Poultry feed conversion efficiency prediction method fusing machine learning

Abstract

The application discloses a kind of poultry feed conversion efficiency prediction methods of fusion machine learning, belong to poultry breeding intelligent technology field.This method obtains the data of forage intake and weight gain in poultry breeding process, and is combined as dynamic index matrix according to preset time window alignment;According to the Euclidean distance between dynamic index matrix row vector, generate behavior distance matrix;Through behavior distance matrix, dynamic index matrix is weighted and reorganized by line, and generate rearranged dynamic feature sequence;Difference decomposition is carried out to the rearranged dynamic feature sequence, extract steady-state component and fluctuation component and calculate current period feed conversion efficiency base value;Obtain the base value of multiple historical periods before current period, generate conversion efficiency time series correlation factor by time series comparison;According to time series correlation factor, correct base value, generate poultry feed conversion efficiency prediction result.The method corrects base value using individual behavior difference and historical time series state, enhances prediction accuracy and time series stability.

Description

Technical Field

[0001] This invention relates to the field of intelligent poultry farming technology, specifically to a method for predicting poultry feed conversion efficiency by integrating machine learning. Background Technology

[0002] Poultry feed conversion efficiency (FCE) is a crucial parameter in poultry farming management used to assess the relationship between feed input and weight gain. Current techniques typically predict FCE by collecting individual or group feed intake and weight gain data, summarizing them over time windows, and then using feed conversion ratio formulas or linear regression models to obtain prediction results. Alternatively, methods can input feed intake, weight gain, and related environmental data into machine learning models for training and prediction to obtain estimates of FCE for the next cycle.

[0003] In actual poultry farming, significant differences in feeding behavior exist among individual poultry. Even when individuals consume similar total amounts of feed, their feeding time distribution, frequency, and weight gain response can vary considerably, impacting feed conversion efficiency. Existing methods fail to adequately represent these inter-individual behavioral differences during feature construction, typically extracting only individual feeding statistics such as mean, cumulative value, or variance. They do not utilize the distance relationships between individual behavioral patterns to reorganize features, making it difficult for prediction models to capture information arising from the similarities and differences in individual behavior within the group. Furthermore, feed conversion efficiency exhibits both continuity and volatility throughout the farming cycle. Historical baseline values ​​contain temporal information such as trend direction, duration, and degree of deviation. Existing methods utilize historical baseline values ​​in a simplistic manner, often employing averaging or direct extrapolation, without quantifying the continuity characteristics of historical states and the deviation characteristics of the current cycle as correction factors. This makes the current cycle prediction results susceptible to single-cycle random fluctuations. These shortcomings limit the accuracy and temporal stability of existing methods in predicting feed conversion efficiency under complex farming environments. Summary of the Invention

[0004] To address the problems of insufficient utilization of individual differences in feeding behavior and inadequate correction of current baseline values ​​based on historical conversion efficiency time-series status in existing poultry feed conversion efficiency prediction methods, this invention provides a poultry feed conversion efficiency prediction method that integrates machine learning. This method constructs a dynamic index matrix from feed intake and weight gain data, performs weighted recombination using inter-individual behavioral distances, performs differential decomposition on the rearranged dynamic feature sequence to obtain baseline values, and corrects these baseline values ​​based on historical periodic conversion efficiency time-series correlation factors to obtain the predicted poultry feed conversion efficiency.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a method for predicting poultry feed conversion efficiency by integrating machine learning, comprising: acquiring feed intake data and weight gain data during poultry farming, and aligning the feed intake data and weight gain data according to a preset time window to form a dynamic index matrix; generating a behavioral distance matrix representing the differences in feeding behavior among individuals based on the Euclidean distance between the row vectors of the dynamic index matrix; performing row-weighted recombination on the dynamic index matrix through the behavioral distance matrix to generate a rearranged dynamic feature sequence; performing differential decomposition on the rearranged dynamic feature sequence to extract steady-state components and fluctuation components, and calculating the baseline value of feed conversion efficiency for the current period based on the steady-state components and fluctuation components; acquiring the baseline values ​​of feed conversion efficiency for multiple historical periods prior to the current period, and performing a time-series comparison between the baseline value of feed conversion efficiency for the current period and the baseline values ​​of feed conversion efficiency for the multiple historical periods to generate a time-series correlation factor for conversion efficiency; correcting the baseline value of feed conversion efficiency based on the time-series correlation factor for conversion efficiency, and generating a poultry feed conversion efficiency prediction result using the corrected value. This method improves the accuracy and stability of poultry feed conversion efficiency prediction by integrating dynamic index matrix construction, inter-individual behavioral distance weighting, difference decomposition, and temporal correlation correction.

[0006] Preferably, the dynamic indicator matrix is ​​constructed as follows: using individual poultry individuals as row indices and continuous sampling times within a preset time window as column indices, feed intake data is filled into the feed intake quantum matrix, and weight gain data is filled into the weight gain submatrix; missing values ​​in the feed intake quantum matrix are linearly imputed, and missing values ​​in the weight gain submatrix are imputed by mean; the imputed feed intake quantum matrix and weight gain submatrix are then concatenated row by row to form the dynamic indicator matrix. This unifies feed intake and weight gain data within the same dynamic indicator framework, reducing the interference of missing data on subsequent behavioral distance calculations and feature rearrangement.

[0007] Preferably, the behavioral distance matrix is ​​generated as follows: The Euclidean distance between every two row vectors in the dynamic index matrix is ​​calculated, and an initial distance matrix is ​​constructed using the calculated Euclidean distances; each row of the initial distance matrix is ​​normalized so that the sum of the elements in each row equals a preset constant; the normalized distance matrix is ​​then used as the behavioral distance matrix. When normalizing each row of the initial distance matrix, the Softmax function is used to convert each row's elements into a probability distribution. The normalized behavioral distance matrix can serve as a stable source of behavioral difference weights, allowing subsequent reorganization to focus more on individuals with similar behaviors.

