Intelligent pig raising and selling platform data analysis system integrated with 4G camera

Abstract

The invention relates to the technical field of artificial intelligence, in particular to an intelligent pig raising and selling platform data analysis system integrated with a 4G camera, and aims to solve the problems of data semantic segmentation and space-time mismatching of an existing heterogeneous sensing system. The system comprises an intelligent sensing node deployed in a breeding place and a cloud data analysis platform. The intelligent sensing node integrates an individual identification unit, a multi-modal visual acquisition unit and an edge calculation unit, and realizes data source fusion and information extraction; the data analysis platform constructs individual digital twins, and performs multi-dimensional intelligent analysis through growth prediction, health anomaly detection, market transaction decision and other models. According to the scheme, accurate prediction and optimization decision making of the live pig individual full life cycle data are realized, and powerful support is provided for fine breeding management, health early warning and intelligent transaction.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and specifically relates to a data analysis system for an intelligent pig farming and selling platform integrating a 4G camera. Background Art

[0002] This invention relates to the field of livestock farming information technology, and more particularly to a system based on machine vision and wireless communication technologies for collecting individualized pig lifecycle data, performing intelligent analysis, and supporting trading decisions. Deeply integrating information technology with traditional farming to develop "smart farming" has become an inevitable trend in enhancing the industry's core competitiveness.

[0003] Against this backdrop, the industry has already conducted numerous beneficial explorations. Early informatization efforts focused primarily on macro-control of the breeding environment and automated management of production processes. Specifically, by deploying environmental sensors for temperature, humidity, and ammonia concentration within piggeries and integrating them with ventilation and temperature control equipment, technicians have achieved automated and intensive control of the breeding environment, significantly improving the pig's growth environment. At the same time, to accurately identify and trace individual pigs, radio frequency identification (RFID) electronic ear tag technology has been introduced. Combined with automatic feeding stations and weighing equipment, it can record key physiological indicators such as feed intake and weight for individual pigs. The application of these technologies has, to a certain extent, transformed the breeding process from one that relied entirely on manual experience to one that relies partially on data recording and automated equipment. This has resolved initial issues in traditional breeding, such as confusion over individual identities and the lack of key growth data, and provided fundamental data support for subsequent production management. In terms of visual monitoring, traditional closed-circuit television (CCTV) or network camera systems are also widely deployed in large-scale farms. However, their main functions are security prevention and remote overview. That is, they provide managers with real-time video streams inside the pig house so that they can remotely view the overall activity status of the pig herd. Their core value lies in "monitoring" and "viewing", rather than in-depth "analysis" and "insight".

[0004] However, as livestock farming continues to expand and the market demands for refined and predictable management become increasingly stringent, the inherent contradictions of this technical architecture, which primarily addresses single functional issues and consists of multiple heterogeneous subsystems, have gradually become apparent. These contradictions are not simply due to functional deficiencies, but rather stem from inherent limitations in the technical principles, leading to data "semantic fragmentation" and "temporal and spatial mismatches." This stems from the inherent isolation of data collected by different subsystems in terms of dimensions, timestamps, and associated entities. For example, an RFID system captures the identity and weight of a pig at a specific moment and location (such as a feeding station), constituting a discrete data point. Meanwhile, a video surveillance system records the behavior of all pigs within the pen over a continuous time dimension. However, this image data itself is not inherently tied to a specific ID. Therefore, even if both weight data and behavioral videos are stored in the backend database, it is impossible to accurately and automatically perform real-time, reliable correlation analysis of abnormal behaviors (such as prolonged lying or aggressive behavior) with the specific individual causing the behavior and its historical health data (such as recent weight loss stagnation) within the massive video frames. This semantic gap between the data makes it extremely difficult to construct a multimodal health and growth model covering the entire lifecycle of a single individual. Managers are left with a large number of fragmented, context-incomplete data silos, rather than a unified view that comprehensively and dynamically reflects the true condition of each individual. Consequently, decisions based on this fragmented data, such as disease warnings and determining the optimal time to market, are inevitably less accurate and timely, leading to the persistent paradox of "data abundance" and "information poverty."

[0005] Therefore, how to break through the technical bottlenecks of data semantic gap and spatiotemporal mismatch caused by the parallel operation of existing heterogeneous systems, and build a technical solution that can realize the native fusion of high-dimensional visual information and individual identity information, and conduct continuous, integrated data collection and deep correlation analysis, so as to provide truly accurate, dynamic and forward-looking decision-making basis for breeding production and market transactions, has become a key challenge and technical problem that needs to be solved urgently by technical personnel in this field. Summary of the Invention

[0006] The purpose of the present invention is to overcome the above-mentioned defects of the prior art and provide a data analysis system for an intelligent pig farming and selling platform integrating 4G cameras, which aims to fundamentally solve the problems of data semantic fragmentation and spatiotemporal mismatch caused by the parallel operation of heterogeneous sensing systems. By constructing a technical system that can realize the native fusion of individual identity information and multimodal visual information at the source of data collection, a continuous, high-dimensional, traceable data closed loop of a single individual throughout its life cycle is formed, thereby providing a deterministic and highly correlated data foundation for the refined management of pig farming, accurate early warning of health status, and intelligent decision-making in market transactions.

[0007] In order to achieve the above-mentioned purpose of the invention, the technical solution adopted by the present invention is: to provide a data analysis system for an intelligent pig farming and selling platform integrated with a 4G camera, the system physically includes at least one intelligent sensing node deployed in the breeding site, and a data analysis platform logically deployed on a cloud server; a data communication link is established between the intelligent sensing node and the data analysis platform through the fourth generation mobile communication (4G) network or other equivalent wireless wide area network.

[0008] Specifically, the intelligent sensing node is an integrated, multimodal sensing device with edge computing capabilities. Its structure consists of a unified, sealed housing with a predetermined protection level, within which resides an individual identification module, a multimodal visual acquisition module, an edge computing processing unit, and a wireless communication module. The individual identification module, multimodal visual acquisition module, and wireless communication module are all electrically connected to the edge computing processing unit to exchange data.

