Early warning and diagnosis and treatment method and system for large-scale pig farm diseases

Abstract

This invention relates to a method and system for early disease warning and diagnosis in large-scale pig farms, belonging to the field of pig farming technology. This method utilizes knowledge graphs and abnormal features for comprehensive disease risk assessment, determining the target disease type and current disease stage of the target pigs. Based on this, an early warning is generated. An initial treatment plan is generated based on the warning information, disease type determination, and disease stage determination. On-site diagnosis and treatment are then performed on the target pigs, and samples are collected and sent for testing, forming diagnostic verification data. This invention unifies and links the initial warning record, initial diagnosis result, on-site actual treatment execution information, re-examination results, and laboratory test results of the target pigs, forming a verification data chain for individual cases. This provides a basis for subsequent verification of warning and diagnosis results, reducing the accumulation of misjudgments caused by the inability to verify model outputs alone.

Description

Technical Field

[0001] This invention relates to the field of pig farming technology, and in particular to methods and systems for early disease warning and diagnosis in large-scale pig farms. Background Technology

[0002] With the development of large-scale and intensive pig farming, modern farms typically need to manage tens of thousands of pigs simultaneously. Because pig diseases are characterized by rapid spread, wide impact, and significant economic losses, how to promptly identify abnormal individuals and take effective intervention measures in pig farming has become a crucial technical challenge in the field of smart farming. Traditional pig health management mainly relies on daily inspections by farmers. They judge whether there is a risk of disease by observing the pigs' mental state, feeding and drinking, breathing, and fecal appearance. This method is highly dependent on the experience of farmers and has problems such as low monitoring efficiency, strong subjectivity, and difficulty in detecting early abnormalities. It is difficult to meet the needs of large-scale farms for early detection, early treatment, and early prevention of diseases.

[0003] In recent years, with the development of artificial intelligence and multi-source sensing technologies, existing technologies have begun to utilize multimodal data such as images, audio, environmental sensors, spatial layout information, and basic pig information to identify abnormal pig behavior, determine disease types, and provide early warnings. However, current technologies are still mainly focused on the stages of anomaly identification, risk warning, or diagnostic result output, and generally lack deep integration with the actual diagnosis and treatment process in the breeding field. Specifically, on the one hand, existing technologies usually only output warning results, disease types, or treatment suggestions, lacking correlation records of the actual treatment execution process, medication information, and post-treatment changes; on the other hand, existing technologies usually lack a unified verification mechanism between follow-up observation results, laboratory test results, and initial warning or diagnostic results, making it difficult to determine whether the initial warning and diagnostic conclusions are accurate, and also difficult to assess the effectiveness of recommended treatment plans in actual production scenarios.

[0004] Furthermore, existing technologies generally lack a feedback mechanism for continuously revising and updating early warning rules, diagnostic results, and treatment strategies based on real-world clinical outcomes. This leads to a disconnect between early warning, diagnosis, treatment, and verification, making it difficult to form a continuous optimization loop tailored to the farming environment. Especially under different breeds, ages, disease stages, and environmental conditions, without feedback learning from historical cases based on follow-up examinations and laboratory confirmations, it becomes difficult to continuously improve the accuracy of early warnings, the reliability of diagnoses, and the rationality of medication for subsequent similar cases.

[0005] Therefore, there is an urgent need in this field for a closed-loop optimization technology solution that can combine early warning of swine diseases, on-site diagnosis and treatment, follow-up testing, laboratory validation, and strategy updates to improve the timeliness, accuracy, and standardization of disease management in large-scale farms. Summary of the Invention

[0006] This invention overcomes the shortcomings of the prior art and provides a method and system for early disease warning and diagnosis for large-scale pig farms.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: This invention provides a method for early disease warning and diagnosis in large-scale pig farms, comprising the following steps: Collect multimodal data information of pigs and extract abnormal features from the multimodal data information of pigs; Obtain typical and atypical features of different diseases in the early, middle and late stages, and construct a knowledge graph based on the typical and atypical features of different diseases in the early, middle and late stages. The knowledge graph and abnormal features are used to conduct a comprehensive disease risk assessment, determine the target disease type and current disease stage of the target pigs, and generate early warning information based on the target disease type and current disease stage of the target pigs. Based on the aforementioned early warning information, disease type judgment, and disease stage judgment, an initial treatment plan is generated, and on-site diagnosis and treatment are performed on the target pigs, and samples are collected and sent for testing to form diagnosis and treatment verification data; Based on the aforementioned diagnostic and treatment validation data, a correlation analysis is performed on the disease early warning structure, diagnostic structure, and treatment structure.

[0008] Furthermore, in the method for early disease warning and diagnosis in large-scale pig farms, multimodal data information of pigs is collected, and abnormal features are extracted from the multimodal data information of the pigs, specifically as follows: Behavioral image data, body surface temperature data, feed and water intake data, environmental parameter data, and aquaculture management data are collected through visible light cameras, infrared thermal imagers, feed and water intake metering devices, and environmental monitoring sensors. The behavioral image data, body surface temperature data, feeding and drinking data, environmental parameter data, and breeding management data are synchronized in time and associated with individuals. The Faster R-CNN target detection algorithm and pose estimation network were used to extract vomiting, limping, kicking, and abnormal agitation behavioral features for early disease identification. LSTM was used to process environmental parameters, abnormal frequency of lying-standing transitions, decreased food and water intake, abnormal appearance of lesions, and information related to the incidence of diseases in the same column.

[0009] Furthermore, in the early disease warning and diagnosis methods for large-scale pig farms, typical and atypical features of different diseases in their early, middle, and late stages are obtained. A knowledge graph is then constructed based on these typical and atypical features of different diseases in their early, middle, and late stages. Specifically: Typical features are clustered in the shallow region near the root node along an exponentially growing divergent path in hyperbolic space, while atypical features are mapped to the edge region far from the root node. A neuron-like ordinary differential equation is introduced, and this equation is used to simulate the state changes of typical and atypical features to form a feature set and construct a time-aware knowledge graph reasoning model. In the time-aware knowledge graph reasoning model, the current feature set of the patient is input, and the position of the feature set in hyperbolic space is calculated through the attention mechanism of the graph. The warning is then issued based on the position of the feature set in hyperbolic space. The training conditional variational autoencoder generates possible atypical features based on disease and stage. When the graph lacks data on preset late-stage atypical features of the disease, the time-aware knowledge graph reasoning model uses counterfactual reasoning to generate an evolution path graph of atypical features.