[0008] Preferably, the row-weighted recombination method is as follows: each row of the behavioral distance matrix is ​​used as a weight vector, and all row vectors of the dynamic index matrix are weighted and summed to obtain the recombined row vector corresponding to each row; all recombined row vectors are arranged sequentially according to their corresponding poultry individual numbers to form an intermediate dynamic feature matrix; the intermediate dynamic feature matrix is ​​subjected to sliding window mean filtering along the column direction, and the filtering result is used as the rearranged dynamic feature sequence. This process introduces information on differences in feeding behavior among individuals and suppresses high-frequency noise in the dynamic feature sequence.

[0009] Preferably, the differential decomposition method is as follows: the rearranged dynamic feature sequence is segmented according to a preset decomposition period; first-order forward differencing is performed on the data points within each segment to obtain the difference sequence corresponding to each segment; the mode of the difference sequence within each segment is taken as the steady-state component of that segment, and the difference between the difference sequence within each segment and the steady-state component is taken as the fluctuation component of that segment; the steady-state component and the fluctuation component are added together and divided by a preset reference coefficient to obtain the conversion efficiency sub-value corresponding to each segment, and the mean of the conversion efficiency sub-values ​​of all segments in the current period is taken as the baseline value of feed conversion efficiency. This method separates the stable change trend in the dynamic features from short-term fluctuations, reducing the impact of occasional fluctuations on the baseline value of feed conversion efficiency.

[0010] As a technical solution of the present invention, the generation method of the time-series correlation factor is as follows: taking the current period as the benchmark, extracting the feed conversion efficiency baseline values ​​corresponding to a consecutive preset number of historical periods to construct a historical baseline value sequence; calculating the difference between two adjacent periods in the historical baseline value sequence to obtain a difference sequence, and marking the difference sequence with symbols to generate a symbol marker sequence; calculating the state persistence factor based on the number of consecutive occurrences of the same symbol in the symbol marker sequence, and dividing the difference between the feed conversion efficiency baseline value of the current period and the last value of the historical baseline value sequence by the standard deviation of the historical baseline value sequence to obtain a deviation indicator value; combining the state persistence factor and the deviation indicator value as the time-series correlation factor for conversion efficiency. The corresponding correction method is as follows: mapping the state persistence factor and the deviation indicator value to a first correction coefficient and a second correction coefficient, respectively; multiplying the first correction coefficient by the feed conversion efficiency baseline value to obtain a first correction value, adding the second correction coefficient to the first correction value to obtain a second correction value; multiplying the second correction value by a preset period adjustment coefficient to obtain the predicted result of poultry feed conversion efficiency. This scheme uses historical trend continuity and current deviation to correct the base value, making the prediction results responsive to both trend continuation and abnormal deviation.

[0011] As another technical solution of the present invention, the generation method of the time-series correlation factor is as follows: Obtain the baseline values ​​of feed conversion efficiency for a predetermined number of historical periods prior to the current period to form a historical baseline value sequence; perform first-order differencing on the historical baseline value sequence to obtain a difference sequence, and statistically analyze the distribution ratio of positive and negative signs in the difference sequence, generating a trend direction indicator value based on the distribution ratio; compare the baseline value of feed conversion efficiency in the current period with the last value of the historical baseline value sequence to generate a deviation direction identifier; combine the trend direction indicator value and the deviation direction identifier as the time-series correlation factor for conversion efficiency. The corresponding correction method is as follows: determine the positive correction step size based on the magnitude of the trend direction indicator value, and determine the negative correction step size based on the value of the deviation direction identifier; add the positive correction step size and the negative correction step size to obtain the comprehensive correction step size; add the baseline value of feed conversion efficiency to the comprehensive correction step size to obtain the corrected baseline value, and use the corrected baseline value as the prediction result for poultry feed conversion efficiency. This solution determines the correction direction and magnitude by combining the historical difference sign distribution and the current deviation direction, ensuring that the prediction result conforms to the historical trend while also taking into account the fluctuations of the current period.

[0012] The technical effects and advantages provided by the present invention in the above technical solution are as follows: A behavioral distance matrix is ​​constructed using the Euclidean distance between row vectors of the dynamic index matrix. Each row of this matrix is ​​normalized so that the sum of its elements equals a preset constant. Each row is then used as a weight vector to sum all row vectors in the dynamic index matrix, resulting in recombined row vectors. The recombined intermediate dynamic feature matrix is ​​then subjected to sliding window mean filtering along its column direction, forming a rearranged dynamic feature sequence. The behavioral distance matrix uses the distance between individual behavioral patterns as the weighting basis, giving higher weights to row vectors with behavioral patterns similar to the target individual and lower weights to those with significantly different patterns. This recombined feature sequence retains useful information from differences in individual feeding behavior, suppresses irrelevant fluctuations, reduces random disturbances during subsequent difference decomposition, and enhances the stability of the baseline feed conversion efficiency calculation.

