Hz, capable of building a point cloud map of the surrounding environment in real time during robot movement, with a ranging accuracy of ±2cm, which is crucial for collision-free path planning in narrow pigsty passageways.

[0041] The mission payload also includes a high-resolution infrared thermal imager 23. The infrared thermal imager 23 uses an uncooled vanadium oxide (VOx) microbolometer detector with a physical resolution of 640x512 pixels, providing richly detailed thermal images. Its thermal sensitivity (NETD) is less than 40 mK, meaning it can resolve extremely small temperature differences, crucial for detecting early fever symptoms. After rigorous blackbody calibration, its absolute temperature measurement accuracy within the target temperature range is ±0.3℃, sufficient for veterinary diagnostic requirements.

[0042] To achieve precise localization of abnormal sounds, the service unit subsystem 20 is equipped with a planar circular microphone array 24. This array consists of eight high signal-to-noise ratio MEMS microphones evenly spaced (e.g., 10 cm) on the same circular plane, mounted on a two-dimensional pan-tilt unit capable of pitch and horizontal rotation. By processing the minute time difference (TDOA) between the signals from the same sound source received by the eight microphones, the system can calculate the precise azimuth and pitch angles of the sound source.

[0043] To uniquely identify each individual pig, the service unit subsystem 20 also integrates a long-range high-frequency (HF) RFID reader 25 conforming to the ISO 11784 / 11785 standard. The HF RFID reader 25 operates at a frequency of 134.2 kHz and, in conjunction with a directional antenna with high gain and good directionality, can reliably read the unique identification code embedded in the RFID electronic ear tag in the pig's ear at a distance of approximately 0.5 to 1 meter.

[0044] Furthermore, to enable more complex interaction and service functions, some of the aforementioned diagnostic sensors are mounted on the end effector of a six-axis collaborative robotic arm 26. The six-axis collaborative robotic arm 26 offers high flexibility, enabling precise movement and alignment of the sensors with target locations. Its end effector also integrates a wireless charging coil and a high-pressure cleaning nozzle. The wireless charging coil is used to non-contactly replenish the power of low-battery static sentinel nodes during robot inspections; while the high-pressure cleaning nozzle uses cleaning fluid to clean the sensor windows of static sentinel nodes or the robot's own sensors that are obscured by dirt, ensuring the long-term autonomous operation capability of the entire system.

