Method for predicting the effect of epidermal growth factor addition based on growth performance data

Abstract

The application discloses an epidermal growth factor adding effect prediction method based on growth performance data, relates to the technical field of time series data analysis, and comprises the following steps: collecting target object data and structuring the data, establishing an original data set containing growth performance time series, environmental variables and epidermal growth factor adding dose records; calculating residual errors after preprocessing, performing wavelet decomposition, and combining input features; inputting a double-track learning model to obtain individualized causal effect estimation and dose-time-growth performance prediction mapping, then performing new observation of the target object to establish a correction factor, and outputting a prediction result after compensation. The application solves the technical problem that traditional epidermal growth factor adding effect prediction lacks consideration of dynamic changes in the time dimension and individual differences of different target objects, and cannot meet the precise evaluation and individualized regulation, and achieves the technical effect of accurate and comprehensive prediction of the epidermal growth factor adding effect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of time series data analysis, and in particular to an epidermal growth factor addition effect prediction method based on growth performance data. BACKGROUND

[0002] Epidermal growth factor (EGF) addition effect prediction is crucial for precise regulation of target object growth, and its results directly affect growth performance optimization and resource utilization efficiency. Existing technologies mostly rely on traditional models to predict the addition effect of epidermal growth factor by combining basic growth data with fixed algorithms. These methods play a certain role in scenarios with sufficient data and stable environment, but as the requirement for prediction accuracy increases, they expose limitations when applied to complex real-world scenarios. SUMMARY

[0003] The present application solves the technical problem that traditional epidermal growth factor addition effect prediction lacks consideration of dynamic changes in the time dimension and individual differences of different target objects, making it difficult to meet the requirements of precise evaluation and individualized regulation.

[0004] To solve the above technical problems, the present application proposes an epidermal growth factor addition effect prediction method based on growth performance data, which includes: collecting and structuring target object data to establish an original data set, which contains growth performance time series records, environmental variable records, and epidermal growth factor addition dose records in chronological order; performing data preprocessing based on the time sequence on the original data set, calculating short-term disturbance residuals and long-term drift residuals in the growth performance time series records in the preprocessing results within multiple time scales of sliding windows, and performing wavelet decomposition of the growth performance time series records in the preprocessing results to construct a frequency band energy spectrum; combining the short-term disturbance residuals, long-term drift residuals, and frequency band energy spectrum with the environmental variable records and epidermal growth factor addition dose records as input features; inputting the input features into a dual-track learning model to output individualized causal effect estimation and dose-time-growth performance prediction mapping; performing new observations of the target object, performing causal significance test of individualized causal effect estimation based on the new observations according to the corresponding sliding windows, and establishing a correction factor; compensating the dose-time-growth performance prediction mapping based on the correction factor to output the prediction results.

[0005] The present application proposes one or more technical solutions, which have at least the following technical effects:

[0006] The application collects and structures the growth performance time sequence, environmental variables and epidermal growth factor additive dose data of a target object to establish an original data set, constructs a frequency band energy spectrum through time sequence preprocessing, multi-time scale sliding window residual calculation and wavelet decomposition, inputs a double-track learning model with multi-dimensional input features, outputs individualized causal effect estimation and dose-time-growth performance prediction mapping, and combines new observations to perform causal significance test to establish a correction factor compensation prediction mapping, so as to accurately predict the epidermal growth factor addition effect, meet the individualized and accurate growth performance regulation needs of the target object, and achieve the technical effects of accurate and comprehensive prediction of the epidermal growth factor addition effect. BRIEF DESCRIPTION OF DRAWINGS

[0007] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0008] Figure 1 is a flowchart of the epidermal growth factor addition effect prediction method based on growth performance data provided by the embodiments of the present application.

[0009] Figure 2 is a flowchart of establishing a dose-time-growth performance prediction mapping in the epidermal growth factor addition effect prediction method based on growth performance data provided by the embodiments of the present application. DETAILED DESCRIPTION

[0010] The present application provides an epidermal growth factor addition effect prediction method based on growth performance data, which solves the technical problem that the traditional epidermal growth factor addition effect prediction lacks consideration of the dynamic changes of time dimension and individual differences of different target objects, and is difficult to meet the accurate evaluation and individualized regulation.

[0011] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor belong to the scope of protection of the present application.

[0012] It is to be understood that any variations of the terms "comprising" and "including" or any other analogous terms are intended to cover a non-exclusive inclusion, for example, a process, method, system, product or device that comprises a list of steps or elements is not necessarily limited to those steps or elements that are expressly listed, but can include additional steps or elements that are not expressly listed or inherent to such process, method, product or device.

[0013] As shown in the figure, the epidermal growth factor addition effect prediction method based on growth performance data, wherein the method comprises: Figure 1

[0014] Step A100: Collect and structure data for target object data, establish original data set, which contains growth performance time series record, environmental variable record and epidermal growth factor addition dose record in time sequence.

[0015] Specifically, when data collection is carried out for the target object, three types of core information need to be obtained to build the basis of the original data set. First, collect growth performance data, record the growth state indicators of the target object at fixed time intervals, form growth performance time series record, for example, for livestock in the breeding scene, record its body weight, body length at fixed time every day, or for cells in the cell culture scene, record its density and activity rate every hour, to ensure that the data can fully reflect the dynamic changes of growth over time. At the same time, the external environmental parameters that affect the growth of the target object are collected synchronously, forming the environmental variable record, such as the temperature and humidity of the breeding environment, or the pH value and dissolved oxygen content of the cell culture environment, and each item of environmental data needs to be corresponded to a specific time node, maintaining the consistency of time dimension with the growth performance data. In addition, the addition information of epidermal growth factor needs to be recorded in detail, forming the epidermal growth factor addition dose record, including the specific dose and corresponding time of each addition, to ensure that the intervention behavior can be clearly traced.

[0016] After completing the data collection, the above three types of data are structured in time sequence, and the scattered information is integrated into the standard original data set. Specifically, the arrangement method of time sequence correlation is adopted, and the growth performance time series record, environmental variable record and epidermal growth factor addition dose record corresponding to the same time node are matched and integrated one by one with the time node as the core axis. Taking the 20-day monitoring data of a batch of experimental cells as an example, a structured data in the form of a table can be constructed: the table rows represent the time nodes of each day, and the columns correspond to the growth performance time series record of cell density and cell activity rate in turn; the environmental variable record of incubator temperature and CO2 concentration; and the epidermal growth factor addition dose record of corresponding epidermal growth factor addition dose, so that each row of data can fully present the growth state of the target object, the environment and the epidermal growth factor intervention at the corresponding time node, and finally form the original data set arranged in time sequence. ​

[0017] By collecting the growth performance, environmental variables, and epidermal growth factor additive dose of the target object in different dimensions and time sequences, and structuring and integrating the three types of data in time sequence, an original data set containing growth performance time sequence records, environmental variable records, and epidermal growth factor additive dose records is established, providing time sequence coherent and information complete data support for subsequent time sequence based data preprocessing, feature extraction, and double track learning model input, and ensuring the orderly progress of the epidermal growth factor addition effect prediction process.

[0018] Step A200: After time sequence based data preprocessing of the original data set, short-term disturbance residuals and long-term drift residuals are calculated in a plurality of time scale sliding windows for the growth performance time sequence records in the preprocessing results, and wavelet decomposition of the growth performance time sequence records in the preprocessing results is performed to construct a frequency band energy spectrum.