[0013] The rearranged dynamic feature sequence is segmented according to a preset decomposition period and subjected to first-order forward differencing. The mode of the difference sequence within each segment is taken as the steady-state component, and the difference between the difference sequence and the steady-state component is taken as the fluctuation component. The baseline value of feed conversion efficiency for the current period is calculated. The baseline values ​​of feed conversion efficiency for multiple consecutive historical periods preceding the current period are obtained to construct a historical baseline value sequence. The differences between adjacent periods are labeled and the state persistence factor is calculated. The difference between the current period's baseline value and the last value of the historical baseline value sequence is divided by the standard deviation of the historical baseline value sequence to obtain the deviation indication value, which is combined into a time-series correlation factor for conversion efficiency. The state persistence factor and the deviation indication value are mapped to the first correction coefficient and the second correction coefficient, respectively. The corrected values ​​are obtained through multiplicative and additive corrections, and then multiplied by a preset period adjustment coefficient to generate the prediction result. This method quantifies the trend persistence state in the historical period and the current deviation amplitude and applies them to the current baseline value, so that the prediction result can follow the trend continuity in the continuous breeding cycle, reduce the interference of single-period random fluctuations on the prediction value, and enhance the time-series adaptability of feed conversion efficiency prediction. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0015] Figure 1 This is a flowchart of a method for predicting poultry feed conversion efficiency that integrates machine learning; Figure 2 This is a flowchart of dynamic feature fusion and feed conversion efficiency calculation based on behavioral distance weighting; Figure 3 It is the Euclidean distance distribution curve between row vectors of poultry individuals in the behavioral distance matrix; Figure 4 This is a schematic diagram of the rearranged dynamic feature sequence, its decomposed components, and the transformation efficiency sub-values; Figure 5 This is a curve showing the change of the baseline value and the revised prediction result of poultry feed conversion efficiency. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] See Figure 1 This invention provides a method for predicting poultry feed conversion efficiency using machine learning. It acquires feed intake and weight gain data during poultry farming, aligns the data according to a preset time window, and combines them into a dynamic index matrix. Continuous sampling moments within the preset time window serve as column indices, and individual poultry individuals as row indices. Feed intake and weight gain data are aligned within the same window. A behavioral distance matrix representing differences in feeding behavior among individuals is generated based on the Euclidean distance between the row vectors of the dynamic index matrix. Elements in the behavioral distance matrix reflect the distances between different poultry individuals on the row vectors of the dynamic index matrix. The dynamic index matrix is ​​then row-weighted and reorganized using the behavioral distance matrix to generate a rearranged dynamic feature sequence. This row-weighted reorganization uses each row of the behavioral distance matrix to weight and sum all row vectors of the dynamic index matrix, ensuring that the rearranged dynamic feature sequence contains fused information about inter-individual behavioral differences. The rearranged dynamic feature sequence is then subjected to differential decomposition to extract steady-state and fluctuating components. The baseline feed conversion efficiency value for the current period is calculated based on the steady-state and fluctuating components. The baseline feed conversion efficiency (FCE) values ​​for multiple historical periods prior to the current period are obtained. The baseline FCE value for the current period is then compared with these historical values ​​over time to generate a time-series correlation factor for FCE. Based on this correlation factor, the baseline FCE value is corrected, and the corrected value is used to generate a predicted FCE result for poultry feed conversion efficiency.

[0018] In practice, feed intake and weight gain data are acquired during poultry farming. Each feed intake record and weight gain record is read from the poultry farming management system. Each feed intake record includes the poultry individual identifier, sampling time, and feed intake value; each weight gain record includes the poultry individual identifier, sampling time, and weight gain value. Feed intake and weight gain data are extracted according to a preset time window, which contains multiple consecutive sampling times. Feed intake and weight gain records for the same poultry individual within the same preset time window are used to construct a matrix. Using the individual poultry individual as the row index and the consecutive sampling times within the preset time window as the column index, feed intake data is filled into the feed intake quantum matrix, and weight gain data is filled into the weight gain submatrix. Each row in the feed intake quantum matrix corresponds to the feed intake value of a poultry individual at each sampling time within the window, and each row in the weight gain submatrix corresponds to the weight gain value of a poultry individual at each sampling time within the window.

[0019] In some embodiments, the number of rows in the feeding quantum matrix equals the total number of poultry individuals, and the number of columns equals the number of consecutive sampling moments within a preset time window. The number of rows and columns in the weight gain submatrix are the same as those in the feeding quantum matrix. The row number of the same poultry individual in the feeding quantum matrix is ​​the same as its row number in the weight gain submatrix.

[0020] Linear imputation is performed on missing values ​​in the feed intake quantum matrix. Linear imputation calculates the imputation value based on the valid sample values ​​before and after the missing position for the same feed intake index in the same poultry individual. For the first missing value in the feed intake quantum matrix... Individual poultry in the first Food intake values ​​at each sampling time point In the case of a missing data point, the most recent valid sampling time before the missing data point is the [number]th [time]. The most recent valid sampling time after the missing position is the sampling time of the nth sampling time. At each sampling time, the linear interpolation value Calculate using the following formula:

[0021] in, This indicates the row number of an individual poultry in the feeding quantum matrix. The value ranges from 1 to the total number of rows in the feeding quantum matrix; This indicates the column number in the feeding quantum matrix at which the sampling time of the missing value occurred. The value ranges from 2 to the total number of columns in the feeding quantum matrix minus 1; This represents the column number difference between the most recent valid sampling time before the missing position and the missing position. This represents the column number difference between the nearest valid sampling time after the missing position and the missing position. and All are positive integers; Indicates the first Individual poultry in the first The effective value of food intake at each sampling time; Indicates the first Individual poultry in the first The effective value of food intake at each sampling time; Indicates the first Individual poultry in the first The feed intake values ​​are obtained by linear interpolation at each sampling time. For missing positions in the feed intake quantum matrix where there are no consecutive valid sample values, the most recent valid feed intake value of the same poultry individual is used to fill the gaps.

[0022] Missing values ​​in the weight gain submatrix were imputed using the mean. Mean imputation was performed based on the average of existing valid values ​​for the same weight gain metric for the same poultry individual. For the first missing value in the weight gain submatrix... Individual poultry in the first Weight gain values ​​at each sampling time In the event of a missing value, obtain the first... For each poultry individual, calculate the average of all non-missing weight gain values ​​in the weight gain submatrix, and fill the average value into the first... Individual poultry in the first The missing locations at each sampling time point are identified. After mean imputation, the missing values ​​in the weighted submatrix are filled.