[0010] Furthermore, in the early disease warning and treatment method for large-scale pig farms, the knowledge graph and abnormal features are used to conduct a comprehensive disease risk assessment, determine the target disease type and current disease stage of the target pigs, and generate warning information based on the target disease type and current disease stage of the target pigs. Specifically: The abnormal features are input into the knowledge graph for knowledge reasoning to determine the target disease type and current disease stage of the target pig. Obtain information on the breed, age, weight, historical disease incidence, disease incidence in the same or adjacent pens, seasonal and climatic factors, indoor environmental conditions, and breeding management events of the target pigs. A comprehensive score is calculated based on the breed, age, weight, historical disease incidence, disease incidence in the same or adjacent pens, seasonal and climatic factors, indoor environmental conditions and management events, the target disease type of the target pig, and its current disease stage. A comprehensive risk score is obtained through comprehensive evaluation. When the comprehensive risk score reaches a preset warning threshold, a warning message corresponding to the target pig is generated.

[0011] Furthermore, in the early disease warning and treatment methods for large-scale pig farms, an initial treatment plan is generated based on the aforementioned warning information, disease type determination, and disease stage determination, specifically including: Based on the aforementioned warning information, disease type judgment, disease stage, and combined with the breed, age, and weight of the target pigs, a structured initial diagnosis result is output. The diagnostic results include the target disease type, disease stage, diagnostic confidence level, and the initial treatment plan corresponding to the current disease stage. The treatment plan includes recommended drug types, drug dosage ranges, administration methods, administration frequency, administration cycle, and information on whether isolation observation, follow-up examinations, and laboratory testing are required.

[0012] Furthermore, in the early disease warning and treatment methods for large-scale pig farms, on-site diagnosis and treatment are carried out on target pigs, and samples are collected and sent for testing to form diagnostic and treatment validation data, specifically: Upon receiving the warning information, the system arrives at the pen where the target pig is located, performs manual verification of the target pig, and conducts on-site diagnosis and treatment based on the output target disease type, disease stage, and initial treatment plan, and uploads the diagnosis and treatment report. Biological samples were collected from the target pigs and sent to the laboratory for testing. The biological samples included at least one or more of the following: blood samples, fecal samples, and secretion samples. Record the sample collection time, sample type, delivery time, delivery items and testing institution information, and uniquely link them with the target pig's early warning information, initial diagnosis results and actual treatment execution information; Receive laboratory test reports, extract etiological results, blood or biochemical test results, confirmed disease type, test time and test conclusions, and link them with the warning information, initial diagnosis results, actual treatment implementation information and re-examination results corresponding to the target pigs to form treatment verification data.

[0013] Furthermore, in the early disease warning and treatment methods for large-scale pig farms, a correlation analysis is performed on the disease warning structure, diagnostic structure, and treatment structure based on the aforementioned diagnostic validation data, specifically as follows: A correlation analysis was performed on the disease early warning structure, diagnostic structure, and treatment structure of the diagnostic and treatment verification data; The disease type identification, feature weights, early warning thresholds, disease stage determination criteria, and treatment plan recommendation priorities were optimized and updated. The updated model weights and the current case data were then written into the case database.

[0014] Furthermore, in the early disease warning and diagnosis methods for large-scale pig farms, LSTM is used to process environmental parameters, abnormal lying-up / standing frequency, decreased feed and water intake, abnormal external lesions, and information related to disease incidence in the same pen. Specifically: An event-driven temporal alignment mechanism is introduced to align continuous environmental parameters to the timestamps of feeding behavior events through piecewise linear interpolation. LSTM is used to calculate the Markov transition entropy of the lying and standing positions to quantify the degree of agitation. Physiological prior constraints are added to the loss function of LSTM, and the weight coefficients of environmental temperature and feed intake are constrained to conform to the known physiological heat stress model. Each individual, water line, feed line and pen are treated as nodes. Individuals in the same column are fully connected, and individuals sharing the same water tank / feed tank are strongly connected. Based on the ventilation direction and column spacing, CFD simulation is used to simplify the connection into dynamic weights, giving downwind individuals a higher connection weight. The spread of pathogens within a column is simulated using gated graph convolution, and a population risk penetration vector is output. This population risk penetration vector is then used to predict the probability that an individual will be affected by the same column within a preset time period.

[0015] A second aspect of the present invention provides a disease early warning and treatment system for large-scale pig farms, including a memory and a processor. The memory includes a method program for disease early warning and treatment for large-scale pig farms. When the method program for disease early warning and treatment for large-scale pig farms is executed by the processor, it implements the steps of any of the methods for disease early warning and treatment for large-scale pig farms described in the present invention.

[0016] This invention addresses the shortcomings of the prior art and has the following beneficial effects: (1) Most existing swine disease early warning schemes end after outputting risk results or diagnostic suggestions, making it difficult to determine whether the early warning and diagnosis are consistent with the actual disease condition. This invention unifies and links the initial early warning record, initial diagnosis result, actual on-site treatment execution information, re-examination results, and laboratory test results of the target pigs to form a verification data chain for individual cases. This provides a basis for subsequent verification of early warning and diagnosis results, which helps to reduce the accumulation of misjudgments caused by the inability to verify the model output alone.

[0017] (2) This invention integrates multimodal data acquisition, on-site personnel diagnosis and treatment and laboratory testing. It can introduce laboratory diagnostic results for high-risk cases, cases with unclear diagnoses and cases with poor treatment results, and compare the laboratory diagnostic results with the initial diagnostic results to provide a basis for subsequent rule correction and improve the precision of diagnostic result evaluation.

[0018] (3) This invention not only records the system-recommended treatment plan, but also records the actual implementation plan, drug administration process and post-treatment status changes on site, and correlates them with the final efficacy results. This enables the identification of better treatment pathways for specific varieties, specific ages and specific disease stages. This mechanism helps to reduce the mismatch problem caused by empirical uniform drug use and enhances the pertinence and feasibility of treatment recommendations.

[0019] (4) Based on diagnostic consistency, treatment matching degree and efficacy results, the present invention updates the risk judgment rules, feature weights, early warning thresholds, disease stage judgment rules and treatment plan recommendation priorities, and writes the update results back to the historical case knowledge base, so that the system has the ability to continuously correct and optimize for subsequent similar cases. Compared with the scheme that only performs multimodal early warning, the technical effect of the present invention is not only reflected in one-time identification, but also in the self-correction ability in subsequent cases.

