Duck feed formula optimization method based on multi-modal neural network

Abstract

The application discloses a duck feed formula optimization method based on a multi-modal neural network and belongs to the field of computer-aided animal nutrition and feed formula optimization. In order to solve the problems of duck feed raw material batch fluctuation, duck group nutrition demand difference and difficulty of a static linear model in depicting a nonlinear response, a heterogeneous exposure graph with edge reliability is constructed, graph coding and duck group time series coding are fused, and group shape preserving calibration, joint risk budget and double residual feedback are combined to solve batch feeding proportion and an applicable time window under the constraints of nutrition, safety, inventory and cost, so that detection uncertainty can be tracked into prediction and formula constraints, the risks of nutrition deficiency and toxin over-limit are reduced, and invalid safety margin is reduced.

Description

Technical Field

[0001] This invention relates to the field of computer-aided animal nutrition and feed formulation optimization, and in particular to a method for optimizing duck feed formulation based on multimodal neural networks. Background Technology

[0002] Duck feed formulation needs to simultaneously consider the nutritional composition of raw materials, toxin safety, the stage-specific needs of the duck flock, production costs, and raw material inventory. Indicators such as crude protein, moisture content, metabolizable energy, and mycotoxins can fluctuate from batch to batch. The age, breed, weight distribution of the flock, temperature and humidity in the shed, and health status of the ducks can also change the nutritional needs of the flock and their susceptibility to risk factors. Therefore, it is difficult for the same static formula to be consistently adapted to different production time windows.

[0003] Existing duck feed formulation methods typically determine ingredient ratios based on static nutritional standards and linear programming models, making it difficult to characterize the nonlinear relationships between ingredients, environment, actual feed intake, daily weight gain, feed conversion ratio, and health risks. Furthermore, there is a lack of a unified mechanism for transmitting uncertainties reflected in test repeatability, sampling coverage, testing time, and missing states, making it easy to misinterpret imputation results as safety data. On the other hand, if manufacturing deviations and duck flock biological response deviations are updated using the same feedback channel, it may lead to incorrect attribution and formulation oscillations, resulting in nutritional deficiencies, excessive toxins, ineffective safety margins, or over-reliance on manual experiments.

[0004] Therefore, there is a need for a method to optimize duck feed formulation that can address the shortcomings of existing technologies. Summary of the Invention

[0005] The purpose of this invention is to provide a method for optimizing duck feed formulation based on multimodal neural networks. This method performs benchmark transformation and time alignment on raw material batch detection data and duck flock time-series data, establishing a heterogeneous exposure map with edge reliability. It generates batch exposure representations and duck flock status representations, respectively. After cross-fusion controlled by reliability and missing masking, it predicts daily weight gain, feed conversion ratio, and health risk. Furthermore, it utilizes group-based conformal calibration to form constraint boundaries, allocating a joint risk budget based on duck flock susceptibility, current health risk, and toxin detection reliability. Under constraints of nutrition, safety, inventory, dosage, and cost, it outputs the feed ratio and applicable time window. The detection edges and calibration boundaries are updated using manufacturing residuals generated from finished product re-inspection and biological response residuals generated from feeding results.

[0006] This invention provides a method for optimizing duck feed formulation based on multimodal neural networks, comprising: S1. Obtain batch data of duck feed raw materials and time series data of target duck flock, perform dry matter baseline conversion, time alignment and generate missing mask to obtain aligned multimodal data; S2. Based on the aligned multimodal data, establish a heterogeneous exposure map including raw material batch nodes, nutrient component nodes, toxin nodes, formula nodes, and duck flock time slice nodes, and determine the edge weights characterizing exposure amount and time lag, and the edge reliability characterizing detection quality. S3. The heterogeneous exposure graph and the duck flock time slice sequence are encoded using a graph encoder and a time sequence encoder respectively, and cross-fused according to the edge reliability and the missing mask to obtain the fused representation; S4. Generate a candidate formulation set based on the aligned multimodal data, input the features of each candidate formulation and the fused representation into the multi-task prediction head to obtain the prediction results of production performance, health risk and toxin exposure, perform duck flock grouping conformity calibration on the prediction results and allocate joint risk budget to obtain the constraint boundary; S5. Based on the constraint boundary, under the staged nutritional and safety constraints, solve the target formula from the candidate formula set, and output the feeding ratio of each batch of raw materials and the applicable time window for the duck flock. S6. Based on the finished product re-inspection results, generate manufacturing residuals and update the heterogeneous exposure map. Based on the actual feeding results, generate biological response residuals and update the conformal calibration samples and constraint boundaries of the duck flock groups for subsequent formula solving.

[0007] Optionally, S1 includes: acquiring the crude protein content, moisture content, estimated metabolizable energy, mycotoxin detection value, price, inventory, detection repeatability, sampling coverage, detection time point, detection validity period, failure time point determined by the detection time point and detection validity period, current evaluation time and missing marker for each batch of raw materials; and acquiring the duck age, breed, group weight distribution, feed intake, health indicators, house temperature, house humidity, formula feeding time and historical daily weight gain and feed conversion ratio of the target duck flock. The wet-based nutrient measurements are converted to dry-based measurements based on moisture content. Information is mapped to duck time slices according to raw material storage time, formula production time, and duck feeding time. Missing markers for each field are encoded as missing masks to obtain the aligned multimodal data and the corresponding missing masks.

[0008] Optionally, S2 includes: establishing a feeding edge between the raw material batch node and the formula node, establishing a detection edge between the raw material batch node and the nutrient component node and the toxin node, and establishing a feeding edge between the formula node and the duck flock time slice node. The edge weights are determined based on the feeding ratio, dry matter basis measurement, actual feed intake, and the time interval from feeding to feeding. The edge reliability is determined based on detection repeatability, sampling coverage, the ratio of the remaining time from the current evaluation time to the failure time to the effective detection time, and the missing marker. When the missing marker indicates a detection failure or the current evaluation time reaches the failure time, the reliability of the corresponding detection edge is set to zero.

[0009] Optionally, S3 includes: the graph encoder aggregating the features of adjacent nodes according to the edge type, and adjusting the aggregation coefficient of each adjacent node feature using the edge weight and edge reliability to form a recipe batch exposure representation; The time encoder encodes the group's weight distribution, feed intake, health indicators, and indoor temperature and humidity according to the chronological order of the duck flock's time slices, forming a representation of the duck flock's status. The edge reliability and the missing mask obtained by the missing tag encoding are used to generate cross-fusion gating coefficients. The cross-fusion gating coefficients are used to control the information transmission between the recipe batch exposure representation and the duck flock state representation to obtain the fused representation. Furthermore, the graph encoder, temporal encoder, multi-task prediction head, and cross-fusion gating coefficient generation network are trained using historical raw material batch data, historical formula data, historical duck flock time slice data, finished product re-inspection results, and actual feeding results. During training, the weighted sum of the daily weight gain prediction loss, feed conversion ratio prediction loss, health risk prediction loss, and toxin exposure prediction loss is used as the multi-task loss. The reliability of detection edges with missing labels or whose current evaluation time has reached the failure time point is set to zero, while the reliability of other detection edges in the same training sample remains unchanged. The reliability of the detection edges associated with the training sample is obtained by normalizing and weighting the sum of the reliability of each batch of raw materials in the training sample according to the feeding ratio of each batch. When the reliability of all detection edges associated with the training sample is zero, the reliability of the sample is set to zero. The reliability of the sample is used as the weight of the multi-task loss for the training sample, and the model parameters are updated based on the weighted multi-task loss.

[0010] Optionally, S4 includes: the conformal calibration process for forming the constraint boundary includes: constructing duck flock groups according to duck age range, breed, and heat stress level divided by temperature and humidity index thresholds; For each duck flock group, scores for daily weight gain, feed conversion ratio, and health risk non-compliance were constructed based on historical prediction results and actual feeding results. Scores for toxin exposure non-compliance were constructed based on the toxin exposure prediction results of candidate formulas and the toxin exposure test results converted from the actual toxin test values ​​of finished product retests, actual feed intake, and applicable time windows of the duck flocks. Crude protein values ​​of candidate formulas were calculated based on the dry matter-based crude protein test values ​​of each batch of raw materials and the candidate feed ratios. Crude protein deficiency non-compliance scores were constructed based on the crude protein values ​​of the candidate formulas and the actual crude protein test values ​​of finished product retests. The metabolizable energy values ​​of candidate formulations are calculated based on the estimated metabolizable energy values ​​of each batch of raw materials and the candidate feed ratios. The metabolizable energy deficiency non-compliance score is constructed based on the metabolizable energy values ​​of the candidate formulations and the metabolizable energy reference values ​​determined by the nutritional component test values ​​of the finished product. The quantile values ​​of various non-compliance scores are determined according to the basic coverage rate, and the basic calibration boundary is formed. When the number of calibration samples of the target duck group does not reach the preset sample threshold, the upper-level merged group containing the target duck group, the global calibration sample, or the preset conservative boundary are used in sequence, and the quantile value of the target duck group is used after the number of samples reaches the preset sample threshold. Furthermore, the allocation of the joint risk budget includes: determining the susceptibility score of the duck flock based on the age, breed, and current health indicators; converting the current health indicators into current health risk values ​​according to the mapping rules consistent with the training labels of the health risk prediction head; and weighting and aggregating the corresponding mycotoxin detection reliability based on the feeding ratio of each batch of raw materials in the candidate formula to obtain the toxin detection reliability. Within the preset upper limit of the joint default probability, default probability budgets are allocated for crude protein deficiency constraint, metabolizable energy deficiency constraint, toxin exposure constraint and health risk constraint, so that the default probability budget of health risk constraint is negatively correlated with the duck flock susceptibility score and the current health risk value, and the default probability budget of toxin exposure constraint is positively correlated with the reliability of toxin detection. The budget coverage rate corresponding to the constraint is obtained by subtracting the budget of each default probability from one, and the larger of the base coverage rate and the budget coverage rate is taken as the target coverage rate. The daily weight gain non-compliance score and feed conversion ratio non-compliance score continue to use the aforementioned baseline coverage and corresponding baseline calibration boundary. The quantile values ​​of the health risk, toxin exposure, crude protein deficiency, and metabolizable energy deficiency non-compliance scores are reselected according to the aforementioned target coverage. The reselected quantile values ​​form the final upper bound of health risk, upper bound of toxin exposure, lower bound of crude protein, and lower bound of metabolizable energy. Together with the lower bound of daily weight gain and upper bound of feed conversion ratio in the baseline calibration boundary, they form the aforementioned constraint boundary.