[0019] Optionally, first, for the growth performance time sequence records in the preprocessing results, a plurality of time scale sliding windows are determined to distinguish short-term and long-term characteristics. Generally, the window size is set according to the growth cycle of the target object and the data collection frequency, for example, for livestock and poultry body weight data collected once a day, the short-term sliding window is set to 3 time nodes, i.e. 3 days, and the long-term sliding window is set to 15 time nodes, i.e. 15 days. Then, in each sliding window, first, the baseline trend curve of the growth performance is fitted by linear regression or time sequence smoothing algorithm, the specific process is as follows: taking the growth performance time sequence records after preprocessing as input, first, the adaptive smoothing window is determined according to the collection frequency and time span of the growth data, for example, in combination with the multi-time scale requirements for subsequent residual calculation, a window of 3-15 time nodes is selected, then in each sliding window, the growth performance data including body weight, daily average weight gain, etc. in the window is calculated by moving average, exponential weighted average, etc., the moving average takes the arithmetic mean of the data in the window, the exponential weighted average gives higher weight to recent data to highlight the time sequence correlation, then the window is slid and the smoothing results of each window are connected in time sequence, while the overall trend of the smoothing results and the original time sequence data is verified, if the window result deviates from the original trend beyond a preset threshold, the window parameters are adjusted to recalculate, finally, the baseline trend curve reflecting the long-term change rule of the growth performance is formed, providing a reference for the calculation of short-term disturbance residuals and long-term drift residuals.

[0020] Then, the actual growth performance data of each time node in the sliding window is subtracted from the corresponding fitting value to obtain the residual of each node. The residual calculated in the short-term sliding window is the short-term disturbance residual, reflecting the fluctuation of growth performance in a short time. For example, the actual weights in a 3-day window are 2.0 kg, 2.1 kg, and 1.9 kg, and the fitting trend is 2.0 kg / day. The short-term disturbance residuals are 0.0 kg, 0.1 kg, and -0.1 kg, respectively. The residual calculated in the long-term sliding window is the long-term drift residual, which reflects the degree of deviation of growth performance from the long-term trend. For example, the fitting long-term trend in a 15-day long-term sliding window is 0.05 kg / day, and the actual weight gain on the 10th day is 0.03 kg. The corresponding long-term drift residual is -0.02 kg.

[0021] Next, after calculating the short-term disturbance residual and the long-term drift residual, wavelet decomposition is performed on the preprocessed growth performance time series record. First, the db4 wavelet basis function suitable for time series signal analysis is selected, and the decomposition layer number is determined according to the time length and fluctuation characteristics of the growth performance data. Generally, it is decomposed into 3-5 layers to fully extract features in different frequency ranges. During the decomposition process, the growth performance time series is gradually decomposed into high-frequency detail signals and low-frequency approximation signals. The high-frequency detail signals correspond to the short-term rapid fluctuations of growth performance, and the low-frequency approximation signals correspond to the long-term slow changes of growth performance. Then, the energy of the high-frequency detail signals and the low-frequency approximation signals of each decomposition layer is calculated. The energy calculation uses the sum of the squares of the values of each sampling point of the signal. Then, the energies of each layer are arranged in order of frequency from high to low to construct a frequency band energy spectrum. This frequency band energy spectrum can clearly show the contribution of different frequency components to the growth performance time series.

[0022] By calculating the short-term disturbance residual and the long-term drift residual in multiple time scale sliding windows and performing wavelet decomposition to construct the frequency band energy spectrum on the preprocessed growth performance time series record, the short-term fluctuations, long-term trend deviations, and multi-frequency characteristics of the growth performance time series are extracted, providing deep time series feature support for subsequent input feature combination and dual-track learning model modeling.

[0023] Step A300: Combine the short-term disturbance residual, long-term drift residual, frequency band energy spectrum, and environmental variable record, epidermal growth factor additive dose record into input features.

[0024] In an embodiment of the present application, before combining the short-term disturbance residual, long-term drift residual, frequency band energy spectrum, environmental variable record, and epidermal growth factor additive dose record into input features, it is necessary to standardize each type of feature. Due to the difference in the dimensions of different types of features, for example, the unit of the short-term disturbance residual may be kilograms, which reflects the fluctuation of growth performance, the unit of temperature in the environmental variable record is degrees Celsius, and the unit of humidity is percentage, and the unit of the epidermal growth factor additive dose is micrograms. The inconsistency of dimensions will cause the model to be overly sensitive to features with a large numerical range, affecting the modeling accuracy. Therefore, the Min-Max standardization method is adopted, which maps the numerical values of each type of feature to the interval [0, 1] by subtracting the minimum value of the feature from the original numerical value of the feature and then dividing by the difference between the maximum and minimum values of the feature. After Min-Max standardization, the numerical values of each type of feature are unified to the same numerical interval, ensuring that each type of feature has equal weight influence in model training.

[0025] After standardization is completed, time series alignment operations are performed on each type of feature based on time sequence. Since the short-term disturbance residual and the long-term drift residual are calculated in multiple time scale sliding windows, the frequency band energy spectrum is generated by wavelet decomposition of the growth performance time series, and the environmental variable record and the epidermal growth factor additive dose record are raw data collected at time nodes, it is necessary to clarify the correspondence between each type of feature and the time node. For example, for growth performance data collected daily, the short-term disturbance residual calculated in a 3-day short-term sliding window corresponds to the time node at the end of the window, such as the residual of the 5-7 day window corresponding to the 7th day, and the long-term drift residual calculated in a 15-day long-term sliding window corresponds to the time node at the end of the window, such as the residual of the 1-15 day window corresponding to the 15th day; if the frequency band energy spectrum is based on a 10-day growth performance subsequence decomposition, it corresponds to each time node of the 10-day period; the environmental variable record and the epidermal growth factor additive dose record are recorded at daily nodes, which directly correspond to the daily time node. In this way, each time node can be matched to a complete set of features, i.e., the short-term disturbance residual, long-term drift residual, frequency band energy spectrum fragment, daily environmental variable data, and daily epidermal growth factor additive dose data corresponding to the time node.

[0026] Subsequently, feature dimension integration is performed to convert the aligned features of each type into a vector form recognizable by the model. For a single time node, the short-term disturbance residual and the long-term drift residual are each a numerical feature; if the frequency band energy spectrum is decomposed into 3 energy values of 3 frequency bands by 3 layers of wavelet decomposition, it is 3 numerical features; if the environmental variable record contains temperature and humidity, it is 2 numerical features; and the epidermal growth factor additive dose record is 1 numerical feature. According to the fixed order of short-term disturbance residual → long-term drift residual → frequency band energy spectrum sorted by frequency from high to low → environmental variable record according to the preset order of temperature, humidity → epidermal growth factor additive dose record, these numerical values are spliced into a single feature vector. Similarly, the feature vectors of multiple time nodes are arranged in chronological order to finally form a feature matrix for inputting into the dual-track learning model.

[0027] By sequentially performing standardization, time alignment and dimension integration on the short-term disturbance residual, the long-term drift residual, the frequency band energy spectrum, the environmental variable record and the epidermal growth factor additive dose record, the multiple types of features are combined into structured input features, achieving the effect of providing the dual-track learning model with time-sequential, dimension-uniform and information-complete input data, and ensuring the model to accurately learn the correlation between features.

[0028] Step A400: inputting the input features into the dual-track learning model to output individualized causal effect estimation and dose-time-growth performance prediction mapping.

[0029] Specifically, the input features are input into the dual-track learning model containing an individualized causal effect estimation track, a dose-time-growth performance prediction track and a dual-track interaction track. After receiving the input features, the individualized causal effect estimation track calculates and outputs an individualized causal effect function and a confidence interval through a representation learning network and historical dose-growth performance corresponding data. The dose-time-growth performance prediction track receives the input features, establishes an initial prediction result through a sequence convolution network, and then sends the initial prediction result, the individualized causal effect function and the confidence interval to the dual-track interaction track. Through causal consistency loss constraint compensation of the initial prediction result, a dose-time-growth performance prediction mapping is finally established and the individualized causal effect estimation and the prediction mapping are output. The specific steps are described in detail in A410-A450.