[0009] In one embodiment of the present invention, the individual identification module is an ultra-high frequency (UHF) radio frequency identification (RFID) reader / writer operating in the 902 MHz to 928 MHz frequency band, with a precisely controlled transmit power of 30 decibel milliwatts (dBm). The reader / writer is connected to a directional antenna designed for high gain and a narrow beam angle. Specifically, it is a microstrip phased array antenna that generates a conical RF energy field with a vertex angle of less than 15 degrees, with the axis of the RF energy field perpendicular to the ground. This creates a precisely defined, high-energy-density identity information reading zone directly below the intelligent sensing node. This configuration ensures that when a pig wearing a UHF RFID electronic ear tag enters the identity information reading zone, its identity code can be instantly, uniquely, and reliably identified. Outside of the identity information reading zone, cross-reading or misreading is prevented, providing a deterministic spatial anchor for subsequent identity-visual information binding.

[0010] Furthermore, the multimodal vision acquisition module consists of a heterogeneous array of image sensors, either co-axial or with a fixed spatial relationship. The image sensor array includes at least one visible light image sensor, a depth information sensor, and a long-wave infrared thermal imaging sensor. The visible light image sensor uses a 1 / 1.8-inch complementary metal oxide semiconductor (CMOS) photosensitive element with a physical resolution of 3840x2160 pixels. It is paired with a fixed-focus lens with a focal length of 4mm to provide a wide-angle, high-resolution color image with a horizontal field of view of 90 degrees, covering the main activity area of ​​the pig pen. The depth information sensor is an active structured light sensor consisting of an infrared laser emitter and an infrared image receiver. The laser emitter projects a specially coded infrared speckle pattern into the detection field of view. The infrared image receiver captures the distorted speckle pattern after reflection from the target object's surface. By calculating the degree of distortion, it determines the three-dimensional coordinates of each point in the field of view, generating a depth image with a resolution of 640x480 pixels. Its effective range is 0.5 to 5 meters, and its depth measurement accuracy is better than 1% of the measured distance. The long-wave infrared thermal imaging sensor utilizes an uncooled vanadium oxide (VOx) microbolometer focal plane array. Operating in the 8 to 14 micron wavelength band, it has a physical resolution of 320x240 pixels and a thermal sensitivity (NETD) below 50 millikelvin (mK). It accurately captures the temperature distribution on the target object's surface and generates a thermal image. These three sensors are triggered synchronously to ensure that at any given moment, the acquired color, depth, and thermal images correspond to the scene state at the same instant.

[0011] The edge computing processing unit (ECU) is the core of the intelligent perception node. Its hardware architecture is based on a system-on-chip (SoC) that integrates a multi-core central processing unit (CPU), a graphics processing unit (GPU), and a dedicated artificial intelligence (AI) acceleration core. Specifically, it includes an 8-core ARM Cortex-A78AE CPU cluster, a Volta architecture GPU with 512 CUDA cores, and a deep learning accelerator (DLA) capable of 20 tera operations per second (TOPS). The unit is equipped with 16 gigabytes (GB) of LPDDR4x unified memory and 64 gigabytes (GB) of eMMC onboard storage. The ECU runs a pre-configured spatio-temporal identity-visual fusion (STIVF) algorithm. The spatiotemporal identity-visual fusion algorithm operates as follows: First, the algorithm continuously monitors the output of the individual identification module. Upon detecting a confirmed electronic ear tag identity code (hereinafter referred to as an "identity anchor event"), the algorithm immediately triggers the multimodal visual acquisition module to perform a synchronous acquisition, acquiring color images, depth images, and thermal images that fully correspond to the identity code in both time and space. The algorithm then extracts a high-dimensional initial feature vector from this set of multimodal images, representing the biometric characteristics of the specific individual. This initial high-dimensional biometric feature vector includes, but is not limited to, convolutional neural network-based surface texture and contour features extracted from the color image, three-dimensional point cloud morphological features (such as body length, height, and backfat thickness estimates) extracted from the depth image, and surface temperature distribution statistical features (such as mean temperature, maximum temperature, and temperature variance) extracted from the thermal image. This data structure containing the identity code and initial feature vector is defined as the "identity anchor object."

[0012] After an identity anchoring event occurs, the continuous tracking module of the STIVF algorithm is activated even if the individual pig leaves the identity information reading area. This module uses a multi-target tracking logic based on a Siamese network and a Kalman filter. Specifically, the Siamese network uses the previously generated initial feature vector as a template, performs sliding window matching in subsequent consecutive video frames (including color images and depth images), and calculates the feature distance between each detected target and the template. Depth information is used in this process to provide accurate three-dimensional position and size information of the target, greatly enhancing the tracking robustness in densely occluded environments. The Kalman filter makes predictions based on the target's motion state (position, velocity, acceleration) and updates the state based on the matching results of the Siamese network, thereby achieving continuous and stable cross-frame tracking of individuals with anchored identities.

[0013] The edge computing processing unit does not directly transmit the raw multimodal video stream to the cloud server. Instead, it encapsulates the STIVF algorithm's processing results into structured data packets. Each packet contains a millisecond-accurate timestamp, a unique individual identity code, the individual's real-time 3D coordinates in the pen coordinate system, a dynamically updated biometric vector, and a series of key physiological and behavioral indicators calculated in real time by the algorithm (e.g., instantaneous weight, respiratory rate, gait parameters, standing / lying time, feeding time, etc., estimated using 3D point cloud voxel integration). These structured data packets are periodically transmitted (e.g., once per second) to the data analysis platform via a wireless communication module (a 4G communication module supporting LTE Cat.6). This edge-based data fusion and information extraction process ensures a high degree of data semantics from the source and significantly reduces the demand for wireless network uplink bandwidth.

[0014] The data analysis platform is deployed in the cloud and consists of a data aggregation and fusion module, an individual digital twin model, a multi-dimensional algorithm analysis engine, and a decision support and visualization interface module.

[0015] The data aggregation and fusion module is responsible for receiving and parsing data packets from one or more intelligent sensing nodes. This module includes a spatiotemporal trajectory splicing submodule. When multiple intelligent sensing nodes are deployed within a farm, this submodule uses the extended Kalman filter algorithm to seamlessly splice and smooth the movement trajectory of the individual throughout the entire farm space based on the data packets with the same individual identity code reported by different nodes, combined with their three-dimensional spatial coordinates and timestamps, to form a unified, global activity path record.

[0016] The individual digital twin model creates an independent, life-cycle digital archive, or "digital twin," for each individual pig in the system (indexed by its unique identity code). This digital twin, stored in a time series database (e.g., InfluxDB), continuously records all structured data related to that individual reported by front-end nodes, including but not limited to estimated weight, temperature fluctuations, activity levels, feeding behavior, and social behavior (e.g., frequency and duration of proximity to other individuals). This provides a comprehensive, multimodal, high-frequency, and long-term depiction of the physiological and behavioral state of a single individual.