[0020] (5) This invention incorporates early warning, initial diagnosis, on-site treatment, re-examination and testing, laboratory verification and strategy update into the same case event chain, reducing information loss caused by the separation of early warning, diagnosis and treatment and verification, which is conducive to improving the standardization, traceability and consistency of disease management in large-scale farms. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained from these drawings without creative effort.

[0022] Figure 1 A flowchart illustrating the overall process of early disease warning and treatment for large-scale pig farms is provided. Figure 2 A system block diagram of an early disease warning and diagnosis system for large-scale pig farms is shown. Detailed Implementation

[0023] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0024] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0025] like Figure 1 As shown, this invention provides a method for early disease warning and diagnosis in large-scale pig farms, including the following steps: Step 1: Collect multimodal data information of pigs and extract abnormal features from the multimodal data information of pigs; Step 2: Obtain the typical and atypical features of different diseases in the early, middle and late stages, and construct a knowledge graph based on the typical and atypical features of different diseases in the early, middle and late stages; Step 3: Utilize the knowledge graph and abnormal features to conduct a comprehensive disease risk assessment, determine the target disease type and current disease stage of the target pigs, and generate early warning information based on the target disease type and current disease stage of the target pigs. Step 4: Generate an initial treatment plan based on the warning information, disease type judgment, and disease stage judgment, and perform on-site diagnosis and treatment on the target pigs and collect and send samples for testing to form diagnosis and treatment verification data; Step 5: Perform correlation analysis on the disease early warning structure, diagnosis structure, and treatment structure based on the diagnostic and treatment verification data.

[0026] The process of acquiring multimodal data and extracting abnormal features of pigs in step 1 is as follows: For target pigs in the farm, multimodal data related to their health status were collected. This included behavioral image data, body surface temperature data, feed and water intake data, environmental parameter data, and farming management data, all acquired through visible light cameras, infrared thermal imagers, feed and water metering devices, and environmental monitoring sensors. The data was synchronized over time and correlated with individual pigs. The Faster R-CNN algorithm, used in target monitoring and pose estimation networks, was used to extract abnormal features for early disease identification, including abnormal behaviors such as vomiting, lameness, kicking, and abnormal agitation; abnormal temperature changes in key areas such as the head, ears, abdomen, and joints; and time series models such as LSTM were used to process environmental parameters, abnormal lying-standing frequency, decreased feed and water intake, abnormal lesions, and information related to disease outbreaks within the same pen.

[0027] Specifically, LSTM was used to process environmental parameters, abnormal frequency of lying-standing transitions, decreased food and water intake, abnormal external lesions, and information related to morbidity in the same column. Because environmental parameters, shifts between lying and standing positions, and external lesions have different temporal granularities, traditional direct splicing can lead to information misalignment. By introducing an event-driven temporal alignment mechanism, continuous environmental parameters are aligned to the timestamps of feeding behavior events through piecewise linear interpolation. LSTM is used to calculate the Markov transition entropy between lying and standing positions to quantify the degree of agitation, thus enabling earlier detection of limping or neurological symptoms than simply counting frequencies.

[0028] Physiological prior constraints are added to the loss function of LSTM, and the weight coefficients of environmental temperature and feed intake are constrained to conform to the known physiological heat stress model. Each individual, water line, feed line and pen are treated as nodes. It should be noted that the weighting coefficients for the constraints of increased ambient temperature and decreased feed intake must conform to known physiological heat stress models, such as a negative correlation between feed intake and temperature after the temperature exceeds a threshold. If the model attempts to learn a path that violates physiological logic, such as increased feed intake due to high temperature, the loss function will penalize it, thereby improving the model's generalization ability and interpretability.

[0029] Individuals in the same column are fully connected, and individuals sharing the same water tank / feed tank are strongly connected. Based on the ventilation direction and column spacing, CFD simulation is used to simplify the connection into dynamic weights, giving downwind individuals a higher connection weight. The spread of pathogens within a column is simulated using gated graph convolution, and a population risk penetration vector is output. This population risk penetration vector is then used to predict the probability that an individual will be affected by the same column within a preset time. When the probability that an individual will be affected by the same column is greater than the preset probability value, it indicates that there is a correlation between the onset of disease in the same column, thus forming correlation information for onset of disease in the same column.

[0030] It should be noted that this solution breaks the closed-loop structure of LSTM by introducing heterogeneous graph convolution as an external regulator of LSTM cell states. This transforms the "in-pen disease correlation" from a simple statistical feature into a structural feature with spatial propagation dynamics, a design not typically employed by those skilled in the art in the field of intelligent livestock farming. It not only processes in-pen disease correlation information but also resolves the spurious temporal interference of environmental and physiological parameters through physical information constraints. By constructing resource competition edges and aerodynamic edges, the implicit propagation paths of ventilation and shared feeding in the farming site are explicitly encoded as adjacency matrices of a graph network, overcoming the limitation of traditional LSTM in handling the disordered arrangement of individuals in the same pen. Furthermore, this solution, through a group risk penetration mechanism, can achieve latent period early warning through gating reset of cell states before the appearance of individual clinical symptoms. This qualitative leap from symptom monitoring to propagation chain early warning constitutes a substantial technological advancement.

[0031] To achieve the above objectives, this invention provides a closed-loop optimization method for early warning and diagnosis of swine diseases, comprising the following steps: 1. Multimodal data acquisition and anomaly feature extraction of pigs For target pigs in the farm, multimodal data related to their health status were collected. This included behavioral image data, body surface temperature data, feed and water intake data, environmental parameter data, and husbandry management data (such as feed feeding data and vaccination data) collected through visible light cameras, infrared thermal imagers, feed and water metering devices, and environmental monitoring sensors. The data was synchronized over time and correlated with individual pigs. The Faster R-CNN algorithm and pose estimation network were used to extract abnormal features for early disease identification, including abnormal behaviors such as vomiting, lameness, kicking, and abnormal agitation; abnormal temperature changes in key areas such as the head, ears, abdomen, and joints; and time series models such as LSTM were used to process environmental parameters, abnormal lying-standing frequency, decreased feed and water intake, abnormal lesions, and disease correlation information within the same pen.