[0023] The padded feed intake quantum matrix and weight gain submatrix are concatenated row-wise to form a dynamic index matrix. During row concatenation, the feed intake row vector and weight gain row vector for the same poultry individual are kept contiguous in the column direction. The feed intake quantum matrix... The row vectors corresponding to the individual poultry individuals are listed first, and the weight gain submatrix contains the first row vector. The row vector corresponding to each poultry individual is placed after the first row vector, and the two row vectors are joined end-to-end in the column direction to form the first row vector in the dynamic index matrix. Each poultry individual corresponds to a row vector. The number of rows in the dynamic index matrix is ​​the total number of poultry individuals, and the number of columns is the sum of the number of columns in the feeding quantum matrix and the number of columns in the weight gain sub-matrix.

[0024] In practical implementation, a dynamic indicator matrix is ​​obtained, where the row vectors of the dynamic indicator matrix are feature vectors arranged according to the individual poultry numbers. For the first row in the dynamic indicator matrix... The row vector corresponding to each individual poultry and the first... Given the row vectors corresponding to each poultry individual, calculate the Euclidean distance between any two row vectors. The Euclidean distance is calculated as follows: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] The value of each column in the row vector corresponding to each individual poultry is related to the value of the first column. Subtract the values ​​in the same column of the row vector corresponding to each poultry individual to obtain the column difference. Square all column differences, sum all the squares, and take the square root of the sum to obtain the first... Individual poultry and the first The Euclidean distance between the row vectors of each poultry individual is calculated. The Euclidean distance is calculated for every two row vectors in the dynamic index matrix, and all Euclidean distance results are arranged into an initial distance matrix according to the poultry individual number. The initial distance matrix is ​​then... Line 1 Column elements represent the first Individual poultry and the first The Euclidean distance between the row vectors of individual poultry. The initial distance matrix is ​​a square matrix, and the number of rows in the initial distance matrix is ​​equal to the number of rows in the dynamic index matrix, and the number of columns in the initial distance matrix is ​​equal to the number of rows in the dynamic index matrix.

[0025] In some embodiments, each row of the initial distance matrix is ​​normalized such that the sum of the elements in each row of the initial distance matrix equals a preset constant, which is 1. The normalization process uses the Softmax function to convert each row of the initial distance matrix into a probability distribution. The Softmax function has the following form: in, This represents the row number in the initial distance matrix. The value range is 1 to ; This indicates the column number in the initial distance matrix. The value range is 1 to ; This indicates that the initial distance matrix is... The column number used when performing a summation. The value range is 1 to ; This indicates the total number of poultry individuals. The value of is equal to the number of rows in the dynamic index matrix; Represents the initial distance matrix. Line 1 Column elements; Represents the initial distance matrix. Line 1 Column elements; Represents a system of terms with the natural constant e as the base and the constant e as the base. The value of the exponential function of the exponent; Represents a system of terms with the natural constant e as the base and the constant e as the base. The value of the exponential function of the exponent; Indicates to From 1 to Perform a summation operation; This represents the elements obtained after normalization, which are used to construct the behavioral distance matrix.

[0026] The normalized distance matrix is ​​used as the behavioral distance matrix. The sum of the elements in each row of the behavioral distance matrix is ​​equal to 1, and the number of rows and columns of the behavioral distance matrix is ​​equal to the total number of poultry individuals. Each row of the behavioral distance matrix represents the weight distribution of the difference in feeding behavior between a poultry individual and all poultry individuals.

[0027] See Figure 3In the figure, the horizontal axis represents the individual poultry number, ranging from 1 to 30, and the vertical axis represents the element values ​​of the behavioral distance matrix, ranging from 0 to 0.30. The values ​​represent the Euclidean distance weights of the row vectors between poultry individuals. The figure shows the changing trends of the row elements of the behavioral distance matrix for all poultry individuals, with the first poultry individual (solid line), the second poultry individual (dashed line), and the third poultry individual (dotted line) as prominent examples. The remaining poultry individuals are shown as thin gray lines.

[0028] As shown in the figure, the behavioral distance matrix elements of the 1st, 2nd, and 3rd poultry individuals exhibit significant peaks at some individual numbers, with element values ​​reaching approximately 0.10 to 0.20. This indicates a larger distance between these poultry individuals and the corresponding numbered poultry in the dynamic indicator matrix row vectors, reflecting a higher degree of difference in feeding behavior. In contrast, the behavioral distance matrix element values ​​of the remaining poultry individuals are more evenly distributed, uniformly within a lower value range, indicating smaller behavioral differences among these individuals. The sum of each row element in the behavioral distance matrix is ​​1, ensuring the probability distribution property of the weights. This figure reflects the specific numerical distribution characteristics of the behavioral distance matrix obtained by calculating the Euclidean distance between the row vectors of the dynamic indicator matrix and then performing Softmax normalization.

[0029] In specific implementation, please refer to Figure 2 Each row of the behavioral distance matrix is ​​used as a weight vector. The row corresponds to the first The weight vector of each poultry individual, the first of the behavioral distance matrices. Line contains Each weight value Let be the total number of poultry individuals. When performing a weighted summation of all row vectors in the dynamic index matrix, the [value] in the behavior distance matrix is ​​[value]. Each weight value in the row is multiplied column-by-column by the row vector corresponding to the row number in the dynamic index matrix to obtain... A weighted row vector, The weighted row vectors are summed column-wise to obtain the distance matrix between the row vectors and the column vectors. The corresponding recombined row vectors of each row in the dynamic index matrix. Each row vector in the dynamic index matrix participates in the row distance matrix. The weighted summation is performed on each row of the row distance matrix to obtain the recombined row vector corresponding to each row of the row distance matrix.

[0030] All recombined row vectors are arranged sequentially according to their corresponding poultry individual numbers to form an intermediate dynamic feature matrix. The first row vector is the row distance matrix. The row corresponds to the first The individual poultry numbers are determined by the behavioral distance matrix. The recombined row vectors obtained from the row calculation are arranged in the middle dynamic feature matrix. The number of rows in the intermediate dynamic feature matrix equals the total number of poultry individuals, and the number of columns in the intermediate dynamic feature matrix equals the number of columns in the dynamic index matrix.

