Laying hen feed raw material sample screening method and system based on multi-source variability

Abstract

The application provides a laying hen feed raw material sample screening method and system based on multi-source variability, and relates to the technical field of data processing and analysis, which comprises obtaining multi-source attribute data and conventional component content data of raw materials, processing through a multi-source variability tensor, mapping attribute data into tensor modal dimensions, taking component data as characteristic components, and obtaining a multi-source variability tensor through decoupling and compression. Variability spectrum decomposition processing is performed, local rank spectrum decomposition is performed along the place of origin, time and component dimension, and a variability spectrum vector set is obtained. Through multi-scale extreme value and sparsity screening, a candidate modeling sample set is identified. Through leave-one-out method and sensitivity analysis, the contribution and sensitivity of the sample to the standard ileal amino acid digestibility prediction equation are evaluated, and the optimal modeling sample is screened out. The application can integrate multi-source variability information of raw materials, comprehensively cover the variability spectrum with the least sample amount, and significantly improve the precision and generalization ability of the prediction model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing analysis, in particular to a laying hen feed raw material sample screening method and system based on multi-source variability. BACKGROUND

[0002] In the field of laying hen feed nutrition research, accurately determining the standard ileal amino acid digestibility of feed raw materials is a key link for optimizing formula and achieving precision nutrition. However, traditional biological determination methods have problems such as high cost, long cycle and strong ethical constraints, prompting in vitro prediction models to become an important research direction.

[0003] The prior art usually determines the content of the conventional chemical components of the raw material, and performs sample screening based on the simple correlation relationship between the content and the digestibility, so as to reduce the sample size required for modeling in vivo. However, this method mainly relies on the difference of a single biochemical index of component dimension, and fails to systematically integrate multi-source variability information such as the origin, batch and processing time sequence of the raw material, resulting in that the sample set screened often cannot fully represent the complex variability structure of the raw material in the actual production system, so that the prediction model built has insufficient generalization ability and poor stability when dealing with multi-source raw materials. It is particularly worth noting that the existing screening methods are mostly based on linear correlation analysis or empirical rules, ignoring the potential influence of complex nonlinear interactions between multi-source attributes on digestibility. This simplified treatment cannot capture the essential characteristics of raw material variability in real scenarios, thereby limiting the accuracy of the prediction model in actual application.

[0004] Therefore, there is an urgent need for a laying hen feed raw material sample screening method and system based on multi-source variability to solve the above technical problems. SUMMARY

[0005] The purpose of the present application is to provide a laying hen feed raw material sample screening method and system based on multi-source variability to improve the above problems. In order to achieve the above purpose, the technical scheme adopted by the present application is as follows:

[0006] In a first aspect, the present application provides a laying hen feed raw material sample screening method based on multi-source variability, comprising:

[0007] obtaining multi-source attribute data and conventional component content data of the laying hen feed raw material sample, wherein the multi-source attribute data includes origin, time, processing plant and batch information, and the conventional component content data includes the content of total energy, crude fat, starch and neutral detergent fiber;

[0008] According to the multi-source attribute data and the conventional component content data, multi-source variability tensor construction processing is performed, the multi-source attribute data is mapped to independent modal dimension of the tensor, and the conventional component content data is decoupled and compressed as a feature component in the modal, to obtain a multi-source variability tensor;

[0009] According to the multi-source variability tensor, variability spectrum decomposition processing is performed, local rank spectrum decomposition is performed along the origin place dimension, the time dimension and the component dimension of the tensor, to obtain a variability spectrum vector set;

[0010] According to the variability spectrum vector set, multi-scale extreme value and sparsity screening processing is performed, extreme points, spectral energy coverage and sparsity centrality of the spectrum vector are calculated in the multi-scale variability space, candidate modeling samples are identified based on the calculation results, and a candidate modeling sample set is obtained;

[0011] According to the candidate modeling sample set, sample screening processing is performed, leave-one-out substitution and sensitivity curve analysis are performed in the multi-source variability spectrum space, the contribution and sensitivity of each candidate sample to the standard ileal amino acid digestibility prediction equation of laying hens are evaluated, and modeling samples are screened based on the evaluation results, to obtain all modeling samples in the laying hen feed raw material sample that meet the requirements.

[0012] In a second aspect, the application also provides a laying hen feed raw material sample screening system based on multi-source variability, comprising:

[0013] An acquisition unit is configured to acquire multi-source attribute data and conventional component content data of a laying hen feed raw material sample, wherein the multi-source attribute data includes origin place, time, processing plant and batch information, and the conventional component content data includes content of total energy, crude fat, starch and neutral detergent fiber;

[0014] A construction unit is configured to perform multi-source variability tensor construction processing according to the multi-source attribute data and the conventional component content data, map the multi-source attribute data to independent modal dimension of the tensor, decouple and compress the conventional component content data as a feature component in the modal, and obtain a multi-source variability tensor;

[0015] A decomposition unit is configured to perform variability spectrum decomposition processing according to the multi-source variability tensor, perform local rank spectrum decomposition along the origin place dimension, the time dimension and the component dimension of the tensor, and obtain a variability spectrum vector set;

[0016] A calculation unit is configured to perform multi-scale extreme value and sparsity screening processing according to the variability spectrum vector set, calculate extreme points, spectral energy coverage and sparsity centrality of the spectrum vector in the multi-scale variability space, identify candidate modeling samples based on the calculation results, and obtain a candidate modeling sample set;

[0017] The screening unit is used to screen samples based on the candidate modeling sample set. By performing leave-one-out substitution and sensitivity curve analysis in the multi-source variability spectrum space, it evaluates the contribution and sensitivity of each candidate sample to the standard ileal amino acid digestibility prediction equation for laying hens. Based on the evaluation results, it screens modeling samples to obtain all modeling samples that meet the requirements in the laying hen feed ingredient samples.