[0011] Optionally, S5 includes: using the feed ratio of each batch of raw materials and the applicable time window for the duck flock as decision variables, and the formula cost and the change in the feed ratio of adjacent production batches as objective function terms; The phased nutritional constraints, safety constraints, inventory constraints, raw material usage constraints, and cost constraints, together with the constraint boundaries, form the formulation feasible region. Within the formulation feasible region, candidate formulations that simultaneously satisfy the constraint boundaries are screened, and the feed ratio and applicable time window for duck flocks corresponding to the candidate formulations whose objective function values ​​satisfy the solution termination condition are output. When the feasible domain of the formulation is empty or the solution termination condition is not met within the preset solution time limit, the toxin exposure constraint and health risk constraint remain unchanged, and the validated formulation that simultaneously meets the current constraint boundary, stage nutrition constraint, safety constraint, inventory constraint and raw material usage constraint after being reviewed by the current data is given priority. If the verified formula does not exist, trigger at least one of the following processes: raw material batch replacement, re-detection, or expansion of the candidate formula set, and re-solve. The target formula is not output until a candidate formula that satisfies the formula feasibility domain is obtained through re-solving. Furthermore, cases where the mycotoxin test value of the raw material batch has missing markers or the current evaluation time has reached the expiration time point are considered as abnormal test branches; In the abnormal detection branch, a conservative value that maximizes the predicted value of toxin exposure is selected from the upper limit of the detection method range, the upper quantile of the effective detection value of similar raw materials, and the safety limit conversion value. The toxin exposure contribution of the raw material batch is determined according to the conservative value and the candidate feeding ratio. When the toxin exposure contribution causes the upper limit of toxin exposure to exceed the safety constraint, the raw material batch is prohibited from being used. Otherwise, the upper limit of the raw material batch is tightened to meet the safety constraint, and a re-detection is triggered. Before obtaining a valid retest value, the edge reliability of the raw material batch shall not be improved or the safety constraints shall be lifted by interpolation results; After obtaining a valid retest value, the corresponding edge weight and toxin exposure contribution are updated using the retest value. Under the condition that the established safety limit remains unchanged, the upper limit of the raw material batch is recalculated and the safety constraint is re-determined. The reliability of the corresponding edge is re-determined based on the repeatability of the retest, sampling coverage, detection time point, failure time point and missing marker.

[0012] Optionally, S6 includes: taking the difference between the measured values ​​of nutritional components and toxins in the finished product re-inspection and the estimated value of the finished product calculated based on the test values ​​of each batch of raw materials and the feeding ratio as the manufacturing residual; correcting the edge weights from the raw material batch node to the nutritional component node and the toxin node based on the manufacturing residual; and correcting the corresponding edge reliability based on the degree of dispersion of repeated measurements of the manufacturing residual. The difference between the actual daily weight gain and actual feed conversion ratio of the duck flock and the corresponding prediction results is used as the production performance residual. The actual health indicators are converted into actual health risk results according to the mapping rules consistent with the training labels of the health risk prediction head. The difference between the actual health risk results and the health risk prediction results is used as the health risk residual. The production performance residual and the health risk residual are combined to form the biological response residual and written into the conformal calibration sample of the corresponding duck flock group. Under the condition of retaining the manufacturing residual update results, the quantile value and constraint boundary of the non-compliance score are re-determined.

[0013] The beneficial effects of this invention are: 1. By establishing heterogeneous exposure maps based on raw material batches, nutrition or toxins, formulations, and duck flock time slices, and converting detection repeatability, sampling coverage, detection time, and missing status into edge reliability, batch detection uncertainty can be incorporated into prediction results and formulation constraints along the model processing chain.

[0014] 2. By using graph coding, time-series coding, and cross-fusion controlled by reliability and missing masking, the nonlinear relationship between raw materials, environment, feed intake and production performance is characterized, reducing the over- or under-nutrition caused by static nutritional standards and linear formulation models.

[0015] 3. By dynamically tightening the feasible domain through cluster-based conformal calibration and joint risk budgeting, and maintaining prohibition, limit or re-inspection constraints when key toxin information is missing or expired, the cost of excessive conservatism is reduced while controlling the actual constraint default rate.

[0016] 4. By distinguishing between the manufacturing residuals corresponding to finished product re-inspection and the biological response residuals corresponding to feeding results, the raw material detection boundary and the duck flock grouping calibration boundary are updated respectively, reducing misattribution and formula oscillations in continuous production. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the duck feed formulation optimization method based on multimodal neural networks according to the present invention.

[0018] Figure 2 This is a flowchart of the S4 prediction, conformal calibration, and joint risk budget allocation of the present invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-2 A method for optimizing duck feed formulation based on multimodal neural networks includes: S1. Obtain batch data of duck feed raw materials and time series data of target duck flock, perform dry matter baseline conversion, time alignment and generate missing mask to obtain aligned multimodal data; S2. Based on the aligned multimodal data, establish a heterogeneous exposure map including raw material batch nodes, nutrient component nodes, toxin nodes, formula nodes, and duck flock time slice nodes, and determine the edge weights characterizing exposure amount and time lag, and the edge reliability characterizing detection quality. S3. The heterogeneous exposure graph and the duck flock time slice sequence are encoded using a graph encoder and a time sequence encoder respectively, and cross-fused according to the edge reliability and the missing mask to obtain the fused representation; S4. Generate a candidate formulation set based on the aligned multimodal data, input the features of each candidate formulation and the fused representation into the multi-task prediction head to obtain the prediction results of production performance, health risk and toxin exposure, perform duck flock grouping conformity calibration on the prediction results and allocate joint risk budget to obtain the constraint boundary; S5. Based on the constraint boundary, under the staged nutritional and safety constraints, solve the target formula from the candidate formula set, and output the feeding ratio of each batch of raw materials and the applicable time window for the duck flock. S6. Based on the finished product re-inspection results, generate manufacturing residuals and update the heterogeneous exposure map. Based on the actual feeding results, generate biological response residuals and update the conformal calibration samples and constraint boundaries of the duck flock groups for subsequent formula solving.

[0021] In this specific embodiment, S1 includes: This specific implementation method takes the continuous production cycle of a batch of 42-day-old Cherry Valley ducks as the execution object. The data processing server uses the raw material batch number, duck flock number and duck flock time slice number as the joint primary key to read the crude protein content, moisture content, estimated metabolizable energy, aflatoxin detection value, price, available inventory, detection repeatability, sampling coverage, detection time point, detection validity period, expiration time point, current evaluation time and missing marker of each batch of corn, soybean meal, wheat bran and vegetable oil from the raw material inspection information system. It also reads the duck age, breed, group weight distribution, daily feed intake, health indicators, indoor temperature, indoor relative humidity, formula feeding time, historical daily weight gain and historical feed conversion ratio of the target duck flock from the breeding management system. The raw material field is stored in the raw material batch table along with the batch record, and the duck flock field is stored in the duck flock time slice table according to the natural day. The raw material inspection information system will form a wet basis determination value by taking the median of three parallel test results for the same batch of the same indicator. The system will also record the coefficient of variation of repeated tests, the ratio of the number of sampled subsamples covering the batch to the number of specified subsamples, the test method number and the unit of measurement. Crude protein and moisture content will be expressed as mass percentages, metabolizable energy as MJ / kg, aflatoxin as μg / kg, price as yuan / t, and inventory as t. Records that lack a legal unit of measurement or batch number will be directly marked as invalid records and will not be included in subsequent numerical conversion. The server performs dry matter-based conversion on the nutrient or toxicity assay fields that need to be converted, using the formula... ,symbol Indicates raw material batch Indicators The corresponding dry basis measurement values ​​have the same units as the original wet basis measurement values, and the symbols are... Indicates the wet basis measurement value read by the testing system, symbol Indicates raw material batch The moisture content is a dimensionless number within the range [0,1), with the symbol […]. Indicates raw material batch index, symbol This indicates the index of crude protein, metabolizable energy, or mycotoxin indicators. When the moisture content is missing, the moisture content reaches 1, or the test record is invalid, the division is not performed and the corresponding dry matter basis field is written as null. In this specific implementation, the time alignment origin is set as 00:00 on the day of formula production. A duck flock time slice is defined as 24 hours. The raw material storage time falls into the most recent valid time slice before the production date. The formula production time is written into the production time slice. The duck flock's feeding and indoor environment records are assigned to the corresponding natural day time slice according to the collection timestamp. For the same time slice, the median weight is taken, the total feed intake is taken, the average of the house temperature and relative humidity is taken according to the 10-minute sampling sequence, and the health index is taken as the most unfavorable level of the day. The sample size and start and end timestamps of each aggregated field are retained. The server generates a missing mask for each raw material batch indicator and duck flock time slice field. The corresponding mask is set to 1 when the original value does not exist, the unit of measurement is invalid, the current evaluation time reaches or exceeds the expiration time point, or the timestamp falls outside the target production cycle. Otherwise, it is set to 0. For continuous variables, the median of the same breed and age range of ducks in the training set is used as the neural network input placeholder value for null values. Categorical null values ​​use independent missing category encoding. The placeholder value does not change the missing mask and is not used as the detection basis for removing toxin safety constraints. For the weight distribution of the flock, the server retains the weight of the 5%, 25%, 50%, 75% and 95th percentiles and the number of birds weighed effectively. When the number of birds weighed is less than 30% of the flock’s stock, the weight distribution missing mask is set to 1. For the temperature and humidity inside the shed, when the effective sampling coverage within the time slice is less than 18 hours, the environmental field missing mask is set to 1. For the feed intake, when the feeding record and the leftover feed weighing record cannot be closed, the feed intake missing mask is set to 1 and the unclosed quality is written in the quality verification field. The server calibrates the maximum permissible deviation threshold for 12 hours based on the 95th percentile of the absolute time difference between the valid business records in the training set and the center of the daily time slice. It constructs the current time slice and its adjacent time slices as a set of candidate time slices. When the absolute time difference between the raw material warehousing, formula production, and duck flock feeding records and the center of the candidate time slice does not exceed 12 hours, it is written into the time slice. When all exceed the threshold, it is marked as unmatched and the corresponding missing mask is set to 1. When multiple candidates meet the threshold, the first candidate is selected in ascending order of absolute time difference. If the time differences are the same, the earlier time slice is selected. The matching results, threshold sources, and conflict rules are written into the alignment parameter file. The server uses the field order frozen during model training to perform robust standardization on continuous variables. The standardization center is the median of the corresponding field in the training set, and the scale is the interquartile range of the corresponding field in the training set. When the scale is zero, it is changed to 1. The standardization result is truncated to [-5,5]. Categorical variables use dictionary encoding with fixed versions. All centers, scales, category dictionaries, units of measurement, and effective times are stored in an alignment parameter file with version number ALIGN-2026-01. A new version is generated only after the training set is updated and verified by leaving time for production. After the transformation and alignment are completed, the server outputs the aligned multimodal dataset according to the composite primary key. The records contain raw material batch feature vectors, duck flock time slice feature vectors, field-level missing masks, original units of measurement, transformed units of measurement, timestamps, expiration timestamps, and alignment parameter version numbers. For duplicate records with the composite primary key, the record with the latest update time and valid verification status is retained. The aligned multimodal dataset is written to the formulation task cache and serves as the sole data input for S2 to construct the heterogeneous exposure map.