[0031] A sliding window mean filter is applied to the intermediate dynamic feature matrix along the column direction. The sliding window moves from the first column to the last column of the intermediate dynamic feature matrix, with a preset sliding window length and a step size of 1 for each movement. At each sliding window position, all column data within the window are collected, and the average of all column data within the window is calculated row-wise to obtain the filtered value corresponding to the center column of the sliding window. For boundary columns of the intermediate dynamic feature matrix that the sliding window cannot completely cover, the nearest calculable filtered value is used to fill the gaps. After calculating the filtered values ​​for all columns of the intermediate dynamic feature matrix, all filtered values ​​arranged in column order are used as the rearranged dynamic feature sequence.

[0032] The rearranged dynamic feature sequence consists of values ​​filtered from the intermediate dynamic feature matrix, and includes the fused dynamic features of all poultry individuals. The rearranged dynamic feature sequence is segmented according to a preset decomposition period, which is set to a fixed positive integer representing the number of data points in each segment. Starting from the first data point, the rearranged dynamic feature sequence is continuously divided into multiple segments according to the preset decomposition period, with each segment containing a preset number of data points. A first-order forward difference is performed on the data points within each segment. The first-order forward difference is calculated as follows: for each pair of adjacent data points within a segment, the difference is obtained by subtracting the previous data point from the latter. The difference values ​​of all adjacent data points within each segment constitute the difference sequence corresponding to that segment.

[0033] The mode of the difference sequence within each segment is taken as the steady-state component of that segment. The mode of the difference sequence is the difference value that appears most frequently in the difference sequence. In practice, if there are multiple difference values ​​that appear most frequently in the difference sequence, the difference value with the smallest value that appears most frequently is taken as the steady-state component. The difference between the difference sequence within each segment and the steady-state component is taken as the fluctuation component of that segment. The fluctuation component is calculated as follows: for each difference value in the difference sequence within each segment, subtract the steady-state component corresponding to that segment from the difference value to obtain the fluctuation component value corresponding to the difference value. All fluctuation component values ​​corresponding to that segment constitute the fluctuation component.

[0034] The conversion efficiency sub-value for each segment is obtained by adding the steady-state component and the fluctuation component and then dividing by a preset reference coefficient. The preset reference coefficient is a fixed constant greater than zero, determined based on the feed conversion efficiency benchmark value for the poultry breed; the same preset reference coefficient is used for the same batch of poultry. The conversion efficiency sub-value for each segment is calculated by adding the steady-state component value and the fluctuation component value corresponding to the segment, and then dividing by the preset reference coefficient. The average conversion efficiency sub-value for all segments within the current period is used as the baseline feed conversion efficiency value. The average is calculated by summing the conversion efficiency sub-values ​​for all segments within the current period and then dividing by the total number of segments.

[0035] See Figure 4 In the figure, the horizontal axis represents the data point number, indicating continuous data sampling points on the time series; the left vertical axis corresponds to the numerical range of the rearranged dynamic feature sequence and steady-state components, and the right vertical axis corresponds to the numerical range of the conversion efficiency sub-values. The legend shows that the blue solid line represents the rearranged dynamic feature sequence, the red dashed line represents the steady-state components, the green dots represent the fluctuation components, and the black squares represent the conversion efficiency sub-values.

[0036] As shown in the figure, the rearranged dynamic feature sequence exhibits a fluctuating trend, with values ​​ranging from approximately 18 to 31, reflecting the dynamic feature changes resulting from the fusion of behavioral differences among multiple poultry individuals. The steady-state component remains consistently low, between approximately -0.2 and 0, indicating a relatively stable baseline level. The fluctuating component is represented by green dots, with values ​​concentrated in the range of 0 to 4, representing short-term fluctuations in the dynamic feature sequence relative to the steady-state component. The conversion efficiency sub-values ​​are presented as black squares, with values ​​fluctuating between -0.15 and 0.2, reflecting the discrete characteristics of the conversion efficiency sub-values ​​calculated for each segment based on the steady-state and fluctuating components.

[0037] The rearranged dynamic feature sequence in the figure shows a significant upward abrupt change at approximately data points 100 and 300, followed by a phased fluctuation trend, reflecting a substantial change in the dynamic characteristics of poultry feeding behavior and weight gain data after fusion at these time points. The steady-state component curve is relatively stable and maintains negative values ​​near each data point, verifying the role of the steady-state component as a baseline level in the difference decomposition. The distribution of the fluctuation component scatter points is dense and the amplitude is stable, indicating that the short-term fluctuation component is relatively evenly distributed throughout the entire period. The scatter points of the conversion efficiency sub-values ​​do not show a significant trend of clustering over time points, reflecting the independence and dispersion of the conversion efficiency sub-values ​​across multiple time windows.

[0038] In practice, using the current period as a benchmark, a historical baseline value sequence is constructed by extracting the baseline feed conversion efficiency values ​​corresponding to a predetermined number of consecutive historical periods. The predetermined number is set to 5, based on the premise that the short-term trend of feed conversion efficiency in poultry farming over 5 consecutive periods can be fully incorporated into the time-series comparison. Each element in the historical baseline value sequence corresponds to the baseline feed conversion efficiency value of one historical period. The historical baseline value sequence is arranged in chronological order from oldest to most recent historical period, with the last element being the baseline feed conversion efficiency value of the most recent historical period preceding the current period.

[0039] Calculate the difference between two adjacent periods in the historical baseline value sequence to obtain the difference sequence. Each element in the difference sequence is the baseline feed conversion efficiency (FCE) value of the later historical period minus the baseline FCE value of the previous historical period. The difference sequence is then labeled to generate a labeled sequence. The generation rules for the labeled sequence are as follows: when an element in the difference sequence is greater than zero, the corresponding position in the labeled sequence is marked with a positive sign; when an element in the difference sequence is less than zero, the corresponding position in the labeled sequence is marked with a negative sign; when an element in the difference sequence is equal to zero, the corresponding position in the labeled sequence is marked with a zero sign. Each element in the labeled sequence maintains the same positional order as its corresponding element in the difference sequence.