[0045] Furthermore, the individual diagnostic protocol executed by the mobile diagnostic and service unit subsystem 20 after entering the on-demand triage mode is a rigorous deterministic procedure consisting of multiple logical steps. Please refer to... Figure 4The workflow diagram shown illustrates that the individual diagnostic protocol is officially activated after the central control and prediction subsystem 30 forwards a scheduling instruction containing the target pen identifier. The first step is path planning and arrival. The robot uses real-time point cloud data collected by its onboard 3D LiDAR 22 to match and locate itself against a high-precision map of the pigsty stored internally. It then calls path planning algorithms such as A or D to plan a collision-free and time-optimal path from its current position to the entrance of the target pen, and autonomously executes the optimal path. The second step is individual segmentation and locking. After entering the target pen, the robot first uses the 3D LiDAR 22 to scan the interior space of the pen once or multiple times, while simultaneously using its onboard RGB-D depth camera to acquire 3D point cloud data and color images of all pigs in the pen. Through point cloud clustering and instance segmentation algorithms, the system can segment each pig as an independent entity in 3D space. At the same time, it activates the planar circular microphone array 24 to continuously listen to and analyze ambient sound. Once an acoustic signal matching the type of abnormality indicated in the triage request (such as a typical cough) is captured, the sound source localization algorithm immediately calculates the direction of the sound's origin. By spatially fusing and calibrating the sound source localization vector with the 3D point cloud segmentation results, the system can pinpoint the specific pig emitting the abnormal sound with a very high probability, identifying it as the primary diagnostic target. The third step involves non-contact, precise measurement of vital signs. Once the target individual is located, the six-axis collaborative robotic arm 26 immediately adjusts its posture, precisely aligning its high-resolution infrared thermal imager 23 with the head and core torso of the targeted individual, particularly areas like the canthus that more accurately reflect core body temperature, and continuously acquires an infrared video stream for at least 5 seconds. Simultaneously with thermal imaging, the robotic arm adjusts its posture, aligning the RFID reader 25 with the individual to read the unique identification code within its ear tag, completing the final confirmation of the individual's identity. The fourth step involves data analysis and report generation. The robot's onboard computing unit (typically an embedded computer with an integrated GPU, such as the NVIDIA Jetson series) processes the newly acquired high-resolution infrared video stream in real time. The algorithm automatically extracts multiple quantified biometrics, such as the highest surface temperature during the 5-second observation period, the average temperature in the corner of the eye area, and the non-uniformity (standard deviation) of surface temperature distribution in the torso. These metrics, along with the individual identification code, diagnosis time, and acoustic feature analysis results (such as cough frequency and spectral characteristics), are integrated to generate a standardized, machine-readable, structured individual diagnostic report. The fifth step is data reporting. After the diagnosis and report generation are complete, the robot uses its onboard high-speed wireless communication module (such as 5G or Wi-Fi 6) to upload this complete individual diagnostic report to the central control and prediction subsystem 30 in real time.After completing the report, the robot will decide, based on the new instructions issued by the central system, whether to continue performing a simplified diagnostic process on other individuals in the enclosure (e.g., only conducting a temperature check), or to end the current triage task and return to the baseline inspection mode.

[0046] As a core technical feature of this invention, the spatiotemporal graph model of pig herd health status deployed in the central control and prediction subsystem 30 is specifically implemented using a spatiotemporal graph neural network (ST-GNN). Please refer to... Figure 5 This schematically illustrates the concept of a spatiotemporal mapping model of swine herd health status. The construction and operation mechanism of the spatiotemporal mapping model of swine herd health status reflects the elevation of swine herd health management from an individual, isolated perspective to a group, interconnected, and systemic perspective.

[0047] During the model initialization phase, the system abstracts all pigs in the entire pig farm into a dynamic graph structure G=(V,E,X). In this structure, V represents the set of nodes in the graph, where each node v_i∈V uniquely corresponds to a pig wearing an RFID electronic ear tag. E represents the set of edges in the graph, and the existence of an edge e_ij∈E indicates a potential disease transmission association between node v_i (pig i) and node v_j (pig j). The initial weights of this association are not set arbitrarily, but are quantified based on the physical pen adjacency relationships of the pigs (e.g., sharing the same pen, being located upstream or downstream of the same airflow path, sharing the same sewage channel, etc.) and the aerosol diffusion path within the pigsty simulated by computational fluid dynamics (CFD). X represents the feature matrix of the nodes, where X_i(t) is an eigenvector describing the state of node v_i at time t. The eigenvector not only contains static or slowly changing information about the individual, such as age, breed, weight, and historical vaccination records, but also reserves space for dynamic health indicators.

[0048] During the real-time update phase of the model, the data fusion module of the central control and prediction subsystem 30 plays a crucial role. The data fusion module performs the following operations: it continuously receives data streams of pen anomaly scores from all static sentinel nodes within the pig farm. When a significant and sustained increase in the anomaly score S_k(t) of a certain pen k is detected, the system performs two update operations: first, it adjusts the weights of the edges e_ij connecting all pig nodes (v_i, v_j ∈ V_k) within pen k in the dynamic augmented graph model, reflecting a sharp increase in the risk of internal transmission due to group anomalies within the pen; second, it appends the anomaly score S_k(t) as a time-varying, shared environmental feature to the feature vector X_i(t) of all pig nodes within that pen. Meanwhile, when the system receives a high-precision individual diagnostic report for a specific pig v_i generated by the mobile diagnostic and service unit subsystem 20, it directly and with high priority updates the latest feature vector X_i(t) of that node with the precise physiological and pathological data contained in the report (e.g., precise body temperature T_i = 40.8℃, cough event frequency C_i = 3 times / minute, etc.). More importantly, this report from the precise diagnosis is regarded as the "GroundTruth" of that node at that moment, which will be used for subsequent model training or calibration.