[0022] In this specific embodiment, S2 includes: The graph construction module reads the aligned multimodal dataset written in S1, establishes a raw material batch node for each raw material batch, establishes nutrient component nodes or toxin nodes for crude protein, metabolizable energy, and various mycotoxin indicators, establishes formula nodes for historical formulas or current benchmark formulas with actual feeding records in S1, and establishes duck flock time slice nodes for each natural day for which observations have been formed up to the current evaluation time for the target duck flock. This step does not read, assume, or generate candidate numbers for S4. The node primary key is formed by concatenating the node type and business number. The node features simultaneously store the numerical vector, missing mask, unit of measurement, data version, and last update time. The graph construction module establishes feeding edges between the raw material batch node and the historical formula or current benchmark formula node, and writes the actual batch feeding ratio, feeding quality and production time. It establishes detection edges between the raw material batch node and the nutrient component node and toxin node, and simultaneously writes the wet basis test value, dry basis test value, moisture value, detection time point, failure time point, detection repeatability and sampling coverage. It establishes feeding edges between the formula node and the observed duck flock time slice node, and writes the actual feed intake, formula feeding time, duck flock number and time slice number. Each edge is deduplicated using the source node primary key, target node primary key, edge type and production batch number. For raw material batches Indicators In time slice The resulting exposed edges are first detected by the graph construction module. Calculate the total dry matter mass ratio of the formulation for this time slice, and then use... Calculate batch The dry matter feed mass ratio, and adopted Calculate edge weights; symbol Indicates time slice The total dimensionless dry matter mass ratio of the historical or current benchmark formulation used, with the symbol [notation missing]. This represents the set of raw material batch indexes that were actually fed into the system during that time slice and passed the primary key integrity check. The summation process iterates through this set, using the symbol... Indicates batch in the set The wet-based feed mass ratio, symbol This indicates the batch stored in S1. Moisture content, symbol Indicates batch The dimensionless dry matter feed ratio within the time slice [0,1] is represented by the symbol. Indicates batch The wet-based feed mass ratio, symbol This indicates the batch stored in S1. Moisture content; symbol This represents the dry matter basis detection exposure edge weight, which includes the dry matter feed ratio and time-delay decay. Its units are the same as the corresponding dry matter basis measurement values, and its symbol is... This represents the dry matter basis measurement value converted and stored by S1, with the symbol... This indicates the time from the start of this batch of material feeding to the time slice. The interval between feeding midpoints, in units of d, symbol Indicators The time-delay decay time constant, in units of days, is obtained by fitting the exponential decay of historical review and feeding records. The function... This represents an exponential function with base e and dimensionless input. The corresponding output is located in (0,1], symbol , , and These represent the target raw material batch index, the summation raw material batch index, the index, and the observed duck flock time slice index, respectively; when If the moisture content is less than or equal to 0, or if the moisture content of any batch used is missing or the corresponding dry matter basis measurement value is empty, the division will not be performed, the reliability of the relevant test exposure edge will be set to 0 and written into the pending verification flag; In this specific embodiment, the time constants for the decay of crude protein and metabolizable energy are both set to 7 days, and for mycotoxins to 14 days. These parameters are stored in the edge parameter file with version number EDGE-2026-01. When the time interval is less than 0, the corresponding feeding association is deleted. When the sum of the feed ratios deviates from 1 by more than 0.0001, the formula node is marked as an abnormal ratio and the subsequent graph coding of the formula is stopped. The detection edge weight retains the original index unit so that it can be normalized according to the edge type in the future. Graph building module according to Calculate the reliability of the detection edges used in the current solution, sign Indicates raw material batch Indicators The reliability of the detected edge at the current evaluation time is a dimensionless number within [0,1], with the sign... This represents the detection repeatability obtained by mapping the coefficient of variation of three parallel detections, and is a dimensionless number within [0,1]. The symbol is... The sampling coverage is represented by dividing the actual number of samples by the specified number of samples and truncating to [0,1]. The symbol is... This represents the timestamp of the test report issuance stored in the laboratory information system, symbol [symbol missing]. This represents the expiration timestamp obtained by adding the validity period registered in the testing method document to the issuance timestamp, with the symbol... This represents the current evaluation timestamp frozen when the current recipe solving task was created. All three time values ​​are uniformly converted to 'd' starting from the same epoch. (Symbol: d) This represents the missing mask generated by S1 at that evaluation time, and it takes the value 0 or 1. The function... This indicates that the effective time proportion is truncated to [0,1]. and The raw material batch index and index in the adjacent definition of the previous detection exposure edge formula are directly used; the reliability of the historical time slice is only used for offline training snapshots and is calculated separately according to their respective frozen evaluation times; S3 and S4 only read the current solution reliability of this formula. when Less than or equal to 0, missing mask is 1, or the same current evaluation time When the failure time point is reached, the graph construction module does not use the division result and directly sets the reliability of the current solved detection edge to 0. When the detection repeatability or sampling coverage field is missing, the reliability is also set to 0 and the detection edge is retained to mark the missing detection position. The placeholder value of the non-toxin field only enters the node feature and its reliability is still 0. The toxin field is simultaneously written to the detection anomaly flag bit as pending re-inspection status. For the feeding edge from the formula node to the duck flock time slice node, the edge weight is the product of the actual feed intake of each duck in that time slice and the feeding ratio corresponding to the formula node, in kg / (d). If the actual feed intake is not obtained, the feeding edge is retained and its reliability is set to 0. The reliability of the feeding edge is the product of the feed scale calibration status, the completeness rate of the weighing record, and the completeness rate of batch traceability. If all three are within [0,1] and any one of them is invalid, the reliability of the feeding edge is set to 0. The graph construction module uses the complete consistency of node business numbers as the main matching rule, and uses the difference between production time and feeding time not exceeding the candidate applicable time window as the matching threshold between formula nodes and duck flock time slice nodes. Records with inconsistent business numbers or time differences exceeding the threshold are not established with feeding edges. When multiple formula nodes are matched in the same time slice, a unique edge is retained according to the formula number in the actual feeding record. If a formula number is missing, the time slice is marked as pending verification and is not included in model training. The graph construction module calculates the 1% and 99% quantile values ​​of the effective edge weights in the training set according to the edge type. The detection edge, feeding edge, and feeding edge are truncated within their respective measurement categories and mapped to [0,1] as the normalized edge weights for neural network input. The original edge weights, original units, normalized edge weights, and edge reliability are stored in the edge record. The quantile value file is generated only from the training data and bound to the model version to avoid directly adding edge weights of different measurement units. After construction, the module uses the duck flock number, production batch number and graph version number as the identifier of the heterogeneous exposure graph. It writes the node table, edge table, edge type dictionary, original edge weight, normalized edge weight, edge reliability, missing mask and anomaly flag into the graph object storage. Abnormal edges are not deleted but participate in subsequent gating with zero reliability. The heterogeneous exposure graph and the duck flock sequence index sorted by time slice are jointly passed to S3.