[0017] The multi-dimensional algorithm analysis engine is the core of this invention's intelligent analysis and decision-making. It integrates a series of interrelated machine learning and operations research models. This engine includes at least: 1. Growth Curve Prediction Model. This model utilizes a long short-term memory (LSTM)-based architecture. It takes the historical weight estimate and feed intake data sequences from a single digital twin as input, combined with static information such as pig breed and age, to predict the individual's future weight growth curve. The model outputs a weight prediction and its confidence interval for a specific point in the future (e.g., the next 30 days). Based on this prediction, it estimates the most likely date for reaching a preset market weight (e.g., 110 kg).

[0018] 2. Health Status Anomaly Detection Model. The health status anomaly detection model is an unsupervised learning model based on a variational autoencoder (VAE). For each individual, the model uses multi-dimensional data from the past seven days of its digital twin (including gait symmetry parameters, body surface temperature distribution entropy, average daily activity, social isolation index, etc.) for training to learn the "normal behavior and physiological patterns" of the individual in a healthy state. In subsequent runs, the model continuously calculates the reconstruction error between newly reported data and normal behavior and physiological patterns. When the reconstruction error continuously exceeds a preset threshold (for example, the 99th percentile of the individual's historical error distribution), the system automatically generates a health warning and points out the multi-dimensional data that caused the anomaly (for example, abnormal gait of the right hind limb or increased local temperature of the eye), providing precise clues for early intervention by veterinarians.

[0019] 3. Market Transaction Optimal Decision Model. The Market Transaction Optimal Decision Model is a decision-making agent based on reinforcement learning. Its core is a policy network that takes a composite state vector as input and outputs a decision: "Immediately market" or "Continue to feed." This composite state vector accurately describes the complete context at the moment of decision. Its components include: the individual pig's current precise weight (provided by the growth curve model), its predicted daily weight gain, its current health score (provided by the anomaly detection model), national and local hog futures and spot prices collected in real time from an external data interface, and the farm's own unit feed cost. The decision-making agent's reward function is designed to maximize the ultimate marginal profit of each pig: reward = (market price at market) - (accumulated feeding costs incurred from the decision moment to market). Through offline training on a large amount of historical data and in a simulated market environment, the model learns an optimal dynamic decision-making strategy that recommends the optimal time to sell each pig that reaches the market weight range, maximizing economic benefits, based on individual growth and market price fluctuations.

[0020] Finally, the decision support and visualization interface module integrates all of the above analysis results and presents them to livestock managers or trading platform users through a web-based graphical user interface (GUI). This interface is not a traditional video surveillance screen, but rather a data-centric management cockpit. Its main view includes a sortable and filterable list of pigs, with each row representing a pig. The columns intuitively display the pig's identity code, current estimated weight, health score, expected market date, and the "recommend to sell" or "recommend to hold" label given by the optimal decision model. Users can click on any pig to enter its individual digital twin details page and view its complete historical growth curve, behavioral heat map, health warning records, and other in-depth information, providing unprecedented data transparency and decision support capabilities for production management and market transactions.

[0021] In summary, this invention, through the design of an intelligent perception node integrating multimodal sensing and edge computing, extracts high-value information and deeply binds it to individual identities at the source of data generation, completely resolving the data fragmentation problem inherent in traditional solutions. Furthermore, by constructing a cloud-based individual digital twin model and a multi-dimensional intelligent analysis engine, it achieves precise prediction and optimized decision-making for individual pigs throughout the entire process, from growth and health to market transactions. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a system block diagram of the present invention.

[0023] Figure 2 It is a structural diagram of the intelligent sensing node in the present invention.

[0024] Figure 3 It is a structural diagram of the multimodal vision acquisition module in the present invention.

[0025] Figure 4 This is a schematic diagram of the application scenario of the intelligent sensing node in the present invention in a breeding site.

[0026] Figure 5 It is a flow chart of data processing and analysis in the present invention.

[0027] The accompanying drawings are marked as follows: 100, intelligent sensing node; 110, sealed shell; 120, individual identity recognition module; 130, multimodal visual acquisition module; 131, visible light image sensor; 132, depth information sensor; 133, long-wave infrared thermal imaging sensor; 140, edge computing processing unit; 150, wireless communication module; 200, data analysis platform; 210, data aggregation and fusion module; 220, individual digital twin modeling module; 230, multi-dimensional algorithm analysis engine; 231, growth curve prediction model; 232, health status anomaly detection model; 233, market transaction optimal decision model; 240, decision support and visualization interface module. DETAILED DESCRIPTION

[0028] To make the objectives, technical solutions, and advantages of the present invention more clear, the following will further describe in detail the "Intelligent Pig Farming and Selling Platform Data Analysis System Integrating 4G Cameras" provided by the present invention in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and do not constitute a limitation of the present invention.

[0029] Reference Figure 1 The present invention provides a data analysis system for an intelligent pig farming and selling platform that integrates 4G cameras. The system is divided into an intelligent sensing node 100 physically deployed at the breeding site and a data analysis platform 200 logically deployed on a cloud server. A stable and reliable data communication link is established between the intelligent sensing node 100 and the data analysis platform 200 via a widely covered fourth-generation mobile communication (4G) network, or other functionally equivalent wireless wide area networks, such as fifth-generation mobile communication (5G) networks or low-power wide-area Internet of Things (LPWAN) technology. This architectural design aims to place the functions of data collection and preliminary processing at the source of data generation, while concentrating the in-depth analysis and decision-making functions that require massive computing resources and a global data view in the cloud, thereby achieving optimal allocation of computing resources and maximizing the overall system efficiency.

[0030] Now combine Figure 2 and Figure 4The internal structure of the intelligent sensing node 100 and its deployment in practical application scenarios are described in detail. The intelligent sensing node 100 is a highly integrated intelligent sensing device, with all its functional components encapsulated within a unified sealed housing 110. This sealed housing 110 is preferably made of polycarbonate (PC) doped with a UV inhibitor and is integrally molded through injection molding. Silicone rubber seals are used at the joints to provide waterproof and dustproof treatment, giving the entire device an IP67 protection rating. This effectively withstands the harsh conditions commonly found in aquaculture environments, such as high humidity, dust, ammonia corrosion, and high-pressure water jet flushing, ensuring the long-term stable operation of the precision electronic components within.