[0018] The beneficial effects of this invention are as follows:

[0019] This invention constructs a multi-source variability tensor by fusing multi-source attribute data with conventional component content data, achieving a structured characterization of multi-dimensional variation information of feed raw materials, including origin, time, processing plant, and batch. Significant variation features under different modalities are extracted through variability spectrum decomposition, effectively capturing key variation patterns affecting amino acid digestibility. Candidate samples with both variation representativeness and spectral spatial coverage are accurately identified based on multi-scale extrema and sparsity screening. Finally, optimization through leave-one-out method and sensitivity analysis ensures that the final modeling sample set maintains variation diversity while possessing optimal predictive robustness, thus comprehensively covering the raw material variation spectrum with the minimum sample size. This lays a solid foundation for establishing a high-precision, highly generalizable standard ileal amino acid digestibility prediction model.

[0020] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of the process for screening sample feed ingredients for laying hens based on multi-source variability, as described in an embodiment of the present invention.

[0023] Figure 2 This is a schematic diagram of the structure of the egg-laying hen feed ingredient sample screening system based on multi-source variability as described in an embodiment of the present invention.

[0024] In the diagram: 701, Acquisition Unit; 702, Construction Unit; 703, Decomposition Unit; 704, Calculation Unit; 705, Filtering Unit. Detailed Implementation

[0025] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some but not all of the embodiments of the present application. The components of the embodiments of the present application described and shown in the drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0026] It should be noted that: similar reference numerals and letters represent similar items in the following drawings, therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. Meanwhile, in the description of the present application, the terms "first", "second" and the like are only used to distinguish description, and cannot be understood as indicating or implying relative importance.

[0027] Embodiment 1

[0028] The embodiment provides a laying hen feed raw material sample screening method based on multi-source variability.

[0029] Referring to Figure 1 , the method includes steps S1, S2, S3, S4 and S5.

[0030] In step S1, multi-source attribute data and conventional component content data of the laying hen feed raw material sample are acquired, wherein the multi-source attribute data includes origin, time, processing plant and batch information, and the conventional component content data includes content of total energy, crude fat, starch and neutral detergent fiber;

[0031] It can be understood that in this step, the multi-source attribute data and the conventional component content data of the laying hen feed raw material are collected by a system, and an initial data set comprehensively reflecting variability characteristics of the raw material is constructed. The multi-source attribute data not only includes environmental factors such as geographical coordinates and climate characteristics of the origin, but also covers specific year and month identifiers of the purchase time, equipment types and process parameters of the processing plant, and traceable information of the raw material batch. These multi-dimensional attribute data are automatically collected through a production record system and Internet of Things sensing equipment, ensuring the authenticity and integrity of the data source. The conventional component data is measured by a near-infrared spectrum analyzer to determine the total energy value, a Soxhlet extraction method is used to determine the crude fat content, an enzyme hydrolysis method is used to determine the starch content, and a fiber analyzer is used to determine the neutral detergent fiber content. These standardized detection methods ensure the accuracy and comparability of the component data.

[0032] In order to ensure the prediction accuracy of the standard ileal amino acid digestibility prediction equation, the present application collects feed raw materials of different origins, different times, different processing plants and different batches as samples in advance, so as to enrich the modeling sample library. In order to further reduce the workload and cost, the present application also screens the feed raw materials of laying hens. Specifically, the contents of several conventional components (such as total energy, crude fat, starch, and neutral detergent fiber) in the poultry feed raw material samples can be determined first, then based on the contents of the above components, the correlation coefficient between the above components and the standard ileal amino acid digestibility is determined, then according to the correlation coefficient, the conventional indicators with higher correlation are selected from the above components, and based on the content of the conventional indicators, the standard ileal amino acid digestibility is calculated, so that the laying hen feed raw material samples with large differences in composition and standard ileal amino acid digestibility are selected as modeling samples for later standard ileal amino acid digestibility determination.

[0033] For example, first, 30 kinds of secondary powder samples of different seasons, different products and different batches are collected, then 10 kinds of secondary powder raw material samples with large differences in composition and standard ileal amino acid digestibility are selected as modeling samples for later standard ileal amino acid digestibility determination, and the 10 kinds of secondary powder feed raw material samples are from different places. Thus, by reducing the number of standard ileal amino acid digestibility determination modeling samples, manpower, material resources, financial resources and time can be saved.