[0030] Step A500: performing new observation of the target object, performing causal significance test of individualized causal effect estimation based on the new observation according to the corresponding sliding window, and establishing a correction factor.

[0031] Optionally, after performing the new observation of the target object, a local growth performance feature representation is established based on the new observation according to the corresponding sliding window, and the local growth performance feature representation is subjected to causal significance verification analysis by using individualized causal effect estimation, significant differences are fused by inverse variance weighting, and then a correction factor is established. The specific steps are described in detail in A510-A520.

[0032] Step A600: Compensate the dose-time-growth performance prediction mapping based on the correction factor, and output the prediction result.

[0033] Optionally, first, the dimension correspondence relationship between the correction factor and the prediction mapping is determined. The correction factor is obtained by causal significance verification and inverse variance weighting of the new observation, and corresponds to the growth performance deviation of a specific epidermal growth factor dose and a time window. Therefore, it needs to be matched to the same dose and time dimension in the prediction mapping, for example, the correction factor is for the growth deviation of 5 micrograms of dose and the 8th-12th day. The prediction value corresponding to the time range under the dose in the prediction mapping needs to be compensated.

[0034] Subsequently, compensation calculation is performed. According to the physical meaning of the correction factor, it reflects the deviation amount of the growth performance from the expectation, and adjusts the growth performance prediction value in the prediction mapping corresponding to the dose and time node. If the correction factor is the comprehensive difference value after inverse variance weighting, the value can be directly added to the prediction value. For example, the body weight of the 12th day under a certain dose is originally predicted to be 3.2 kg, the correction factor is 0.08 kg, which represents that the actual growth is better than the expectation, and then the prediction value of the time node after compensation is adjusted to 3.2+0.08=3.28 kg. For multiple time nodes and dose combinations, this operation is performed in turn to complete the preliminary compensation of the prediction mapping.

[0035] After performing the compensation, the prediction mapping also needs to be verified for time sequence consistency to ensure that the prediction results under different time nodes and different doses conform to the growth law. By checking whether the change amplitude of the prediction values of adjacent time nodes is smooth, such as whether the daily weight gain fluctuation is within a reasonable range, and whether the overall trend of the prediction values under the same dose matches the growth stage of the target object, such as linear growth of body weight in the juvenile period and slow growth rate in the growth period, if an abnormality is found, such as a sudden change in the prediction value of a certain time node, the correction factor matching and compensation process needs to be traced back to find and correct the problem to ensure that the time sequence logic of the compensated prediction mapping is self-consistent.

[0036] Finally, after the above verification is completed, the final prediction result is output based on the compensated dose-time-growth performance prediction mapping. The prediction result covers the growth performance indicators of the target object at each time node in the future under each epidermal growth factor dose, including body weight, daily weight gain, etc., and can be accompanied by prediction uncertainty information such as confidence interval, providing data support for optimization of epidermal growth factor addition scheme. For example, if the output prediction result shows that under 3 micrograms, 5 micrograms, and 7 micrograms of dose, the predicted daily weight gain of the target object in the next 5 days is 0.035 kg, 0.048 kg, and 0.042 kg respectively, the prediction result can help technicians to select an addition dose that better meets the growth needs.

[0037] By matching the correction factor with the prediction mapping dimension, performing compensation calculation, verifying the timing consistency, and outputting the result, the effect of correcting the prediction mapping with the newly observed actual deviation is achieved, and the prediction accuracy and dynamic adaptability of the epidermal growth factor addition effect are improved.

[0038] Further, the step A100 in the method provided by the embodiment of the present application comprises:

[0039] A110: evaluating the data volume of the original data set and establishing an evaluation result.

[0040] A120: if the evaluation result fails to meet the preset threshold, generating a similar acquisition instruction.

[0041] A130: after performing feature extraction on the target object according to the similar acquisition instruction, performing similar matching of the feature extraction result, establishing an additional data set, and compensating the original data set according to the additional data set.

[0042] Specifically, after the original data set is established, it is first necessary to evaluate the data volume to determine whether the data set can support the subsequent preprocessing and modeling work. The evaluation process focuses on checking the time series length and record completeness of the data, such as setting the preset threshold to include 25 days of growth performance time series records of the target object, 25 days of environmental variable records, and the dose records of each epidermal growth factor addition during the period; during the evaluation, the effective days and missing conditions of the three types of records in the original data set are counted one by one, if the original data set only contains 12 days of growth performance time series records, and the environmental variable records have 6 days of blank, the evaluation result is determined as failing to meet the preset threshold.

[0043] Next, when the evaluation result does not reach the preset threshold, a similar acquisition instruction is generated, which clearly requires the type of target object features to be extracted to ensure matching accuracy. According to the similar acquisition instruction, the target object is executed for feature extraction, covering the variety, initial growth indicators such as initial body weight, body height, environmental preferences including suitable humidity range, light demand and other characteristics; then the extracted features are matched with historical data, and the historical object data with matching degree meeting the requirements is selected by calculating the feature similarity, specifically:

[0044] The variety, initial body weight, initial body height, suitable humidity range, suitable light duration of the extracted target object features and the corresponding features of the historical objects are standardized, and the feature values of different dimensions are uniformly mapped to the interval [0, 1] to eliminate the interference of dimension difference on similarity calculation. The process is the same as in step A300, and will not be repeated here. Then, the feature vectors of the target object and the historical objects are constructed. Taking the target object feature vector as the reference, the cosine similarity or Euclidean distance method is used to calculate the similarity between it and each historical object feature vector. If cosine similarity is used, the cosine value of the angle between the two vectors is calculated to measure the directional similarity, and the value closer to 1 indicates that the features are more similar. If Euclidean distance is used, the straight line distance between the vectors is calculated to measure the numerical difference, and the smaller the value, the more similar the features. Then, a preset similarity threshold of 0.8 is set, and the historical object data with a similarity value meeting the threshold requirement is selected, i.e. the cosine similarity is greater than or equal to 0.8 or the Euclidean distance is less than or equal to 0.2. At the same time, abnormal historical data with a similarity far below the threshold is removed, and finally the historical object data with a matching degree meeting the requirements of the target object features is obtained, laying a foundation for subsequent construction of additional data sets.

[0045] For example, if the target object is a certain strain of broiler chicken, its initial body weight is 0.5 kg and the suitable humidity is 50%-60%, and after matching, 2 groups of historical broiler chicken data with the same strain, initial body weight of 0.48-0.52 kg and suitable humidity of 48%-62% are matched. These historical data are arranged into an additional data set containing growth performance time series records, environmental variable records, and epidermal growth factor dosage records.

[0046] After completing the construction of the additional data set, it is integrated and compensated with the original data set in chronological order. For example, the original data set in the above example lacks 13 days of growth performance and environmental variable records, so the corresponding records in the additional data set for 13 days are supplemented to the original data set according to the time node, and the epidermal growth factor dosage records during the period are checked and supplemented, so that the original data set finally has complete records of the three types for 25 consecutive days, meeting the data quantity requirements for subsequent processing.

[0047] The data quantity evaluation is carried out on the original data set, similar acquisition instructions are generated when the threshold is not reached, and the additional data set is established by extracting the target object feature matching, and the original data set is compensated by the additional data set, so as to ensure that the data quantity of the original data set is sufficient and meets the requirements of subsequent data preprocessing and model input.

[0048] Further, the step A200 in the method provided by the embodiment of the application comprises:

[0049] A210: the data preprocessing comprises performing missing value interpolation, denoising and batch identification in time sequence.