[0049] Finally, in the model prediction phase, this constructed and real-time updated ST-GNN model performs a complete forward propagation computation at a fixed time step (e.g., every 10 minutes). In each computation, the model first propagates and aggregates node information in the spatial dimension through graph convolutional operations such as Graph Convolutional Networks (GCNs) or Graph Attention Networks (GATs). This means that the feature vector of each node is fused with the health status information of all its neighboring nodes (i.e., pigs with a risk of propagation). Next, the model models the sequence of feature vectors of each node in the temporal dimension through a Gated Recurrent Unit (GRU) or Long Short-Term Memory (LSTM) network unit to capture the inherent trends and patterns of its health status evolution over time. Through this spatiotemporal joint modeling, the final output of the model is a prediction matrix P(t+Δt), where each element P_i(t+Δt) of the matrix is ​​a prediction vector. The prediction vector contains a series of quantitative indicators such as the probability of a specific health event occurring in a pig v_i within a preset time window Δt (e.g., the next 12 or 24 hours), the probability of body temperature exceeding the clinical high fever threshold (e.g., 40.5℃), the probability of experiencing severe cough symptoms, and the comprehensive probability of being identified as a high-risk individual for infection.

[0050] Furthermore, to make the entire system operate more intelligently and efficiently, the central control and prediction subsystem 30 also integrates a dynamic threshold and scheduling decision module. The scheduling decision module has two core functions. The first is to calculate and issue a private, time-varying trigger threshold for each static sentinel node. The calculation of the dynamic threshold is a multi-factor decision-making process, based on: the statistical distribution (mean, variance, percentiles, etc.) of the pen's historical anomaly scores over the past few weeks or months; the overall activity rhythm of the pig herd inferred from information provided by the pig farm production management system. For example, the system knows that 8:00 to 9:00 AM is the peak feeding period, during which pig activity and noise levels are naturally higher, so it will appropriately increase the trigger threshold during this time period to reduce the false alarm rate, while significantly lowering the threshold during deep sleep from midnight to 4:00 AM to maximize the sensitivity to detect subtle anomalies; and global environmental parameters obtained from the pig house environment control system, such as the overall average temperature and ventilation volume of the pig house, which also affect the baseline behavior of the pig herd. Through this adaptive dynamic threshold adjustment, the system can effectively separate genuine abnormal signals from normal physiological and environmental fluctuations.

[0051] The second core function of the scheduling decision module is to intelligently prioritize these requests when the system receives triage requests from multiple different pens simultaneously, generating an optimal scheduling sequence and instructing the mobile diagnostic and service unit subsystem 20 to execute tasks according to this optimal sequence. This prioritization decision is not based on a single dimension, but rather comprehensively evaluates multiple factors: first, the magnitude and duration of the anomaly score carried by the request itself; second, the economic value of the pigs in the pen as determined by the background database (for example, the diagnostic priority of a pen of gilts awaiting farrowing is obviously higher than that of a pen of fattening pigs about to be sold); and most importantly, the ST-GNN model's prediction of the future disease spread risk level for that pen. A pen with a moderate anomaly score but predicted by the model as a high-risk source of spread may have its diagnostic priority raised above that of a pen with a high anomaly score but predicted as an isolated event.

[0052] Example 1

[0053] This embodiment describes the specific application of the invention in a modern pig farm containing 20 independent farrowing pens for sows. Each pen houses one sow and 10 suckling piglets.