[0034] This step establishes a multi-dimensional data set that can simultaneously represent the external source characteristics and internal nutritional characteristics of the raw materials, providing a reliable data foundation for subsequent variability analysis. In particular, through the fine collection of multi-source attributes, the limitations of traditional methods focusing only on component data are broken through, enabling subsequent analysis to capture the comprehensive effects of potential influencing factors such as production environment and processing technology on amino acid digestibility.

[0035] Step S2, multi-source variability tensor construction processing is performed according to the multi-source attribute data and the conventional component content data, the multi-source attribute data is mapped to the independent modal dimension of the tensor, and the conventional component content data is decoupled and compressed as the feature component within the modal to obtain a multi-source variability tensor;

[0036] It can be understood that this step firstly maps the multi-source attribute data such as origin, time, processing factory, etc. into independent modal dimensions of the tensor, so that each modal can maintain its unique variation characteristics; at the same time, the conventional component content data is taken as the feature component inside each modal, the nonlinear coupling relationship between the modes is decoupled through the self-adaptive kernel entropy mapping algorithm, and data compression is realized by using Tucker decomposition. This step constructs a multi-dimensional feature representation that can maintain data integrity and reduce computational complexity, laying a structural foundation for subsequent variation pattern mining, and can better reflect the essential characteristics of raw material variation in actual production compared with traditional methods. In this step, step S2 includes step S21, step S22 and step S23.

[0037] Step S21, according to the multi-source attribute data, carries out multi-source attribute coding processing, and through embedding and mapping based on the geographical information entropy of the laying hen feed raw material origin and the periodicity of the processing time, the multi-source attribute data is converted into a continuous numerical feature vector, and a coded multi-source attribute feature set is obtained.

[0038] It can be understood that this step quantifies the comprehensive differences of different regional environmental factors based on the geographical information entropy of the origin, and uses the periodic embedding and mapping of the processing time to capture the regular characteristics of the raw material changing with time during production and storage. Discrete origin, time, factory, etc. are converted into continuous and calculable feature vectors. This method breaks through the limitations of traditional simple processing methods such as one-hot encoding, and in the feed raw material scene, geographical information entropy can effectively represent the cumulative influence of different production area climate, soil and other environmental factors on the formation of raw material nutrients, and time periodic embedding can reflect the variation law in the time dimension such as crop growth period and processing season.

[0039] The quantization formula of the geographical information entropy of the origin is as follows:

[0040] ;

[0041] wherein, E g (l geo ) represents the geographical information entropy of the origin l geo , which is a continuous scalar value, w i is the weight of the i-th environmental factor, N e represents the number of environmental factor categories, K i represents the number of possible states of the i-th environmental factor, p i,j (l geo ) represents the frequency or probability of the i-th environmental factor appearing in the j-th state within the range of the origin l geo .

[0042] The periodic embedding formula of the processing time is as follows:

[0043] ;

[0044] where T(t) is the embedding vector at time point t, a t and β t are learnable scaling parameters or fixed amplitudes set according to domain knowledge, used to adjust the influence strength of different periodic components on the final feature, and π is the circular constant, is the period length.

[0045] Step S22, the encoded multi-source attribute feature set and the conventional component content data are subjected to multi-modal tensor initialization processing, each multi-source attribute feature vector is taken as an independent modal dimension of a tensor, and the conventional component content data is taken as a feature component within a modal to perform dimension alignment and splicing, to obtain an initial multi-modal tensor;

[0046] It can be understood that this step first divides the multi-source attribute feature set subjected to the encoding processing into modal dimensions, and defines the origin feature vector, the time feature vector, and the processing plant feature vector as different modal dimensions of a tensor. Then, the feature standardization processing is performed on the conventional component content data to eliminate the dimensional difference between different component indicators. Then, the tensor expansion technique is adopted to expand the dimension of the feature vector corresponding to each attribute modal, so that it can accommodate the component data features. When performing the dimension alignment, the system solves the problem of inconsistent dimensions between different modals by constructing a block circulant matrix, and realizes the unified mapping of the feature space by using the Kronecker product operation. Finally, the multi-dimensional array splicing technique is adopted to assemble the feature matrices of each modal according to the tensor structure, wherein each tensor element corresponds to all component data of the raw material sample of a specific origin, a specific time, and a specific processing plant, so as to form a complete initial multi-modal tensor. This processing process guarantees the structural integrity of the multi-source data through strict mathematical transformation, and lays a solid foundation for the subsequent deep tensor analysis.

[0047] It can be understood that the construction formula of the feature matrix in this step is as follows:

[0048] ;

[0049] wherein, is the feature matrix corresponding to the i-th category of modal A, is all rows with index belonging to selected from the standardized component data matrix X std , to form a submatrix. Each row of the submatrix is a sample, and each column is a standardized component, W c is an optional component weight diagonal matrix based on domain knowledge.

[0050] Step S23, the initial multi-modal tensor is subjected to nonlinear decoupling and compression processing, the nonlinear coupling between modes is decoupled through an adaptive kernel entropy mapping algorithm, and dimension reduction compression is performed by using a Tucker decomposition method, to obtain a multi-source variability tensor representing source heterogeneity and component difference.