[0023] In this specific embodiment, S3 includes: The model inference module reads the heterogeneous exposure map and duck flock time slice sequence output by S2. In this specific implementation, a two-layer relational graph attention network is selected as the graph encoder. An independent linear transformation matrix is ​​set for each edge type. The input features of raw material batch, nutrient components, toxins, formula and duck flock time slice nodes are first mapped to 64-dimensional vectors. When the number of nodes exceeds the training limit, the normalized edge weight of the formula node is retained by multiplying it by the 128 detection edges with the highest edge reliability. The remaining edges are recorded as pruned edges and do not participate in this aggregation. opposite edge type Next node Pointing to node The adjacent edges of the graph encoder are used. The formation of unnormalized attention scores, symbolic This represents the dimensionless attention score of the adjacent edge, with the symbol [symbol missing]. Representing edge type The corresponding 130-dimensional trainable attention column vector, symbol Represents the transpose of a vector or matrix, symbol Representing edge type The corresponding 64×64-dimensional trainable feature transformation matrix, symbol and symbols Representing the target node respectively and adjacent nodes The 64-dimensional input feature column vector, symbol The normalized edge weights of [0,1] obtained by S2 are represented by the symbol. This represents the reliability of the [0,1] edge obtained by S2, after... The two 64-dimensional vectors transformed are concatenated with two scalars to form a 130-dimensional column vector, since the features of each node have been standardized according to S1 and and All of these are neural network training parameters. The vectors, matrices, and attention scores mentioned are not expressed in physical units, and the symbols are... Indicates vector concatenation, symbol , and These represent the target node index, adjacent node index, and edge type index, respectively. The graph encoder performs softmax on the unnormalized scores within the adjacency set of the same target node and the same edge type, and then multiplies the obtained coefficients by the corresponding edge reliability before renormalizing. When the sum of all reliability weighted coefficients under the edge type is 0, the aggregate vector of the edge type is set to a zero vector. The aggregate vectors of each edge type are concatenated after residual connection, layer normalization and ReLU activation. The second layer aggregates according to the same rules and reads the 64-dimensional output of the recipe node as the recipe batch exposure representation. Therefore, the detection edge with a reliability of 0 retains the graph structure but does not transmit the measured value. The time encoder uses a two-layer gated recurrent unit network. The input is a 48-dimensional vector of the 14 most recent time slices of the target duck flock arranged in chronological order. The vector includes the quintiles of body weight, feed intake, health indicators, temperature and humidity, historical daily weight gain, historical feed conversion ratio, time slice missing mask, and effective sampling coverage. Continuous fields are standardized using the alignment parameter S1. The placeholder value and mask of missing fields are input simultaneously. Each layer has a hidden dimension of 64 and uses the output of the last effective time slice as the duck flock state representation. When there are fewer than 7 valid feed intake records in 14 time slices, the time encoder still outputs the flock state representation but writes a low coverage flag into the state record. When all time slices are invalid, the state prototype vector of the training sample of the same breed and age range is used instead and the prototype source version is written into the record. The low coverage flag and prototype flag are then used as input to the cross-fusion gating network instead of being treated as real observations. The cross-fusion network first normalizes and weights the reliability of all incoming edges of the current historical recipe or current benchmark recipe node according to the corresponding feed ratio to obtain the recipe-level average reliability, and then calculates... and ,symbol Indicates the observed time slice The 64-dimensional cross-fusion gating coefficients have each component located in (0,1), and the sign is... This represents a 64×130 dimensional gated training weight matrix, with the symbol... Represents a 64-dimensional biased column vector, with the symbol... The output of the graph encoder represents a 64-dimensional historical or baseline recipe exposure representation, with symbols. This represents the 64-dimensional duck flock state representation output by the timing encoder, symbol... Represents the formulation-level average reliability, which is a dimensionless number within [0,1], with the symbol... This represents the average of the missing mask of the current input field and is a dimensionless number within [0,1]. Two 64-dimensional representations are concatenated with two scalars to form a 130-dimensional input. (Symbol) This represents the 64-dimensional baseline exposure-duck state fusion representation of the output. The above representation vectors, training matrices, and biases are all formed from standardized features or obtained through network training, and none of them use physical units of measurement. The symbols... Represents the element-wise Sigmoid function, symbol Indicates element-wise multiplication, index symbol This indicates the time slice index of the observed duck flock; To ensure that gating direction aligns with data quality, a gating monotonicity regularization term is added during training. This ensures that when the recipe-level average reliability increases while other inputs remain constant, the gating mean corresponding to the recipe batch exposure representation does not decrease. During inference, bimodal availability is assessed first. When the recipe-level average reliability is 0 and all temporal inputs are invalid, the highest priority bimodal unavailability branch is executed. Conflicting gating cutoffs are not calculated, and no new gating branches are generated. The fusion representation is marked as unavailable and the graph version, prototype version, and reason for the missing information are sent to S4. S4 then prohibits the output of the target recipe and triggers static conservative boundary and manual safety review. The upper limit of the gating coefficient is truncated to 0.05 only when the average reliability at the recipe level is 0, and the lower limit of the gating coefficient is set to 0.95 only when all timing inputs are invalid. The priority, truncation rules, and unavailable status are fixed with the model version and written into the fusion record. During training, the graph encoder, temporal encoder, multi-task prediction head, and cross-fusion gating network used raw material batches, historical formulas, duck flock time slices, finished product re-inspection, and actual feeding records from the past 18 months. The production batches were divided into 60% training set, 15% validation set, 10% test set, and 15% independent calibration set according to the production time and with each production batch as an indivisible unit. The production batches in the four sets were mutually exclusive. Model parameters and hyperparameters were determined only using the training and validation sets. The test set was used only for publication evaluation. The independent calibration set only received predictions frozen and saved by the corresponding model version before the results were generated. Daily weight gain and feed conversion ratio used Huber loss, health risk used four-category cross-entropy loss with class weights, and toxin exposure used Huber loss after logarithmic transformation. The training procedure uses the sum of weighted multi-task loss and gated monotonicity regularization term as the optimization objective function. Each batch contains 64 production samples and uses the Adam optimizer to iterate with an initial learning rate of 0.001. Training stops when the validation loss does not decrease for 10 consecutive rounds. The model version MODEL-2026-01 is only released when all four tasks on the validation set are not inferior to the current production version and there are no new violations in the toxin safety constraint replay in the last 90 days. After release, if the error of any task exceeds the 95th percentile of the test set in two consecutive production batches, the model will revert to the previous production version and freeze the input of the new version. After the revert, at least 3 production batch replays must be completed and the errors of all four tasks must return to within the 95th percentile of the test set before the freeze is lifted and the new version inference is resumed. For training samples The training program adopts Calculate sample reliability, sign Indicates sample The dimensionless reliability of [0,1], symbol Indicates training samples The actual set of raw material batch indexes that have passed batch primary key integrity verification, both summations traverse this set, and the symbol... Indicates sample Chinese raw material batches The feeding ratio is dimensionless, and the symbol is... The minimum value of toxin and nutritional indicators representing the reliability of batch-related detection of this raw material is a dimensionless number within [0,1]. The symbol is... and symbols These represent the training sample index and the raw material batch index, respectively. When the sum of the feeding ratios of each raw material batch in the training sample is 0 or the reliability of all associated detection edges is 0, the sample reliability is directly set to 0. The training program adopts Form a multi-task loss for samples, with sign Indicates sample The dimensionless weighted multi-task loss, the symbols in this formula We directly adopt the sample reliability formula from the previous sample reliability formula and the sample reliability defined in its immediate adjacent definition, with the symbol... , , and These represent daily weight gain, feed conversion ratio, health risk, and task loss due to toxin exposure, all normalized to the training set. (Symbols) , , and The four tasks are represented by their dimensionless weights, and in this specific implementation, they are all set to 0.25. The reliability of the detection edge with missing markers or that has reached the failure time point is set to 0 before calculating the sample reliability. The reliability of the other edges in the same sample remains unchanged. The inference module writes the baseline exposure-duck state fusion representation, model version, recipe-level average reliability, time series low coverage flag, and node pruning list into the fusion representation record. The fusion representation record is entered into S4 with the duck number and the currently observed time slice number as the primary key, and does not include candidate numbers that have not yet been generated. The multi-task prediction head only reads the fusion representation generated by the same model version. When the model input field versions are inconsistent, prediction stops and rolls back to the previous production version that has been validated.