[0031] Inside the sealed housing 110, four core functional modules are housed: an individual identification module 120, a multimodal visual acquisition module 130, an edge computing processing unit 140, and a wireless communication module 150. These four modules are physically integrated and electrically interconnected via a custom-designed printed circuit board (PCB). Specifically, the individual identification module 120, the multimodal visual acquisition module 130, and the wireless communication module 150 are each connected as submodules to the edge computing processing unit 140, the system's main controller, via standardized interfaces (such as USB, MIPI-CSI, and PCIe). The edge computing processing unit 140 provides unified power management, clock synchronization, and data scheduling.

[0032] As a preferred embodiment of the present invention, the individual identity recognition module 120 is embodied as a radio frequency identification (RFID) reader / writer operating in the ultra-high frequency (UHF) band. The operating frequency range of the reader / writer is precisely set within the public frequency band of 920 MHz to 925 MHz specified by the Ministry of Industry and Information Technology of the People's Republic of China, and its radio frequency output power is precisely calibrated and stably controlled at 30 decibel milliwatts (dBm), or 1 watt. The reader / writer is connected to a specially designed directional antenna via a coaxial cable with a 50-ohm impedance. This directional antenna is not an ordinary dipole antenna, but an 8-unit phased array antenna using microstrip patch technology. By precisely controlling the feeding phase of each antenna unit, the antenna can generate a radio frequency energy field with high gain (greater than 10dBi) and extremely narrow beam width. In actual deployment, such as Figure 4As shown, the intelligent sensing node 100 is mounted in the center of the ceiling of a pig pen (e.g., a typical fattening pen), approximately 3 meters above the ground, with its main axis maintained strictly perpendicular to the ground. In this configuration, the RF energy field generated by the directional antenna projects onto the ground as a circular area with a diameter of approximately 1.5 meters, forming a precisely defined identity information reading zone. The extremely high energy density of the identity information reading zone ensures that when the head or body of an individual pig wearing an ISO18000-6C-compliant UHF RFID ear tag enters this zone, its unique electronic identity code (EPC) is instantly (read in less than 50 milliseconds), accurately, and uniquely captured. Furthermore, due to the sharp narrowing of the beam, the identity information reading zone is clearly defined, effectively preventing cross-reading or misreading of ear tags of pigs in adjacent areas. This provides a crucial, deterministic spatial anchor point for the subsequent precise and unambiguous binding of individual identity information with multimodal visual information.

[0033] Further, refer to Figure 3 , the composition of the multimodal vision acquisition module 130 is explained. This module is essentially a heterogeneous image sensor array, whose design goal is to synchronously acquire rich dimensional information about the target scene at a single moment. The image sensor array is fixed to a precision-machined aluminum alloy bracket to ensure the permanent relative spatial position relationship between each sensor. Its calibration parameters are pre-written into the non-volatile memory of the edge computing processing unit 140. The image sensor array includes at least three core sensors: a visible light image sensor 131, a depth information sensor 132, and a long-wave infrared thermal imaging sensor 133.

[0034] Specifically, the visible light image sensor 131 uses the Sony IMX586 1 / 1.8-inch back-illuminated complementary metal oxide semiconductor (CMOS) photosensitive element, which has a physical effective pixel count of up to 48 million. Using four-in-one pixel technology, it can output high-quality 12-megapixel images in low-light environments, and a video stream with a physical resolution of 3840x2160 pixels (i.e., 4K ultra-high-definition) under normal lighting conditions. The sensor is paired with a distortion-corrected, wide-angle fixed-focus lens with a focal length of 4mm, providing a wide horizontal field of view of 90 degrees that covers most of the active area of ​​the pig pen, allowing it to capture high-resolution visual details such as the surface pattern, color, and body outline of individual pigs.

[0035] The depth information sensor 132 is a three-dimensional imaging device based on active structured light technology. It integrates an infrared laser emitter and an infrared image receiver. The laser emitter is an 850nm vertical-cavity surface-emitting laser (VCSEL). Through a diffractive optical element (DOE), it projects a pseudo-randomly coded infrared speckle pattern consisting of tens of thousands of discrete light spots into the forward detection field of view. When these speckles are projected onto the surface of a pig or other object, they are distorted by the object's three-dimensional shape. An infrared image receiver (a 640x480 pixel infrared CMOS sensor) located coaxially or at a fixed baseline distance captures this modulated speckle pattern. A dedicated algorithm within the edge computing processing unit 140 performs pixel-by-pixel matching and triangulation of the captured distorted pattern with a pre-stored reference pattern to determine the three-dimensional spatial coordinates (X, Y, Z) of each pixel in the field of view relative to the sensor. This ultimately generates a 640x480 pixel depth image aligned with the visible light image. The effective range of the depth information sensor 132 is optimized to 0.5 to 5 meters, completely covering the distance from the pigpen floor to the sensor. Its absolute accuracy of depth measurement at a distance of 2 meters is better than 20 mm, or less than one percent of the measured distance. This depth map provides critical geometric information for subsequent individual segmentation, 3D morphological analysis, and behavior recognition.

[0036] The long-wave infrared thermal imaging sensor 133 utilizes an uncooled vanadium oxide (VOx) microbolometer focal plane array. Its detection spectrum falls within the atmospheric window of 8 to 14 microns, effectively eliminating interference from visible and near-infrared light and directly sensing the heat radiated by an object. The sensor has a physical resolution of 320x240 pixels, and its thermal sensitivity, or noise equivalent temperature difference (NETD), is less than 50 millikelvin (mK), meaning it can discern subtle temperature changes of less than 0.05 degrees Celsius. This sensor enables the system to obtain a thermal image accurately reflecting the temperature distribution of the pig's body surface and the surrounding environment, providing direct physiological indicators for health status monitoring (such as localized high temperatures caused by fever and inflammation).

[0037] Crucially, these three sensors are connected to the general-purpose input / output (GPIO) pins of the edge computing processing unit 140 via hardware trigger signal lines, achieving microsecond-level synchronized triggering. When the edge computing processing unit 140 issues an acquisition command, the visible light image sensor 131, depth information sensor 132, and long-wave infrared thermal imaging sensor 133 are exposed at the exact same time. This ensures that the acquired color image, depth image, and thermal image form a multimodal data frame that is perfectly aligned in time, laying a solid foundation for subsequent cross-modal information fusion.