[0040] The state persistence factor is calculated based on the number of consecutive occurrences of the same symbol in the symbol-marked sequence. The calculation uses the maximum consecutive identical symbol count method. The specific calculation process is as follows: Starting from the first element of the symbol-marked sequence, each element in the sequence is scanned sequentially, maintaining the current consecutive identical symbol count and the maximum consecutive identical symbol count. The initial value of both the current and maximum consecutive identical symbol counts is 1. When an element in the symbol-marked sequence is scanned, if it is the same as the previous element, the current consecutive identical symbol count is incremented by 1; if it is different, the current consecutive identical symbol count is compared with the maximum consecutive identical symbol count. If the current consecutive identical symbol count is greater than the maximum consecutive identical symbol count, the maximum consecutive identical symbol count is updated to the current consecutive identical symbol count, and then the current consecutive identical symbol count is reset to 1. After scanning all elements in the symbol-marked sequence, the current consecutive identical symbol count is compared with the maximum consecutive identical symbol count again. If the current consecutive identical symbol count is greater than the maximum consecutive identical symbol count, the maximum consecutive identical symbol count is updated to the current consecutive identical symbol count. The final maximum consecutive identical symbol count is used as the state persistence factor.

[0041] The deviation indicator value is obtained by dividing the difference between the current period's baseline feed conversion efficiency value and the last value in the historical baseline value sequence by the standard deviation of the historical baseline value sequence. The standard deviation of the historical baseline value sequence is calculated from all elements within the sequence. The calculation process is as follows: first, calculate the arithmetic mean of all elements in the historical baseline value sequence; then, for each element, calculate the square of the difference from the arithmetic mean; sum all the squares and divide by the number of elements in the historical baseline value sequence minus 1; take the square root of the quotient to obtain the standard deviation of the historical baseline value sequence. The deviation indicator value is calculated using the following formula:

[0042] in, This represents the current period number, and its value is a positive integer. This indicates the deviation indicator value corresponding to the current period; This represents the baseline value of feed conversion efficiency for the current period; The last value in the historical baseline value sequence represents the baseline value of feed conversion efficiency from the most recent historical period before the current period. This represents the standard deviation of the historical base value sequence. The standard deviation is calculated based on all elements in the historical base value sequence, and the calculation process is as described above.

[0043] The combined state persistence factor and deviation indicator value are used as the time-series correlation factor of conversion efficiency. The time-series correlation factor of conversion efficiency includes the values ​​of the state persistence factor and the deviation indicator value. These two values ​​are the two components of the time-series correlation factor of conversion efficiency.

[0044] The state persistence factor and deviation indicator value in the time-series correlation factor of conversion efficiency are mapped to a first correction coefficient and a second correction coefficient, respectively. The value of the first correction coefficient is set to the value of the state persistence factor, and the value of the second correction coefficient is set to the value of the deviation indicator value. Both the first and second correction coefficients maintain numerical consistency with their corresponding components in the time-series correlation factor of conversion efficiency.

[0045] The first correction value is obtained by multiplying the first correction coefficient by the baseline feed conversion efficiency value. The calculation process for the first correction value is as follows: obtain the baseline feed conversion efficiency value for the current period, multiply the baseline feed conversion efficiency value for the current period by the first correction coefficient, and use the product as the first correction value. The second correction value is obtained by adding the second correction coefficient to the first correction value. The calculation process for the second correction value is as follows: obtain the first correction value, add the first correction value to the second correction coefficient, and use the sum as the second correction value.

[0046] The second correction value is multiplied by a preset period adjustment coefficient to obtain the predicted poultry feed conversion efficiency. The preset period adjustment coefficient ranges from 0.5 to 1.5. When the current period and the historical period have the same time length, the preset period adjustment coefficient is set to 1.0. This setting is based on the fact that the current period and the historical period use the same time window length, ensuring consistency in the time scale of the baseline feed conversion efficiency value, thus eliminating the need for period length scaling. When the current period and the historical period have different time lengths, the preset period adjustment coefficient is set as the ratio of the current period's time length to the historical period's time length. The second correction value is then multiplied by the preset period adjustment coefficient, and the product is used as the predicted poultry feed conversion efficiency.

[0047] In practice, the baseline feed conversion efficiency values ​​for a predetermined number of historical periods preceding the current period are obtained to form a historical baseline value sequence. The predetermined number is set to 4, based on the premise that 4 historical periods can cover a complete short-term fluctuation window of poultry feed conversion efficiency. The historical baseline value sequence contains 4 elements, and the sequence is arranged from oldest to most recent historical period. The last element of the historical baseline value sequence is the baseline feed conversion efficiency value of the most recent historical period preceding the current period.

[0048] Perform a first-order difference on the historical base value sequence to obtain a difference sequence. The first-order difference process is as follows: starting from the second element of the historical base value sequence, subtract the previous element from the next element to obtain the difference value. Subtract adjacent elements of the historical base value sequence in turn to form the difference sequence. The number of elements in the difference sequence is the number of elements in the historical base value sequence minus 1, that is, the number of elements in the difference sequence is 3.

[0049] The distribution ratio of positive and negative signs in the difference sequence is statistically analyzed, and a trend direction indicator value is generated based on the distribution ratio. The number of positive and negative signs in the difference sequence is counted. Positive signs are those corresponding to difference values ​​greater than zero, and negative signs are those corresponding to difference values ​​less than zero. Difference values ​​equal to zero are not included in the count of positive or negative signs. The trend direction indicator value is calculated using the following formula: in, Indicates the trend direction indicator value. The value ranges from -1 to 1; This indicates the number of positive signs in the difference sequence. The value of is an integer ranging from 0 to 3; This indicates the number of negative signs in the difference sequence. The value of is an integer ranging from 0 to 3; This represents the number of elements in the historical base value sequence. The value of is 4; This represents the number of elements in the difference sequence, with a value of 3.