[0050] Optionally, when the original data set is subjected to time sequence-based data preprocessing, the missing value interpolation operation is first performed. Since the original data set comprises growth performance time series records, environmental variable records and epidermal growth factor additive dose records in time sequence, missing values can occur at a specific time node of any type of record, and therefore the data needs to be checked one by one along the time axis to locate the missing time point and the corresponding record type. In the interpolation process, the time sequence correlation is first constructed according to the effective data before and after the missing time point, for example, if the body weight data of the 5th day in the growth performance time series record is missing, and the body weight of the 4th day is 2.1 kg and the body weight of the 6th day is 2.3 kg, the linear interpolation method is used to calculate the body weight of the 5th day as 2.2 kg; if the humidity data of the 8th day in the environmental variable record is missing, and the humidity of the previous three days and the next three days is stable at about 55%, the adjacent mean value method is used to supplement the humidity of the 8th day as 55%, so as to ensure that the interpolated data meets the growth and environmental change law in the time dimension and avoids destroying the time sequence continuity.

[0051] After the missing value interpolation is completed, the denoising processing is continued in time sequence. The original data can have abnormal values due to detection equipment errors, temporary environmental interference and the like, and such abnormal values can interfere with the accuracy of subsequent feature extraction, and therefore the time sequence smoothing algorithm is used to filter the noise. The specific steps are as follows: a reasonable abnormal value determination threshold is first set, for example, the data deviating from the mean value of the adjacent 3 time nodes by ±15% in the growth performance time series is determined as an abnormal value, and the data exceeding the normal fluctuation range, for example, the temperature fluctuation ±2℃ in the environmental variable record is determined as an abnormal value; then the sliding window average method is used to correct the abnormal values, for example, the mean value of the effective data in the sliding window is used to replace the abnormal value. Taking the epidermal growth factor additive dose record as an example, if the additive dose of the 12th day is mistakenly recorded as 10 micrograms, the additive doses of the adjacent time points are all about 5 micrograms, and there is no dose adjustment record, the mean value of 5 micrograms of the 11th day, 5 micrograms of the 13th day and 5 micrograms of the 14th day in the sliding window of 3 time nodes is calculated as 5 micrograms, and the value is used to replace the abnormal data of 10 micrograms, so as to ensure that the denoised data can truly reflect the actual situation of the epidermal growth factor addition and the growth trend of the target object.

[0052] Finally, the batch identification operation is performed in chronological order, and the batch is divided and identified according to the time range of data collection. Specifically, the time span of data collection is combed, and a continuous and consistent experimental condition time period is divided into a batch, for example, the data collected from the 1st to the 20th day for the same batch of young animals is divided into the first batch, and the data collected from the 21st to the 40th day for another batch of young animals is divided into the second batch. Then, a batch identification field is added to the original data set, and each piece of data is labeled with the corresponding batch number in chronological order, so that the data of different batches can be clearly distinguished in the subsequent processing process, laying a foundation for the accurate adaptation of the subsequent individualized causal effect estimation and prediction model.

[0053] By sequentially performing the preprocessing steps of missing value imputation, denoising and batch identification on the original data set in chronological order, the effects of repairing data integrity, filtering interference noise and distinguishing data batches are achieved, providing high-quality time series data for subsequent short-term disturbance residual calculation, wavelet decomposition and double-track learning model input.

[0054] Further, the method provided in the embodiment of the application comprises the following steps:

[0055] A410: The double-track learning model comprises an individualized causal effect estimation track and a dose-time-growth performance prediction track, and a double-track interaction track.

[0056] A420: After receiving the input features, the individualized causal effect estimation track learns the target object feature embedding through a representation learning network.

[0057] A430: Based on the historical dose-growth performance corresponding data, the individual causal effect and uncertainty of the target object at each dose are calculated, and an individualized causal effect function and a confidence interval are output.

[0058] A440: After receiving the input features, the dose-time-growth performance prediction track performs growth performance prediction through a sequence convolution network to establish an initial prediction result.

[0059] A450: The initial prediction result, the individualized causal effect function and the confidence interval are sent to the double-track interaction track, and the initial prediction result compensation is performed through causal consistency loss constraint to establish a dose-time-growth performance prediction mapping.

[0060] Specifically, first, for the construction of the individualized causal effect estimation track, the short-term disturbance residual, the long-term drift residual, the frequency band energy spectrum, and the environmental variable record, and the epidermal growth factor additive dose record are determined as input characteristics, a representation learning network is built to convert the input characteristics into target object feature embeddings, and historical dose-growth performance corresponding data is accessed to calculate the individual causal effect and uncertainty under each dose through causal inference methods, and finally the individualized causal effect function and confidence interval are output; for the construction of the dose-time-growth performance prediction track, the above input characteristics are also used as the basis to build a sequence convolution network with multi-scale convolution kernels, and the convolution and pooling operations are used to capture the time correlation features, and the initial prediction results of the growth performance corresponding to each time node and each dose are output; the construction of the double-track interactive track requires the design of a data receiving module to synchronously access the output data of the previous two tracks, a causal consistency loss constraint module is built to calculate the difference between the effect and the causal effect function and determine the compensation strength, and a compensation calculation unit is accessed to adjust the initial prediction results, and finally the dose-time-growth performance prediction mapping is established.

[0061] Next, when the input characteristics are input into the double-track learning model, first, the individualized causal effect estimation track is entered, which receives the integrated input characteristics, and then the characteristics of the target object are extracted and embedded through the representation learning network. The input characteristics include the short-term disturbance residual, the long-term drift residual, the environmental variable characteristics, and the epidermal growth factor additive dose characteristics in the growth performance time series characteristics of the target object. The representation learning network will filter and strengthen the hidden information related to the individual attributes of the target object through a multi-layer neural network structure, such as extracting the growth stability characteristics of the target object from the growth performance residual and associating the adaptation characteristics of the target object to specific environments from the environmental variable record. Finally, the high-dimensional and multi-type input characteristics are converted into low-dimensional and strongly representative target object feature embedding vectors, so that the individual differences of the target object can be accurately captured by the model.

[0062] After completing the feature embedding, the individualized causal effect estimation track calls the historical dose-growth performance corresponding data to calculate the individual causal effect and uncertainty of the target object under each epidermal growth factor dose. The historical dose-growth performance corresponding data needs to be similarity matched with the feature embedding vector of the current target object, and the historical target object data of the same type and similar growth base conditions is selected, and then the growth performance differences of these historical target objects under different epidermal growth factor doses are compared, such as the daily average weight gain difference of different microgram dose groups in the same feeding period, and the confounding factors such as environmental variables are excluded by combining the propensity score matching method of statistical analysis, to determine the individual causal effect of the current target object under each dose.

[0063] Subsequently, after the feature embedding vector accurately depicts the characteristics of the growth baseline and environmental adaptability of the individual attributes of the target object, the corresponding database of historical dose-growth performance is called, and historical samples with high similarity to the current target object feature embedding and small environmental variable differences are screened out to exclude interference data caused by large differences in feeding environment and initial growth state. Then, using causal inference methods such as propensity score matching, the influence of environmental fluctuations and initial body weight on growth performance is eliminated, and the change in growth performance indicators of the target object under each different epidermal growth factor dose relative to the non-addition dose, i.e., the individual causal effect value, is calculated. Subsequently, the Bootstrap sampling method is used to repeatedly calculate the causal effect value under each dose multiple times to obtain the standard error of the effect value; combining the preset confidence level of 95%, the standard error and the critical value of the corresponding statistical distribution (such as t distribution), the confidence interval of each individual causal effect value is calculated to quantify the uncertainty of the effect estimate. Finally, the corresponding relationship between the target object feature vector-epidermal growth factor dose-individual causal effect value-confidence interval is constructed into a mathematical function with epidermal growth factor dose as the independent variable and individual causal effect value as the dependent variable by piecewise linear regression or nonlinear fitting method. This function is the individualized causal effect function that reflects the association between the three, and the confidence interval corresponding to each dose is output as a reference for uncertainty.