[0032] Step 2 specifically includes: Based on the multidimensional health status data of the target pigs obtained in Step 1, disease type identification, disease stage determination, and initial treatment plan generation are performed for the target pigs. Specifically, this includes: The abnormal behavioral characteristics, local temperature abnormalities, appearance abnormalities, feed and water intake abnormalities of the target pigs, as well as environmental and management context information, are matched with a pre-set database of full-cycle characteristics of swine diseases. This database is constructed based on a knowledge graph and includes at least typical and atypical characteristics of different diseases at early, middle, and late stages, along with their associated weights. A comprehensive disease risk assessment is conducted by combining the target pig's breed, age, weight, historical disease incidence, disease incidence in the same or adjacent pens, seasonal climate factors, indoor environmental conditions, and management events to determine the target disease type and current disease stage of the target pig. When the comprehensive risk score reaches a pre-set warning threshold, a warning message corresponding to the target pig is generated.

[0033] Specifically, in acquiring the typical and atypical features corresponding to the early, middle, and late stages of different diseases, and constructing a knowledge graph based on these typical and atypical features, the process involves: To address the issue of atypical features being easily overwhelmed by noise in a flat Euclidean space, this approach employs a hyperbolic graph convolutional network for knowledge graph embedding. Typical features are clustered in shallow regions near the root node along an exponentially growing divergent path within the hyperbolic space, while atypical features are mapped to edge regions far from the root node. When a feature is atypical in the early stages of a disease but becomes typical in later stages, the graph captures its centripetal motion within the hyperbolic space. This trajectory constitutes the subtype evolution path of the disease, which traditional graphs cannot represent.

[0034] A neuron-like ordinary differential equation is introduced, and this equation is used to simulate the state changes of typical and atypical features to form a feature set and construct a time-aware knowledge graph reasoning model. It should be noted that if a patient's characteristics fall within the typical area but the time label shows an early stage, it is considered to be rapidly progressing; if it falls on the edge of atypical, an atypical risk alert is triggered, indicating that there may be a misdiagnosis and that the disease is progressing in a non-standard way.

[0035] In the time-aware knowledge graph reasoning model, the current feature set of the patient is input, and the position of the feature set in hyperbolic space is calculated through the attention mechanism of the graph. The warning is then issued based on the position of the feature set in hyperbolic space. The training conditional variational autoencoder generates possible atypical features based on disease and stage. When the graph lacks data on preset late-stage atypical features of the disease, the time-aware knowledge graph reasoning model uses counterfactual reasoning to generate an evolution path graph of atypical features.

[0036] In this step, addressing the challenge of utilizing atypical features, the hierarchical nature of hyperbolic geometry is used to forcibly separate typical and atypical features spatially. This makes the knowledge graph not only a knowledge base but also an anomaly detector, enabling it to automatically identify correct features appearing at the wrong stage or incorrect features appearing at the right stage.

[0037] It should be noted that typical characteristics are defined as those that occur at a significantly higher frequency in the diseased population than in the non-diseased population (OR value > threshold) at a specific stage and have high consensus. Atypical characteristics are defined as those that occur at a low frequency (< 5%) at a specific stage, but have a weak or occult association with disease progression in terms of clinical pathophysiological mechanisms.

[0038] After determining the target disease type and current disease stage, a structured initial diagnosis is generated based on the breed, age, and weight of the target pigs. This structured initial diagnosis includes at least the target disease type, disease stage, diagnostic confidence level, and the initial treatment plan corresponding to the current disease stage. The treatment plan includes recommended drug types, dosage ranges, administration methods, frequency of administration, administration cycle, and information indicating whether isolation observation, follow-up examinations, and laboratory testing are required.

[0039] Step 3 specifically includes: After generating early warning information, disease type judgment, disease stage judgment, and initial treatment plan in step S2, on-site diagnosis and treatment are performed on the target pigs, and samples are collected and sent for testing to form diagnosis and treatment verification data, specifically including: Upon receiving the warning information, on-site personnel arrive at the pen where the target pig is located, manually verify the pig's condition, and implement on-site diagnosis and treatment based on the target disease type, disease stage, and initial treatment plan output in step 2, uploading a diagnosis and treatment report. Simultaneously, on-site personnel collect biological samples from the target pig and send them to the laboratory for testing. Biological samples include at least one or more of the following: blood samples, fecal samples, and secretion samples. Blood samples can be used for routine blood tests, biochemical indicator tests, inflammatory marker tests, or etiological tests.

[0040] The system records the sample collection time, sample type, delivery time, delivery items, and testing institution information, and uniquely identifies and binds these information with the target pig's early warning information, initial diagnosis results, and actual treatment execution information.

[0041] During and after on-site diagnosis and treatment and sample delivery, the system continuously collects information on the status changes of the target pigs. The status change information includes at least the recovery of feed intake, recovery of water intake, changes in body surface temperature, improvement of behavior, persistence of symptoms, worsening of condition or death. At the same time, on-site personnel will conduct follow-up treatment and observe the status of the target pigs at regular intervals and upload subsequent reports.

[0042] The system receives laboratory test reports, extracts etiological results, blood or biochemical test results, confirmed disease type, test time and test conclusions, and links them with the warning information, initial diagnosis results, actual treatment execution information and re-examination results corresponding to the target pigs to form treatment verification data.

[0043] In step 4, based on the case validation data generated in step 3, a correlation analysis is performed on the disease warning structure, diagnostic structure, and treatment structure. The recommended treatment plan, the actual on-site implementation plan, and the final diagnosis and treatment structure are compared to obtain the diagnosis-treatment matching degree evaluation results. The disease type judgment, feature weights, warning thresholds, disease stage determination criteria, and treatment plan recommendation priorities are optimized and updated. The updated model weights and the current case data are written into the case database for use in subsequent warnings, diagnoses, and treatment recommendations for similar cases, achieving closed-loop optimization of early warning, on-site diagnosis and treatment, follow-up testing, laboratory validation, and strategy updates.

