An AI-based early warning system for laying hen diseases
By combining the evidence attribution module, the state inference module, and the attribution verification module, the problem of unclear sources of health observation in early warning of diseases in laying hens is solved, and the effective use of health data and the accuracy of disease risk assessment are realized.
Patent Information
- Application Number
- CN202611627637.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-21
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies cannot effectively handle situations where the source of health observations is unclear in early warning of diseases in laying hens, resulting in discontinuous health data recording, which affects the accuracy of disease risk assessment and the continuity of individual health status.
The system uses an evidence attribution module to establish candidate attribution relationships, a state deduction module to infer the range of health status and disease risk results, and an attribution verification module to verify and eliminate unreasonable candidate attribution relationships. Finally, the risk warning module outputs a disease warning.
This ensures the usability of health data even when the source of health observation is unclear, avoids unconfirmed data from affecting health records, and improves the accuracy of disease risk assessment and the continuity of individual health status.
Smart Images

Figure CN123120329A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of intelligent processing technology for animal health information, and in particular to an AI early warning system for diseases in laying hens. Background Technology
[0002] Early warning of diseases in laying hens involves the continuous collection of animal health information, assessment of individual health status, and prediction of disease risk. It falls under the application field of using information and communication technology to process health data. With the application of health monitoring equipment and intelligent analysis technology in large-scale farming, existing systems can establish individual health records based on health information such as body temperature changes, feed intake, and activity status of laying hens during continuous observation periods. The system can also identify the trend of health status changes relative to the existing baseline through machine learning models. When the continuous health status of a laying hen shows changes that are incompatible with the existing health pattern, the system can further calculate its disease risk to assist in early health management.
[0003] The above processing relies on a fundamental condition: that continuously obtained health observations can be stably associated with the same laying hen. In actual continuous monitoring, this condition may sometimes temporarily fail. For example, if two adjacent laying hens obstruct each other or their positions overlap within a short period of time, the data collection system can confirm that a certain abnormal body temperature originated from one of the laying hens, but it cannot immediately determine the specific source. If one laying hen has previously shown slight changes in its health status, while the other remains stable, then attributing the same health observation to two different laying hens will have different impacts on subsequent disease risk assessments. Existing systems can complete individual association based on location relationships, identity matching results, or source credibility. Such methods are beneficial for maintaining the continuity of health data processing. However, when there is still reasonable ambiguity regarding the source, prematurely forming a unique association result may cause a genuine health observation to be included in the health record of an individual whose actual source is not present.
[0004] This issue is particularly relevant in the early stages of disease, when health changes typically haven't yet formed a clear and stable disease phenotype. A single observation has a significant impact on subsequent health status calculations. Once a health observation from an unconfirmed source is included in a laying hen's historical health trajectory, this record will participate in subsequent health baseline, status evolution, and disease risk calculations. Newly acquired health observations will then be evaluated based on the already changed historical trajectory, potentially leading to the continued propagation of early individual association biases. Conversely, if the health observation is discarded directly when its source is uncertain, the early disease change information it may contain cannot continue to participate in risk analysis. Furthermore, retaining only coarse-grained population risk results reduces the ability of disease risk to converge towards specific individuals.
[0005] Furthermore, a health observation generated by a single laying hen is exclusive among multiple candidate individuals. When the observation is attributed to one laying hen, it also means that other candidate laying hens did not obtain this health evidence at that moment. Therefore, judging the reasonableness of the source is not only related to whether the health status of the individual who obtained the health observation is continuous, but also to whether the subsequent health status formed by other candidate individuals under the condition of not obtaining the health observation can be consistent with the actual observation. The existing processing method based on the results of a single identity matching or the credibility of the association of a single individual still has room for improvement in the utilization of disease risk information formed by the mutual constraints of individual health status.
[0006] Therefore, in the processing of health information for early warning of diseases in laying hens, it is still necessary to study how to maintain the usability of health observation for disease risk assessment when the health observation is real and effective but the individual source cannot be uniquely determined at the moment, while avoiding premature changes to the health trajectory of unconfirmed individuals, and gradually improving the certainty of disease risk results for specific laying hens by combining the health status of the subsequently identified source. Summary of the Invention
[0007] This application proposes an AI-based early warning system for laying hen diseases to address the problems mentioned in the background art.
[0008] To achieve the above objectives, this application adopts the following technical solution: an early warning system for laying hen diseases using AI, comprising an evidence attribution module, a status deduction module, an attribution verification module, and a risk warning module; The evidence attribution module acquires health data to be attributed and a set of candidate laying hens that have been verified as having a single source but lack independent source evidence. It reads the confirmed health status of the candidate laying hens before the formation of the health data to be attributed and establishes candidate attribution relationships with the constraint that the health data to be attributed corresponds to only one candidate laying hen in each candidate attribution relationship. The status inference module only accesses the health data of the candidate laying hens to be assigned under each candidate attribution relationship, and infers the subsequent health status of the candidate laying hens based on the disease health status evolution model, forming the range of allowable health status and disease risk results; The attribution verification module obtains subsequent health data with a clear source, compares the corresponding health status with the range of allowed health statuses under the candidate attribution relationship, and excludes the corresponding candidate attribution relationship when the health status exceeds the range of allowed health statuses. It then connects the subsequent health data with a clear source to the remaining candidate attribution relationships and updates the disease risk results. The risk warning module compares the disease risk results of the same laying hen under each remaining candidate attribution relationship, and outputs disease warnings for laying hens that all meet the disease risk judgment conditions; after obtaining the independent source credentials corresponding to the health data to be assigned, the health data to be assigned is written into the confirmed health trajectory of the corresponding laying hen according to the original collection time.
[0009] The beneficial effects of this invention are as follows: This invention addresses situations where the source of individual health data cannot be uniquely determined temporarily. It establishes candidate attribution relationships for each data point, ensuring that health data to be assigned only participates in health status deduction within the corresponding candidate attribution relationship and remains isolated from the confirmed health trajectories of each candidate laying hen. This restricts data with unconfirmed sources from entering the formal health record prematurely.
[0010] This invention utilizes health data with clearly defined sources to jointly verify the health status of candidate laying hens that have obtained health data to be assigned and those that have not obtained health data to be assigned under each candidate attribution relationship. This eliminates candidate attribution relationships that are incompatible with actual health evolution, thus creating a continuous constraint relationship between the determination of the source of health data and the subsequent changes in the disease status of candidate laying hens.