[0054] The system deployment is as follows: One static sentinel node is installed on the wall approximately 1.5 meters above the feed trough in each farrowing pen. The entire farrowing area is covered by a single mobile diagnostic and service unit subsystem. The central control and prediction subsystem is deployed in the central control room of the pig farm, communicating with 20 static sentinel nodes via a LoRaWAN network and with the mobile diagnostic unit via a Wi-Fi 6 network.

[0055] At 2:30 a.m. one day, the pigs were generally at rest. Based on historical circadian rhythm data, the central control and prediction subsystem lowered the trigger threshold of static sentinel nodes in all farrowing pens from 0.65 during the day to 0.40.

[0056] At the static sentinel node located in enclosure 12, the MEMS accelerometer inside began detecting weak but rhythmic atypical vibrations transmitted from the enclosure's metal fence. The end-side computing module performed FFT analysis on the atypical vibration signal, finding a persistent peak in the 50Hz–200Hz frequency band with energy 3 standard deviations higher than the normal background noise. Based on this, the module calculated a vibration anomaly index of 0.75. Simultaneously, data from its 8x8 infrared thermopile sensor showed that 5 out of 64 pixels had a temperature consistently slightly higher than surrounding pixels, causing the thermal anomaly index to rise from the usual 0.1 to 0.35. After weighted summation (0.7 for nighttime vibration and 0.3 for the thermal map), the instantaneous enclosure anomaly score for this node reached 0.63. Over the following 300 seconds, the moving average of this score stabilized at 0.55, consistently above the dynamically adjusted trigger threshold of 0.40.

[0057] At 2:35 AM, the static sentinel node of hut 12 automatically constructed and sent a triage request. The data packet content was: {Hut ID: 0x000C, Timestamp: 1678822500, Trigger Reason: 0x01 (mainly vibration), Abnormal Score: 0.55}.

[0058] The central control and prediction subsystem received a triage request. With no other requests at this time, the system immediately issued a dispatch instruction to the mobile diagnostics and service unit subsystem, which was waiting at the charging station.

[0059] After receiving instructions, the mobile diagnostic unit autonomously planned its path and arrived at the entrance of pen number 12 within 2 minutes. Upon entering the pen, it initiated an individual segmentation and locking process. The microphone array quickly located the continuous coughing sound emanating from a piglet in a corner of the pen, with the sound source localization vector pointing towards the target. Simultaneously, a 3D LiDAR and depth camera scanned the pen, segmenting the piglet from its sow and litter into an independent point cloud instance. After fusing the sound source and point cloud data, the target was uniquely locked.

[0060] The robotic arm aimed a high-resolution infrared thermal imager at the piglet's head and captured a 5-second video stream. Analysis by the onboard computer revealed that the piglet's highest surface temperature was 40.8℃, and its eye temperature was 40.6℃. Simultaneously, an RFID reader successfully read the ear tag ID as "3201-12-07".

[0061] At 2:41 AM, the mobile diagnostic unit generated and reported an individual diagnostic report. Upon receiving the report, the central control and prediction subsystem immediately updated the node status representing piglet "3201-12-07" in the spatiotemporal graph model, marking it as a confirmed case of fever and cough. The ST-GNN model performed forward transmission prediction, and the results showed that the probability of the other 9 piglets in the same litter who had close contact with this piglet developing fever symptoms within the next 12 hours increased to 75%, while the probability of infection for piglets in the adjacent pen No. 11 was 20%.

[0062] The system then pushed a high-priority warning to the mobile terminal of the pig farm manager: "Warning: In farrowing house No. 12, piglet ID3201-12-07 has been diagnosed with high fever (40.8℃) and cough. The model predicts that the risk of infection within the same litter is extremely high. It is recommended to immediately isolate and observe the pig and intervene in the entire pen."

[0063] Comparative Example 1

[0064] This comparative example employs a traditional monitoring scheme that relies solely on mobile robots for periodic inspections. This scheme was deployed in the same pig farm environment as Example 1, but without any static sentinel nodes. The mobile robots were programmed to perform a routine inspection of all 20 farrowing pens every 4 hours.