[0024] In this specific embodiment, S4 includes: The candidate generation module first reads the aligned multimodal data of S1, the current inventory, and the phased nutritional standards of the duck flock. With a minimum step size of 0.5 mass percentage for the feed ratio, it generates feed ratio vectors for the allowed raw material batches and assigns candidate numbers. First, it ensures that each ratio is non-negative and the sum is 1. Then, it performs step-by-step filtering according to the available inventory, the upper and lower limits of single raw material addition, the lower limit of crude protein screening, the lower limit of metabolizable energy screening, and the upper limit of toxin concentration screening. After the candidate generation is completed, it reads the baseline exposure-duck flock status fusion representation of S3 according to the duck flock number and the currently observed time slice number. Therefore, S2 and S3 do not depend on the candidate number. The candidate record includes the candidate number, the feed ratio of each batch, the expected production quality, the start time slice, and the end time slice. In this specific implementation, the applicable time window for duck flocks is set to 3 to 7 consecutive natural days. The start time slice is no earlier than the day after the formula production and the end time slice is no later than the end of the current breeding stage. The same feed ratio vector is combined with each legal time window to form a candidate formula. When the number of candidates exceeds 20,000, the top 15,000 are retained in ascending order of initial screening cost. Then, the top 5,000 are retained in ascending order of the sum of the absolute differences between the feed ratios of the previous production batch and the unselected candidates. If the cost or distance is the same, the candidate number is selected in lexicographical order. The final set is fixed as the union of the two levels of results. The multi-task prediction head consists of a shared two-layer fully connected network and four task branches. The inputs of the shared layer are candidate feed ratios, candidate time window statistics, the 64-dimensional baseline fusion representation of S3, the formula-level average reliability, the current flock time slice features, and candidate prospective exposure features. The prospective exposure features are calculated using candidate feed ratios and the planned feed intake of candidate time windows given by the frozen feed intake prediction model before the results are generated. When the predicted feed intake is missing or has low coverage, the conservative planned feed intake fixed according to breed, duck age, and heat stress level is used and a low confidence flag is written. The future actual feed intake that has not yet occurred is not used. The dimensions of the hidden layers are 128 and 64 respectively. The four task branches output the predicted daily weight gain, predicted feed conversion ratio, four-category probability vector of health risk, and predicted toxin exposure, respectively. The daily weight gain unit is g / (bird·d), the feed conversion ratio is a dimensionless number, the sum of the probabilities of the four categories is 1, and the toxin exposure is expressed as μg / (kg body weight·d). The health risk training labels are formed by mutually exclusive categories of four levels of status in veterinary records: normal, attention, intervention, and severe. The health risk branch first outputs the probabilities of the four categories in a fixed order through softmax and trains them with four-class cross-entropy with category weights. Then, the four probabilities are weighted and summed with 0, 0.33, 0.67, and 1 as category risk weights, respectively, to obtain a unique health risk scalar in [0,1]. Isochronous regression only performs calibration on this scalar. The current health indicators and the actual health results of S6 use the same category order, weights, and mapping versions. The category probabilities, scalar results, and version numbers are written into the candidate prediction record. The toxin exposure prediction uses the measured exposure obtained by multiplying the actual feed intake by the wet basis concentration of the finished product retested toxin and dividing by the median weight of the duck flock in the time slice as the training target. The calibration module groups duck flocks according to age range, breed, and heat stress level. The age ranges are 1 to 14 days, 15 to 28 days, and 29 days to slaughter day, respectively. The heat stress level is determined by... Calculation, symbol This represents the dimensionless temperature and humidity index obtained according to THI-2026-01 version, with the symbol... This represents the average temperature in Celsius over the time window after the sensor passes quality control, with the unit being °C. (Symbol: ) This represents the average relative humidity percentage within the range [0,100] within the same time window. Constants 1.8, 32, 0.55, 0.0055, and 26 are fixed conversion factors for this version; calculations are rounded to two decimal places. Classified into the low-heat stress group, Classified into the moderate heat stress group, Ducks were classified into the high heat stress group, while those with boundary values ​​of 72 and 78 were classified into the middle group. No grouping was performed and manual review was triggered when the input was missing or exceeded the limit. The grouping primary key was formed by concatenating three fields: duck age range, breed, and heat stress level. Calibration samples used only the independent calibration set defined by S3 or the out-of-bounds predictions obtained strictly according to the production batch. Calibration batches were not allowed to enter the parameter training, validation, or parameter tuning of the corresponding model version. Only predictions that had been frozen and saved before the actual results were generated were accepted. The model version, prediction time, result time, and dataset attribution were recorded. When overlapping sets were found or when recalculation was performed afterward, the record was isolated. For calibration samples and prediction tasks The calibration module adopts Calculate the one-sided non-compliance score, sign Indicates sample In the mission The function has a non-compliance score and the unit is the same as the task output. This indicates that the value of the parameter in parentheses is the largest. This represents the task orientation coefficient. The lower bound for daily weight gain, crude protein, and metabolizable energy is set to 1, while the upper bound for feed conversion ratio, health risk, and toxin exposure is set to -1. The symbol is... This indicates a historical predicted value frozen and stored before the result was generated, or a formulation estimate calculated based on the raw material test values ​​and candidate proportions of the same wet-basis conversion version. The symbol is... This indicates the corresponding actual feeding results, wet-based finished product retest value, or metabolizable energy reference value based on the same standard, with the symbol... and symbols These represent the calibration sample index and the task index, respectively. For candidate formulations The candidate generation module adopts Calculate the total mass proportion of candidate dry matter, and use... Calculate the dry matter feed ratio for each batch, and then use the following methods respectively: and Calculate the dry matter-based crude protein estimate and dry matter-based metabolizable energy estimate for candidate samples; (symbols) Indicates candidate The total dimensionless dry matter mass ratio, symbol Indicates candidate The actual set of raw material batch indexes that have passed primary key integrity checks, and the three summations all iterate through this set, with the symbols... Indicates candidate Medium batch The wet-based feed mass ratio, symbol This indicates the batch stored in S1. Moisture content, symbol Indicates candidate Medium batch The dimensionless dry matter feed ratio within [0,1], symbol Indicates candidate The dry matter-based crude protein estimate in mass percentage, symbol [symbol missing]. Indicates candidate The estimated dry matter-based metabolizable energy is given in MJ / kg dry matter, with the symbol [symbol missing]. and symbols These represent the batches stored in S1. Dry matter-based crude protein and dry matter-based metabolizable energy values, symbols and symbols These represent crude protein and metabolizable energy indicators, respectively, with symbols... and symbols These represent the candidate index and the raw material batch index, respectively; when When the moisture content is less than or equal to 0, missing, or the corresponding dry matter basis field of any batch used is empty, division is not performed and the corresponding nutrient task of the candidate is marked as uncalibrable; the prediction of crude protein and metabolizable energy, the finished product re-inspection value and the demand threshold are all converted to the same dry matter basis and the same unit before constructing non-compliance and nutrient constraints; the mycotoxin safety calculation still uses the same wet basis concentration, wet basis feed ratio and wet basis safety limit; the calibration record saves the source of moisture value, original benchmark, converted benchmark, conversion rule version and abnormal branch; The calibration module is only used in groups. Number of valid calibration samples If the value is greater than or equal to 1, the group will be placed in the task. The non-compliance scores are sorted in ascending order and then... and Select quantile value, sign Indicates grouping of ducks In the mission and coverage The calibration quantile value is below, and the unit is the same as the task output, with the sign... Indicates grouping Medium task The A number of non-compliance scores in ascending order, with symbols... Indicates the quantile position index, symbol Indicates grouping Number of valid calibration samples, symbol Represents the target coverage within (0,1), symbol Indicates the duck flock grouping index, index symbol The prediction task index from the previous one-sided non-compliance formula and its interpretation is directly adopted, with the symbol... This indicates taking the smallest value, symbol Indicates rounding up; The baseline coverage is set to 0.90. The calibration module backs down layer by layer in a unique order: when the number of valid samples in the target duck flock group reaches 60, the target group quantile value is used; when it is less than 60, the upper-level groups of the same breed and adjacent age range with the same heat stress level are merged; when the number of valid samples in the upper-level group reaches 60, the upper-level group quantile value is used; otherwise, it is transferred to the global sample; when the global sample reaches 60, the global quantile value is used; when the global sample is 1 to 59, the maximum non-compliance of the corresponding task in the independent playback data of the past two years is used; when the global sample is 0, regardless of whether there are 1 to 59 records remaining in the target group or the upper-level group, it enters the static boundary branch. The static conservative boundary that has been manually reviewed and version-fixed is read first according to the task, breed applicable scope and unit of measurement; when the playback data is empty or the static boundary is missing, the task is set to calibration unavailable, the target formula is prohibited from being output, and supplementary testing and manual safety review are triggered. The boundary record saves the source, validity period and review responsibility. The next production batch after the number of samples in each layer reaches 60 will switch to the corresponding level quantile value. The joint risk budget module first calculates the susceptibility score of the duck flock based on the age range of ducks, the breed baseline susceptibility coefficient and the current health risk value, and then cuts it to [0,1]. Then, it normalizes and weights the reliability of mycotoxin detection for each batch according to the candidate feed ratio to obtain the reliability of toxin detection. When the sum of all ratios is 0, the reliability of toxin detection is set to 0. The susceptibility mapping table and the breed baseline coefficient are fitted from the veterinary intervention records of the past 18 months and stored in the model parameter library in the RISK-2026-01 version. For the four constraints of insufficient crude protein, insufficient metabolizable energy, toxin exposure, and health risks, the budget module adopts... , , and Allocate default probability budget, symbol , , and The positive weights of the constraints of health risk, toxin exposure, insufficient crude protein, and insufficient metabolizable energy are respectively represented by the function. This represents an exponential function with base e and a dimensionless input corresponding to a positive dimensionless output, denoted by the symbol e. The susceptibility score of the duck flock within the range [0,1] is represented by the symbol. Represents the current health risk value within the range [0,1], with the symbol... The symbol represents the reliability of endotoxin detection in the range [0,1]. , and symbols These represent the dimensionless sensitivity coefficients obtained from the historical constraint replay grid search, and in this specific implementation, they are all taken as 1. Representing constraints Budget for the probability of default, symbol This represents the upper limit of the joint default probability, set to 0.10, with the symbol... Representing constraints The corresponding positive weights, signs Represents a four-term constraint index; The above allocation ensures that the sum of the four default probability budgets equals the upper limit of the joint default probability. Increased flock susceptibility or current health risk reduces the health risk constraint budget and increases its coverage, while increased toxin detection reliability increases the toxin exposure constraint budget. The budget module adopts... Determine the target coverage, sign Representing constraints Target coverage, function The maximum value function defined in the immediate definition of the one-sided non-compliance formula in this step is directly adopted, with the sign... Directly apply the previous default probability budget formula and the constraints defined in its immediate adjacent interpretation. Default probability budget, index symbol Using a four-item constraint index within the same definition, symbols This indicates a base coverage rate of 0.90. The base coverage rate is always used for daily weight gain and feed conversion ratio tasks. The calibration module uses the task prediction value minus the corresponding quantile value to form the lower bound of daily weight gain, crude protein, and metabolizable energy. It uses the task prediction value plus the corresponding quantile value to form the upper bound of feed conversion ratio, health risk, and toxin exposure. The upper bound of health risk is truncated to [0,1]. Each boundary record saves the candidate number, duck flock grouping primary key, sample quantity, coverage, quantile value, unit, calibration sample cutoff batch, and calibration version. When the fusion representation of candidate records carries a low-coverage time-series flag, the health risk and production performance tasks additionally use the maximum value between the upper-level merged group quantile value and the current group quantile value. When the toxin detection reliability is 0, the toxin exposure constraint directly uses the global conservative boundary and the candidate record is written to the pending re-inspection flag. All prediction results, basic calibration boundaries, joint risk budgets, and final constraint boundaries are written to the candidate constraint table and passed to S5.