[0051] It can be understood that this step first calculates the mutual information entropy between different modal feature vectors through a Gaussian kernel function, identifies the nonlinear coupling relationship between multi-dimensional attributes such as origin-time-factory, constructs a modal correlation matrix, separates the independent variability components of each mode by using kernel principal component analysis, and at the same time, retains significant interaction effects. Then, the Tucker decomposition method is used, and the core tensor and factor matrix are solved by the alternating least squares algorithm, and the high-dimensional data is projected to a low-dimensional subspace while maintaining the integrity of the tensor structure. In this process, the step will automatically determine the retained dimensions of each mode according to the eigenvalue contribution rate, and realize intelligent filtering of redundant information. This processing is particularly important in the feed raw material screening scene, because it can effectively solve the complex problem of interweaving of component variability and processing technology caused by environmental factors such as climate and soil of raw materials from different origins. The technical effect of this step is that both noise interference in multi-source data is eliminated and key variability features are retained, so that the obtained multi-source variability tensor can accurately represent the source heterogeneity of raw materials and reflect the essential characteristics of their component differences, providing a high-quality data basis for subsequent variability pattern analysis.

[0052] It can be understood that the calculation formula of mutual information entropy in this step is as follows:

[0053] ;

[0054] Wherein, MI(A;B) is the mutual information entropy between modal A and modal B, (a i ,b j ) represents the i-th category of modal A and the j-th category of modal B, I is the total number of categories of modal A, J is the total number of categories of modal B, p(a i ,b j ) is the joint probability, and p(a i )p(b j ) represents the marginal probability.

[0055] Wherein, the regularization Tucker decomposition objective function is as follows:

[0056] ;

[0057] Wherein, T is the initial multi-modal tensor, G is the core tensor, U (1) is the factor matrix of the origin mode, U (2) is the factor matrix of the time mode, and U (3)is a factor matrix of the component modalities, denotes the Frobenius norm of a matrix, λ reg is a regularization parameter, MI(m, n) denotes mutual information entropy between modalities m and n, R(U (m) , U (n) ) denotes a regularization term.

[0058] Step S3, performing a variability spectrum decomposition process according to the multi-source variability tensor, obtaining a set of variability spectrum vectors by performing local rank spectrum decomposition along the provenance dimension, the time dimension and the component dimension of the tensor respectively;

[0059] It can be understood that this step can effectively separate the systematic variability caused by inherent factors such as provenance environment and seasonal change, and filter out the noise caused by random fluctuations. This step converts high-dimensional tensor data into a set of spectrum vectors with clear physical meaning. These vectors not only retain the multi-dimensional variability characteristics of the original data, but also significantly improve the efficiency of subsequent calculations through dimension reduction processing, providing a feature basis for accurately identifying representative samples. In this step, step S3 includes step S31, step S32 and step S33.

[0060] Step S31, performing a single-modality local rank analysis process according to the multi-source variability tensor, calculating the local rank distribution of each modality along the provenance dimension, the time dimension and the component dimension respectively, and obtaining the variability components of each modality based on a preset dynamic rank variation rate threshold;

[0061] It can be understood that this step first divides the samples into several regional subsets based on the geographical clustering results in the origin dimension, and then calculates the eigenvalue distribution of the covariance matrix in each subset through singular value decomposition, and adaptively determines the local rank order combined with the sample density, effectively capturing the systematic variation characteristics caused by different production areas due to environmental factors such as soil and climate. Then, in the time dimension, the system adopts a sliding time window strategy to divide the continuous time series into segments with biological significance (such as crop growing season, processing period, etc.), and in each time window, the cumulative contribution rate of the eigenvalue is calculated through principal component analysis, and when the contribution rate changes rate exceeds the dynamic threshold, the rank transition point is identified, so as to accurately identify the seasonal variation pattern caused by seasonal change, storage period and other time factors. Then, for the component dimension, the correlation network analysis is used to construct the correlation map between the nutritional components, and the weight distribution of each component in the local area is determined based on the node centrality index, and the dominant component combination is determined by combining the eigenvalue decay gradient analysis. The unique feature of this step is the introduction of a dynamic rank change rate threshold mechanism. This threshold is not a fixed value, but is adaptively adjusted according to the distribution characteristics of each modal data. For the modal with larger variation (such as the origin dimension), a more relaxed threshold is used to retain more detailed features, while for the relatively stable modal (such as the component dimension), a strict threshold is used to filter noise. In the actual application of feed raw material screening, this method can effectively distinguish between inherent variation caused by geographical differences and accidental fluctuations caused by random factors, such as effectively separating the stable component characteristics of raw materials from different production areas from the accidental abnormal values of individual batches. This step accurately extracts the significant variation components with statistical significance in each modal, providing clean and biologically meaningful input data for subsequent cross-modal coupling analysis.

[0062] wherein the time window cumulative contribution rate is as follows:

[0063] ;

[0064] wherein CumVar(r,t) is the cumulative contribution rate of the first r eigenvalues at time t, r is the number of eigenvalues, k is the kth eigenvalue, ω (k,t) is the kth eigenvalue at time t, is the total number of eigenvalues.