[0038] As the "brain" of the intelligent sensing node 100, the edge computing processing unit 140 boasts extremely powerful hardware. It's based on NVIDIA's Jetson AGX Xavier series industrial-grade system-on-chip (SoC). This SoC integrates an 8-core 64-bit ARM Cortex-A78AE CPU cluster for general control logic and serial tasks; a Volta architecture graphics processing unit (GPU) with 512 CUDA cores for parallel acceleration of traditional computer vision algorithms and image pre-processing; and two dedicated deep learning accelerators (DLAs). With a combined computing power of up to 20 teraflops (20 TOPS INT8), it provides the hardware foundation for efficiently running complex neural network models. The unit also features 16GB of LPDDR4x unified memory, enabling the CPU and GPU to share physical memory, significantly reducing data copy latency between processors. Furthermore, the onboard 64GB of eMMC5.1 flash memory is used to store the operating system (a tailored Linux distribution), applications, and cache data during network outages.

[0039] This hardware platform pre-deploys and runs one of the core algorithms of the present invention: the Spatio-Temporal Identity-Visual Fusion (STIVF) algorithm. The STIVF algorithm's execution flow is designed as a state machine, with rigorous logic and interconnected processes. First, the algorithm's main thread continuously polls the output buffer of the individual identification module 120 at a high frequency. Upon detecting the successful reading of a valid, new electronic ear tag identity code, this constitutes an "identity anchor event." This event immediately triggers an interrupt, and the algorithm records the identity code (for example, 3B7A-0012C4E6) along with the current millisecond-accurate timestamp. It then immediately issues a synchronization acquisition command to the multimodal visual acquisition module 130. After execution, the algorithm obtains a set of raw data that fully corresponds to the identity code in both time and space: a 3840x2160 pixel color image, a 640x480 pixel depth map, and a 320x240 pixel thermal image.

[0040] The algorithm then enters the feature extraction phase. Using the depth map information, it first accurately segments the pixel regions corresponding to the identified pig in both the color image and the thermal image (generating a mask for the pig). Then, for this segmented region, the algorithm concurrently extracts a high-dimensional initial feature vector from the three modalities that comprehensively characterizes the individual's current biological state. This initial high-dimensional biological feature vector is constructed from multiple dimensions. Specifically, a pre-trained ResNet-50 convolutional neural network running on a GPU is used to extract a 2048-dimensional depth feature of the surface texture and contour from the pig region in the color image. Principal component analysis (PCA) and geometric fitting are used to calculate the pig's 3D morphological features, such as preliminary estimates of body length, height, chest circumference, and backfat thickness, estimated by interpolating the point cloud from a specific dorsal region, using the 3D point cloud data recovered from the depth map. Finally, statistical features of the pig's surface temperature, such as mean, maximum, minimum, and temperature distribution variance and entropy, are extracted from the aligned thermal images. This composite data structure, which includes identity code, timestamp and initial high-dimensional feature vector, is defined as an "identity anchor object" within the system. It constitutes the first "digital portrait" of the individual pig at a specific point in time and space.

[0041] Once an "identity anchor object" is successfully established, the STIVF algorithm's continuous tracking module seamlessly takes over, even if the individual pig subsequently leaves the identification information reading area. The core of this module is a multi-target tracking logic that combines deep learning and classical filtering theory. Specifically, it employs a Siamese network fine-tuned for a specific task. This network uses the 2048-dimensional deep feature vector previously generated for each pig as a unique, personalized "template." In the video stream continuously captured at 10 frames per second by the multimodal vision acquisition module 130, the tracking algorithm first uses a lightweight object detection network (such as YOLOv5-s) to identify candidate bounding boxes for all pigs in each frame. Then, for each candidate bounding box, the Siamese network extracts its feature vector and calculates cosine similarity with the pig's "template" vector. The candidate bounding box with the highest similarity, exceeding a preset threshold (e.g., 0.9), is considered the location of the pig in the current frame. Depth information plays a crucial role in this process: it provides accurate 3D spatial position and size information for each candidate frame. When pigs are physically occluded, even if visual features are temporarily blurred, the system can still significantly improve tracking accuracy and robustness based on the 3D position and motion continuity of the unobstructed parts. Simultaneously, a Kalman filter is instantiated for each tracked individual. This filter predicts the position of the next frame based on the target's historical motion state (position, velocity, and acceleration in 3D space). This prediction is then weighted and fused with the matching results of the Siamese network, enabling continuous, stable, and long-term cross-camera tracking of individuals with anchored identities.

[0042] A key technical feature is that the edge computing processing unit 140 does not upload the raw, data-heavy multimodal video stream directly to the cloud. Instead, it executes the aforementioned STIVF algorithm locally in real time, converting the raw sensory data into highly condensed, semantically rich, structured information. This information is encapsulated into standardized JSON-formatted data packets. Each packet contains a millisecond-accurate UTC timestamp synchronized by an NTP service, a unique hexadecimal individual identity code, the individual's real-time coordinates (X, Y, Z) within the three-dimensional coordinate system of the pigpen, a biometric vector dynamically updated based on the individual's latest visual appearance, and a series of key physiological and behavioral indicators calculated in real time by specialized sub-algorithms. These indicators directly reflect the value of the present invention, such as instantaneous weight estimated by voxel integration of the tracked individual's three-dimensional point cloud (with an accuracy of up to ±1.5%), respiratory rate calculated by analyzing the periodic fluctuations of the chest and abdomen on the depth map, gait parameters (such as stride length, cadence, and symmetry) extracted by analyzing the three-dimensional motion trajectory of the limb joints, and standing / lying time, eating time, and drinking time calculated through long-term series analysis. These structured data packets are typically around a few kilobytes (KB) in size and are efficiently transmitted to the cloud-based data analysis platform 200 at a preset frequency (e.g., once per second) via wireless communication module 150, a 4G communication module supporting the LTE Cat. 6 standard with uplink and downlink peak rates of 300 / 50 Mbps. This "edge intelligence" processing paradigm, which completes data fusion, feature extraction, and information refinement at the edge of the network, fundamentally ensures the "identity-time-space-state" integration of uploaded data, and reduces the system's demand for wireless network uplink bandwidth by several orders of magnitude, greatly improving the system's deployment flexibility and economy.

[0043] Now let's turn to the data analysis platform 200, logically deployed in the cloud. This platform is typically built on a virtual server cluster within a large public cloud (such as Alibaba Cloud or Tencent Cloud), leveraging the elastic computing, distributed storage, and high availability services provided by the cloud platform. Its internal logical functions can be divided into a data aggregation and fusion module 210, an individual digital twin modeling module 220, a multi-dimensional algorithm analysis engine 230, and a decision support and visualization interface module 240.