[0025] In this specific embodiment, S5 includes: The formula solving module reads the candidate constraint table of S4, takes the feed ratio of each raw material batch and the applicable start and end time slices of the duck flock in the candidate records as the decision objects, first reads the lower limit of crude protein and metabolizable energy requirements, the range of essential amino acid ratios and the maximum allowable feed conversion ratio according to the duck flock stage, reads the upper limit of exposure to each mycotoxin according to the regulatory safety parameter library, reads the number of available batches and production batches according to the inventory system, and writes the parameter library version, regulatory effective time and inventory snapshot time into the solving task; For candidate formulations The solution module uses the formula cost and the change in the feed ratio between adjacent production batches as the objective function terms, and employs... Calculate the objective function, symbol Indicates candidate formulations The objective function value is given in units of yuan / t, and the symbol is... This represents the union of the currently allowed batches and the batches used in the previous production batch. Both summations iterate through this set. The proportion of batches not present in a given recipe is counted as 0. Indicates raw material batch The factory gate price is expressed in yuan / ton, with the symbol... Indicates candidate formulations Medium batch The wet basis feed mass ratio is a dimensionless number within [0,1] and sums to 1 within the valid batch set, with the sign... This represents the penalty coefficient for changes in the proportion of adjacent production batches. It is a dimensionless number and is set to 0.08. This represents the median price of currently available raw material batches, expressed in yuan / ton, with the symbol... Indicates the batch number in the previous produced batch. The wet basis feed mass ratio is a dimensionless number within [0,1] and sums to 1 within its valid batch set, with the symbol... This indicates taking the absolute value of the value within the parentheses, with the sign... , and These represent the candidate formula index, raw material batch index, and previous production batch identifier, respectively. The solution module checks each candidate in turn: the proportion of each feed is non-negative, the deviation of the sum of the proportions from 1 does not exceed 0.0001, the planned production quality multiplied by the feed proportion does not exceed the available inventory of the corresponding batch, the feed proportion of a single raw material is at the upper or lower limit of the process addition, and the expected formula cost does not exceed the upper limit of the task cost. If any hard constraint is not met, the candidate is deleted from the feasible region and the reason for deletion is recorded. For candidates that pass the basic constraints, the solution module requires that the lower bounds of crude protein and metabolizable energy given by S4 be no less than the requirements of the duck flock stage, the lower bound of daily weight gain be no less than the lower limit of the production task, the upper bound of feed conversion ratio be no more than the upper limit of the production task, and the upper bounds of health risk and toxin exposure be no more than the corresponding upper limits of regulations and the farm safety parameter database. All comparisons are performed under the same unit of measurement. If the parameter units are different, they are converted according to the fixed conversion factor registered in the parameter database before comparison. For branches with abnormal detection, when raw material batches When the mycotoxin detection value has a missing marker or the current evaluation time has reached the expiration time, the solution module uses... Determine the concentration of conserved toxins, symbol Indicates batch The conserved toxin wet basis concentration in μg / kg, function The maximum value function defined in the immediate vicinity of the S4 one-sided non-compliance formula is directly adopted, with the sign... Indicates the upper limit of the measurement range of this detection method, symbol This represents the 95th percentile empirical value of the effective wet basis test results for similar raw materials over the past 12 months, with the symbol [symbol missing]. This indicates that the established safety limit is converted to the same wet-based feed concentration based on the median weight of the candidate duck flock, the planned or predicted daily feed intake, and the applicable time window. All three are in the same unit, and the conservative value is recorded and saved with the data source, benchmark, and calculation time. Abnormal batches for candidate formulations The contribution of toxin concentration was adopted Calculation, symbol Indicates abnormal batch For candidate formulations The contribution of the finished product to the wet basis concentration of toxins, expressed in μg / kg, symbol. This indicates the dimensionless wet-based feed ratio in the candidate sample, denoted by [symbol]. The batch size is directly defined in the preceding conservative toxin concentration formula and its immediately adjacent definition. Conservative toxin wet basis concentration; if the sum of this contribution and the contributions of other batches, after conversion of the planned or predicted feed intake and the median weight of the duck flock frozen before the results are generated, causes the upper limit of toxin exposure to exceed the safe limit, then the feed ratio of the abnormal batch in the candidate is set to the prohibited value of 0 and the candidate that does not meet the conditions is deleted. When there is still positive safety margin for inclusion in the abnormal batch, the solver module first uses... Convert the regulatory daily exposure limit to the finished product wet basis concentration limit corresponding to the candidate time window, sign Indicates candidate The upper limit of the wet basis concentration of the finished product toxin, in μg / kg, symbol This indicates the daily exposure limit under regulations, expressed in μg / (kg body weight·d), with the symbol... This represents the median body weight of the candidate duck flock frozen within the time window before the results were generated, in kg. (symbol) This indicates the planned or predicted daily feed intake for each duck frozen within the same time window, expressed in kg / d; when Greater than 0 and If it is greater than 0, then use Calculate the upper limit of admixture in abnormal batches, sign Indicates candidate Abnormal batches The upper limit of the dimensionless wet matrix doping ratio, symbol Indicates the upper limit of process admixtures in this batch, symbol Indicates candidate The total number of valid raw material batches is summed and iterated from 1 to... Number and Index Not equal to Other valid batches, symbols Represents the batch index of other raw materials, function sum function These represent taking the minimum and maximum values ​​of the parameters, respectively. Other symbols directly adopt the adjacent interpretations of the aforementioned conservative toxin concentration and objective function formulas. Output is disabled when the value is less than or equal to 0. When the value equals 0, only the upper limit of the process is used; the solver module locks the proportion of abnormal batches to no higher than 0. The remaining quality proportion is allocated among other batches according to the original ratio and each process and inventory limit is met. If allocation is not possible, the candidate is deleted. After reconstruction, all constraints are re-verified from the prediction, calibration and constraint boundary process of S4, and the toxin exposure upper limit of the original candidate is not directly used. The feed intake, body weight, time window, regulatory version and parameter version used for conversion are saved together. Before a valid retest value is obtained, the abnormal batch record remains in a state of pending retest. Any interpolation value will not improve the detection edge reliability of S2, reduce the conservative toxin concentration, or remove the safety constraints. After the retest value is obtained, the corresponding detection edge weight, reliability, and toxin contribution will only be updated when repeatability, sampling coverage, detection time point, failure time point, and missing marker all pass the S1 verification. The established regulatory safety upper limit remains unchanged and the doping upper limit is re-determined. The solution module will sort all candidates by objective function value in ascending order. If the difference in objective function value does not exceed 0.01 yuan / t, the candidate with the smaller upper bound of toxin exposure will be selected. If they are still the same, the candidate with the smaller upper bound of health risk will be selected. If they are still the same, the candidate with the earlier lexicographical order of candidate number will be selected. The first candidate will be used as the target formula and the module will output the batch feeding ratio of each raw material, the planned feeding quality, the applicable start and end time slices for the duck flock, the constraint margin, the parameter version, and the solution time. The solution status record includes the total number of candidates, the number of constraints deleted, the current minimum objective function value, the time elapsed, and the termination status. When all candidate filtering is completed and the first candidate is selected according to the predetermined sorting rules, the termination status is written as "Optimal candidate determined". When there are still no candidates after the traversal ends, it is written as "Recipe feasible region is empty". When the 2-second time limit is reached and filtering has not been completed, it is written as "Solution timeout". "Optimal candidate determined" is the only driving output. "Recipe feasible region is empty" and "Solution timeout" are both the only driving outputs. The verified recipe verification is driven first. Only after the verification finds no usable recipe will the raw material replacement, re-detection, candidate set expansion, and re-solution branches be entered according to the reasons described below. In this specific implementation, the single solution time limit is set to 2 seconds. If the candidate feasible domain is empty, the constraint combination is infeasible, or the solution termination condition is not met within 2 seconds, the toxin exposure constraint and health risk constraint remain unchanged. First, the validated formulas that have been produced in the last 30 days and have finished product re-inspection records are read. The current inventory, current test validity, and current constraint boundaries are used for re-verification. Only validated formulas that simultaneously meet the stage-specific nutrition, safety, inventory, and raw material usage constraints can be used as the output for this time. When the deletion reason shows that the hard constraint of toxin or health risk conflicts with all candidates, the constraint combination is marked as an infeasible constraint set and the safety threshold is not relaxed. When the inventory constraint conflicts, the corresponding raw material batch is requested to be replaced. When abnormal conflict is detected, a re-detection task is created. When the coverage is insufficient due to the candidate step size, the scale step size is reduced from 0.5 mass percentage to 0.25 mass percentage. The processing command carries the conflict constraint, candidate number and data version for the next solution to read. If no verified formula is available, the solution task will not output the target formula. Instead, it will perform the selected raw material batch replacement, re-testing, or candidate set expansion. After the processing is completed, it will be re-executed from the S1 data snapshot and will remain in the unsolved state until a feasible candidate is obtained. The successfully output target formula and its actual production batch number will be written into the production execution table and used as the upstream record for the association between the S6 manufacturing residual and the biological response residual.