[0011] This invention further compares the disease risk results of the same laying hen under each remaining candidate attribution relationship, outputs a disease warning when each relationship meets the disease risk judgment conditions, and writes the health data to be assigned to the corresponding laying hen into the confirmed health trajectory of the corresponding laying hen according to the original collection time after obtaining independent source evidence, thereby realizing conditional processing of disease risk warning and individual health record update. Attached Figure Description
[0012] 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 embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort: Figure 1 This is a flowchart illustrating the overall processing of the present invention; Figure 2 This is a flowchart illustrating how the present invention verifies candidate attribution relationships based on subsequently sourced, clearly defined health data and forms remaining candidate attribution relationships. Detailed Implementation
[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0014] like Figure 1 and Figure 2 As shown, this invention discloses an early AI warning system for laying hen diseases, including an evidence attribution module, a status inference module, an attribution verification module, and a risk warning module; the core object of processing is the health data to be attributed that has been verified to be generated by a single laying hen but for which independent source evidence has not yet been obtained.
[0015] The evidence attribution module establishes multiple candidate attribution relationships for the same health data to be attributed; the status inference module infers the subsequent health status of candidate laying hens under each candidate attribution relationship and forms the allowable health status range and disease risk results; the attribution verification module verifies the candidate attribution relationships using health data from subsequent sources and forms the remaining candidate attribution relationships; the risk warning module forms disease warnings based on the disease risk results corresponding to each remaining candidate attribution relationship, and after obtaining independent source evidence, writes the health data to be attributed into the confirmed health trajectory of the corresponding laying hen according to the original collection time.
[0016] In this implementation, the confirmed health status is formed from confirmed health data from an individual source. The predicted health status formed by the disease health status evolution model retains the attributes of the calculation results. The candidate attribution relationship limits the use of the health data to be attributed among candidate laying hens without changing the actual physiological process of the laying hens. The disease risk result represents the model score of the target disease obtained independent confirmation within a fixed disease risk prediction time window. The disease warning is used to output individual health risk information and is not directly used as an independent confirmation conclusion of the target disease. The independent confirmation of the target disease is formed by veterinary diagnostic records that are independent of the disease risk model. The system's health status inference, attribution verification, and disease risk comparison are all performed by the information processing device.
[0017] The evidence attribution module, state inference module, attribution verification module, and risk warning module operate sequentially in the order of health data attribution, state inference, attribution verification, and risk warning. The individual source attributes formed at the collection end serve as the health data input information. The identity recognition technology only plays a role in forming source information. The disease health status evolution model, calibration parameters for the allowable health status range, and disease risk judgment conditions are established and fixed before the system goes online. During online processing, the same version is called to complete the individual health risk calculation. Health data to be attributed is only used within the candidate attribution relationship before obtaining independent source evidence and is kept isolated from the confirmed health trajectory.
[0018] The system continuously receives health data generated by the acquisition terminal. Each health data has a unique record identifier, original acquisition time, receiving time, original sampling number, and sampling support period. The original acquisition time indicates the time when the health event occurred, and the receiving time indicates the time when the record entered the system. Time-summary health data also saves the start time of the statistical window, so that the summary results can be traced back to the original sample that actually participated in the calculation.
[0019] The health data to be assigned retains the original health data values that have not been converted for reference to specific individual laying hens, as well as the valid health dimension markers. When the valid health dimension marker is 1, it means that the corresponding original sample meets the data quality conditions. When it is 0, it means that the corresponding health dimension is missing in the current health data to be assigned. Missing health dimensions remain missing and are not filled with normal values.
[0020] The acquisition end first completes the individual source verification. Taking the individual temperature measurement port as an example, if a valid sampling session has continuous individual occupancy records and the original samples that constitute the temperature health data are all located in the same occupancy period, it is determined that the health data is generated by a single laying hen. If the acquisition end fails to eliminate multiple overlaps or has not yet completed the source quantity verification, the corresponding health data is saved according to the original source status and does not enter the candidate attribution relationship establishment process.
[0021] All valid health dimensions in the same health data to be assigned are traced back to the same individual sampling session, or confirmed by continuous individual occupancy records to come from the same physical subject. Health data with the same collection time, the same collection device, and adjacent collection locations are saved separately and are not merged into one health data to be assigned. In this implementation, the body surface temperature health data formed by the individual temperature measurement session is used as the health data to be assigned, and other health dimensions are mainly formed by individual health data whose source has been confirmed.
[0022] The candidate laying hen set is formed based on the valid individual registration records within the sampling support period of the health data to be assigned. One implementation uses a closed breeding unit with entry and exit registration records. Registered laying hens that are qualified to contact the corresponding collection port within the sampling support period are included in the candidate laying hen set. When there are unregistered individuals, registration interruptions, or failure to verify the integrity of the candidate range, the system retains the original record and stops the automatic assignment and screening of the record. The candidate laying hen set is not truncated based on identity matching scores.
[0023] When the source of an individual hen has been verified and an independent source credential indicating the unique identity of the laying hen has been obtained, the corresponding health data is processed according to the health data of the confirmed source. When the source of an individual hen has been verified, the candidate laying hen set contains at least 2 laying hens and an independent source credential has not yet been obtained, the system establishes a candidate affiliation relationship, and the group environment data is saved according to the group health auxiliary data and is not included in the individual candidate affiliation relationship.
[0024] The system maintains confirmed health trajectories according to the laying hen identification. The confirmed health trajectory saves the original health data of the individual whose source has been independently confirmed and its corresponding confirmed health status. When only one candidate attribution relationship remains after attribution verification, the system saves the unique remaining candidate attribution relationship and its verification record. This record represents the candidate attribution relationship screening result. The independent source evidence is still formed by the collection facts corresponding to the original sampling record.
[0025] One reproducible implementation adopts a 4-dimensional health status. The 4-dimensional structure is determined by the type of health data that can be continuously acquired in the current acquisition system, has individual source confirmation records, and can form an effective prediction error scale in independently calibrated data. The number of this dimension belongs to the model input structure and is not used as a threshold for disease risk determination.
[0026] Individual body surface temperature was measured as the median of temperature samples within the same valid temperature measurement session, in °C; behavioral status was characterized by feeding access frequency and activity intensity; egg production status was characterized by the time interval between two consecutive egg production events that confirmed the source, in h; feeding access frequency was measured in times / h; and activity intensity was measured in m / min.
[0027] The behavioral statistics window length is used as a model version parameter, with an optimal range of 15-60 minutes. Candidate values are set to 15 minutes, 30 minutes, 45 minutes, and 60 minutes in 15-minute increments. For each candidate window, the system calculates the percentage of effective windows in the parameter tuning and verification data that meet the requirements of complete data collection coverage, confirmed original event sources, and the ability to form the corresponding behavioral health dimension, and calculates the corresponding health status prediction bias. Candidate windows with higher effective window percentages are prioritized, and when effective window percentages are the same, candidate windows with lower health status prediction biases are selected.
[0028] Once the window length is determined, the corresponding allowable health status range is established using independent calibration data. The final test data is only used to verify the fixed version. In this implementation, 30 minutes is used as the preferred value, and the same statistical window is maintained during the training, calibration and online processing stages.