[0044] The data aggregation and fusion module 210 is the platform's data entry point. It receives and parses data packets reported by hundreds or even thousands of intelligent sensing nodes 100 within one or more farms via a high-concurrency message queue (such as Apache Kafka). When multiple intelligent sensing nodes 100 are deployed within a large farm to achieve comprehensive coverage, a key submodule within this module—the spatiotemporal trajectory splicing submodule—comes into play. When different nodes report data packets containing the same individual's identity code (e.g., "3B7A-0012C4E6"), this submodule uses the extended Kalman filter (EKF) algorithm to seamlessly stitch and smooth these segmented trajectory data, which may contain slight measurement errors, based on the three-dimensional spatial coordinates and high-precision timestamps contained in each packet. Ultimately, this generates a unified, global, and highly accurate record of the individual's movement path throughout the entire farm space.

[0045] Next, the individual digital twin modeling module 220 creates an independent, persistent digital profile in the cloud for each individual pig monitored by the system (using its unique electronic ear tag as a globally unique identifier). This profile serves as the "digital twin" of the physical entity. The module's underlying storage architecture preferably utilizes a high-performance time series database, such as InfluxDB. For each digital twin, the database continuously appends all structured time series data related to the individual, reported by the front-end nodes and processed by the aggregation module. This includes, but is not limited to, weight estimates, average surface temperature changes, daily activity levels (total distance moved in meters), total feeding duration, and social behavior parameters (e.g., frequency and duration of interaction with other specific individuals within a 2-meter radius). In this way, the system constructs a comprehensive, multimodal, high-frequency (down to seconds), and long-term (up to several months) profile of each pig's entire lifecycle, from entry to exit. This creates a gold mine of data for in-depth analysis and precise prediction.

[0046] The multi-dimensional algorithm analysis engine 230 is the core of the present invention for achieving advanced intelligent analysis and automated decision-making. It integrates a series of interrelated and collaborative machine learning and operations research models. This engine includes at least the following three key models: First, there's the Growth Curve Prediction Model 231. This model aims to accurately predict future pig growth trends. Its technical implementation utilizes a deep learning architecture based on a long short-term memory (LSTM) network. For each individual digital twin, the Growth Curve Prediction Model 231 takes as primary input a series of historical weight estimates and average daily feed intake data from the past 30 days. It also incorporates static physiological information such as the pig's breed (e.g., Duroc), age, and sex as auxiliary inputs. Through supervised training on growth data from millions of historical pigs, the LSTM network learns the complex, nonlinear patterns of weight growth under varying conditions. During the inference phase, the Growth Curve Prediction Model 231 predicts the weight of a given individual at a specific point in the future (e.g., each day for the next 30 days) and provides a 95% confidence interval for the predicted value. Based on this predicted curve, the model can further accurately estimate the most likely date for the individual to reach a preset market weight (e.g., 110 kg), providing a scientific basis for the farm's marketing and feed procurement plans.

[0047] Second, there's the Health Anomaly Detection Model 232. Its purpose is to provide early, automated warnings of pig health issues. Given the diversity and difficulty of pre-labeling "abnormal" states, the Health Anomaly Detection Model 232 employs an unsupervised learning paradigm based on a variational autoencoder (VAE). For each individual, the Health Anomaly Detection Model 232 first extracts multi-dimensional data from their digital twin over the past seven days (serving as a baseline health period). This data forms a high-dimensional vector, including, but not limited to, the left-right limb movement symmetry index calculated from gait parameters, the surface temperature distribution entropy (a measure of body temperature uniformity) calculated from thermal imaging, average daily activity levels, and the social isolation index (the degree to which the frequency of interaction with other pigs is below the group average) derived from social network analysis. The Health Anomaly Detection Model 232 is trained using this data to learn to encode and reconstruct a low-dimensional latent representation of the individual's "normal behavioral and physiological patterns" in a healthy state. During subsequent daily operations, the health anomaly detection model 232 continuously feeds newly reported daily data into the trained VAE and calculates the reconstruction error between the input data and the data reconstructed by the health anomaly detection model 232. When this reconstruction error significantly exceeds a dynamically set threshold (e.g., the 99th percentile value calculated based on the individual's historical error distribution) for several consecutive periods (e.g., six consecutive hours), the system automatically generates a health alert. This alert not only indicates that a particular pig may have a health issue but also, by analyzing the contribution of the reconstruction error across various input dimensions, identifies the primary cause of the anomaly (e.g., "a significant decrease in the gait symmetry index of the right hind limb" or "an abnormally elevated local temperature in the periorbital area"), providing invaluable and precise clues for rapid diagnosis and early intervention by veterinarians.

[0048] Third, there is the Market Transaction Optimal Decision Model 233. This Market Transaction Optimal Decision Model 233 aims to address the core economic question of "when to sell pigs" to maximize farming profitability. Its technical core is a decision-making agent based on deep reinforcement learning. At the heart of this agent is a policy network, which learns a mapping from "state" to "action." At each decision moment (e.g., once a day), the network receives as input a composite state vector that accurately describes the complete contextual information required for the decision. The dimensions of this composite state vector include: the current precise weight estimate of the individual pig provided by the growth curve model 231; the predicted average daily weight gain for the next seven days; the current health score output by the Health Status Anomaly Detection Model 232; national and local hog futures and spot prices and price volatility, as well as feed costs per unit weight recorded in the farm's own ERP system, obtained in real time from external data sources (such as the National Pig Market Network and the Dalian Commodity Exchange) via APIs. The action space of this decision-making agent is very simple, consisting of only two discrete options: "sell immediately" or "keep raising for another day." Its reward function is carefully designed to maximize the ultimate marginal profit of each pig, calculated as follows: Reward = (market price at slaughter) - (accumulated feeding costs incurred from the decision moment to the actual slaughter moment in the future). Through millions of rounds of offline training in a simulated environment that includes historical growth data for tens of thousands of pigs and market price fluctuations from past years, the decision-making agent has learned a highly optimized decision-making strategy that is adaptable to market dynamics. In practice, for each pig that reaches a preset slaughter weight range (e.g., 105-125 kg), the model will provide a clear daily recommendation: "Sell" or "Hold," thereby helping managers find the optimal selling time to maximize economic benefits in the complex game between individual growth potential and market price fluctuations.