[0026] In this specific embodiment, S6 includes: After the target formula is produced and the finished product re-inspection results are obtained, the production feedback module associates the feeding ratio output by S5, the raw material test values ​​of S1, and the finished product re-inspection records according to the production batch number. The re-inspection items include at least crude protein, metabolizable energy reference values, and each controlled mycotoxin. Three independent samples are taken from the same finished product batch, and the original measured value, unit of measurement, test method, re-inspection time, and sample number of each sample are retained. All three samples are entered into the manufacturing residual calculation only after the integrity of the sample primary key, test method, unit, moisture value, and conversion version is verified. No sample is deleted. If any sample is missing, invalid, the unit cannot be converted, or there is a test conflict defined by the laboratory re-testing rules among the three samples, no manufacturing residual is formed and the process enters manual review. For production batches Re-inspection indicators The feedback module first unifies the finished product retest values, moisture values, and raw material test values ​​of the three valid samples to the wet basis according to the conversion version recorded in S4. Then, it sorts the three converted finished product retest values ​​in ascending order and takes the median as the final value. Then adopt Formation of manufacturing residuals, symbols Indicates wet-based manufacturing residuals and units and indicators The retest values ​​are the same, the symbols are the same. This represents the median of the converted wet basis measured values ​​of the three valid independent samples above, with the sign... Indicates production batch The actual set of raw material batch indexes that have passed primary key integrity verification is used, and the summation and traversal of this set is performed, with the symbols... Indicates production batch Chinese raw material batches Dimensionless wet basis feed ratio, symbol S1 indicates that it is saved or controlled by S1. Reduced raw material batch In indicators The wet basis measurement value, symbol , and These represent the production batch index, re-inspection index, and raw material batch index, respectively. When the moisture value or conversion relationship is missing, there are not enough three valid samples, or there is a test conflict, the manufacturing residual of this index is not formed and enters manual review. The three original conversion values ​​are still completely retained for the calculation of the degree of dispersion. The feedback module does not directly rewrite the detection exposure edges formed by S2 according to time slices, but first adopts... Correction of raw material batch indicators based on wet basis determination status; symbol and symbols These represent the batches before and after the correction, respectively. index Basic wet basis test values ​​and units and indicators Same, function This indicates truncating the first parameter to the closed interval formed by the second and third parameters, with the symbol... The manufacturing residual update rate is represented by a dimensionless number and set to 0.2, with the symbol... and symbols These represent the 1% and 99% historical quantiles of the effective basis wet-basis measured values ​​for the same raw materials and indicators, respectively, with units corresponding to the indicator. The same symbols in this formula , and index symbols , and Directly adopt the previous manufacturing residual formula and its immediately adjacent interpretation; when batch When the moisture content is effective, the feedback module then uses formula S1 to... Converted to the new dry basis determination values, for all time slices that referenced this batch and whose evaluation time was earlier than the corresponding failure time point. Re-execute S2 using the original feed ratio, original moisture content, and original time delay parameters for each time slot. , and The formula is used to re-normalize the training quantile files of the same type, unit, and time scale bound by S2; the affected time slice set, original basic measurement value, correction amount, recalculated edge weights before and after, normalization results and parent graph version are all retained in the version record. If the water content is missing or the quantile file version is inconsistent, no update is released and it enters manual review. The feedback module uses the three valid wet-basis retest original values ​​of the same group that form the median to calculate the sample standard deviation as the manufacturing repeated measurement dispersion. Only when there are no less than 20 valid historical values ​​of the same index and the same type of raw material and their interquartile range is greater than 0, the standard deviation is divided by the interquartile range to form a dimensionless dispersion ratio and truncated to [0,1]. The corrected detection edge reliability is the original reliability multiplied by one minus 0.5 times the dispersion ratio. If there are fewer than 20 historical samples, missing interquartile ranges, or interquartile ranges less than or equal to 0, the division is not performed, the original reliability is maintained and cannot be increased, and the abnormal branches of the interquartile range, the number of samples, and the historical data version are written into the manual review record. The detection edge reliability of missing or invalid detections remains 0. The edge reliability can only be recalculated after obtaining a valid retest value that meets the S1 verification and cannot be increased by manufacturing residuals alone. The edge weights and reliability are updated and written to the new version of the heterogeneous exposure graph. The version record includes the parent graph version, production batch number, manufacturing residual, re-inspection sample number, update rate, cutoff boundary and update time. The new version only replaces the detection edges associated with the production batch. The remaining nodes, edges and graph structure determined by S2 remain unchanged. Subsequent formulation tasks read this new graph version. After obtaining the actual feeding results of the target duck flock, the feedback module aligns the actual daily weight gain, actual feed conversion ratio and actual health indicators according to the candidate number and the duck flock time slice. The actual health indicators continue to be converted into [0,1] health risk values ​​using the fixed S4 order of normal, attention, intervention and severe categories and risk weights of 0, 0.33, 0.67 and 1 and the mapping version is recorded. Records that do not reach the end of the candidate applicable time window or whose effective weighing coverage is less than 30% of the duck flock stock will not form biological response residuals for the time being. Grouping of ducks Time slice and production performance or health tasks The feedback module adopts Formation of biological response residuals, symbol Represents the biological response residual with the same units as the task output, and the symbol is... Indicates actual daily weight gain, actual feed conversion ratio, or actual health risk outcome, with the following symbols: S4 represents the corresponding prediction result stored in S4, with the symbol... , and These represent the duck flock grouping index, time slice index, and task index, respectively. The feedback module uses the historical predictions and corresponding actual results frozen before the results are generated to recalculate the one-sided non-compliance score defined by S4. It writes the predicted values, measured values, orientation coefficients, non-compliance scores, residuals, candidate numbers, duck flock grouping primary keys, model versions, fusion representation versions, target coverage, prediction time, and result acquisition time of the daily weight gain, feed conversion ratio, and health risk tasks into the conformal calibration sample library of the corresponding group. Only the records that pass the integrity check for the first time are retained for duplicate candidate numbers and time slice combinations. Records that are manually modified afterward are entered into an independent correction version and the audit chain is retained. For tasks involving insufficient crude protein, insufficient metabolizable energy, and toxin exposure, the feedback module reads candidate estimated or predicted values ​​and actual measured values ​​of finished product retested wet basis according to the same wet basis conversion rules as S4. For toxin tasks, the retested toxin wet basis concentration, actual feed intake, median weight of ducks, and applicable time window are used to form actual toxin exposure values. Then, the one-sided non-compliance is calculated according to each task direction, and the candidate number, production batch number, duck group, task type, predicted or estimated value, measured value, direction, unit, non-compliance, model version, and conversion version are written into the calibration sample library. If any field, unit, or conversion version is incomplete, the record is isolated. When a duck flock group adds 10 new valid calibration samples or when the time since the last boundary update reaches 7 days, the calibration module reads the records that have passed the integrity check for each task, reselects the corresponding non-compliance quantile value, and updates the constraint boundary according to the same basic coverage, joint risk budget and group rollback rules as S4. The production batch playback of the most recent 90 days is performed before and after the update. If the new boundary leads to an increase in the number of regulatory toxin safety violations, the release is rejected and the previous boundary version is retained. Before releasing a new version of the heterogeneous exposure map, the feedback module performs playback on the 20 most recent production batches with finished product re-inspection. It requires that the median absolute error of the corrected finished product estimate is not higher than the parent version and that the 95th percentile of the error of any toxin index does not increase. If these conditions are not met, the new map version will be rejected and the manufacturing residual record will be retained for manual review. If the review confirms that the testing method or sample association is incorrect, the residual will be revoked but the original re-inspection record will not be deleted. Before releasing a new constraint boundary version, the feedback module checks that the model version and task unit of the newly added calibration sample are consistent with the current production version. If the units are inconsistent, they are converted according to the registered conversion coefficients. If the conversion relationship is missing, the sample is isolated. If the health label cannot be mapped or the applicable time window is not completed, it remains in the pending completion state and is not included in the quantile value sorting. The isolated sample is added to the next update trigger only after the fields are completed and the integrity check is passed. Manufacturing residual updates and biological response residual updates are committed using two independent transactions. The manufacturing residual transaction only modifies the detection edge weights and edge reliability of the heterogeneous exposure graph, while the biological response residual transaction only adds duck flock grouping calibration samples and updates constraint boundaries. If the biological response transaction fails, the manufacturing residual update that has passed the integrity check will not be rolled back, and the biological response residual will not be used to replace the raw material detection edge correction if the manufacturing transaction fails. When any feedback transaction fails to commit, the task scheduler will retry after 10 minutes with the same production batch number and idempotent key. If it fails three times in a row, the release of the corresponding graph version or boundary version will be frozen and the previous production version will be retained. The freeze will be lifted after the fault field is repaired and a complete replay is passed. The unfreezing will only continue the incomplete transactions and will not write the already committed residual records again. After feedback is completed, the system outputs a feedback record containing the new heterogeneous exposure map version, the new calibration sample version, the new constraint boundary version, the manufacturing residual summary, the biological response residual summary, and the release status. The next recipe solution must simultaneously record the map version and boundary version read, so that the finished product manufacturing deviation and the duck flock biological response deviation enter the subsequent S2 to S5 processing along their respective channels and remain traceable.

[0027] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0028] This invention organizes raw material batch detection, duck flock status, house environment, and historical feeding results into time-trackable multimodal information, and uses edge reliability and missing mask to control the influence of each modality on cross-fusion and constraint boundaries, so that the output raw material batch feeding ratio and applicable duck flock time window correspond to the actual data quality.

[0029] This invention provides feedback on manufacturing deviations and biological response deviations separately, and does not release safety constraints with interpolation values ​​when key toxin modes are missing or expired, thereby improving the stability and safety of formulation optimization under batch fluctuations, duck flock differences, and heat stress changes.

Claims

1. A method for optimizing duck feed formulation based on multimodal neural networks, characterized in that, include: S1. Obtain batch data of duck feed raw materials and time series data of target duck flock, perform dry matter baseline conversion, time alignment and generate missing mask to obtain aligned multimodal data; S2. Based on the aligned multimodal data, establish a heterogeneous exposure map including raw material batch nodes, nutrient component nodes, toxin nodes, formula nodes, and duck flock time slice nodes, and determine the edge weights characterizing exposure amount and time lag, and the edge reliability characterizing detection quality. The edge weights are determined based on the feeding ratio, dry matter basis measurement, actual feed intake, and the time interval from feeding to feeding; the edge reliability is determined based on detection repeatability, sampling coverage, the ratio of the remaining time from the current evaluation time to the failure time to the effective detection time, and missing markers. S3. The heterogeneous exposure map and the duck flock time slice sequence are encoded using a graph encoder and a temporal encoder respectively, and cross-fused according to edge reliability and missing mask to obtain a fused representation. The graph encoder aggregates the features of adjacent nodes according to edge type, and adjusts the aggregation coefficient of each adjacent node feature using the edge weight and edge reliability to form a formula batch exposure representation. The temporal encoder encodes the group weight distribution, feed intake, health indicators and indoor temperature and humidity according to the order of the duck flock time slices to form a duck flock status representation. The cross-fusion gating coefficient is generated by the edge reliability and the missing mask obtained by missing marker encoding. The cross-fusion gating coefficient controls the information transmission between the formula batch exposure representation and the duck flock status representation to obtain the fused representation. S4. Generate a candidate formulation set based on aligned multimodal data, input the features and fusion representation of each candidate formulation into the multi-task prediction head to obtain the prediction results of production performance, health risk and toxin exposure, perform duck flock grouping conformity calibration on the prediction results and allocate joint risk budget to obtain the constraint boundary; S5. Based on the constraint boundaries, under the staged nutritional and safety constraints, solve the target formula from the candidate formula set, and output the feeding ratio of each batch of raw materials and the applicable time window for the duck flock. S6. Based on the finished product re-inspection results, generate manufacturing residuals and update the heterogeneous exposure map. Based on the actual feeding results, generate biological response residuals and update the conformal calibration samples and constraint boundaries of the duck flock groups for subsequent formula solving.