[0044] In summary, most existing swine disease early warning systems end after outputting risk results or diagnostic suggestions, making it difficult to determine whether the early warning and diagnosis are consistent with the actual disease condition. This invention unifies and links the initial early warning record, initial diagnostic result, actual on-site treatment execution information, re-examination results, and laboratory test results for the target pigs, forming a verification data chain for individual cases. This provides a basis for subsequent verification of early warning and diagnostic results, reducing the accumulation of misjudgments caused by the inability to verify model outputs alone. Furthermore, this invention unifies multimodal data collection, on-site personnel treatment, and laboratory testing, enabling the introduction of laboratory confirmation results for high-risk cases, cases with unclear diagnoses, and cases with poor treatment outcomes. The consistency of laboratory confirmation results with initial diagnostic results is compared, providing a basis for subsequent rule revisions and improving the precision of diagnostic result evaluation. Moreover, this invention not only records the system-recommended treatment plan but also the actual on-site implementation plan, drug administration process, and post-treatment status changes, and correlates these with the final efficacy results. This allows for the identification of better treatment pathways for specific breeds, ages, and disease stages. This mechanism helps reduce the mismatch problem caused by empirically standardized medication, enhancing the specificity and feasibility of treatment recommendations. Furthermore, based on diagnostic consistency, treatment matching, and efficacy results, this invention updates risk assessment rules, feature weights, early warning thresholds, disease stage determination rules, and treatment plan recommendation priorities, and writes the updated results back to the historical case knowledge base, thereby enabling the system to continuously correct and optimize for subsequent similar cases. Compared with solutions that only perform multimodal early warning, the technical effect of this invention is not only reflected in one-time identification, but also in its self-correction capability in subsequent cases. Finally, this invention incorporates early warning, initial diagnosis, on-site treatment, follow-up testing, laboratory verification, and strategy updates into the same case event chain, reducing information loss caused by the separation of early warning, diagnosis, and verification, and is conducive to improving the standardization, traceability, and consistency of disease management in large-scale farms.

[0045] In addition, a comprehensive score can be obtained by combining the breed, age, weight, historical disease incidence, disease incidence in the same or adjacent pens, seasonal climate factors, indoor environmental conditions and breeding management events, the target disease type and current disease stage of the target pigs using techniques such as hierarchical analysis and grey relational analysis.

[0046] In addition, this method also includes: A discrete spatiotemporal propagation kernel based on the SEIR model is established. For airborne diseases, a simplified alternative to the indoor airflow simulation model is introduced. Indoor temperature, humidity, wind speed, and ventilation mode are used to construct an aerosol propagation intensity field and dynamically correct the spatial adjacency relationship. The discrete spatiotemporal propagation kernel based on the SEIR model is used to simulate the propagation probability of the target disease on a pen-level grid, and to obtain the disease stage, estimated viral shedding, and contact frequency of pigs in the same pen. The cumulative infection pressure of the pigs in the same pen is calculated based on the disease stage, estimated viral shedding, contact frequency, and transmission probability of the target disease. An early warning model and a clinical progression model were constructed. The disease stage, estimated viral shedding, contact frequency, and transmission probability of the target disease in pigs in the same pen were used as model inputs, and the cumulative infection pressure of pigs in the same pen was used as model output. The early warning model and the clinical progression model were trained. For the incubation period / prodromal period, the focus was on historical onset, risk of transmission within the same column, and environmental factors. The cumulative infection pressure value risk index of the early warning model was used to output the data. For the onset period / recovery period, the focus was on individual clinical manifestations, duration of illness, and risk of complications. The cumulative infection pressure value severe illness risk index of the clinical progression model was used to output the data.

[0047] It should be noted that for airborne diseases (such as porcine reproductive and respiratory syndrome and swine influenza), a simplified alternative, an indoor airflow simulation model, is introduced. This model utilizes indoor temperature, humidity, wind speed, and ventilation patterns to construct an aerosol transmission intensity field, dynamically adjusting spatial adjacency relationships: even if pens are not adjacent, if a pen is downstream of the airflow, the transmission risk weight increases. Furthermore, the comprehensive score does not set a fixed threshold but dynamically generates tiered intervention recommendations based on the pig's disease stage, the target disease type, and current farming resources. These recommendations may include measures such as enhanced observation, pen relocation and isolation, and immediate sampling and testing, enabling the scoring system to have a closed-loop decision-making capability.

[0048] In addition, the Faster R-CNN target monitoring algorithm and pose estimation network were used to extract vomiting, lameness, kicking, and abnormal agitation behavioral features for early disease identification in pigs, specifically including: HRNet is used as a shared bottom layer, and high-resolution feature maps are output to two branches. A cascaded RPN is connected at the end of HRNet for dense pig scenes. The unique color difference features of the pig's head and buttocks are used as attention priors. A spatial transformer network is introduced as middleware. Based on the pig's head orientation information fed back in real time by the detection branch, the feature map is subjected to affine transformation to correct the side-facing pig to a front-facing view, and then key points are extracted. The two branches interact bidirectionally through the feature pyramid alignment module. The detection branch provides the pose branch with an accurate individual region attention mask, while the pose branch provides the detection branch with a key point heatmap to help distinguish adhered individuals. In the pose estimation network, deformable convolution is introduced. The deformable convolution is used to adaptively adjust the sampling position according to the current pose of the pig and construct the hind leg-abdomen distance constraint. When the pose network detects that the Euclidean distance between the hind hoof key point and the abdominal key point is periodically narrowing and the movement speed exceeds the threshold, the kicking behavior candidate is triggered. The target monitoring algorithm Faster R-CNN and the pose estimation network were used to extract early disease identification features in pigs, including vomiting, lameness, kicking, and abnormal agitation.

[0049] It should be noted that in traditional methods, when Faster R-CNN detection boxes are provided to the pose estimation network, the boxes often contain multiple pigs or a large amount of background noise due to pigs jostling and lying down, leading to severe distortion in pose estimation. By employing HRNet (High Resolution Network) as a shared bottom layer and simultaneously outputting high-resolution feature maps to both branches, a closed loop of detection-guided pose correction and pose-assisted detection segmentation is achieved, significantly improving the individual parsing accuracy in densely populated pig scenes. Furthermore, deformable convolutions are used to solve the keypoint localization problem under non-rigid deformation, providing a physically interpretable feature basis for fine behaviors such as kicking.

[0050] like Figure 2 As shown, the second aspect of the present invention provides a disease early warning and treatment system for large-scale pig farms, including a memory and a processor. The memory includes a method program for disease early warning and treatment for large-scale pig farms. When the method program for disease early warning and treatment for large-scale pig farms is executed by the processor, it implements the steps of any of the methods for disease early warning and treatment for large-scale pig farms.