[0064] Subsequently, the input features are synchronized and input into the dose-time-growth performance prediction track for processing. This track models the time series properties in the input features, such as the residual changes in the growth performance time series, the fluctuations of environmental variables over time, and the time distribution of epidermal growth factor addition doses, using a sequence convolution network. By setting different sizes of convolution kernels, the sequence convolution network can capture local correlation features of input features in the time dimension. For example, a 3-time node convolution kernel is used to extract the influence of short-term environmental variable fluctuations on growth performance, and a 7-time node convolution kernel is used to capture the medium and long-term change trend of growth performance after the addition of epidermal growth factor. Through multiple convolution and pooling operations, the model gradually learns the time series mapping relationship between input features and growth performance, and finally outputs the initial prediction results of the growth performance of the target object under different epidermal growth factor doses and different time nodes.

[0065] Finally, the dose-time-growth performance prediction track sends the generated initial prediction results, i.e., the growth performance numerical sequence under each time-dose combination, to the dual-track interaction track together with the individualized causal effect function and confidence interval output by the individualized causal effect estimation track, to prepare data for subsequent compensation of the initial prediction results and establishment of accurate dose-time-growth performance prediction mapping through causal consistency loss constraints. This process is detailed in steps A451-A454.

[0066] The individualized causal effect estimation track extracts individual characteristics of the target object and calculates causal effects, the dose-time-growth performance prediction track captures time sequence correlation and outputs initial prediction, and the double-track interaction track receives double-track output data, so as to provide core data support for subsequent optimization of prediction results and establishment of an epidermal growth factor addition effect prediction mapping with individualization and time sequence accuracy.

[0067] Further, the step A410 in the method provided by the embodiment of the application includes:

[0068] A411: The individualized causal effect estimation track and the dose-time-growth performance prediction track in the double-track learning model interact through a dynamic attention gate unit, and the dynamic attention gate unit adaptively adjusts a constraint weight of the individualized causal effect estimation track on the dose-time-growth performance prediction track according to time sequence stability of the target object.

[0069] Specifically, during running of the double-track learning model, the dynamic attention gate unit first needs to obtain a time sequence stability index of the target object, and the index is calculated based on the preprocessed growth performance time sequence record. Specifically, statistical characteristics of short-term disturbance residuals in the growth performance time sequence are extracted, and the time sequence stability is quantified by calculating a variance or a coefficient of variation of the short-term disturbance residuals. The smaller the variance, the smaller the fluctuation of the growth performance of the target object in the short term, and the more stable the time sequence. The larger the variance, the more frequent the fluctuation of the growth performance, and the weaker the time sequence stability.

[0070] Next, the dynamic attention gate unit adaptively adjusts the constraint weight of the individualized causal effect estimation track on the dose-time-growth performance prediction track according to a preset weight mapping rule. When the time sequence stability is high, it means that the sequence convolution network relied on by the dose-time-growth performance prediction track can accurately capture the growth performance change through a stable time sequence rule, and at this time, the constraint weight of the individualized causal effect estimation track needs to be reduced to reduce its intervention on the prediction track, so that the prediction track can more autonomously output results based on time sequence characteristics. When the time sequence stability is low, the sequence convolution network is difficult to accurately predict only by the time sequence data with large fluctuations, and the constraint weight of the individualized causal effect estimation track needs to be increased, so that the individualized causal effect function output by the track participates more in the prediction process to correct the deviation caused by the time sequence fluctuation.

[0071] Subsequently, the dynamic attention gate unit builds an interactive bridge between the individual causal effect estimation track and the dose-time-growth performance prediction track based on the adjusted constraint weight. On the one hand, the gate unit receives the individual causal effect function and the confidence interval output by the individual causal effect estimation track, and extracts the causal effect information corresponding to the current prediction time node and the epidermal growth factor dose; on the other hand, it receives the initial prediction result generated by the dose-time-growth performance prediction track, and analyzes the degree of fit between the result and the causal effect information. According to the constraint weight, the causal effect information is integrated into the modeling process of the prediction track in proportion to the weight. If the weight is high, the causal effect function will directly participate in the intermediate correction of the initial prediction result, for example, correcting the convolution kernel parameters of the sequence convolution network, so that the prediction is more in line with the individual causal law of the target object; if the weight is low, only the causal effect information is used as an auxiliary reference to ensure the dominance of the time series modeling of the prediction track. Taking the cell culture scene as an example, when the cell growth environment is stable, the gate unit allows the prediction track to mainly capture the time series changes of cell density through the sequence convolution network; when the environmental pH value fluctuates and causes the cell growth time series to be unstable, the gate unit increases the weight of the causal track, and uses the inherent proliferation effect of epidermal growth factor on the cells to correct the possible deviation of the prediction track.

[0072] Through the dynamic attention gate unit, the target object time series stability is calculated first, then the constraint weight of the individual causal effect estimation track on the dose-time-growth performance prediction track is adaptively adjusted, and finally the interaction between the two tracks is realized, so that the dual-track learning model can adapt to different growth time series states of the target object, and the dose-time-growth performance prediction mapping accuracy is improved.

[0073] Further, as shown in Figure 2 , the step A450 in the method provided by the embodiment of the present application comprises:

[0074] A451: aggregate the initial prediction result into an effect quantity under the corresponding dose.

[0075] A452: perform difference analysis using the effect quantity and the individual causal effect function, and establish a difference analysis result.

[0076] A453: perform weighted fusion based on the difference analysis result and the confidence interval, and establish a compensation strength.

[0077] A454: perform initial prediction result compensation according to the compensation strength.

[0078] Specifically, after the double-track interactive track receives the initial prediction result of the dose-time-growth performance prediction track output, the first step is to group the initial prediction result according to the epidermal growth factor additive dose, where the initial prediction result is the growth performance prediction value corresponding to different time nodes and different doses. The prediction values of multiple time nodes under the same dose collectively reflect the overall impact of the dose on the growth performance, and need to be converted into the effect amount of the corresponding dose through aggregation operation. During the aggregation process, according to the type of the growth performance index, a suitable statistical method is selected. If the growth performance index is daily average weight gain, the arithmetic mean of the initial prediction daily average weight gain of all time nodes under the same dose is calculated; if it is cumulative weight, the cumulative value of the initial prediction weight within a specified period under the same dose is calculated. For example, for the epidermal growth factor 5 microgram dose group, the initial prediction result contains daily average weight gain data for 7 consecutive days, which are 0.03 kg, 0.04 kg, 0.03 kg, 0.05 kg, 0.04 kg, 0.03 kg, and 0.04 kg, respectively. Through arithmetic mean calculation, the effect amount corresponding to this dose is (0.03+0.04+0.03+0.05+0.04+0.03+0.04) / 7≈0.037 kg, completing the conversion of the initial prediction result to the effect amount.

[0079] Then, after obtaining the effect amount of each dose, difference analysis is performed in combination with the individualized causal effect function of the individualized causal effect estimation track output. The individualized causal effect function has determined the individual causal effect value of each epidermal growth factor dose corresponding to the target object, and the difference analysis calculates the difference value between the effect amount and the individual causal effect value under the same dose, and judges the significance of the difference value in combination with the confidence interval corresponding to the individualized causal effect function. For example, under the epidermal growth factor 5 microgram dose, the individual causal effect value output by the individualized causal effect function is 0.04 kg, and the corresponding 95% confidence interval is 0.035 kg to 0.045 kg. The effect amount of this dose is 0.037 kg, the difference value is 0.003 kg, and the difference value falls within the confidence interval, indicating that the difference between the initial prediction result and the individual causal law is small; if the effect amount is 0.032 kg, the difference value is 0.008 kg, which is close to the lower limit of the confidence interval, indicating that the difference between the two is relatively significant, and attention should be paid to subsequent compensation.