[0065] Similarly, at 2:30 a.m. on a certain day, piglet “3201-12-07” in pen number 12 began to show the same early symptoms as in Example 1.

[0066] However, since there are no static sentinel points for continuous monitoring, the system is completely unaware of this. The robot's last inspection was completed at midnight, and the next scheduled inspection time is 4:00 AM.

[0067] During this 90-minute "monitoring blind spot," the diseased piglet's condition worsened, and it transmitted the pathogen to the other three piglets in the same litter through direct contact and aerosols.

[0068] At 4:00 AM, the mobile robot arrived at pen number 12 as planned to conduct a routine inspection. During the inspection, its onboard infrared thermal imager scanned the entire pen and detected multiple hot spots. The robot then initiated a detailed diagnostic procedure, measuring and confirming that, including the first batch of piglets, a total of four piglets had a body temperature exceeding 40.5℃.

[0069] At 4:05 AM, the robot completed the diagnosis and reported the results. The system then sent an alert to the administrators.

[0070] Data Comparison and Analysis

[0071] To more intuitively demonstrate the beneficial effects of the technical solution of the present invention, the key performance indicators of Example 1 and Comparative Example 1 are quantitatively compared, and the specific data are shown in the table below.

[0072]

[0073] As can be seen from the detailed description and data comparison of the above embodiments and comparative examples, the pig farm robot health monitoring and early warning system provided by the present invention for pig farming platforms significantly shortens the delay time from the occurrence of disease to the detection and early warning through the collaborative working mechanism of static sentinels and mobile robots, greatly moves the intervention window forward, effectively controls the spread of disease within the population, and greatly improves the utilization efficiency of high-value diagnostic resources.