[0049] Finally, the decision support and visualization interface module 240 is responsible for integrating and presenting all of the above analysis results in a user-friendly manner. This interface is provided to livestock managers or traders connected to the platform via a graphical user interface (GUI) based on modern web technologies (such as React.js and D3.js). The core design concept of this interface is "data-centric," completely abandoning the traditional video surveillance wall that only provides viewing functions. Its main view is a powerful management dashboard, typically displaying a list of pigs that can be sorted and filtered by multiple criteria. Each row in the list represents a pig, and the columns intuitively display its core management indicators: identification code, current estimated weight, 30-day weight gain, health score, expected date to reach 110 kg, and a "sell" or "hold" decision label determined by the market trading optimal decision model 233. Users can simply click on any pig entry to drill down to its individual digital twin's detailed information page. This page visually displays the individual's complete historical growth curve, daily activity level changes, temperature fluctuations, social network diagrams, and all historical health warning records in the form of charts and timelines. This unprecedented data transparency and deep insight provide solid technical support for the refined and standardized management of pig farming and intelligent, high-profit decision-making in market transactions.

[0050] Example: To verify the effectiveness of the technical solution of the present invention, a 120-day deployment test was conducted at a modern commercial pig fattening farm in Zhucheng City, Shandong Province. The farm housed 500 Duroc-Landrace-Yorkshire hybrid pigs of similar age in a standard pen measuring 20 meters long and 10 meters wide. All pigs were fitted with UHF RFID ear tags upon entry.

[0051] Two intelligent sensing nodes 100 of the present invention are symmetrically installed along the centerline of the pigpen ceiling, 3.5 meters above the ground and 10 meters apart. This ensures that their respective identity information reading and visual acquisition areas effectively cover the entire pigpen area, with some overlap for track splicing. The nodes connect to China Unicom's 4G network via a built-in 4G wireless communication module 150. The cloud-based data analysis platform 200 is deployed on the Alibaba Cloud North China node.

[0052] The system operated continuously during the test. For example, consider a fattening pig numbered "A8E4-00F1B3D9." On the 65th day after entering the pen, the individual identification module 120 of the intelligent sensing node (Node-01) located on the east side of the pen first read the pig's ear tag ID near the water trough below it. The STIVF algorithm was immediately triggered, capturing multimodal images of the pig, creating its "identity anchor object," calculating its initial feature vector, and estimating its weight at the time to be 92.7 kg and its average body surface temperature to be 38.8 degrees Celsius. Subsequently, even after the pig left the water trough and walked toward the resting area in the center of the pen, Node-01's continuous tracking module maintained stable tracking of the pig using the Siamese network and Kalman filter. As the pig continued to move into the west side of the pen, its position was captured by Node-02. The cloud-based spatiotemporal trajectory stitching submodule seamlessly reconstructed its complete trajectory based on the data alternately reported by the two nodes.

[0053] Over the next 20 days, the pig's digital twin was continuously updated with data. Based on the pig's actual weight increase from 92.7 kg to 108.5 kg, combined with its stable feeding behavior, the growth curve prediction model 231 predicted that the pig would reach its target market weight of 115 kg on day 91. On day 78, the health status anomaly detection model 232 issued a low-level warning, indicating that the pig's gait symmetry index had been below 2 standard deviations below its individual baseline for 12 consecutive hours, suggesting possible mild discomfort in its left forelimb. Upon receiving this alert, the on-site veterinarian conducted a detailed observation and discovered that the pig was indeed experiencing mild lameness. He intervened promptly, preventing the problem from worsening.

[0054] Starting on day 85, when the pig's weight exceeded 105 kg, Market Transaction Optimal Decision Model 233 began daily assessing the timing of its sale. At the time, the local spot price of live pigs was 16.5 yuan / kg and trending downward. Combining the price trend with the prediction that the pig would continue to experience high daily weight gain, the model repeatedly issued "hold" decisions. By day 90, the pig's weight had reached 114.8 kg. The price rebounded slightly to 16.8 yuan / kg, but external data indicated a large number of pigs would be sold in the future, putting downward pressure on prices. After comprehensive assessment, the decision model changed the decision label to "sell" for the first time that day. The farm manager adopted this recommendation and sold the pig the following day, achieving a satisfactory profit margin.

[0055] Comparison: In another pen on the same farm, identical conditions were met, using traditional management methods. This pen also housed 500 commercial pigs from the same batch. Four standard HD network cameras were installed in corners for monitoring. Video streams were transmitted to a National Video Recorder (NVR) in the central control room for storage, primarily for manual inspections. Pig identification and weight measurement were performed manually: once a week, two workers entered the pen, scanned each pig's ear tag with a handheld RFID reader, and then weighed each pig on a mobile electronic scale. The data was manually recorded in an Excel spreadsheet.

[0056] The comparative operation revealed numerous problems. First, data fragmentation. The real-time surveillance video from the central control room could not be directly linked to the weekly weight and identity data. When a pig was observed to be listless or lying down for an extended period, managers were unable to immediately identify it, nor could they determine its recent weight changes or feeding habits, making assessments extremely difficult. Second, the manual weighing process placed significant stress on the pigs, impacting their feed intake and weight gain. It was also inefficient, with weighing 500 pigs taking nearly half a day. Third, the extremely infrequent data collection (once a week) made it impossible to develop an accurate growth model, much less provide early and precise warnings of health issues. During the test period, a pig in the comparative pen developed arthritis due to a streptococcal infection. This was not detected until lameness became apparent and weight loss continued for several days, by which time the optimal treatment opportunity had passed. Finally, when it comes to pig selling decisions, managers can only decide the timing of selling the entire batch based on experience and rough average weight estimates, combined with a macro-judgment of market prices. They are unable to optimize for individual differences and short-term market fluctuations, and their economic benefits are highly uncertain.

[0057] Data comparison and conclusion: In order to quantify the advantages of the embodiments of the present invention over the comparative examples, we conducted statistics and comparisons on the key performance indicators during the test period. The results are shown in the following table:

[0058] In summary, through the above detailed description, specific embodiments and quantitative data comparison, it can be clearly seen that the present invention's "intelligent pig farming and selling platform data analysis system integrating 4G cameras" realizes the native fusion of identity and multimodal information at the perception source, and on this basis builds an individual digital twin and multi-dimensional intelligent analysis engine throughout the entire life cycle, fundamentally solving the core pain points in traditional farming management.