[0080] Then, based on the difference analysis result and the confidence interval, the compensation intensity is determined by weighted fusion. The width of the confidence interval directly reflects the uncertainty of the individualized causal effect estimation. The narrower the interval, the higher the reliability of the individual causal effect value, and a higher weight is given in the weighting. The difference value in the difference analysis result reflects the deviation degree of the initial prediction result from the individual causal law. The larger the difference value, the more the initial prediction needs to be corrected, and a higher weight is also given in the weighting. In specific operation, first, the width of the confidence interval is converted into a weight coefficient, for example, the interval width and the weight are negatively correlated, then the difference value is converted into a deviation coefficient, for example, the difference value and the deviation coefficient are positively correlated, and finally the weight coefficient and the deviation coefficient are multiplied to obtain the compensation intensity. For example, the difference value is 0.008 kg at a certain dose, that is, the deviation coefficient is 0.8, the confidence interval width is 0.01 kg, that is, the weight coefficient is 1.0, and then the compensation intensity is 0.8*1.0=0.8. If the difference value is 0.003 kg at another dose, the deviation coefficient is 0.3, the confidence interval width is 0.02 kg, and the weight coefficient is 0.5, then the compensation intensity is 0.3*0.5=0.15, and the difference of the compensation intensity is realized.

[0081] Finally, the initial prediction result of the dose-time-growth performance prediction track output is compensated according to the intensity. In the compensation calculation, the initial prediction value of each time node is added to the result of the compensation intensity*the difference value, so that the initial prediction value is close to the individualized causal effect value, and the close degree is determined by the compensation intensity. For example, at an epidermal growth factor 5 microgram dose, the initial prediction daily weight gain of a certain time node is 0.03 kg, the compensation intensity is 0.8, and the difference value is 0.008 kg, and then the compensated prediction value is 0.03+(0.8*0.008)=0.0364 kg. If the initial prediction value of another time node is 0.05 kg, the compensation intensity is 0.15, and the difference value is 0.003 kg, then the compensated prediction value is 0.05+(0.15*0.003)=0.05045 kg, and the prediction result of each time node is ensured to be reasonably corrected according to the individual causal law.

[0082] By aggregating the initial prediction result into the effect amount corresponding to the dose, using the effect amount and the individualized causal effect function to perform difference analysis, combining the difference analysis result and the confidence interval to establish the compensation intensity by weighted fusion, and performing the initial prediction result compensation according to the compensation intensity, the dose-time-growth performance prediction mapping is made to fit the individualized causal law of the target object, and the prediction accuracy of the epidermal growth factor addition effect is improved.

[0083] Further, the step A450 in the method provided in the embodiments of the present application comprises:

[0084] A455: .

[0085] A456: wherein, is the compensated dose-time-growth performance prediction value, is the initial prediction result, x represents the target object feature vector, d is the epidermal growth factor additive dose, and t is the time node, represents the individualized causal effect estimate, is the effect size, is the uncertainty of the individualized causal effect estimate, determined by the confidence interval, is the uncertainty of the initial prediction result, is the time weight function.

[0086] In one embodiment, the compensated dose-time-growth performance prediction value is calculated First, the initial prediction result, the individualized causal effect estimate, the effect size, and the uncertainty of the individualized causal effect estimate, the uncertainty of the initial prediction result are determined, and the time weight function is determined, which is used to reflect the weight difference of different time nodes in compensation. Specifically: the initial prediction result receives input features from the dose-time-growth performance prediction track, and the growth performance prediction is established by a sequence convolution network; the individualized causal effect estimate receives input features from the individualized causal effect estimate track, learns the target object feature embedding through a representation learning network, and then calculates the individual causal effect of the target object at each dose by combining historical dose-growth performance corresponding data; the effect size is formed by aggregating the initial prediction result according to the corresponding dose; the uncertainty of the individualized causal effect estimate is derived by calculating the standard error of the individual causal effect value combined with the preset confidence level, which can be determined by the confidence interval; the uncertainty of the initial prediction result is calculated by analyzing the fluctuation of the initial prediction result at different time nodes and different doses; the time weight function is set according to the importance of different time nodes in compensation, and the closer to the current time node, the higher the weight is given, so as to reflect the difference in its influence on compensation.

[0087] Next, the weight distribution coefficient is calculated This step is based on the logic that the smaller the uncertainty of the result, the higher the reliability, and the greater the weight should be allocated. The individualized causal effect estimate and the initial prediction result are allocated according to their respective reliability, and the influence in the compensation process. If the uncertainty of the individualized causal effect estimate is lower, the coefficient will be closer to 1, meaning that the individualized causal effect estimate has a larger proportion in the compensation; on the contrary, if the uncertainty of the initial prediction result is lower, the coefficient will be more biased towards the value determined by the uncertainty of the initial prediction result.

[0088] Then, the difference between the individualized causal effect estimate and the effect size is calculated The difference reflects the deviation between the initial prediction result aggregated as an effect quantity and the individualized causal law, and is the core deviation source of subsequent compensation. Then, the weight distribution coefficient, the difference item, and the time weight function are multiplied to obtain a compensation item: The time weight function gives different weights according to the characteristics of the time node, such as the closer the time is to the current, the stronger the reference of the prediction, so that the compensation intensity of different time nodes matches their time sequence importance.

[0089] Finally, the initial prediction result is added to the above-mentioned compensation item to obtain the compensated dose-time-growth performance prediction value. For example, when the initial prediction result of a certain target object at a specific dose and time node deviates due to growth timing fluctuations, and the individualized causal effect estimation is more reliable, that is, the uncertainty is small, the compensation item will push the initial prediction value to adjust in the direction of the individualized causal effect, so that the final prediction result not only integrates the dynamic characteristics of the timing modeling, but also fits the inherent causal response law of the target object.

[0090] By sequentially determining the parameters, calculating the weight distribution coefficient and the difference item, generating the compensation item and adding it to the initial prediction result, the compensation of the initial prediction result is realized by using the causal consistency loss constraint, so that the dose-time-growth performance prediction value simultaneously fits the timing dynamics and the individualized causal law, and the effect prediction accuracy of epidermal growth factor addition is improved.

[0091] Further, the method provided in the embodiments of the present application comprises the following steps A500:

[0092] A510: Establishing a local growth performance feature representation based on the new observation according to the sliding window.

[0093] A520: Using the individualized causal effect estimation to analyze the causal significance of the local growth performance feature representation, and fusing the significant differences by inverse variance weighting to establish a correction factor.

[0094] Optionally, after performing the new observation of the target object, first, according to the sliding window used in the previous processing of the growth performance time series record, that is, the multiple time scale windows consistent with the calculation of the short-term disturbance residual and the long-term drift residual, the short-term 3 time nodes, and the long-term 15 time nodes window, the growth performance data in the new observation is segmented and feature extracted to establish a local growth performance feature representation. The new observation contains the latest growth performance time series record of the target object. According to the time scale of the sliding window, the new data is intercepted into several continuous local segments, each corresponding to a sliding window. Then, for the growth performance data in each window, statistical features that can reflect the local growth state are extracted, such as the mean, maximum, and fluctuation amplitude of the growth performance indicators in the window, and these features are combined to form the local growth performance feature representation corresponding to each window. For example, when using a 3-day short-term sliding window, the 10-day new observation data of livestock and poultry body weight can be divided into three local segments: the first 1-3 days, the fourth to sixth days, and the seventh to tenth days. If the window end is less than 3 days, it is merged into one window, and the daily average weight and weight fluctuation range of each segment are calculated to form three groups of local growth performance feature representations, ensuring that the local features remain consistent with the previous time series processing logic.