2. The method for optimizing duck feed formulation based on multimodal neural networks according to claim 1, characterized in that, S1 includes: Obtain the crude protein content, moisture content, estimated metabolizable energy, mycotoxin detection value, price, inventory, detection repeatability, sampling coverage, detection time point, detection validity period, failure time point determined by the detection time point and detection validity period, current evaluation time and missing marker for each batch of raw materials; and obtain the duck age, breed, group weight distribution, feed intake, health indicators, house temperature, house humidity, formula feeding time, historical daily weight gain and feed conversion ratio of the target duck flock. The wet-based nutrient measurements are converted to dry-based measurements based on moisture content. Information is mapped to duck time slices according to raw material storage time, formula production time, and duck feeding time. Missing markers for each field are encoded as missing masks to obtain the aligned multimodal data and the corresponding missing masks.

3. The method for optimizing duck feed formulation based on multimodal neural networks according to claim 2, characterized in that, S2 includes: establishing a feeding edge between the raw material batch node and the formula node; establishing a detection edge between the raw material batch node and the nutrient component node and the toxin node; and establishing a feeding edge between the formula node and the duck flock time slice node. When a missing marker indicates a missing detection or the current evaluation time has reached the failure point, the reliability of the corresponding detection edge is set to zero.

4. The method for optimizing duck feed formulation based on multimodal neural networks according to claim 3, characterized in that, In step S4, the conformal calibration process for forming the constraint boundary includes: constructing duck flock groups according to age range, breed, and heat stress level divided by temperature and humidity index thresholds; For each duck flock group, scores for daily weight gain, feed conversion ratio, and health risk non-compliance were constructed based on historical prediction results and actual feeding results. Scores for toxin exposure non-compliance were constructed based on the toxin exposure prediction results of candidate formulas and the toxin exposure test results converted from the actual toxin test values ​​of finished product retests, actual feed intake, and applicable time windows of the duck flocks. Crude protein values ​​of candidate formulas were calculated based on the dry matter-based crude protein test values ​​of each batch of raw materials and the candidate feed ratios. Crude protein deficiency non-compliance scores were constructed based on the crude protein values ​​of the candidate formulas and the actual crude protein test values ​​of finished product retests. The metabolizable energy values ​​of candidate formulations are calculated based on the estimated metabolizable energy values ​​of each batch of raw materials and the candidate feed ratios. The metabolizable energy deficiency non-compliance score is constructed based on the metabolizable energy values ​​of the candidate formulations and the metabolizable energy reference values ​​determined by the nutritional component test values ​​of the finished product. The quantile values ​​of various non-compliance scores are determined according to the basic coverage rate, and the basic calibration boundary is formed. When the number of calibration samples for the target duck group does not reach the preset sample threshold, the upper-level merged group containing the target duck group, the global calibration sample, or the preset conservative boundary are used in sequence, and the quantile value of the target duck group is used after the number of samples reaches the preset sample threshold.

5. The method for optimizing duck feed formulation based on multimodal neural networks according to claim 4, characterized in that, In step S4, allocating the joint risk budget includes: determining the susceptibility score of the duck flock based on the age, breed, and current health indicators; converting the current health indicators into current health risk values ​​according to the mapping rules consistent with the training labels of the health risk prediction head; and weighting and aggregating the corresponding mycotoxin detection reliability based on the feeding ratio of each batch of raw materials in the candidate formula to obtain the toxin detection reliability. Within the preset upper limit of the joint default probability, default probability budgets are allocated for crude protein deficiency constraint, metabolizable energy deficiency constraint, toxin exposure constraint and health risk constraint, so that the default probability budget of health risk constraint is negatively correlated with the duck flock susceptibility score and the current health risk value, and the default probability budget of toxin exposure constraint is positively correlated with the reliability of toxin detection. The budget coverage rate is obtained by subtracting the budget of each default probability from the base coverage rate, and the larger of the base coverage rate and the budget coverage rate is taken as the target coverage rate. The daily weight gain non-compliance score and feed conversion ratio non-compliance score continue to use the aforementioned baseline coverage and corresponding baseline calibration boundary. The quantile values ​​of the health risk, toxin exposure, crude protein deficiency, and metabolizable energy deficiency non-compliance scores are reselected according to the aforementioned target coverage. The reselected quantile values ​​form the final upper bound of health risk, upper bound of toxin exposure, lower bound of crude protein, and lower bound of metabolizable energy. Together with the lower bound of daily weight gain and upper bound of feed conversion ratio in the baseline calibration boundary, they form the aforementioned constraint boundary.

6. The method for optimizing duck feed formulation based on multimodal neural networks according to claim 5, characterized in that, S5 includes: The feed ratio of each batch of raw materials and the applicable time window for duck flocks are used as decision variables, and the formula cost and the change in feed ratio of adjacent production batches are used as objective function terms. The phased nutritional constraints, safety constraints, inventory constraints, raw material usage constraints, and cost constraints, together with the constraint boundaries, form the formulation feasible region. Within the formulation feasible region, candidate formulations that simultaneously satisfy the constraint boundaries are screened, and the feed ratio and applicable time window for duck flocks corresponding to the candidate formulations whose objective function values ​​satisfy the solution termination condition are output. When the feasible domain of the formulation is empty or the solution termination condition is not met within the preset solution time limit, the toxin exposure constraint and health risk constraint remain unchanged, and the validated formulation that simultaneously meets the current constraint boundary, stage nutrition constraint, safety constraint, inventory constraint and raw material usage constraint after being reviewed by the current data is given priority. If the verified formulation does not exist, trigger at least one of the following processes: raw material batch replacement, re-detection, or expansion of the candidate formulation set, and re-solve. The target formulation is not output until a candidate formulation that satisfies the formulation feasibility domain is obtained through re-solving.

7. The duck feed formulation optimization method based on multimodal neural network according to claim 6, characterized in that, S6 include: The difference between the actual measured values ​​of nutritional components and toxins in the finished product re-inspection and the estimated values ​​of the finished product calculated based on the test values ​​of each batch of raw materials and the feeding ratio is taken as the manufacturing residual. The edge weights from the raw material batch node to the nutritional component node and the toxin node are corrected based on the manufacturing residual, and the corresponding edge reliability is corrected based on the degree of dispersion of repeated measurements of the manufacturing residual. The difference between the actual daily weight gain and actual feed conversion ratio of the duck flock and the corresponding prediction results is used as the production performance residual. The actual health indicators are converted into actual health risk results according to the mapping rules consistent with the training labels of the health risk prediction head. The difference between the actual health risk results and the health risk prediction results is used as the health risk residual. The production performance residual and the health risk residual are combined to form the biological response residual and written into the conformal calibration sample of the corresponding duck flock group. Under the condition of retaining the manufacturing residual update results, the quantile value and constraint boundary of the non-compliance score are re-determined.

8. The method for optimizing duck feed formulation based on multimodal neural networks according to claim 6, characterized in that, The case where the mycotoxin test value of the raw material batch has missing markers or the current evaluation time has reached the expiration time point is regarded as the test abnormal branch; In the abnormal detection branch, a conservative value that maximizes the predicted value of toxin exposure is selected from the upper limit of the detection method range, the upper quantile of the effective detection value of similar raw materials, and the safety limit conversion value. The toxin exposure contribution of the raw material batch is determined according to the conservative value and the candidate feeding ratio. When the toxin exposure contribution causes the upper limit of toxin exposure to exceed the safety constraint, the raw material batch is prohibited from being used. Otherwise, the upper limit of the raw material batch is tightened to meet the safety constraint, and a re-detection is triggered. Before obtaining a valid retest value, the edge reliability of the raw material batch shall not be improved or the safety constraints shall be lifted by interpolation results; After obtaining a valid retest value, the corresponding edge weight and toxin exposure contribution are updated using the retest value. Under the condition that the established safety limit remains unchanged, the upper limit of the raw material batch is recalculated and the safety constraint is re-determined. The reliability of the corresponding edge is re-determined based on the repeatability of the retest, sampling coverage, detection time point, failure time point and missing marker.

9. The method for optimizing duck feed formulation based on multimodal neural networks according to claim 3, characterized in that, The graph encoder, temporal encoder, multi-task prediction head, and cross-fusion gating coefficient generation network are trained using historical raw material batch data, historical formula data, historical duck flock time slice data, finished product re-inspection results, and actual feeding results. During training, the weighted sum of daily weight gain prediction loss, feed conversion ratio prediction loss, health risk prediction loss, and toxin exposure prediction loss is used as the multi-task loss. The reliability of detection edges with missing labels or whose current evaluation time has reached the failure time point is set to zero, while the reliability of other detection edges in the same training sample remains unchanged. The reliability of the detection edges associated with the training sample is obtained by normalizing and weighting the sum of the reliability of each batch of raw materials in the training sample according to the feeding ratio of each batch. When the reliability of all detection edges associated with the training sample is zero, the reliability of the sample is set to zero. The reliability of the sample is used as the weight of the multi-task loss for the training sample, and the model parameters are updated based on the weighted multi-task loss.

Images