[0029] Adjacent windows use a front-open, back-closed boundary. The feeding access frequency is formed by dividing the number of valid access events with confirmed sources by the window duration. The activity intensity is formed by dividing the valid movement distance with confirmed sources by the window duration. When the collection coverage is incomplete or there are original events within the window that have not yet been confirmed by sources and the candidate laying hen set contains the corresponding laying hen, the corresponding health dimension remains missing.
[0030] Data cleaning is performed based on the collection quality. Duplicate reports with the same original sampling number retain only one valid record. Records that exceed the effective range of the equipment, have invalid equipment quality markings, or whose time sequence cannot be verified are kept as invalid health dimensions. Abnormal health data within the effective range of the equipment are retained and are not deleted based on their degree of deviation from the population center state. Equipment quality judgment and target disease monitoring labels are formed separately.
[0031] The system establishes an individual health reference for each laying hen. When establishing an individual health reference, it reads historical health data that has been sourced and independently verified before the reference cutoff time. The independent health verification results are formed by veterinary examination records that are independent of the disease risk model. During the observation period corresponding to the health reference, the veterinary examination records clearly record the hen's health status and save the records of disease detection, immunization and treatment during the same period as the basis for verification. Records of treatment processes for the target disease or health status that cannot be clearly verified are not included in the health reference.
[0032] The preferred range for the individual reference history length is 21-35 days, corresponding to 3-5 complete 7-day cycles; the preferred range for the intraday segment length is 1-4 hours, which is used to maintain the effective sample size in each segment while preserving intraday rhythm differences. The history length uses 7 days as the adjustment step size, and the intraday segments use 1 hour, 2 hours, and 4 hours as candidate values.
[0033] Each candidate combination forms an individual reference based on parameter tuning and verification data with confirmed sources and health verification results. Combinations with higher effective reference coverage are given priority. When the effective reference coverage is the same, combinations with lower health status prediction deviation are selected. In this implementation, 28 days and 2 hours are used as preferred values.
[0034] The preferred range for the effective day coverage ratio is 70%-85%, and this implementation adopts 75%. Therefore, the 28-day reference period requires at least 21 effective day values. This value ensures that the reference data covers four complete 7-day cycles, and that at least 3 / 4 of the effective days form an individual reference. The parameters are only used for health status conversion and do not constitute a condition for disease risk determination.
[0035] For each health dimension, the system arranges the reference samples in ascending order, uses symmetrical quantiles to determine the individual offset scale, and optimizes the range of unilateral tail quantile probability to 0.05-0.2. The central quantile is fixed at 0.5, and the upper quantile is 1 minus the unilateral tail quantile probability. During the model building phase, different tail quantile probabilities are compared with a step size of 0.05. The median absolute change of health status between adjacent reference windows is calculated on parameter tuning and validation data with confirmed sources and health verification results. Only candidate settings that can form non-zero directional scales for each health dimension are retained, and the setting with the smallest median absolute change is selected. When there are ties, the smaller unilateral tail quantile probability is used.
[0036] In this implementation, 0.1 is used as the one-sided tail quantile probability. Therefore, the lower reference value is the 10th quantile, the central reference value is the 50th quantile, and the upper reference value is the 90th quantile. This setting allows the central reference interval to cover 80% of the records in the middle of the reference sample, without requiring the reference sample to meet a normal distribution.
[0037] The health status quantity is calculated according to the following textual logic: The system first subtracts the central reference value of the corresponding health dimension from the current original health data value to obtain a directional offset; when the original health data value is not lower than the central reference value, the difference between the upper reference value and the central reference value is used as the directional scale; when the original health data value is lower than the central reference value, the difference between the central reference value and the lower reference value is used as the directional scale. The system takes the larger value between the directional scale and the minimum resolvable step size of the corresponding health dimension as the normalization scale, then divides the offset by the normalization scale, and performs hyperbolic tangent mapping on the resulting dimensionless result.
[0038] The minimum resolvable step size for the corresponding health dimension is kept in the same unit as the original health data. The minimum resolvable step size for the feeding access frequency is determined by dividing one access by the statistical window duration. The minimum resolvable step size for the egg production time interval is determined by the event clock resolution. The mathematical range of the health status quantity is -1 to 1. The actual storage resolution is determined by the model input data type and the numerical precision at the deployment end and recorded in the model version. When there is no valid data for the health dimension, a missing marker is saved synchronously. The numerical placeholder only serves a storage function and does not represent a normal health status.
[0039] When the health data to be assigned is processed, the system fixes the cutoff time of the individual health reference before the sampling support period of the health data to be assigned, and freezes the individual health reference version of each candidate laying hen. The same frozen reference version is used for historical replay, conversion of health data to be assigned, and subsequent assignment verification. Late health review results are entered into the new health record version according to the actual available time and are not recorded as references that were available before the formation of the health data to be assigned.
[0040] The same health data to be assigned is converted using the frozen individual health reference of each candidate laying hen. The same original body surface temperature corresponds to the individual central reference state of different candidate laying hens, forming a health status input corresponding to the individual reference of each candidate laying hen. When a certain health dimension lacks common conversion conditions within the candidate laying hen set, that health dimension remains missing in all candidate assignment relationships corresponding to the current health data to be assigned.
[0041] The system reads the most recent actual health data of each candidate laying hen that has obtained independent source credentials and contains at least one valid health dimension before the sampling period of the health data to be assigned. It then uses the currently frozen individual health reference to convert it into a confirmed health status. This confirmed health status serves as the starting point for the current candidate assignment relationship of the corresponding candidate laying hen and remains read-only during the current processing period.
[0042] Before the system goes online, an effective transfer time limit is established for the disease health status evolution model. When the time span between the original collection time of the health data to be assigned and the starting point of the status inference of each candidate laying hen is within the effective transfer time limit, the current health data to be assigned enters the automatic candidate assignment process.
[0043] If any candidate laying hen that still has eligibility for sourcing lacks a valid confirmed health status, the system retains the original health data to be assigned and the set of candidate laying hens, and stops the automatic assignment screening of the current record, keeping the set of candidate laying hens intact.
[0044] The system establishes a candidate attribution relationship for each candidate laying hen in the candidate laying hen set. Each candidate attribution relationship only connects the same health data to one candidate laying hen. The remaining candidate laying hens continue to participate in the subsequent health status inference along the health data sequence corresponding to their respective confirmed health trajectories. The same health data to be assigned is only assigned once in a candidate attribution relationship. If the health data to be assigned is not connected, it means that the candidate laying hen has not obtained the observation in the corresponding candidate attribution relationship. The health status of the candidate laying hen is still determined by its actual confirmed health data.