[0095] Then, after completing the establishment of the local growth performance feature representation, the individualized causal effect function output by the individualized causal effect estimation is used to perform causal significance check analysis on each local feature representation. First, according to the time node corresponding to each local growth performance feature representation and the epidermal growth factor dosage, the expected value of the growth performance of the target object at that dosage is extracted from the individualized causal effect function, that is, the growth performance reference range based on individual causal rules; then the actual growth performance indicators in the local growth performance feature representation, such as the local daily average weight, are compared with the expected value, and a statistical test (such as a t-test) is used to determine whether the difference between the two is significant. If the actual value falls outside the confidence interval corresponding to the expected value, it is determined that there is a significant difference, and the specific value of the difference is recorded. If the actual value is within the confidence interval, it is determined that the difference is not significant and is not included in the subsequent calculation. At the same time, during the check process, the variance corresponding to each significant difference is recorded simultaneously. The variance is calculated from the fluctuation of the local growth performance data and reflects the reliability of the difference result. The smaller the fluctuation, the smaller the variance, and the more reliable the difference. For example, a certain local feature representation corresponds to a 5 microgram dosage of epidermal growth factor, and the expected daily average weight given by the individualized causal effect function is 0.04 kg, with a confidence interval of 0.035 kg to 0.045 kg. The actual daily average weight of this local feature is 0.032 kg, which is outside the lower limit of the confidence interval, and is determined to be a significant difference with a difference value of -0.008 kg. Moreover, the local data fluctuation is small, and the calculated variance is 0.000064.

[0096] Finally, the correction factor is established in the manner of inverse variance weighted fusion, wherein the core logic of inverse variance weighting is that the more significant the difference in significance, the higher the reliability of the result, and the higher the weight should be given in the fusion process. In specific operation, first, the weight value of each significant difference is calculated, which is the inverse of the variance of the difference, i.e. weight = 1 / variance; then each significant difference is multiplied by the corresponding weight to obtain the weighted difference value; subsequently, the sum of all weighted difference values is divided by the sum of all weight values to obtain the integrated difference value after fusion; the integrated difference value is the correction factor used to compensate the dose-time-growth performance prediction mapping. For example, two significant differences are obtained through verification analysis: the first difference is -0.008 kg, the variance is 0.000064, and the weight = 1 / 0.000064 = 15625; the second difference is -0.006 kg, the variance is 0.0001, and the weight = 1 / 0.0001 = 10000. After weighted fusion, the integrated difference value = (-0.008 x 15625 + (-0.006) x 10000) ÷ (15625 + 10000) = (-125 - 60) ÷ 25625 ≈ -0.00722 kg, which is the final established correction factor.

[0097] By establishing local growth performance feature representation from new observations according to corresponding sliding windows, performing causal significance verification analysis on local features using individualized causal effect estimation, and inversely variance weighting and fusing significant differences, a correction factor reflecting the deviation of new observations from historical causal rules is established, thereby providing accurate adjustment basis for subsequent compensation of dose-time-growth performance prediction mapping, and improving the dynamic adaptability of epidermal growth factor addition effect prediction.

[0098] Further, the step A520 in the method provided in the embodiments of the present application comprises:

[0099] A521: if the causal significance verification analysis is a significant pass result, a compensation instruction is generated.

[0100] A522: according to the compensation instruction, the significant differences are inversely variance weighted and fused to establish a correction factor.

[0101] A523: if the causal significance verification analysis is a significant fail result, an abnormality warning is generated, and abnormality reporting management is performed according to the abnormality warning.

[0102] Optionally, first determine the passing and failing criteria of the verification, which is based on the confidence interval of the individualized causal effect estimation output. Specifically, extract the local growth performance feature representation from the individualized causal effect function, which represents the expected value and confidence interval of the growth performance under the corresponding epidermal growth factor dose, such as an expected daily average weight gain of 0.04 kg under a certain dose, with a 95% confidence interval of 0.035 kg to 0.045 kg. Then compare the actual growth performance indicator (such as local daily average weight gain) in the local growth performance feature representation with the expected value and confidence interval. If the actual indicator falls within the confidence interval and the difference is not statistically significant by statistical methods such as t-test (i.e. P value > 0.05), it is determined that the causal significance verification is passed. If the actual indicator is outside the confidence interval or the statistical test shows that the difference is statistically significant (i.e. P value ≤ 0.05), it is determined that the causal significance verification is failed.

[0103] Next, when the causal significance verification analysis result is significant, the system will automatically generate a compensation instruction. The compensation instruction needs to include key information: first, the time window and epidermal growth factor dose corresponding to the local growth performance feature representation, to ensure that the subsequent processing can accurately match the corresponding dimensions of the prediction mapping; second, the significant difference value of each passing verification local feature and the expected value (such as the difference of 0.002 kg between the actual daily average weight gain of 0.042 kg and the expected value of 0.04 kg); third, the variance corresponding to each significant difference (such as the variance of 0.00004 calculated from the local data fluctuation), which provides data support for inverse variance weighted fusion. After generating the compensation instruction, the difference value and variance in the instruction are used to calculate the weight according to the inverse variance weighted rule (weight = 1 / variance, the smaller the variance, the higher the weight), then the difference value and the corresponding weight are multiplied and summed, and the sum is divided by the sum of all weights to obtain the fused comprehensive difference value, which is the correction factor. For example, two passing verification local differences are 0.002 kg (variance 0.00004, weight 25000) and 0.003 kg (variance 0.00009, weight 11111.11), the weighted fusion correction factor = (0.002 x 25000 + 0.003 x 11111.11) ÷ (25000 + 11111.11) ≈ 0.00231 kg.

[0104] If the result of the causal significance check analysis is that the significance fails, an abnormal early warning is immediately generated, which needs to be clearly marked with abnormal key information: specifically including the local growth performance characteristic expression corresponding to the time node, the corresponding epidermal growth factor dose, the specific difference between the actual growth performance index and the expected value, and the P value of the statistical test result, to ensure accurate abnormal positioning. After generating the abnormal early warning, abnormal reporting management is performed according to the warning information: on the one hand, the abnormal details are recorded to the corresponding system log, including the abnormal occurrence time, the involved parameters, the difference data, etc., to facilitate subsequent traceability analysis; on the other hand, the relevant operating personnel are informed through the preset notification mechanism including system pop-up window, email reminder, etc., and the preliminary troubleshooting suggestions are triggered, such as checking whether the data acquisition equipment is malfunctioning, whether the actual addition amount of epidermal growth factor is consistent with the record.

[0105] By first determining the causal significance check criteria, then generating compensation instructions and establishing correction factors for the check passing results, and generating abnormal early warnings and performing abnormal reporting management for the non-passing results, the reliability of the correction factor construction is ensured, abnormal situations affecting the prediction accuracy are identified and handled in a timely manner, and the stability of the epidermal growth factor addition effect prediction process is maintained.

[0106] Further, the step A600 in the method provided in the embodiments of the present application comprises:

[0107] A610: Establish a calibration effect curve, and perform deviation comparison of the prediction result based on the calibration effect curve as a comparison standard, and establish a deviation comparison result, the deviation comparison result comprising a deviation node and a deviation amplitude.

[0108] A620: Perform deviation abnormal reporting management according to the deviation comparison result.

[0109] Optionally, after outputting the prediction result, a calibration effect curve can be established, which is based on the growth performance actual data of the same type of target object under different epidermal growth factor doses, and can be generated by a non-linear regression statistical fitting method by those skilled in the art, representing the growth response law of the target object in an ideal state. For example, the daily average weight gain data of multiple batches of healthy piglets under the doses of 3 micrograms, 5 micrograms and 7 micrograms of epidermal growth factor is collected, and a smooth curve of the daily average weight gain with time change corresponding to each dose is fitted as the standard for subsequent comparison.