[0051] Example 2 In the implementation process, the non-infectious intestinal disease of piglets occurring in the nursery of a large-scale pig farm is used as an example to illustrate the implementation process of this invention. This embodiment mainly focuses on scenarios where there is no group outbreak and the optimization of early individual abnormality identification and treatment closed-loop is the primary approach. In this embodiment, each pen in the nursery is equipped with a visible light camera, an infrared thermal imager, an environmental temperature and humidity sensor, a drinking water monitoring module, and a feeding monitoring module. These devices are electrically connected to an edge computing gateway, which in turn communicates with the farm server. The server contains a built-in database of full-cycle characteristics of swine diseases, a case knowledge base, an early warning model module, a staged diagnosis module, and a strategy update module. The system continuously collects data on the behavior, feeding duration, watering frequency, local body surface temperature, abdominal contour changes, and fecal image features of the 36-day-old piglet (P-1027) over the past 6 hours, and simultaneously receives feed change records and environmental parameter records for that pen over the past 24 hours. System analysis revealed that P-1027's feeding time decreased from the daily average of 42 minutes to 16 minutes within the past 6 hours, water intake decreased from 11 times to 5 times, and the frequency of lying down / standing transitions decreased. It also exhibited persistent lying down, occasional abdominal kicking, and loose stool residue around the anus. Infrared thermal imaging showed that the local temperature at the base of its ears and abdomen was 0.8℃ higher than its historical baseline. However, no other pigs in the same pen showed synchronous diarrhea spread. After comparing these characteristics with the disease's full-cycle characteristic database, the system initially provided two candidate results: the first candidate disease was "early stage of bacterial diarrhea in piglets," with a comprehensive score of 0.79; the second candidate disease was "early stage of weaning stress-related intestinal dysfunction," with a comprehensive score of 0.68. Because the system set a relatively high weight for "localized temperature rise + loose stool image" in this identification, the "early warning for bacterial diarrhea" was triggered first. A structured initial diagnostic result was generated on the computer, including: target disease type, early stage determination, diagnostic confidence level, recommended review window, and a three-stage initial treatment plan.

[0052] The system-generated three-stage initial treatment plan is as follows: (1) The first stage is the control stage from 0 to 12 hours after the warning. It is recommended to use oral rehydration salts 80 mL / time, twice a day, montmorillonite powder 0.3 g / kg, twice a day, and enrofloxacin injection 2.5 mg / kg, once a day. (2) The second stage is a 12 to 36-hour follow-up and adjustment stage. If the frequency of diarrhea decreases but the food intake is not restored, continue to replenish fluids and reduce the dosage of antidiarrheal drugs. At the same time, decide whether to maintain antibiotics based on laboratory results. (3) The third stage is a recovery stage of 36 to 72 hours. If the laboratory does not support bacterial infection, the medication in the third stage will be adjusted to a milder intestinal repair and flora regulation program.

[0053] Upon arrival, the personnel first isolated and observed the pig individually. The results were manually verified via a mobile terminal, confirming that the pig was lethargic, mildly dehydrated, and had a palpable abdomen but no obvious signs of group transmission. Subsequently, the personnel administered medication according to the system's first-phase protocol, simultaneously collecting venous blood and fecal samples, which were sent to the laboratory. The system synchronously recorded the actual execution information, including drug name, dosage, administration time, administerer, sample collection time, sample type, and submitted test items, and uniquely linked it to the pig's early warning record.

[0054] Laboratory test results returned after 6 hours. Blood biochemistry results showed mild dehydration and elevated stress indicators, but the abnormal increase in white blood cells was not significant. Fecal pathogen testing did not detect major bacterial enteropathogenic factors, and fecal microscopy results suggested indigestion and gut microbiota imbalance. Based on this, the system determined that this case was more consistent with "early stage of weaning stress-related intestinal dysfunction" than "early stage of bacterial diarrhea." Therefore, the system activated a closed-loop correction mechanism, downgrading the first candidate disease in the original diagnosis, increasing the comprehensive score of "early stage of weaning stress-related intestinal dysfunction" from 0.68 to 0.84, and automatically adjusting the subsequent treatment plan. The revised second-stage regimen is as follows: oral rehydration salts 80 mL / time, twice daily; montmorillonite powder 0.15 g / kg, twice daily; enrofloxacin discontinued; probiotic preparation 1 g / head, twice daily added. The revised third-stage regimen is as follows: probiotic preparation 1 g / head, twice daily; electrolyte multivitamins 0.5 g / head, once daily, for 2 consecutive days, as a gentler restorative treatment plan.

[0055] The system continuously monitored the pig's condition at three follow-up points: 24 hours, 48 ​​hours, and 72 hours. Results showed that within 12 hours of the start of the second phase of treatment, the pig's water intake returned to 8 times per day, and the frequency of diarrhea significantly decreased. At the 48-hour mark, its feed intake recovered to 82% of its daily average, the local temperature difference in its abdomen fell below 0.2℃, and the kicking behavior disappeared. At the 72-hour mark, its fecal morphology returned to near normal, and the system determined the treatment effect to be "good recovery." Therefore, the system added this case to the historical case database and marked it in the case label as a typical sample "initially misjudged but corrected by laboratory verification."

[0056] More importantly, the system performed two types of optimizations based on this case: First, it optimized the early warning identification rules, namely, in scenarios such as "within 24 hours after feed change, no spread in the piglet population, local temperature rise of less than 1°C, and symptoms mainly of sudden drop in feed intake and kicking," it reduced the weight of "early stage of bacterial diarrhea" and increased the weight of "early stage of weaning stress-related intestinal dysfunction." Second, it optimized the treatment strategy, namely, for piglets aged 30 to 45 days, weighing 10 to 15 kg, and in the early stage of intestinal dysfunction, it prioritized the three-stage treatment plan of "fluid resuscitation + adsorption antidiarrheal + probiotic regulation," and only retained antimicrobial drugs as the preferred option when the etiological results supported bacterial infection. Subsequently, when another 38-day-old piglet in the same pen showed similar symptoms, the system directly listed "early stage of weaning stress-related intestinal dysfunction" as the first candidate disease, increasing the diagnostic confidence to 0.83, and no longer defaulted to setting the third stage of treatment to antimicrobial maintenance, thus demonstrating the closed-loop effect of this invention in optimizing early warning accuracy and medication moderation.

[0057] In another implementation, taking a scenario of a highly contagious viral febrile syndrome occurring in a fattening shed as an example, the process of implementing the present invention in the context of group-based, rapidly spreading diseases is illustrated. This embodiment mainly demonstrates the application of the present invention in group early warning, isolation and treatment, laboratory confirmation, age- and disease-stage stratified treatment, and group rule optimization.