[0045] After the candidate attribution relationship is established, the system will permanently associate the health data to be assigned, the set of candidate laying hens, each candidate attribution relationship, and the confirmed health status of each candidate laying hen. The status inference module will only call the corresponding health data within each candidate attribution relationship. The source confirmation result will remain unchanged, and the provisional data in the candidate attribution relationship will remain isolated from the confirmed health trajectory.
[0046] The disease health status evolution model is built using historical health sequences that have been independently confirmed. Disease supervision labels are derived from independent disease confirmation records that are independent of the model's early warning results. Each model version has a fixed target disease and its confirmation criteria.
[0047] This implementation method uses infectious bronchitis in chickens as the target disease. The independent disease confirmation record adopts the veterinary diagnostic record that can be associated with the sampling identifier of a specific laying hen. The veterinarian forms a final diagnosis based on the laboratory nucleic acid test results, the corresponding clinical review record, and the concurrent immunization record. The model supervises the label to read the final diagnosis, and does not directly use a single laboratory test result as a positive disease label.
[0048] When using other target diseases, an independent model version is established according to the independent confirmation rules of the corresponding disease, and the training, calibration and disease risk determination conditions are re-completed. The model supervision target uses the moment when the independent disease confirmation result is first available to the system as the confirmation event. Health data after the target disease has been confirmed will not continue to be used as early risk training segments for the same disease.
[0049] The disease risk prediction time window is determined based on the independent disease review cycle during the model development phase, preferably 1-2 complete review cycles. Under the condition that the model development data is independently reviewed once a day, the preferred range of the disease risk prediction time window is 24-48 hours. This implementation adopts 24 hours as the preferred value, so that each training segment corresponds to one complete review cycle, while reducing the impact of the new health treatment after crossing the second review cycle on the observability of the outcome.
[0050] Training segments that obtain the first independent confirmation result of the target disease within 24 hours after the observation time are assigned a positive supervision label; training segments that do not obtain the first independent confirmation result of the target disease within 24 hours after the observation time, and have a valid independent negative review result at the end of the time window, are assigned a negative supervision label. Training segments with incomplete follow-up, uncertain review conclusions, and observable changes in disease outcomes due to health management are left without disease labels, and their valid health data continue to participate in health status prediction training.
[0051] Historical data are grouped according to common exposure relationships. Laying hens in the same breeding unit that share the same main rearing environment during the same observation period are grouped into the same data group. The complete health sequence of the same case is included in only one data set. Training data, parameter tuning and verification data, independent calibration data, and final test data are isolated from each other according to data groups. The preferred range for the proportion of training data is 50%-70%, the preferred proportions for parameter tuning and verification data and independent calibration data are 10%-20% and 10% respectively, and the preferred proportion for final test data is 10%-20%. In this implementation, 60%, 15%, 15%, and 10% are preferred configurations.
[0052] When using this configuration, the parameter tuning and validation data must include both positive and negative supervision labels, and the independent calibration data must meet the minimum number of complete records for the corresponding calibration group. If any supervision label is missing from the parameter tuning and validation data, no disease risk assessment threshold will be established. If the independent calibration data is insufficient, the corresponding calibration group will be deactivated. The data set will remain complete, and samples will not be supplemented by splitting the same common exposure data set. The proportion of the data set only serves as data isolation during the model development phase and will not be used as a condition for disease risk assessment.
[0053] A reproducible disease-health state evolution model employs a single-layer gated recurrent unit network. The candidate range for the recurrent hidden state dimension is 32-128, with 32, 64, and 128 used as candidate values. After completing two-stage training for each candidate dimension, the candidate rankings are determined from smallest to largest based on the health state prediction loss and disease risk loss on the parameter tuning validation data. The two rankings are added together, and the candidate dimension with the smallest ranking sum is selected. If the ranking sums are the same, the candidate value with the smaller dimension is used.
[0054] This implementation uses 64 dimensions. The model input consists of 4-dimensional health status, 4-dimensional health dimension valid labels, time interval between adjacent valid health data, laying hen age, and sine and cosine periodic components of intraday time, for a total of 12 dimensions.
[0055] The time interval between adjacent valid health data is normalized according to the upper limit of the valid transition time of the current model version, and the laying hen age is normalized according to the maximum age in the training data of the current model version. The 12-dimensional input is connected to a single-layer gated recurrent unit, and the hidden state of the gated recurrent unit is connected to the health status prediction output layer and the disease risk output layer respectively. The health status prediction output layer adopts a 4-dimensional hyperbolic tangent output, which corresponds to the 4-dimensional predicted health status at the next query time. The disease risk output layer adopts a 1-dimensional sigmoid output, which corresponds to the disease risk result within a fixed disease risk prediction time window.
[0056] The model input dimension, recurrent hidden state dimension, output layer connection relationship, and output dimension are fixed in the same model version and are not adjusted based on the current health data to be assigned during online processing.
[0057] The model training is performed in two stages. In the first stage, the gated recurrent unit and the health status prediction output layer are trained using health sequences with confirmed sources. The smoothed absolute error loss is used, the inflection point is fixed at 1, and only data with effective health dimensions are counted. In the second stage, the parameters of the gated recurrent unit obtained in the first stage are fixed, and the disease risk output layer is trained using data with clear disease supervision labels. The binary cross-entropy loss is used.
[0058] This training sequence ensures that health status prediction and disease risk output each have a defined training target, avoiding reliance on undisclosed multi-task weights. The model training employs an adaptive moment estimation optimization method, with candidate values for the learning rate set to 0.0005, 0.001, and 0.002, candidate values for the batch size set to 16, 32, and 64, candidate values for the maximum number of training epochs set to 100, 200, and 300, and candidate values for the number of consecutive epochs without improvement set to 10, 20, and 30. Parameter tuning and validation are performed using candidate combinations at each training stage, selecting the combination with the minimum validation loss for the corresponding stage. When validation losses are the same, the combination with a smaller learning rate, smaller batch size, fewer maximum training epochs, and fewer consecutive epochs without improvement is selected.
[0059] This implementation uses a learning rate of 0.001, a batch size of 32, a maximum number of training rounds of 200, and a stop training method when there is no improvement after 20 consecutive rounds as the preferred configuration. The input weights are initialized uniformly using Xavier, the cyclic weights are initialized orthogonally, the bias is initialized to 0, the random seed is saved with the model version, the training parameters are used to determine the fixed model version, and the disease risk determination conditions are formed separately from the parameter tuning and verification data.
[0060] The model state initialization adopts the historical calculation state that is consistent with the current model version, individual health reference version, and confirmed health trajectory version. If the confirmed health trajectory has already saved the calculation state of the same version model, the corresponding state is read directly. If the versions are inconsistent or the historical records are backfilled, the available starting point of the confirmed health trajectory of the same version is replayed in real time order to the starting point of the current state deduction.