Claims

1. A data analysis system for an intelligent pig farming and selling platform integrating a 4G camera, comprising an intelligent sensing node (100) deployed at a breeding site and a data analysis platform (200) in the cloud, the two communicating via a 4G network; characterized in that: The intelligent sensing node (100) is an integrated sensing device, comprising an individual identity recognition module (120), a multimodal visual acquisition module (130), and an edge computing processing unit (140); the edge computing processing unit (140) performs the following operations: when the individual identity recognition module (120) recognizes the pig identity code, the multimodal visual acquisition module (130) is triggered to synchronously acquire multimodal visual information corresponding to time and space; based on the identity code and the visual information, a structured data packet containing the identity code, an accurate timestamp, and a biometric feature vector is locally generated; The structured data packet is sent to a cloud-based data analysis platform (200) via a 4G network; the cloud-based data analysis platform (200) receives and parses the structured data packet, constructs an individual digital twin modeling module (220) throughout the life cycle for each identity code, and analyzes the individual's growth status, health status, and market transaction opportunities through a multi-dimensional algorithm analysis engine (230).

2. The system according to claim 1, wherein: The individual identification module (120) is an ultra-high frequency radio frequency identification reader / writer, and the ultra-high frequency radio frequency identification reader / writer is connected to a directional antenna; the directional antenna is configured to form a radio frequency energy field with a narrow beam angle, and the axis of the radio frequency energy field is perpendicular to the ground, thereby forming an identity information reading area with a precisely defined spatial range directly below the intelligent sensing node (100), so as to instantly and uniquely identify the identity code of the individual pig wearing the electronic ear tag that enters the identity information reading area, and avoid cross-reading of individuals outside the identity information reading area.

3. The system according to claim 1 or 2, characterized in that The multimodal vision acquisition module (130) is composed of a group of image sensor arrays with a fixed spatial position relationship. The image sensor array realizes synchronous acquisition through a hardware synchronization trigger signal, and the image sensor includes at least: a visible light image sensor (131) configured to collect high-resolution color images covering the main activity area of ​​the pig pen to obtain surface texture and contour features of individual pigs; A depth information sensor (132), which is an active structured light sensor, is configured to calculate the three-dimensional spatial coordinates of objects in the field of view by projecting coded infrared speckles and capturing their reflected distortion to generate a depth image for obtaining three-dimensional morphology and spatial position information of individual pigs; and A long-wave infrared thermal imaging sensor (133) is configured to collect thermal imaging images reflecting the surface temperature distribution of an object, so as to obtain the surface temperature distribution characteristics of individual pigs.

4. The system according to claim 3, characterized in that The edge computing processing unit (140) is further configured to execute a spatiotemporal identity-visual fusion algorithm, and the execution logic of the spatiotemporal identity-visual fusion includes: Identity anchoring: When the individual identity recognition module (120) outputs a confirmed identity code, it is defined as an identity anchoring event, and the multimodal visual acquisition module (130) is immediately triggered to perform a synchronous acquisition; Initial feature extraction: Accurately segment the area corresponding to the identified individual from the color image, depth image, and thermal image acquired during this simultaneous acquisition, and extract an initial high-dimensional biometric feature vector from the identity information reading area. The initial high-dimensional biometric feature vector includes at least body surface visual features based on a convolutional neural network, morphological features based on three-dimensional point cloud data, and statistical features of body surface temperature distribution, thereby generating an identity anchor object; and Continuous tracking: After an identity anchoring event occurs, the initial high-dimensional biometric feature vector is used as a template to track the individual with the anchored identity across frames in subsequent continuous visual data frames through a tracking logic based on a Siamese network and a Kalman filter.

5. The system according to claim 4, characterized in that The structured data packet generated by the edge computing processing unit (140) also includes a series of key physiological and behavioral indicators calculated in real time by the spatiotemporal identity-visual fusion algorithm; Key physiological and behavioral indicators include at least: instantaneous weight estimated based on the voxel integration method of the three-dimensional point cloud of the tracked individual, respiratory rate calculated based on analysis of the periodic fluctuations of its chest and abdomen on the depth map, gait parameters extracted based on analysis of the three-dimensional motion trajectory of its limb joints, and standing time, lying time and eating time obtained through statistical analysis of long-term tracking sequences.

6. The system according to claim 1, wherein: The data aggregation and fusion module (210) in the data analysis platform (200) includes a spatiotemporal trajectory splicing submodule; when multiple intelligent sensing nodes (100) are deployed in the farm, the spatiotemporal trajectory splicing submodule is configured as follows: based on multiple structured data packets containing the same individual identity code reported by different nodes, combined with the three-dimensional spatial coordinates and timestamps in each data packet, the extended Kalman filter algorithm is used to seamlessly splice and smooth the movement trajectory of the individual in the entire farming space to form a unified global activity path record, and store it in the individual digital twin modeling module (220) of the individual.

7. The system according to claim 1, wherein: A growth curve prediction model (231) is integrated within the multi-dimensional algorithm analysis engine (230); the growth curve prediction model (231) is a regression model based on a long short-term memory network, which is configured to: use the historical weight estimation sequence and feed intake data sequence recorded in the individual digital twin modeling module (220) of a single individual as time series input, and combine the static information of pig breed and age to predict the future weight growth curve of the individual, and output the predicted date when the individual reaches the preset market weight.

8. The system according to claim 1, wherein: A health status anomaly detection model (232) is integrated within the multi-dimensional algorithm analysis engine (230); the health status anomaly detection model (232) is an unsupervised learning model based on a variational autoencoder, which is configured as follows: first, multi-dimensional data from a specific past period in each individual digital twin modeling module (220) is used for training to learn the normal behavior and physiological pattern of the individual in a healthy state; then, during continuous operation, the reconstruction error between the newly reported data and the normal behavior and physiological pattern is calculated. When the reconstruction error continuously exceeds the threshold dynamically set for the individual, a health warning is automatically generated, and the physiological or behavioral dimension that causes the anomaly is pointed out.

9. The system according to claim 7, wherein: The multi-dimensional algorithm analysis engine (230) also integrates a market transaction optimal decision model (233); the market transaction optimal decision model (233) is a decision agent based on reinforcement learning, and its core is a policy network. The policy network is configured to: at each decision moment, receive a real-time composite state vector of the market pig price and the breeding feed cost as input, and output a decision of "immediately sell" or "continue breeding".

10. The system according to claim 9, characterized in that The system also includes a decision support and visualization interface module (240), which is configured to present a data-centric management cockpit to the user through a graphical user interface; the main view of the management cockpit is a sortable and filterable pig list, each row of the sortable and filterable pig list represents a pig, and the column data intuitively displays its identity code, current estimated weight, health score, expected market release date, and the selling or holding recommendation given by the market transaction optimal decision model (233); the user can view the detailed historical data of the individual digital twin modeling module (220) of any individual through the graphical user interface.

Images