[0065] wherein the correlation coefficient matrix of the nutritional components is as follows:

[0066] ;

[0067] wherein r gq is the correlation coefficient between components g and q, x ig represents the value of component g of sample i, represents the average value of component g, and n represents the number of samples.

[0068] Step S32, according to the variation component of each mode, a cross-modal coupling analysis process is performed, by introducing the physiological parameters of laying hens as constraint conditions, the synergistic and antagonistic relationship of the variation components between different modes is analyzed, and a coupling strength matrix between modes is obtained;

[0069] It can be understood that this step first constructs a coupling detection framework based on canonical correlation analysis, respectively calculates the canonical correlation objective function between the modes of origin-component, time-component, factory-component, and introduces the physiological parameters such as digestive enzyme activity and intestinal retention time of laying hens as regularization constraint terms. The least square method is used to optimize the synergism of the digestive efficiency when solving the correlation weight between modes, and the conditional mutual information entropy is calculated to distinguish the real biological coupling from the false statistical correlation. For the significant coupling relationship identified, further division of synergistic and antagonistic types is performed through structural equation model—when the variation directions of two modes jointly promote the improvement of the digestibility, it is marked as synergistic relationship, otherwise, if it produces offsetting effect, it is marked as antagonistic relationship. In the specific application of feed raw material screening, this method can effectively identify cross-modal interactions such as "high fiber raw materials have a negative impact on the digestibility under the processing conditions of a specific production area", and avoid misjudgment caused by single-dimensional analysis. This step can generate a coupling matrix that can quantify the interaction strength between modes. The matrix not only contains statistical significance information, but also reflects the biological directionality of the interaction, providing a correlation network basis verified by physiology for subsequent fusion of variation spectrum vectors.

[0070] The physiological constraint canonical correlation objective function is as follows:

[0071] ;

[0072] wherein w A and w B are weight vectors, represents a covariance matrix, represents the transpose of w A , λ cca is a regularization parameter of physiological constraint, Corr is a correlation coefficient, U A and U B are characteristic matrices, and Z is a digestive physiological parameter.

[0073] The formula of the synergistic and antagonistic relationship determination function is as follows:

[0074] ;

[0075] wherein η AB is a synergistic and antagonistic coefficient, and MI(A;B|Z) denotes conditional mutual information entropy for the derivative of standard ileal amino acid digestibility (SID) to the characteristic.

[0076] Step S33, according to the inter-modal coupling strength matrix, a spectral vector fusion process is performed, the variation components and coupling relationship of each mode are fused through a nonlinear tensor contraction operation, and a variation degree spectrum vector set representing the energy distribution of the sample in the multi-source variation degree space is generated.

[0077] It can be understood that in this step, the variation components of each mode are first organized into a high-order tensor structure, wherein the feature vector of each mode is taken as a dimension of the tensor, and the coupling strength matrix is embedded in the feature space of the tensor in the form of a block diagonal matrix. By using a nonlinear tensor contraction operation based on Tucker decomposition, the core tensor and the factor matrix are iteratively solved by an alternating least squares method, and in this process, the coupling strength is introduced as a regularization constraint term to ensure that the fusion process maintains the biological correlation characteristics between modes.

[0078] In the specific application of feed raw material screening, this step will automatically adjust the fusion weight according to the digestion physiological characteristics of laying hens, for example, a higher fusion coefficient is given to the mode with stronger correlation with amino acid digestibility. The variation degree spectrum vector set generated in this step not only retains the independent variation information of each mode, but also encodes the interaction mode between modes, and the energy distribution characteristics directly reflect the position and importance of the sample in the overall variation space. This step realizes the deep integration of multi-source variation characteristics, overcoming the limitation of traditional methods that can only process single-dimensional variation; through nonlinear fusion, the representation ability of the characteristics is significantly improved, so that the subsequent screening process can more accurately evaluate the representativeness of the sample; the generated spectrum vector has clear biological interpretability, providing a new perspective for understanding the influence mechanism of raw material variation on digestibility.

[0079] It can be understood that in this step, the objective function of the nonlinear tensor contraction operation is as follows:

[0080] ;

[0081] wherein U is an input tensor, V is an approximate tensor, , and are the transposes of the transformation matrix, B is a weight matrix, λ is a regularization parameter, tr is a trace operation, W is a matrix before transformation, and W T is the transpose of the matrix before transformation.

[0082] Step S4, according to the variation degree spectrum vector set, a multi-scale extreme value and sparsity screening process is performed, the extreme points, spectral energy coverage and sparsity centrality of the spectrum vector are calculated in the multi-scale variation degree space, and based on the calculation results, candidate modeling samples are identified to obtain a candidate modeling sample set.

[0083] It can be understood that this step realizes intelligent screening of candidate samples by establishing a multi-level feature analysis framework, which can take into account both the extremeness and the uniformity of the variation characteristics. For example, in the screening of feed raw materials, not only can special samples from special production areas with abnormal nutritional components be identified, but also can the coverage of different variation types be ensured. Through multi-scale analysis, the limitations of single-scale screening are avoided, and the robustness of sample selection is improved. The spectral energy coverage evaluation ensures the complete representation of the feature space. The sparsity optimization maximizes the reduction of sample quantity under the premise of ensuring representation, and provides a high-quality candidate set for subsequent fine screening. In this step, step S4 includes steps S41, S42 and S43.