Claims

1. A pig farm robot health monitoring and early warning system for pig farming platforms, characterized in that, The system includes: a static sentinel node subsystem (10), consisting of at least one static sentinel node deployed in each independent pig pen, used to collect pen environment and pig activity data at high frequency and continuously, generate pen abnormality scores in real time based on the end-side computing model, and send a triage request containing pen identifier when the pen abnormality score continues to exceed the dynamically adjusted trigger threshold. The static sentry node is an integrated electronic device whose hardware consists of a central microcontroller unit (11), a wireless communication transceiver module (12), a power supply module (13), and a set of sensor arrays (14). The central microcontroller unit (11) is electrically connected to the wireless communication transceiver module (12), the power supply module (13) supplies power to all hardware units, the central microcontroller unit (11) is the central control chip, and the set of sensor arrays (14) is electrically connected to the central microcontroller unit (11). The static sentry node is encapsulated in a corrosion-resistant and bite-resistant shell (15) with an IP68 protection rating, made of polycarbonate injection molding and covered with 304 stainless steel mesh. The sensor array (14) specifically includes: a triaxial microelectromechanical system accelerometer, whose sampling frequency is set to be no less than 4096Hz and whose measurement range is ±16g, used to continuously capture the weak vibration signals of the pen structure caused by pigs' coughing, panting and abnormal limb movements, and the weak vibration signals of the pen structure caused by pigs' coughing, panting and abnormal limb movements are the original vibration data. A temperature and humidity composite sensor with a temperature measurement accuracy of ±0.2℃ and a humidity measurement accuracy of ±2%RH is used to monitor changes in the microenvironment within the pen; and an 8x8 pixel array infrared thermopile sensor with a field of view of 60 degrees and a temperature measurement range covering 20℃ to 100℃ is used to acquire a low-resolution thermal map of the overall body surface temperature distribution of the pigs in the pen. The mobile diagnosis and service unit subsystem (20) is a pig house inspection robot with autonomous navigation capabilities. It executes baseline inspection mode and on-demand triage mode: In the baseline inspection mode, it performs full-domain inspection and obtains baseline health data according to the low-frequency scheduling strategy. In the on-demand triage mode, it responds to the triage request, autonomously navigates to the target pen, executes the individual diagnosis protocol to obtain the precise physiological and pathological data of the target pig and generates an individual diagnosis report. The hardware platform of the mobile diagnostic and service unit subsystem (20) includes a four-wheel differential drive chassis (21), an inertial measurement unit and a battery management system. The four-wheel differential drive chassis (21) is the transmission part of the pig house inspection robot and is used to provide power to the pig house inspection robot. The inertial measurement unit is used to compensate for the inertia of the pig house inspection robot when it is inspecting. The core mission payload includes: A 3D lidar (22) for autonomous navigation and three-dimensional environment perception has a scanning frequency of 15Hz and a ranging accuracy of ±2cm. A high-resolution infrared thermal imager (23) with an uncooled vanadium oxide microbolometer as its detector, a resolution of 640x512 pixels, a thermal sensitivity of less than 40mK, and a temperature measurement accuracy of ±0.3℃ after calibration; a planar circular microphone array (24) consisting of 8 MEMS microphones spaced 10cm apart and mounted on a two-dimensional gimbal, used for locating the sound source. A long-range high-frequency RFID reader (25) conforming to ISO11784 / 11785 standards and operating at a frequency of 134.2kHz, with a directional antenna; And a six-axis collaborative robotic arm (26); The central control and prediction subsystem (30) communicates bidirectionally with the static sentinel node subsystem (10) and the mobile diagnosis and service unit subsystem (20), receives and integrates the continuous pen abnormality scores of the static sentinel node and the individual diagnosis reports of the mobile diagnosis and service unit subsystem (20), constructs a spatiotemporal map model of the pig herd health status, predicts the disease transmission path and health evolution trend and generates early warning information; The spatiotemporal graph model of swine herd health status is specifically a spatiotemporal graph neural network, and the model is constructed as follows: Abstract the entire pig farm into a dynamic graph structure G=(V,E,X); Where V is the set of nodes in the graph, and each node v_i∈V uniquely corresponds to a pig wearing an RFID ear tag; E is the set of edges in the graph. Edge e_ij∈E indicates that there is a potential disease transmission association between nodes v_i and v_j. The initial weight of the association is set based on the physical pen adjacency relationship of the pigs and the fixed air circulation simulation path. X is the feature matrix of a node, and X_i(t) is the feature vector of node v_i at time t. The feature vector contains the historical health data of the pigs.

2. The system according to claim 1, characterized in that, The fixed end-side calculation model within the static sentinel node is specifically a pen anomaly scoring calculation module, and its execution flow is as follows: A fast Fourier transform is performed on the raw vibration data collected by the triaxial microelectromechanical system accelerometer within a continuous 1-second time window to convert the time-domain signal into a frequency-domain signal. In the frequency domain, the energy of the first characteristic frequency band and the second characteristic frequency band are integrated to obtain the vibration anomaly index, wherein the frequency band range of the first characteristic frequency band is 50Hz to 200Hz, and the frequency band range of the second characteristic frequency band is 400Hz to 800Hz. The mean and standard deviation of the temperature of 64 pixels output by the infrared thermopile sensor of the 8x8 pixel array are calculated and compared with a dynamic baseline temperature issued by a central control and prediction subsystem (30) to obtain the thermal map anomaly index. The vibration anomaly index and the heat map anomaly index are weighted and summed. The weight coefficients are dynamically configured by the central control and prediction subsystem (30) based on historical data and the daily activity rhythm of the pig herd. The summation result is the pen anomaly score at the current moment.