[0050] The baseline feed conversion efficiency (FCE) value for the current period is compared with the last value of the historical baseline value sequence to generate a deviation direction indicator. The deviation direction indicator is set to 1 when the baseline FCE value for the current period is greater than the last value of the historical baseline value sequence; -1 when the baseline FCE value for the current period is less than the last value of the historical baseline value sequence; and 0 when the baseline FCE value for the current period is equal to the last value of the historical baseline value sequence.

[0051] The combined trend direction indicator and deviation direction indicator are used as the time-series correlation factor for conversion efficiency. The time-series correlation factor for conversion efficiency includes the value of the trend direction indicator and the value of the deviation direction indicator. The two values ​​are used as components of the time-series correlation factor for conversion efficiency.

[0052] The positive correction step size is determined based on the magnitude of the trend direction indicator value. The positive correction step size is calculated by multiplying the trend direction indicator value by the positive correction baseline step size; the product is used as the positive correction step size. The positive correction baseline step size is set to 0.05 times the current period's baseline feed conversion efficiency. This setting is based on the fact that the trend direction indicator value ranges from -1 to 1, and the maximum absolute value of the positive correction step size after multiplying by 0.05 does not exceed 5% of the current period's baseline feed conversion efficiency, thus limiting the correction range. When the trend direction indicator value is positive, the positive correction step size is positive; when the trend direction indicator value is negative, the positive correction step size is negative; and when the trend direction indicator value is zero, the positive correction step size is zero.

[0053] The negative correction step size is determined based on the value of the deviation direction indicator. When the deviation direction indicator is 1, the negative correction step size is set to 0; when the deviation direction indicator is -1, the negative correction step size is set to the negative correction baseline step size, which is set to 0.03 times the baseline value of feed conversion efficiency for the current period. The basis for setting the negative correction baseline step size to 0.03 times the baseline value of feed conversion efficiency for the current period is that when the deviation direction indicator is -1, it means that the baseline value of feed conversion efficiency for the current period is lower than the last value in the historical baseline value sequence. A negative correction of 3% of the baseline value is introduced to make the correction direction consistent with the deviation direction; when the deviation direction indicator is 0, the negative correction step size is set to 0.

[0054] The positive and negative correction step sizes are added together to obtain the overall correction step size. The overall correction step size is the algebraic sum of the positive and negative correction step sizes. The baseline feed conversion efficiency value is added to the overall correction step size to obtain the corrected baseline value, and this corrected baseline value is used as the prediction result for poultry feed conversion efficiency.

[0055] See Figure 5 In the graph, the horizontal axis represents the period number, the left side of the vertical axis represents the baseline feed conversion efficiency value, and the right side of the vertical axis represents the comprehensive correction step size. The legend includes the baseline feed conversion efficiency value (blue scatter dots), the corrected baseline value (red solid line), the comprehensive correction step size (green dashed line), and the predicted poultry feed conversion efficiency result (purple diamond).

[0056] The blue scatter plots in the graph represent the baseline feed conversion efficiency values ​​obtained for each period. The distribution of the scatter plots shows the fluctuation trend of the baseline values ​​with the period number, generally exhibiting a trend of first high and then low, reflecting the dynamic changes in poultry feed conversion efficiency. The red solid line represents the value after correction of the baseline values ​​based on the trend direction indicator and deviation direction indicator. The solid line is relatively smooth, following the fluctuation trend of the blue scatter plots but reducing the noise influence in the baseline values, demonstrating the effective adjustment effect of trend correction and deviation correction on the baseline values. The green dashed line represents the calculated comprehensive correction step size, which fluctuates periodically, with an amplitude range of approximately -0.1 to 0.15, reflecting the combined influence of the trend direction indicator and deviation direction indicator on the correction step size. The fluctuation trend of the comprehensive correction step size and the red corrected baseline value curve has a certain correspondence within multiple periods, indicating that the correction step size effectively reflects the short-term fluctuation characteristics of the baseline values. The purple diamond in the lower right corner of the graph represents the predicted poultry feed conversion efficiency. This point is located next to the corrected baseline value of the most recent period, indicating that the prediction result is generated based on the corrected baseline value of the current period.

[0057] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for predicting poultry feed conversion efficiency by integrating machine learning, characterized in that, The method includes: Data on feed intake and weight gain during poultry farming are acquired, and the data are aligned and combined into a dynamic index matrix according to a preset time window. A behavioral distance matrix representing differences in feeding behavior among individuals is generated based on the Euclidean distance between the row vectors of the dynamic index matrix. The dynamic indicator matrix is ​​row-weighted and reorganized using the behavioral distance matrix to generate a rearranged dynamic feature sequence. The rearranged dynamic feature sequence is subjected to differential decomposition to extract steady-state components and fluctuation components, and the basic value of feed conversion efficiency for the current period is calculated based on the steady-state components and fluctuation components. Obtain the baseline values ​​of feed conversion efficiency for multiple historical periods prior to the current period, and perform a time-series comparison between the baseline value of feed conversion efficiency for the current period and the baseline values ​​of feed conversion efficiency for the multiple historical periods to generate a time-series correlation factor for conversion efficiency. The baseline value of feed conversion efficiency is corrected based on the time-series correlation factor of conversion efficiency, and the corrected value is used to generate a prediction result of poultry feed conversion efficiency.