[0084] Step S41, multi-scale extreme value detection processing is performed according to the variation spectrum vector set, and by using an adaptive threshold mechanism based on the preset correlation of the standard ileal amino acid digestibility, extreme points with significant variation characteristics are identified at different scales in the spectrum space, and a multi-scale extreme sample set is obtained.

[0085] It can be understood that this step first uses wavelet transform to perform multi-scale decomposition on the variation spectrum vector, forming a multi-layer subspace containing different granularity features in the time-frequency domain. On this basis, a scale-adaptive threshold function is constructed based on the correlation pattern between the standard ileal amino acid digestibility and the spectral characteristics in the historical database. At the macro scale, structural variation characteristics such as raw material production area and batch are emphasized, and a relatively loose threshold is used to capture the overall variation trend. At the micro scale, the focus is on the subtle fluctuations of nutritional components, and a strict threshold is used to filter specific variation patterns. During the processing, this step dynamically evaluates the correlation significance of the extreme points at each scale and the digestibility. When it is detected that certain spectral characteristics are stably correlated with the digestibility, the detection sensitivity of the corresponding scale is automatically enhanced. This method has unique value in feed raw material screening, such as effectively identifying abnormal nutritional component distribution of raw materials in some areas due to special processing technology, while avoiding misjudgment of accidental fluctuations as significant variations. This step constructs an intelligent detection system closely related to biological effects, which not only ensures the statistical significance of variation feature extraction, but also ensures that these variations have actual nutritional significance, providing high-quality basic data for subsequent sample screening.

[0086] The scale-adaptive threshold function is as follows:

[0087] ;

[0088] wherein, is the threshold, μ l is the mean, σ l is the standard deviation, δ l is the scaling factor, f(rl ) is a correlation coefficient.

[0089] wherein the extreme point detection formula is as follows:

[0090] ;

[0091] wherein Extreme(s, k) is an extreme point indication function, s is a scale, is a wavelet coefficient, is a p value, γ l is a significance threshold.

[0092] Step S42, performing spectral space coverage analysis processing according to the multi-scale extreme sample set, obtaining a coverage optimized sample subset by calculating the coverage contribution degree of each extreme sample in the spectral energy space;

[0093] It can be understood that this step first constructs the feature space topology of the variability spectrum vector, divides the influence range of each extreme sample using a Voronoi diagram, and quantifies the local coverage contribution degree by calculating the energy spectrum density integral in the Voronoi unit corresponding to each sample. This step will pay special attention to the radiation ability of the sample point in different variation mode directions, and use radial basis function interpolation to estimate the spectral energy distribution of the uncovered area, so as to identify the spectral space blind area that the current sample set cannot fully represent. When processing the multi-source variation characteristics specific to the feed raw material, this method can effectively solve the problem of uneven space coverage in traditional screening, for example, discovering that some raw material samples with special ingredient combinations, although not belonging to extreme variation points, have important supplementary value for fully representing the variation spectrum. Through the quantitative analysis of the energy spectrum space, this step ensures that the selected sample subset can achieve the maximum coverage of the variation feature space while retaining the significant extreme points, avoiding the incomplete representation of the spectrum caused by excessive attention to extreme values in traditional methods, and laying a spatial distribution foundation for constructing a sample set with comprehensive representation.

[0094] wherein the calculation method of the coverage contribution degree is:

[0095] ;

[0096] wherein C i is the coverage of the sample, E i is the spectral energy of the sample, is the average coverage radius, A i is the area, is the maximum spectral energy of the sample, is the maximum average coverage radius of the sample, is the maximum area, α c , β c and γ c represent weight parameters.

[0097] The calculation formula of the spectral energy of the uncovered area is as follows:

[0098] ;

[0099] wherein, is the spectral energy of the node v, m is the number of points of the uncovered area, w i is the weight of the uncovered area, Φ is a radial basis function, is the difference between the node v and the i-th node, P(v) is a polynomial term.

[0100] Step S43, perform a sparsity centrality evaluation process according to the coverage-optimized sample subset, identify the key sample points that can minimize the number of samples while maintaining the integrity of the spectral features by constructing a sparsity distribution map of the variability spectrum space, and construct a candidate modeling sample set by taking all the key sample points as candidate modeling samples.

[0101] It can be understood that this step first maps the coverage-optimized sample subset into a network structure in the spectral feature space, taking samples as nodes and the spectral feature similarity between samples as edge weights, to construct a weighted undirected graph model; on this basis, this step uses a pre-set betweenness centrality algorithm to calculate the frequency of each sample point appearing in all shortest paths, and introduces spectral feature fidelity as a constraint condition to ensure that significant collapse of the feature space will not occur when removing redundant samples. When dealing with the specific multi-dimensional variability characteristics of feed raw materials, this method can accurately identify key samples at the intersection of multiple variability modes, which often have unique value in representing multiple variability characteristics. This step uses network centrality analysis to systematically select key sample points that can best represent overall variability characteristics while ensuring the integrity of the spectral feature space, achieving an optimal balance between sample quantity and representation ability, and providing an optimized sample basis for establishing an efficient and accurate prediction model.