[0058] Example 3 In this embodiment, each pen in the fattening house is equipped with a visible light camera, an infrared thermal imaging device, an audio collector, a water meter, and an environmental monitoring sensor. All front-end devices are electrically connected to an edge gateway, which in turn is connected to the farm server. During continuous monitoring, the system detected that three pigs aged 90 to 108 days in pens 6 to 8 of fattening house A exhibited obvious abnormalities: rapid increase in localized body surface temperature, simultaneous decrease in feed and water intake, prolonged lying near the ventilation opening, lethargy, reluctance to get up, accompanied by short-term screaming and abnormal respiratory rhythm. Since the first cases initially presented with high fever and decreased appetite, and there was no clear evidence of group spread, the system initially classified the first cases as "severe febrile inflammatory syndrome" and issued an individual red alert and a housewide orange alert.

[0059] Upon arrival, on-site personnel manually verified the first three sick pigs, confirming they exhibited persistent high fever, dry nasal mucus, lethargy, and significant refusal to drink. The system's initial treatment measures were then implemented: the abnormal pigs were isolated, a buffer zone was established between adjacent pens, and personnel access routes, equipment, and floors were disinfected for the first time. Simultaneously, on-site personnel collected blood samples, nasal and oral swab samples, and environmental swab samples for testing. The system spatially and temporally linked the on-site isolation time, disinfection time, personnel entry time, sample collection time, and the pigs targeted in the initial treatment.

[0060] In this embodiment, the initial unified treatment plan provided by the system for the first batch of cases is divided into three stages: (1) The first stage is the emergency control stage of 0 to 12 hours. All pigs with abnormal high fever are given flunixin meglumine 1.1 mg / kg once a day to reduce fever and relieve inflammatory response. At the same time, compound electrolyte solution 150 mL / time, twice a day is given for fluid replacement. For cases with the risk of secondary bacterial infection, florfenicol 15 mg / kg is given at the same time, once every 48 hours. (2) The second stage is the disease differentiation stage of 12 to 36 hours. For pigs with persistent high fever but still able to stand and drink a small amount of water, increase the immune support preparation by 0.2 mL / kg once a day and continue to replenish fluids. (3) The third stage is the outcome observation stage of 36 to 72 hours. Pigs whose body temperature drops and whose feed intake recovers are transferred to the maintenance program and given 0.5g / head of electrolyte multivitamins once a day. Pigs with persistent high fever, lying down and unable to get up and with poor peripheral circulation are marked as high-risk late cases, and the priority of repeated administration is reduced, while the priority of isolation, disinfection and protection of surrounding pig herds is increased.

[0061] The system also provides review windows for 6 hours, 12 hours, 24 hours, and 48 hours.

[0062] After the laboratory test results returned, it showed that the first batch of samples tested positive for a highly infectious viral factor, while common bacterial septicemia indicators did not support this as the primary cause. The system thus confirmed that this scenario was not ordinary sporadic bacterial fever, but rather the early stage of a highly infectious viral febrile syndrome outbreak. The following strategies were automatically triggered: First, pigs in the same and adjacent pens were stratified and managed according to age, contact history, and fever status; second, inter-pens transfers and mixed-group operations were suspended; third, the frequency of disinfection was increased, and pigs exhibiting prolonged lying down near ventilation openings, persistent high fever, and a continuous presence of pigs refusing to drink water within the group were identified as key group characteristics.

[0063] Over the following 48 hours, the system tracked and found significant differences in efficacy among individuals under the same medication conditions: early cases, aged 90 to 95 days, identified and isolated within 12 hours of onset, showed a decrease in body temperature within 24 hours and resumed small amounts of feed within 48 hours after receiving antipyretics, fluid replacement, and immune support; while late cases, aged 105 to 108 days, who had a prolonged period of high fever before onset and were found to be bedridden for an extended period, responded poorly to the same treatment plan, with two dying within 48 hours. The system correlated this difference with age, disease stage, duration of initial symptoms, and the initial decrease in feed intake. The disease stage was found to be a greater determinant of prognosis and medication effectiveness than fever alone, and late cases should not be treated with the same treatment plan as early cases.

[0064] Based on the above results, the system completed two layers of closed-loop optimization. The first layer is the optimization of the group early warning rules: the spatiotemporal combination of "at least 3 pigs with high fever appearing consecutively in the same pen within 2 hours, all of which are not drinking water, lying down near the ventilation opening, and showing depression" is set as a strong trigger condition for highly infectious diseases; at the same time, two group parameters are added: "interval time between onset in adjacent pens" and "proportion of pigs of the same age with concentrated abnormalities". The second layer is the optimization of stratified treatment and management strategies: for early cases, "antipyretics + fluid replacement + immune support + control of conditional secondary infections" is recommended first; for mid-stage cases, the frequency of follow-up examinations and monitoring of secondary infection risks are increased; for late-stage high-risk cases, resources are prioritized for isolation, prevention of spread, and protection of healthy pigs in the same pen, instead of mechanically continuing the same uniform dosing regimen.

[0065] After completing this closed-loop learning process, the system writes the laboratory confirmation results, mortality outcomes, recovery outcomes, age differences, isolation timing, and treatment response of the first batch of cases into the historical case database. Two days later, when only two pigs with mild high fever and one pig that did not yet show obvious fever but exhibited decreased feed intake, reduced water consumption, and abnormal lying down near the ventilation opening appeared in the adjacent fattening pens in Building B, the system directly issued an "early group warning for highly infectious viral febrile syndrome" based on the updated group rules, without waiting for widespread high fever before issuing an alarm. This demonstrates that the present invention not only achieves a closed loop of isolation, testing, verification, stratified treatment, and strategy updates after an outbreak, but also enables the system to reverse-engineer the real diagnosis and confirmation results from an outbreak for subsequent scenarios, allowing the system to identify, isolate, and allocate treatment resources more rationally in the next similar outbreak.