2. The method for predicting poultry feed conversion efficiency by incorporating machine learning according to claim 1, characterized in that, The specific implementation method of acquiring feed intake data and weight gain data during poultry farming, and aligning the feed intake data and weight gain data according to a preset time window to combine them into a dynamic indicator matrix includes: Using individual poultry individuals as row indexes and continuous sampling times within a preset time window as column indexes, the feed intake data is filled into the feed intake quantum matrix, and the weight gain data is filled into the weight gain submatrix. The missing values ​​in the feeding quantum matrix are linearly imputed, and the missing values ​​in the weight gain submatrix are imputed by mean imputation. The supplemented feeding quantum matrix and weight gain sub-matrix are then concatenated row by row to form the dynamic index matrix.

3. The method for predicting poultry feed conversion efficiency by incorporating machine learning according to claim 2, characterized in that, The specific implementation of generating a behavioral distance matrix representing differences in feeding behavior among individuals based on the Euclidean distance between the row vectors of the dynamic index matrix includes: Calculate the Euclidean distance between every two row vectors in the dynamic index matrix, and construct an initial distance matrix using the calculated Euclidean distances; Normalize each row of the initial distance matrix so that the sum of the elements in each row is equal to a preset constant; The normalized distance matrix is ​​used as the behavior distance matrix.

4. The method for predicting poultry feed conversion efficiency by incorporating machine learning according to claim 3, characterized in that, The specific implementation of row-weighted rearrangement of the dynamic index matrix using the behavioral distance matrix to generate the rearranged dynamic feature sequence includes: Each row of the behavior distance matrix is ​​used as a weight vector, and all row vectors of the dynamic index matrix are weighted and summed to obtain the recombined row vector corresponding to each row. All recombined row vectors are arranged sequentially according to their corresponding poultry individual numbers to form an intermediate dynamic feature matrix; The intermediate dynamic feature matrix is ​​subjected to sliding window mean filtering along the column direction, and the filtering result is used as the rearranged dynamic feature sequence.

5. The method for predicting poultry feed conversion efficiency by incorporating machine learning according to claim 4, characterized in that, The specific implementation method for performing differential decomposition on the rearranged dynamic feature sequence, extracting steady-state components and fluctuation components, and calculating the baseline value of feed conversion efficiency for the current period based on the steady-state components and fluctuation components includes: The rearranged dynamic feature sequence is segmented according to a preset decomposition period, and the data points in each segment are subjected to first-order forward difference to obtain the difference sequence corresponding to each segment. The mode of the difference sequence within each segment is taken as the steady-state component of that segment, and the difference between the difference sequence within each segment and the steady-state component is taken as the fluctuation component of that segment. The steady-state component and the fluctuation component are added together and then divided by a preset reference coefficient to obtain the conversion efficiency sub-value corresponding to each segment. The average of the conversion efficiency sub-values ​​of all segments in the current period is used as the basic value of the feed conversion efficiency.

6. The method for predicting poultry feed conversion efficiency by incorporating machine learning according to claim 5, characterized in that, The specific implementation methods for generating the time-series correlation factor of conversion efficiency include: Using the current cycle as a benchmark, extract the baseline values ​​of feed conversion efficiency corresponding to a preset number of consecutive historical cycles to construct a historical baseline value sequence; Calculate the difference between two adjacent periods in the historical base value sequence to obtain a difference sequence, and then mark the difference sequence with symbols to generate a symbol-marked sequence; The state persistence factor is calculated based on the number of consecutive occurrences of the same symbol in the symbol tag sequence. The difference between the current period's feed conversion efficiency baseline value and the last value of the historical baseline value sequence is divided by the standard deviation of the historical baseline value sequence to obtain the deviation indicator value. The state persistence factor and the deviation indicator value are combined as the conversion efficiency time-series correlation factor.

7. The method for predicting poultry feed conversion efficiency by incorporating machine learning according to claim 6, characterized in that, The specific implementation method for correcting the baseline value of feed conversion efficiency based on the time-series correlation factor of conversion efficiency, and generating the predicted result of poultry feed conversion efficiency using the corrected value includes: The state persistence factor and deviation indication value in the time-series correlation factor of the conversion efficiency are mapped to the first correction coefficient and the second correction coefficient, respectively. Multiply the first correction coefficient by the base value of feed conversion efficiency to obtain the first correction value, and add the second correction coefficient to the first correction value to obtain the second correction value; Multiply the second correction value by a preset periodic adjustment coefficient to obtain the predicted result of the poultry feed conversion efficiency.

8. The method for predicting poultry feed conversion efficiency by incorporating machine learning according to claim 1, characterized in that, The specific implementation methods for generating the time-series correlation factor of conversion efficiency include: Obtain the baseline values ​​of feed conversion efficiency for a preset number of historical periods preceding the current period, forming a historical baseline value sequence; The historical base value sequence is subjected to first-order difference to obtain a difference sequence, and the distribution ratio of positive and negative signs in the difference sequence is statistically analyzed. A trend direction indicator value is generated based on the distribution ratio. The baseline value of feed conversion efficiency for the current period is compared with the last value of the historical baseline value sequence to generate a deviation direction identifier. The trend direction indicator and the deviation direction identifier are combined as the time-series correlation factor of the conversion efficiency.

9. The method for predicting poultry feed conversion efficiency by incorporating machine learning according to claim 8, characterized in that, The specific implementation method for correcting the baseline value of feed conversion efficiency based on the time-series correlation factor of conversion efficiency, and generating the predicted result of poultry feed conversion efficiency using the corrected value includes: The positive correction step size is determined based on the magnitude of the trend direction indicator value, and the negative correction step size is determined based on the value of the deviation direction indicator. The positive correction step size is added to the negative correction step size to obtain the comprehensive correction step size; The baseline value of feed conversion efficiency is added to the comprehensive correction step size to obtain the corrected baseline value, and the corrected baseline value is used as the predicted result of poultry feed conversion efficiency.

10. The method for predicting poultry feed conversion efficiency by incorporating machine learning according to claim 3, characterized in that, When normalizing each row of the initial distance matrix, the Softmax function is used to convert each row element into a probability distribution.

Images