[0102] The calculation formula of the frequency of appearing in the shortest path is as follows:

[0103] ;

[0104] wherein, BC(v i ) represents the betweenness centrality of the i-th node, v s and v t represent the source node and the target node, σ st represents the number of shortest paths from the source node to the target node, and f represents a function.

[0105] Step S5, sample screening processing is performed according to the candidate modeling sample set, the contribution degree and sensitivity of each candidate sample to the standard ileal amino acid digestibility prediction equation of laying hens are evaluated by performing leave-one-out method substitution and sensitivity curve analysis in the multi-source variability spectrum space, and all modeling samples in the laying hen feed raw material samples that meet the requirements are screened based on the evaluation results to obtain the modeling samples.

[0106] It can be understood that the leave-one-out cross-validation framework is adopted, the prediction equation is reconstructed after removing a single sample from the candidate set in turn, and the contribution degree of each sample is quantified by comparing the change amplitude of the model parameters and the fluctuation degree of the prediction error; At the same time, the sensitivity curve analysis is introduced, the change gradient of the prediction result is observed by gradually changing the variability spectrum characteristic value of the sample, so as to evaluate the dependence degree of the model on the sample. In this step, step S5 includes step S51, step S52 and step S53.

[0107] Step S51, the contribution degree analysis processing is performed according to the candidate modeling sample set, the influence degree of each candidate sample on the preset standard ileal amino acid digestibility prediction equation of laying hens is calculated by removing a single candidate sample in turn in the variability spectrum space, and the prediction stability disturbance matrix of each sample is obtained;

[0108] It can be understood that this step adopts an iterative calculation mode, after removing each candidate sample from the training set in turn, the standard ileal amino acid digestibility prediction equation is retrained (the prediction equation is: ileal amino acid digestibility is equal to the sum of the concentration, pH value and other parameters multiplied by the coefficient), and the sample contribution degree is evaluated by comparing the change amplitude of the model parameters before and after removal. Among them, the Euclidean distance change of the prediction equation coefficient vector, the fluctuation amplitude of the determination coefficient and the variability of the prediction error are monitored, and the disturbance entropy concept is introduced to comprehensively measure the influence degree of sample removal on the stability of the model. In the actual application of laying hen feed raw material screening, this step can effectively identify those special samples that have a lever effect on the model, such as some raw material samples with unique nutrient ingredient combination, which although rare in quantity, play a key role in establishing an accurate prediction equation.

[0109] The calculation formula of disturbance entropy is as follows:

[0110] ;

[0111] Wherein, H k represents the disturbance entropy of the sample, p k represents the disturbance probability.

[0112] The step generates a prediction stability disturbance matrix through quantitative analysis, which not only reflects the influence strength of a single sample on the prediction equation, but also reveals the interaction relationship between different samples, providing data support for subsequent sensitivity optimization, so as to ensure that the finally screened sample set can support the construction of a prediction model with good generalization ability.

[0113] In step S52, sensitivity optimization processing based on digestive physiological characteristics is performed according to the prediction stability disturbance matrix. By introducing the biological coefficient of variation of the standard ileal amino acid digestibility of laying hens as a weight factor, the disturbance matrix is weighted and reconstructed to obtain a sensitivity optimized sample set.

[0114] It can be understood that this step first constructs a database of biological coefficients of variation of different amino acid digestibilities based on laying hen digestive physiological experimental data, which serves as the calculation basis for the weight factor. This step adopts a hierarchical weighting strategy to weight and reconstruct the prediction stability disturbance matrix: first, the basic weight is allocated according to the variation degree of the amino acid digestibility (for example, higher weight is given to amino acids with greater variation), and then the actual contribution of different raw material samples in digestion and metabolism is dynamically adjusted.

[0115] The formula of the biological coefficient of variation of the standard ileal amino acid digestibility is as follows:

[0116] ;

[0117] wherein BCV a represents the biological coefficient of variation of the amino acid, σ a represents the standard deviation of the amino acid, and μ a represents the mean of the amino acid.

[0118] Then, the weight matrix and the disturbance matrix are operated element by element through Hadamard product, so that the prediction sensitive points corresponding to the highly variable amino acids obtain greater evaluation weight. In the special scenario of laying hen feed raw material screening, this step effectively solves the problem of ignoring the biological significance in traditional statistical optimization, such as giving priority to the variation characteristics related to essential amino acids (such as lysine and methionine) of laying hens, to ensure that the screened samples can better reflect the key nutritional variations in actual feeding. This step converts the simple statistical disturbance information into a sensitivity index with physiological significance, making the subsequent screening process more in line with the actual digestive characteristics of laying hens, and significantly improving the biological representativeness and prediction practicability of the final modeling sample set.