[0066] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0067] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0068] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0069] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0070] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0071] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A disease early warning and diagnosis method for large-scale pig farms, characterized in that, Includes the following steps: Collect multimodal data information of pigs and extract abnormal features from the multimodal data information of pigs; Obtain typical and atypical features of different diseases in the early, middle and late stages, and construct a knowledge graph based on the typical and atypical features of different diseases in the early, middle and late stages. The knowledge graph and abnormal features are used to conduct a comprehensive disease risk assessment, determine the target disease type and current disease stage of the target pigs, and generate early warning information based on the target disease type and current disease stage of the target pigs. Based on the aforementioned early warning information, disease type judgment, and disease stage judgment, an initial treatment plan is generated, and on-site diagnosis and treatment are performed on the target pigs, and samples are collected and sent for testing to form diagnosis and treatment verification data; Based on the aforementioned diagnostic and treatment verification data, a correlation analysis was performed on the disease early warning results, diagnostic results, and treatment results; Multimodal data of pigs is collected, and abnormal features are extracted from the multimodal data of the pigs, specifically as follows: Behavioral image data, body surface temperature data, feed and water intake data, environmental parameter data, and aquaculture management data are collected through visible light cameras, infrared thermal imagers, feed and water intake metering devices, and environmental monitoring sensors. The behavioral image data, body surface temperature data, feeding and drinking data, environmental parameter data, and aquaculture management data are synchronized over time and associated with individuals. The Faster R-CNN object detection algorithm and pose estimation network were used to extract vomiting, limping, kicking, and abnormal agitation behavioral features for early disease identification. LSTM was used to process environmental parameters, abnormal frequency of lying-standing transitions, decreased food and water intake, abnormal appearance of lesions, and information related to the incidence of diseases in the same column. LSTM was used to process environmental parameters, abnormal frequency of lying-standing transitions, decreased food and water intake, abnormal physical lesions, and correlation information of morbidity within the same column. Specifically: An event-driven temporal alignment mechanism is introduced to align continuous environmental parameters to the timestamps of feeding behavior events through piecewise linear interpolation. LSTM is used to calculate the Markov transition entropy of the lying and standing positions to quantify the degree of agitation. Physiological prior constraints are added to the loss function of LSTM, and the weight coefficients of environmental temperature and feed intake are constrained to conform to the known physiological heat stress model. Each individual, water line, feed line and pen are treated as nodes. Individuals in the same column are fully connected, and individuals sharing the same water tank / feed tank are strongly connected. Based on the ventilation direction and column spacing, CFD simulation is used to simplify the connection into dynamic weights, giving downwind individuals a higher connection weight. The spread of pathogens within a column is simulated using gated graph convolution, and a population risk penetration vector is output. This population risk penetration vector is then used to predict the probability that an individual will be affected by the same column within a preset time period.

2. The method for early disease warning and diagnosis in large-scale pig farms according to claim 1, characterized in that, The typical and atypical features of different diseases in their early, middle, and late stages are obtained. A knowledge graph is then constructed based on these typical and atypical features. Specifically: Typical features are clustered in the shallow region near the root node along an exponentially growing divergent path in hyperbolic space, while atypical features are mapped to the edge region far from the root node. A neuron-like ordinary differential equation is introduced, and this equation is used to simulate the state changes of typical and atypical features to form a feature set and construct a time-aware knowledge graph reasoning model. In the time-aware knowledge graph reasoning model, the current feature set of the pig is input, and the position of the feature set in hyperbolic space is calculated through the attention mechanism of the graph. Warning is then given based on the position of the feature set in hyperbolic space. The training conditional variational autoencoder generates possible atypical features based on disease and stage. When the graph lacks data on preset late-stage atypical features of the disease, the time-aware knowledge graph reasoning model uses counterfactual reasoning to generate an evolution path graph of atypical features.

3. The method for early disease warning and diagnosis in large-scale pig farms according to claim 1, characterized in that, Using the knowledge graph and abnormal features, a comprehensive disease risk assessment is performed to determine the target disease type and current disease stage of the target pigs. Based on the target disease type and current disease stage, an early warning is generated, specifically as follows: The abnormal features are input into the knowledge graph for knowledge reasoning to determine the target disease type and current disease stage of the target pig. Obtain information on the breed, age, weight, historical disease incidence, disease incidence in the same or adjacent pens, seasonal and climatic factors, indoor environmental conditions, and breeding management events of the target pigs. A comprehensive score is calculated based on the breed, age, weight, historical disease incidence, disease incidence in the same or adjacent pens, seasonal and climatic factors, indoor environmental conditions and management events, the target disease type of the target pig, and its current disease stage. A comprehensive risk score is obtained through comprehensive evaluation. When the comprehensive risk score reaches a preset warning threshold, a warning message corresponding to the target pig is generated.

4. The method for early disease warning and diagnosis in large-scale pig farms according to claim 3, characterized in that, An initial treatment plan is generated based on the aforementioned warning information, disease type determination, and disease stage determination, specifically including: Based on the aforementioned warning information, disease type judgment, disease stage, and combined with the breed, age, and weight of the target pigs, a structured initial diagnosis result is output. The diagnostic results include the target disease type, disease stage, diagnostic confidence level, and the initial treatment plan corresponding to the current disease stage. The treatment plan includes recommended drug types, drug dosage ranges, administration methods, administration frequency, administration cycle, and information on whether isolation observation, follow-up examinations, and laboratory testing are required.

5. The method for early disease warning and diagnosis in large-scale pig farms according to claim 1, characterized in that, On-site diagnosis and treatment were performed on the target pigs, and samples were collected and sent for testing to generate diagnostic and treatment validation data, specifically: Upon receiving the warning information, the system arrives at the pen where the target pig is located, performs manual verification of the target pig, and conducts on-site diagnosis and treatment based on the output target disease type, disease stage, and initial treatment plan, and uploads the diagnosis and treatment report. Biological samples were collected from the target pigs and sent to the laboratory for testing. The biological samples included at least one or more of the following: blood samples, fecal samples, and secretion samples. Record the sample collection time, sample type, delivery time, delivery items and testing institution information, and uniquely link them with the target pig's early warning information, initial diagnosis results and actual treatment execution information; Receive laboratory test reports, extract etiological results, blood or biochemical test results, confirmed disease type, test time and test conclusions, and link them with the warning information, initial diagnosis results, actual treatment implementation information and re-examination results corresponding to the target pigs to form treatment verification data.

6. The method for early disease warning and diagnosis in large-scale pig farms according to claim 5, characterized in that, Based on the aforementioned diagnostic and treatment verification data, a correlation analysis is performed on the disease early warning results, diagnostic results, and treatment results, specifically as follows: A correlation analysis was performed on the disease early warning results, diagnostic results, and treatment results of the aforementioned diagnostic and treatment verification data; The disease type identification, feature weights, early warning thresholds, disease stage determination criteria, and treatment plan recommendation priorities were optimized and updated. The updated model weights and the current case data were then written into the case database.

7. A disease early warning and diagnosis system for large-scale pig farms, characterized in that, The system includes a memory and a processor. The memory stores a program for early disease warning and diagnosis methods for large-scale pig farms. When the processor executes the program for early disease warning and diagnosis methods for large-scale pig farms, it implements the steps of the method for early disease warning and diagnosis methods for large-scale pig farms as described in any one of claims 1-6.

Images