[0061] Replay only uses health data that has obtained independent source credentials and maintains the same health dimension missing rules as online processing. This initialization method does not set a fixed historical truncation length to avoid deleting candidate laying hens that still have source eligibility due to insufficient historical data in the fixed replay window.
[0062] The effective transfer time limit is determined based on the historical health data processing records with final independent source evidence. The system calculates the effective time span of each historical record from the earliest state deduction starting point to the last confirmed source of health data within the disease risk prediction time window before the health intervention. The preferred range for the quantile probability used to determine the time limit is 0.9-0.99, and this implementation adopts 0.95 as the preferred value. The 0.95 quantile value ensures that the effective transfer time limit covers 95% of the effective time span in the complete historical record, while keeping the longest sparse tail span outside the automatic attribution verification range.
[0063] The system sorts all valid time spans in ascending order, takes the 95th percentile value, and compares it with the disease risk prediction time window. The smaller value is taken as the upper limit of the valid transfer time for the current model version. This parameter only limits the time range applicable to automatic attribution verification. Candidate attribution relationships outside this range remain in an undetermined state.
[0064] For each online health data to be assigned, the system adds the effective transfer time limit to the starting point of the status projection of each candidate laying hen, and takes the earliest time as the assignment verification deadline for the current candidate assignment processing. Subsequent assignment verification is performed within the time range after the original collection time of the health data to be assigned and no later than the assignment verification deadline. New subsequent health data with a clear source will keep the assignment verification deadline unchanged.
[0065] After the model training is completed, the health status prediction bias is generated on the parameter tuning and validation data in the same order as online processing: first predict, then read the actual health data. For each health dimension, the 50th percentile of all effective absolute prediction biases is determined as the robust prediction error benchmark for that health dimension. It is then compared with the actual storage resolution of the health status quantity, and the larger value is taken as the prediction error scale for that health dimension.
[0066] The 50th percentile corresponds to the median of the absolute prediction bias. It does not require the assumption that the prediction error follows a normal distribution. Furthermore, the prediction error scale only serves to unify the error magnitudes of different health dimensions, allowing the final boundary of the health status range to continue to be determined by independent calibration data. Therefore, the prediction error scale itself is not used as a threshold for excluding candidate attribution relationships.
[0067] The range of health statuses can be independently calibrated using historical health data processing records with final independent source credentials. Each historical record used for calibration contains the candidate laying hen set formed at that time, the final independent source result, and subsequent health data with a clear source formed before the end of the attribution verification deadline.
[0068] The calibration process freezes individual health references, establishes candidate attribution relationships, and replays candidate attribution relationships consistent with the final independent source results according to online processing rules. Each subsequent verification first uses inputs that are strictly earlier than the health data to be tested to form a prediction, then reads the actual health data to calculate the prediction bias, and updates the model calculation status after the comparison is completed.
[0069] For each valid test position in each historical health data processing record used for calibration, the system calculates the absolute difference between the health status quantity with a clear source and the predicted health status quantity formed before reading the health data. This difference is then divided by the prediction error scale of the corresponding health dimension to obtain the normalized prediction bias. The system takes the maximum normalized prediction bias among all candidate laying hens, all valid test times, and all valid health dimensions in the same historical health data processing record and uses this value as the value for ranking the calibration boundary of the historical record. This event-level calculation uses the same test granularity as the rule that excludes the entire candidate attribution relationship if any candidate laying hen exceeds the allowable health status range during the online phase.
[0070] Independent calibration establishes calibration groups based on collection type and the number of candidate laying hens. Within the same calibration group, only one historical health data processing record that meets the complete observation conditions is selected for the same common exposure data group. The preferred range for the event-level nominal misclassification ratio is 1%-10%, which represents the nominal proportion of candidate attribution relationships corresponding to the true source that are allowed to be excluded under the event-level calibration rules. In this implementation, 5% is used as the preferred value, corresponding to a nominal retention ratio of 95% for the allowable health status range. The system sorts the maximum normalized prediction deviation of each historical record in the same calibration group in ascending order, adds 1 to the number of complete calibration records, multiplies it by 0.95, and rounds the resulting value up to obtain the sorting position. When the sorting position does not exceed the number of complete calibration records, the maximum normalized prediction deviation at the corresponding sorting position is determined as the calibration boundary of the allowable health status range.
[0071] When using a 5% event-level nominal missort ratio, at least 19 complete calibration records are required for the sorting position to fall within the range of existing calibration records. If the number of records is less than 19, the corresponding calibration group remains in a deactivated state. The 19 records are directly obtained from the finite sample sorting conditions and are not used as an empirical threshold for deployment performance adequacy.
[0072] Event-level calibration requires that historical records within the calibration group and online health data to be assigned use the same collection type, candidate laying hen number grouping, health dimension statistical rules, and record selection rules. The system separately calculates the laying hen age range in training data, parameter tuning and verification data, and independent calibration data. The intersection of the age ranges corresponding to the three types of data is taken as the online applicable age range for the current model version. Online records are required to continue to use the fixed collection type and health dimension statistical rules for the corresponding model version. When the online laying hen age exceeds the online applicable age range, the current record stops automatically excluding candidate attribution relationships and retains the original health data. When the collection device type, health dimension statistical caliber, and candidate laying hen number grouping change, the system re-forms calibration groups based on the changed conditions and performs independent calibration. The nominal error tolerance only represents the allowable proportion of candidate attribution relationships corresponding to the final independent source that are excluded under this calibration rule, and does not represent the source accuracy rate or disease warning accuracy rate.
[0073] Under each candidate attribution relationship, the system starts by accessing health data in real time order from the confirmed health status of each candidate laying hen and the corresponding initial calculation status of the model. When the original collection time of the health data to be assigned is reached, the system only accesses the health data to be assigned to the candidate laying hen corresponding to that candidate attribution relationship. The remaining candidate laying hens continue to advance along the confirmed health data sequence from their respective sources. Thus, the same health data to be assigned forms different candidate laying hen health data input structures under different candidate attribution relationships.
[0074] For any subsequent query time, the state inference module first uses health data strictly earlier than that query time to form the predicted health status and disease risk results of each candidate laying hen. The allowable health status range is determined according to the following textual calculation logic: For each valid health dimension of the health status to be compared, the absolute difference between the health status quantity and the corresponding predicted health status quantity is calculated, and the absolute difference is compared with the product of the prediction error scale of the corresponding health dimension and the calibration boundary of the allowable health status range; when the absolute difference of each valid health dimension does not exceed the corresponding product, the health status to be compared is within the allowable health status range of the corresponding candidate laying hen, and missing health dimensions do not participate in the current comparison.