[0110] Then, the deviation comparison of the prediction result and the calibration effect curve is performed. The difference value between the growth performance prediction value of each epidermal growth factor dose in the prediction result and the calibration value of the same dose and time node of the calibration effect curve is calculated to obtain the deviation amplitude, and the corresponding deviation node is recorded. For example, the daily weight gain prediction value of a certain prediction result at the 5 microgram dose on the 7th day is 0.045 kg, and the calibration value of the calibration effect curve at the same dose on the 7th day is 0.05 kg. At this time, the deviation amplitude is-0.005 kg, and the deviation node is the 7th day. After the calculation of all the predicted time-dose combinations is completed, the deviation comparison results including the deviation node and the deviation amplitude are summarized.

[0111] Subsequently, the deviation abnormality report management is performed according to the deviation comparison results. The deviation threshold can be pre-set according to the actual test or project requirements by those skilled in the art. For example, the daily weight gain deviation threshold is ±0.01 kg. If the deviation amplitude exceeds the threshold, it is determined that the deviation is abnormal and the abnormality report is triggered. When the abnormality report is triggered, the deviation node, the deviation amplitude and the corresponding dose information are recorded, and the related personnel are informed or reported through the system. At the same time, the deviation amplitude and the number of nodes involved are combined to start the processing flow. The slight deviation is prompted for subsequent observation, and the serious deviation is checked for the influencing factors of the prediction model or the actual growth process.

[0112] By establishing the calibration effect curve as the comparison standard, performing the deviation comparison of the prediction result and based on the results, the abnormality report management is performed. The deviation between the prediction result and the historical ideal rule is identified in time, and the reliability of the epidermal growth factor addition effect prediction is ensured.

[0113] In summary, the epidermal growth factor addition effect prediction method based on growth performance data provided by the embodiments of the present application has the following technical effects:

[0114] The application records, collects and structures the growth performance time series record, the environmental variable record and the epidermal growth factor additive dose data of the target object, establishes an original data set, obtains short-term disturbance residuals, long-term drift residuals and frequency band energy spectrum through time sequence-based data preprocessing, short-term disturbance residual and long-term drift residual calculation in a multi-time scale sliding window and wavelet decomposition of the growth performance time series record, combines the short-term disturbance residuals, the long-term drift residuals and the frequency band energy spectrum with the environmental variable record and the epidermal growth factor additive dose record as input features to input a double-track learning model, calculates individualized causal effect estimation and dose-time-growth performance prediction mapping, performs causal significance test of individualized causal effect estimation according to corresponding sliding windows combined with new observations of the target object to establish a correction factor, compensates the dose-time-growth performance prediction mapping based on the correction factor, thereby accurately predicting the influence of different epidermal growth factor doses on the growth performance of the target object at different time nodes, making the epidermal growth factor addition effect prediction result more accurate and reliable, meeting the demand for accurate evaluation and individualized regulation of the growth performance of the target object, and achieving the technical effect of accurate and comprehensive prediction of the epidermal growth factor addition effect.

[0115] The above description of disclosed embodiments enables one of ordinary skill in the art to make or use the application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein can be applied to other embodiments without departing from the spirit or scope of the application. Thus, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0116] Obviously, various modifications and changes can be made to the present application by those skilled in the art without departing from the spirit and scope of the present application. Accordingly, the present application intends to include all such modifications and changes as fall within the scope of the present application and its equivalents.

Claims

1. A method for predicting the effect of epidermal growth factor addition based on growth performance data, characterized by, The method comprises: Collecting and structuring data of a target object, establishing an original data set, which contains growth performance time series records, environmental variable records and epidermal growth factor additive dose records in chronological order; After time sequence-based data preprocessing of the original data set, short-term disturbance residuals and long-term drift residuals are calculated in a sliding window of multiple time scales for the growth performance time series records in the preprocessing results, and wavelet decomposition of the growth performance time series records in the preprocessing results is performed to construct a frequency band energy spectrum; The short-term disturbance residuals, long-term drift residuals and frequency band energy spectrum are combined with the environmental variable records and epidermal growth factor additive dose records as input features; The input features are input into a double-track learning model to output individualized causal effect estimation and dose-time-growth performance prediction mapping; Performing new observations of the target object, performing causal significance test of individualized causal effect estimation based on the new observations according to the corresponding sliding window, and establishing a correction factor; Compensating the dose-time-growth performance prediction mapping based on the correction factor to output a prediction result; The input features are input into a double-track learning model to output individualized causal effect estimation and dose-time-growth performance prediction mapping, comprising: The double-track learning model comprises an individualized causal effect estimation track and a dose-time-growth performance prediction track, and a double-track interaction track; After receiving the input features, the individualized causal effect estimation track learns the target object feature embedding through a representation learning network; Based on the corresponding historical dose-growth performance data, the individual causal effect and uncertainty of the target object at each dose are calculated, and an individualized causal effect function and a confidence interval are output; The dose-time-growth performance prediction track is used to receive the input features, perform growth performance prediction through a sequence convolution network, and establish an initial prediction result; The initial prediction result, the individualized causal effect function and the confidence interval are sent to the double-track interaction track, and initial prediction result compensation is performed through causal consistency loss constraint to establish a dose-time-growth performance prediction mapping; The initial prediction result compensation is performed through causal consistency loss constraint to establish a dose-time-growth performance prediction mapping, comprising: The initial prediction result is aggregated into an effect amount under the corresponding dose; Difference analysis is performed using the effect amount and the individualized causal effect function to establish a difference analysis result; Based on the difference analysis result and the confidence interval, weighted fusion is performed to establish a compensation strength; According to the compensation strength, initial prediction result compensation is performed; The new observations of the target object are performed, the causal significance test of individualized causal effect estimation is performed based on the new observations according to the corresponding sliding window, and the correction factor is established, comprising: Local growth performance feature representation is established based on the new observations according to the sliding window; The causal significance of the local growth performance feature representation is checked and analyzed using the individualized causal effect estimation, significant differences are inversely weighted and fused, and a correction factor is established.

2. The method of claim 1, wherein the method is based on growth performance data. The individualized causal effect estimation track and the dose-time-growth performance prediction track in the double-track learning model interact through a dynamic attention gate unit that adaptively adjusts the constraint weight of the individualized causal effect estimation track on the dose-time-growth performance prediction track according to the time series stability of the target object.

3. The method of claim 1, wherein the method is based on growth performance data. The initial prediction result compensation is performed by constraining the causal consistency loss, which is calculated as follows: ； wherein, is the compensated dose-time-growth performance prediction, is the initial prediction result, x characterizes the target object feature vector, d is the epidermal growth factor additive dose, t is the time node, characterizes the individualized causal effect estimate, is the effect size, is the uncertainty of the individualized causal effect estimate, determined by the confidence interval, is the uncertainty of the initial prediction result, is the time weight function.

4. The method of claim 1, wherein the method is characterized by: The causal significance check analysis of the individualized causal effect estimation on the local growth performance feature representation further includes: If the causal significance check analysis is a significant pass result, a compensation instruction is generated; According to the compensation instruction, the significant difference is inversely weighted and fused to establish a correction factor; If the causal significance check analysis is a significant fail result, an abnormality warning is generated, and abnormality reporting management is performed according to the abnormality warning.

5. The method of claim 1, wherein the method is based on growth performance data. The data preprocessing includes missing value imputation, denoising and batch identification in chronological order.

6. The method of claim 1, wherein the method is based on growth performance data. After the output prediction result, it includes: Establishing a calibration effect curve, taking the calibration effect curve as a comparison standard, performing deviation comparison of the prediction result, establishing a deviation comparison result, and the deviation comparison result includes deviation nodes and deviation amplitude; According to the deviation comparison result, deviation abnormality reporting management is performed.

7. The method of claim 1, wherein the method is based on growth performance data. After the establishment of the original data set, it includes: Performing data volume evaluation on the original data set to establish an evaluation result; If the evaluation result fails to meet the preset threshold, a similar collection instruction is generated; After performing feature extraction on the target object according to the similar collection instruction, the similarity matching of the feature extraction result is performed, an additional data set is established, and the original data set is compensated according to the additional data set.

Images