3. The system according to claim 1, characterized in that, The wireless communication between the static sentinel node subsystem (10) and the central control and prediction subsystem (30) adopts the LoRaWAN protocol. The static sentinel node is used as the terminal device, and a LoRaWAN gateway is deployed on the central control and prediction subsystem (30). The conditions for generating a triage request are: the moving average of the pen anomaly score of a static sentinel node is continuously higher than the trigger threshold for dynamic adjustment set by the central control and prediction subsystem (30) for the pen within a continuous 300 seconds. The triage request is a structured uplink data packet whose payload always includes: a 16-bit column unique identifier, a 32-bit timestamp, an 8-bit trigger reason code, and a 16-bit floating-point number representing the current anomaly score.

4. The system according to claim 1, characterized in that, An individual diagnostic protocol consists of multiple steps, including: Step 1, Path planning and arrival: The mobile diagnostic and service unit subsystem (20) uses the scanning data of the 3D LiDAR (22) and the pre-stored high-precision map to plan and execute a collision-free optimal path to reach the entrance of the target enclosure; The second step is individual segmentation and locking: After the mobile diagnosis and service unit subsystem (20) enters the pen, it uses 3D lidar (22) to scan the space inside the pen and combines it with the RGB-D depth camera on board to perform three-dimensional point cloud instance segmentation of all pigs in the pen; at the same time, it starts the planar circular microphone array (24) to locate the abnormal acoustic signal indicated in the triage request; by fusing the direction vector obtained by the sound source localization with the three-dimensional point cloud instance segmentation results, it locks the individual pig that made the abnormal sound as the primary diagnostic target; The third step is to measure vital signs in a non-contact manner: the six-axis collaborative robotic arm (26) adjusts its posture and precisely aligns the high-resolution infrared thermal imager (23) with the head and core area of ​​the locked pig individual to collect a continuous 5-second video stream, and uses a long-range high-frequency RFID reader (25) to read the unique identification code in its ear tag. The fourth step is data analysis and report generation: The onboard computing unit of the mobile diagnosis and service unit subsystem (20) processes the collected infrared video stream, extracts the highest body surface temperature, the corner of the eye temperature, and the quantitative indicators of uneven body surface temperature distribution, and generates a structured individual diagnosis report by combining the read individual identity code. The fifth step is data reporting. After the diagnosis and report are generated, the robot uploads the complete individual diagnostic report to the central control and prediction subsystem (30) in real time through its onboard high-speed wireless communication module.

5. The system according to claim 1, characterized in that, The six-axis collaborative robotic arm (26) is equipped with a service end effector that integrates a wireless charging coil for non-contact energy replenishment of static sentinel nodes and a high-pressure cleaning nozzle for cleaning the sensors of the static sentinel nodes or the mobile diagnostics and service unit subsystem (20).

6. The system according to claim 1, characterized in that, The real-time update mechanism of the spatiotemporal mapping model of swine herd health status includes: We continuously receive data streams of fence anomaly scores from all static sentinel nodes. When the anomaly score S_k(t) of fence k increases, we enhance the weight of the edge e_ij connecting all nodes v_i∈V_k within fence k. At the same time, we attach S_k(t) as a time-varying fence-level environmental feature to the feature vector X_i(t) of all nodes within the fence. Furthermore, upon receiving an individual diagnostic report for pig v_i generated by the mobile diagnostic and service unit subsystem (20), the precise physiological and pathological data in the individual diagnostic report are directly updated to the latest feature vector X_i(t) of this node, and used as the true state label of this node at this moment.

7. The system according to claim 1, characterized in that, The central control and prediction subsystem (30) also includes a dynamic threshold and scheduling decision module, which is used for: For each static sentinel node, a private, time-varying trigger threshold is calculated and issued. The calculation of the trigger threshold is based on: the historical abnormal score statistical distribution of this pen, the overall activity rhythm of the pig herd, and global environmental parameters. Furthermore, when multiple triage requests arrive simultaneously, the scheduling decision module is responsible for prioritizing the tasks of the mobile diagnosis and service unit subsystem (20) based on the abnormal score of the request, the value of the pigs in the pen, and the disease spread risk level of the pen predicted by the spatiotemporal graph model, and generating the optimal scheduling sequence.

Images