[0075] Each candidate laying hen in the same candidate attribution relationship forms an allowable health status range and disease risk result. The condition for retaining the candidate attribution relationship is that each candidate laying hen that has obtained subsequent health data with a clear source meets the corresponding allowable health status range. Therefore, both candidate laying hens that have obtained health data to be assigned and those that have not obtained health data to be assigned retain the corresponding status inference results. This joint judgment is not based on the premise that the health data of the candidate laying hens are independent of each other.
[0076] When new health data to be assigned involves laying hens in the current candidate laying hen set, the system temporarily stores the health data according to the original collection time and maintains its pending assignment status during the current candidate assignment processing. The current pending health data ends when it obtains independent source credentials, reaches the assignment verification deadline, all candidate assignment relationships are excluded, and the current collection type, candidate laying hen number grouping or record selection rules are no longer consistent with the corresponding calibration group. The system then processes the temporarily stored health data according to the original collection time.
[0077] If the processing of temporary health data has exceeded the deadline for attribution verification determined by the original collection time, the temporary health data will only be used for historical source record revision and health status replay, and will not generate retrospective early disease warnings. Health data that has been obtained during processing but can only be used after the original collection time of the temporary health data will retain the actual available time and will not be moved forward to be input that can be used at the original collection time. The health data to be assigned of candidate laying hens that are not intersecting will be processed independently.
[0078] When health intervention changes the evolutionary conditions of the non-intervention state corresponding to the disease health status evolution model, the system ends the current attribution verification after the time of the health intervention. The disease warning that has been formed retains its original assessment time. When there is a conflict between the collection mechanism, candidate range and original record version, the current automatic attribution screening stops, and the original health data continues to be saved.
[0079] The attribution verification module processes subsequent health data with clearly defined sources in the order of first verifying and then updating. At a certain observation time, the system first uses health data that is strictly earlier than that observation time to generate and seal the allowable health status range corresponding to each candidate attribution relationship.
[0080] Subsequently, the health data with clear sources at the observation time is read, converted into a health status according to the current frozen individual health reference, and then compared with the already sealed range of allowed health statuses. After completing the comparison of each candidate attribution relationship within the current observation batch, the health data with clear sources is connected to the reserved candidate attribution relationship and used for subsequent status updates.
[0081] The individual source of subsequent health data with a clear origin is determined by independent association records at the collection end. The identity inference results output by the disease health status evolution model are not used. The system also checks the original sampling number on which the health data to be examined depends. Duplicate data formed from the original sampling of the health data to be assigned, re-aggregated data, and original samples that have entered the current prediction state retain their original uses and are not repeated as new health data with a clear origin. Aggregated health dimensions containing duplicate original samples remain missing in the current attribution verification.
[0082] For subsequent health data from multiple candidate laying hens at the same observation time with a clear source, the system first completes the source and quality verification of the same observation batch, then reads the predicted health status of each candidate attribution relationship that has been sealed, and uniformly updates the retained candidate attribution relationships after completing the candidate attribution relationship judgment for the entire observation batch. This execution order ensures that each health data item in the same observation batch corresponds to the predicted health status previously formed in that batch of data.
[0083] For each candidate attribution relationship, the system accumulates all valid verification positions from the first valid verification of the current candidate attribution processing to the current observation time. For each valid verification position, the absolute difference between the subsequent health status quantity with a clear source and the corresponding sealed predicted health status quantity is first calculated, and then divided by the prediction error scale of the corresponding health dimension to form a normalized prediction bias. The system takes the maximum normalized prediction bias among all valid verification positions of the current candidate attribution relationship and compares it with the calibration boundary of the allowable health status range.
[0084] If the maximum normalized prediction deviation does not exceed the calibration boundary of the allowable health status range, the current candidate attribution relationship is retained; if the maximum normalized prediction deviation exceeds the calibration boundary of the allowable health status range, the current candidate attribution relationship is terminated. If any candidate laying hen with subsequent health data of a clearly defined source has an out-of-bounds result, the entire candidate attribution relationship is terminated. Candidate laying hen with subsequent health data of a clearly defined source remain in an untested state. The termination of a candidate attribution relationship indicates that the relationship is incompatible with the health data already obtained under the current fixed model and calibration rules. The true source is confirmed based on independent source evidence.
[0085] Candidate laying hens that have not been connected to the health data to be assigned also participate in the candidate attribution relationship verification. The candidate laying hen forms an allowed health status range along the health data sequence of the unconnected health data to be assigned health data. When the health status of the subsequent source is clearly found to exceed the allowed health status range, the entire candidate attribution relationship is withdrawn. When the subsequent health status falls into the allowed health status range corresponding to both the connected and unconnected health data to be assigned health data, the corresponding candidate attribution relationship is retained.
[0086] After completing the attribution verification of the current observation batch, the system will connect subsequent health data with clear sources to the candidate attribution relationships that have not been withdrawn, update the corresponding status inference results, and re-obtain the disease risk results at the same observation time. The system will determine the candidate attribution relationships that have not been withdrawn as the remaining candidate attribution relationships, and ensure that the remaining candidate attribution relationships, the current model version, the current individual health reference version, the current observation time, and the updated disease risk results are consistent.
[0087] In one implementation scenario, the candidate set of laying hens includes laying hen A and laying hen B. The system establishes candidate attribution relationships for laying hen A and laying hen B respectively, based on the health data to be assigned. The candidate attribution relationship for laying hen A connects the health data to laying hen A, and laying hen B performs state deduction along its confirmed health trajectory. The candidate attribution relationship for laying hen B adopts the opposite data access method.
[0088] When hen A and hen B subsequently obtain health data with clear sources, the attribution verification module first reads the allowed health status range of the two candidate attribution relationships that have been sealed according to the same observation batch, and then compares the confirmed health status of hen A and hen B.
[0089] If any candidate laying hen exhibits a result exceeding the corresponding permissible health status range, the entire candidate attribution relationship that resulted in that combined state consequence is excluded; candidate attributions that remain compatible continue to participate in disease risk outcome updates.
[0090] When there are still multiple remaining candidate attribution relationships, the system will continue to participate in subsequent state deductions for each remaining candidate attribution relationship. It will not select a unique attribution based on disease risk results, prediction bias, and identity matching scores. When there is only one candidate attribution relationship left, the system will save the unique remaining candidate attribution relationship and the basis for the withdrawal of other candidate attribution relationships. This record will retain the candidate attribution relationship screening result attributes, and the independent source evidence will still be formed by the collected facts.
[0091] When all candidate attribution relationships are withdrawn, the system will keep the current health data to be attributed in an unresolved source state, end the cross-candidate attribution disease risk comparison triggered by the health data to be attributed, and retain the original health data and the already formed attribution verification records.
[0092] If a laying hen that is still in the candidate range lacks an effective health dimension for comparison of the range of acceptable health statuses at the current observation time, exceeds the upper limit of the effective transfer time, or the corresponding model version cannot output an effective disease risk result, the candidate attribution relationship remains in an undetermined state, and the current cross-candidate attribution disease risk comparison remains inactive. The system does not form a consistent disease risk result for the remaining candidate attribution relationships by deleting candidate attribution relationships that cannot be calculated.
[0093] Health data with a clear source that arrives later than the end time of the corresponding observation batch is included in the health record version correction as a late record. Late records are used to revise the actual health records. No previously published historical predictions are inserted to form new attribution verification samples. Duplicate original samples are deduplicated according to the original sample number. The same original sample performs one status update and one attribution verification function in the same processing version.
[0094] The risk warning module performs cross-candidate attribution disease risk comparison when the remaining candidate attribution relationships are not empty, at least one valid attribution verification has been completed, the current observation time is within the attribution verification deadline, and each remaining candidate attribution relationship has a valid disease risk result at the same observation time.
[0095] When any remaining candidate attribution lacks a valid disease risk result, cross-candidate attribution disease risk comparison at the current observation time remains disabled. If an independent confirmation result for the target disease has been obtained during the current processing period, the generation of new early disease warnings for the same target disease will cease from the moment the independent confirmation result becomes available to the system. Existing disease warnings will retain their original assessment time, and subsequent health data will continue to be processed according to the health records for which the target disease has been confirmed.
[0096] The disease risk determination criteria are implemented using a fixed disease risk determination threshold. The disease risk output value of the disease health status evolution model ranges from [0,1]. The disease risk determination threshold is greater than 0 and less than 1. The preferred value is determined by the parameter tuning and validation data according to fixed rules. The system sequentially uses different disease risk results formed by parameter tuning and validation data with confirmed sources and clear disease supervision labels as candidate thresholds. The system calculates the sum of sensitivity and specificity and then subtracts 1. The candidate threshold with the largest value is determined as the disease risk determination threshold of the current model version. When there are candidates with the same maximum value, the candidate threshold with the smaller value is used. When the parameter tuning and validation data have both positive and negative supervision labels, the corresponding disease risk determination threshold of the model version is established.
[0097] For the same laying hen, the system reads the disease risk results of the hen under each of the remaining candidate attribution relationships and compares them using the same disease risk judgment threshold. When the disease risk results under each remaining candidate attribution relationship are not lower than the disease risk judgment threshold, the laying hen is identified as a warning object and a disease warning is output. When only some of the remaining candidate attribution relationships meet the disease risk judgment conditions, the system keeps the corresponding disease risk record and continues to perform subsequent attribution verification.
[0098] The disease risk comparison across candidate attribution relationships adopts the method of each remaining candidate attribution relationship satisfying the disease risk judgment condition one by one. It does not use the average value of the disease risk result, nor does it use the highest single disease risk result to replace other remaining candidate attribution relationships. The resulting disease warning indicates that each remaining candidate attribution relationship that is still valid supports the same laying hen to meet the disease risk judgment condition. The disease warning maintains the individual health risk information attribute.
[0099] When writing health data to be assigned to a confirmed health trajectory, the independent source condition is adopted. The independent source credential establishes a correspondence with the unique record identifier, original sampling number and sampling session of the health data to be assigned, and clearly indicates the unique laying hen identity. When only the current identity of a laying hen is confirmed but no correspondence is established with the original sampling session of the health data to be assigned, the source information retains the current identity record attribute.
[0100] One implementation uses identity association messages cached at the acquisition end. When the original temperature measurement record of an individual arrives first, it carries the sampling session number and the candidate laying hen set. The identity association message that arrives later carries the same sampling session number and a unique laying hen identifier. After the system verifies that the sampling session number is consistent, the individual occupying record is valid, and the individual identifier is valid, it determines the independent source credential. This verification reads the existing collection facts, and the disease risk result does not participate in the formation of identity information.
[0101] Before obtaining independent source credentials, the health data to be attributed only participates in the calculation of subsequent health status and disease risk results within each candidate attribution relationship, and remains isolated from the confirmed health trajectory of the candidate laying hens. When only one candidate attribution relationship remains and a disease warning has been issued, the health data to be attributed continues to remain isolated until an independent source credential is obtained.
[0102] After obtaining the independent source credential, the system writes the health data to be assigned to the confirmed health trajectory of the laying hen indicated by the independent source credential according to the original collection time, and performs idempotent deduplication using the original record identifier. When the laying hen identity indicated by the independent source credential is inconsistent with the candidate attribution relationship screening result, the system updates the health data attribution according to the independent source credential and retains the difference record between the candidate attribution relationship screening result and the collection fact. The compatibility result formed by the disease health status evolution model maintains the attributes of the model calculation result.
[0103] When the independent source credential arrives later than the original collection time of the health data to be assigned, the system will recalculate the affected subsequent health status according to the real time sequence after writing the credential. The system will start from the most recent confirmed health status before the original collection time of the health data to be assigned, and reprocess the health data to be assigned and the subsequent health data whose sources have been confirmed in sequence. When the recalculation path exceeds the applicable time range of the current model or crosses the health intervention that changes the state evolution conditions, the system will save the confirmed original health data and establish a new calculation state after it has a qualified confirmed history again.
[0104] For health dimensions formed by fixed time windows, after the assigned health data is backfilled, the original event set is revised first, and then the affected time window is recalculated according to the same statistical rules. The original events and the summary data that already contains the original events are deduplicated based on the data dependency identifier, so that the same original event only participates in the health status update once. The predicted status formed by the recalculation continues to maintain the attributes of the model calculation result. The confirmed health status used in the next candidate attribution process is still formed by the actual health data that has been independently confirmed from the source.
[0105] Independent source credentials determine the eligibility of health data to be included in the corresponding confirmed health trajectory of laying hens. The inclusion of this health data in the subsequent normal health reference also requires the independent health review results during the corresponding observation period. Health data from confirmed sources that show abnormalities in disease are retained in the confirmed health trajectory and kept outside the normal health reference sample. Disease risk assessment criteria are used for disease early warning. Health reference updates use independent health review results as the eligibility basis.
[0106] The disease health status evolution model remains fixed during online processing. Once the health data to which it belongs obtains independent source credentials, it is included in the candidate historical health data set for subsequent model versions. Health data that simultaneously meet the source confirmation conditions and the corresponding disease supervision label conditions are included in the model training data. When a new model version is established, data grouping, parameter tuning and validation, independent calibration and final verification are re-executed, and the online disease risk results maintain the model output attributes.
[0107] Through continuous processing, the health data to be assigned only changes the available health data input within each candidate attribution relationship during the period when the source is not confirmed, and remains isolated from the confirmed health trajectory of each candidate laying hen; subsequently, when the source of the health data is clear, both the candidate laying hen that has obtained the health data to be assigned and the candidate laying hen that has not obtained the health data to be assigned are verified, so that the candidate attribution relationship that is incompatible with the actual health status is withdrawn; each of the remaining candidate attribution relationships supports the formation of a disease warning when the same laying hen meets the disease risk judgment conditions; after the independent source certificate is obtained, the health data to be assigned enters the confirmed health trajectory of the corresponding laying hen according to the actual collection time.
[0108] This continuous processing binds the risk calculation purpose in candidate attribution relationships to the permanent writing purpose of confirmed health trajectories to disease risk determination conditions and independent source credentials, respectively. This allows the disease risk information jointly supported by the remaining candidate attribution relationships under the condition that the source has not yet been confirmed to continue to be used for individual health risk assessment, while avoiding the direct conversion of model screening results into individual historical source facts.
[0109] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. An AI-based early warning system for laying hen diseases, characterized in that, It includes an evidence attribution module, a status deduction module, an attribution verification module, and a risk warning module; The evidence attribution module acquires health data to be attributed and a set of candidate laying hens that have been verified as having a single source but lack independent source evidence. It reads the confirmed health status of the candidate laying hens before the formation of the health data to be attributed and establishes candidate attribution relationships with the constraint that the health data to be attributed corresponds to only one candidate laying hen in each candidate attribution relationship. The status inference module only accesses the health data of the candidate laying hens to be assigned under each candidate attribution relationship, and infers the subsequent health status of the candidate laying hens based on the disease health status evolution model, forming the range of allowable health status and disease risk results; The attribution verification module obtains subsequent health data with a clear source, compares the corresponding health status with the range of allowed health statuses under the candidate attribution relationship, and excludes the corresponding candidate attribution relationship when the health status exceeds the range of allowed health statuses. It then connects the subsequent health data with a clear source to the remaining candidate attribution relationships and updates the disease risk results. The risk warning module compares the disease risk results of the same laying hen under each remaining candidate attribution relationship, and outputs disease warnings for laying hens that all meet the disease risk judgment conditions. After obtaining the independent source credentials corresponding to the health data to be assigned, the health data to be assigned is written into the confirmed health trajectory of the corresponding laying hen according to the original collection time.
2. The AI early warning system for laying hen diseases according to claim 1, characterized in that, The specific steps for establishing candidate attribution relationships in the evidence attribution module are as follows: A candidate laying hen set is formed based on the individual source verification results of the health data to be assigned and the valid individual registration records within the collection period. For each candidate laying hen in the candidate laying hen set, a candidate attribution relationship is established. Each candidate attribution relationship only maps the health data to be assigned to one candidate laying hen, and determines that other candidate laying hens will participate in the subsequent health status inference along their respective confirmed health trajectories.
3. The AI early warning system for laying hen diseases according to claim 2, characterized in that, The specific steps for the evidence attribution module to generate health status input are as follows: Read the confirmed health status of each candidate laying hen from the confirmed health trajectory before the formation of the health data to be assigned, and convert the same health data to be assigned according to the confirmed health trajectory of each candidate laying hen to obtain the health status input corresponding to each candidate laying hen; Confirmed health status is formed from confirmed health data from an individual source.
4. The AI early warning system for laying hen diseases according to claim 1, characterized in that, The specific operations of the state projection module to generate the allowable range of health status and disease risk results are as follows: Under each candidate attribution relationship, the confirmed health status of each candidate laying hen is used as the starting point for state deduction. The health data to be assigned is connected to the candidate laying hen corresponding to the candidate attribution relationship, and the subsequent health status is deduced based on the health data obtained by each candidate laying hen under the candidate attribution relationship. The acceptable range of health status and disease risk results are formed based on the subsequent health status at the corresponding observation time.
5. The AI early warning system for laying hen diseases according to claim 4, characterized in that, The specific steps for calibration within the permitted health status range are as follows: Calibration samples are formed based on historical health data from confirmed sources. The prediction deviation of the health status of candidate laying hens in each calibration sample during the inspection period is calculated. The prediction deviation of the health status is normalized according to the prediction error scale of the corresponding health dimension. The calibration boundary of the allowable health status range is determined by the maximum normalized prediction deviation in the same calibration sample.
6. The AI early warning system for laying hen diseases according to claim 1, characterized in that, The specific operations of the attribution verification module in comparing subsequent health data with clearly defined sources are as follows: Before accessing subsequent health data with clearly defined sources, the permissible health status range formed by the status inference module for the same observation time is read, and the health status corresponding to the subsequent health data with clearly defined sources is compared with the permissible health status range. After the comparison is completed, subsequent health data with clearly defined sources will be integrated into the candidate attribution relationships that have passed the comparison.
7. The AI early warning system for laying hen diseases according to claim 6, characterized in that, The specific operations of the attribution verification module to form the remaining candidate attribution relationships are as follows: By comparing the subsequent sources of each candidate laying hen in the same candidate attribution relationship, the health status corresponding to the health data and the corresponding allowable health status range can be determined. If the health status of any candidate laying hen exceeds the corresponding allowable health status range, the candidate attribution relationship is excluded. The candidate attribution relationships where the health status of all candidate laying hens is within the corresponding allowable health status range are determined as the remaining candidate attribution relationships, and the corresponding disease risk results are updated.
8. The AI early warning system for laying hen diseases according to claim 1, characterized in that, The specific steps for the risk warning module to output disease warnings are as follows: The disease risk results of the same laying hen under each of the remaining candidate attribution relationships were compared using the same disease risk assessment threshold. When all disease risk results reach the disease risk judgment threshold, the laying hen is identified as a warning target and a disease warning is issued. The disease risk judgment threshold constitutes the disease risk judgment condition.
9. The AI early warning system for laying hen diseases according to claim 1, characterized in that, The specific steps for the risk warning module to update confirmed health trajectories based on independent source credentials are as follows: Read the original sampling record corresponding to the health data to be attributed and indicate the unique source credential for the laying hen's identity; After obtaining independent source credentials, the health data to be assigned is written into the confirmed health trajectory of the corresponding laying hen according to the original collection time of the health data to be assigned, and the subsequent health status is re-inferred from the original collection time in chronological order.
10. The AI early warning system for laying hen diseases according to claim 1, characterized in that, The specific steps for controlling the access permissions to health data to be assigned in the risk warning module are as follows: Before obtaining independent source credentials, the health data to be attributed is used in the calculation of subsequent health status and disease risk results within each candidate attribution relationship, and is kept isolated from the confirmed health trajectory of each candidate laying hen; A disease warning is issued when the same laying hen meets the disease risk assessment criteria under all remaining candidate attribution relationships. After obtaining independent source credentials, the health data to be attributed will be written into the confirmed health trajectory of the corresponding laying hen.