[0119] In step S53, multi-objective screening processing is performed according to the sensitivity optimized sample set. By establishing the Pareto optimal frontier between variation coverage and prediction sensitivity, the prediction error risk is minimized while ensuring the integrity of the spectral space, and all modeling samples meeting the requirements are obtained.

[0120] It can be understood that this step first constructs an optimization function targeting at maximizing the variability coverage and minimizing the prediction sensitivity, and adopts the non-dominated sorting genetic algorithm to find the Pareto optimal solution set in the solution space. In the processing, this step generates multiple combination schemes of the candidate sample set, and evaluates the comprehensive performance of each scheme in the target space by calculating the hyper volume index thereof.

[0121] In particular, for the special requirements of screening of feed raw materials for laying hens, the process further introduces a biological constraint condition to ensure that the finally selected sample set is not only mathematically optimal, but also meets the nutritional evaluation requirements in actual feeding. This step balances the representativeness and practicality of the sample set through a multi-objective optimization framework, ensuring comprehensive coverage of variability characteristics while controlling the uncertainty risk of the prediction model, so that the obtained modeling sample set has optimal prediction stability while maintaining the integrity of the spectrum space.

[0122] The optimization function is as follows:

[0123] ;

[0124] Among them, F1(X) is the coverage function, F2(X) is the sensitivity function, Coverage(X) represents the variability coverage, Sensitivity(X) represents the sensitivity, X represents the sample subset, represents the optimized sample set, N min ,N max represents the sample number range.

[0125] The hyper volume index is as follows:

[0126] ;

[0127] Among them, HV(X) represents the hyper volume, P represents the Pareto solution set, and represent the reference points, X i represents the i-th solution in the Pareto solution set, F1(X i ) represents the coverage function of the i-th solution in the Pareto solution set, F2(X i ) represents the sensitivity function of the i-th solution in the Pareto solution set, and Volume represents the volume calculation function.

[0128] The selection formula of the optimal solution set is as follows:

[0129] ;

[0130] Among them, M is the final modeling sample set, X * is the dominant solution set in the Pareto optimal frontier, Coverage AA Coverage AA (X * ) represents the amino acid coverage of the original solution set, Coverage d after adding sample v Coverage

[0131] Embodiment 2

[0132] As Figure 2 shown, the embodiment provides a laying hen feed raw material sample screening system based on multi-source variability, see Figure 2 The system comprises an acquisition unit 701, a construction unit 702, a decomposition unit 703, a calculation unit 704 and a screening unit 705.

[0133] The acquisition unit 701 is configured to acquire multi-source attribute data and conventional component content data of laying hen feed raw material samples, wherein the multi-source attribute data comprises origin, time, processing plant and batch information, and the conventional component content data comprises content of total energy, crude fat, starch and neutral detergent fiber;

[0134] The construction unit 702 is configured to perform multi-source variability tensor construction processing according to the multi-source attribute data and the conventional component content data, map the multi-source attribute data to independent modal dimensions of the tensor, and decouple and compress the conventional component content data as feature components within the modal to obtain a multi-source variability tensor;

[0135] The decomposition unit 703 is configured to perform variability spectrum decomposition processing according to the multi-source variability tensor, perform local rank spectrum decomposition along the origin dimension, time dimension and component dimension of the tensor respectively, and obtain a variability spectrum vector set;

[0136] The calculation unit 704 is configured to perform multi-scale extreme value and sparsity screening processing according to the variability spectrum vector set, calculate extreme points, spectral energy coverage and sparsity centrality of the spectrum vector in the multi-scale variability space, identify candidate modeling samples based on the calculation results, and obtain a candidate modeling sample set;

[0137] The screening unit 705 is configured to perform sample screening processing according to the candidate modeling sample set, perform leave-one-out substitution and sensitivity curve analysis in the multi-source variability spectrum space, evaluate the contribution and sensitivity of each candidate sample to the standard ileal amino acid digestibility prediction equation of laying hens, and screen modeling samples based on the evaluation results to obtain all modeling samples in the laying hen feed raw material samples that meet the requirements.

[0138] It should be noted that the specific manner in which the various modules perform operations in the system of the above embodiments has been described in detail in the embodiments related to the method, and will not be described in detail here.

[0139] The above only describes the preferred embodiments of the present application, and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

[0140] The above only describes the preferred embodiments of the present application, and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

[0140] The above only describes the preferred embodiments of the present application, and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

[0140] The above only describes the preferred embodiments of the present application, and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for screening feedstock samples of laying hens based on multi-source variability, characterized in that, The method comprises the following steps: obtaining multi-source attribute data and conventional component content data of the layer feed raw material sample, wherein the multi-source attribute data comprises origin, time, processing plant and batch information, and the conventional component content data comprises content of total energy, crude fat, starch and neutral detergent fiber; performing multi-source variability tensor construction processing according to the multi-source attribute data and the conventional component content data, mapping the multi-source attribute data into independent modal dimensions of a tensor, and decoupling and compressing the conventional component content data as feature components within the modal to obtain a multi-source variability tensor; performing variability spectrum decomposition processing according to the multi-source variability tensor, performing local rank spectrum decomposition along the origin dimension, the time dimension and the component dimension of the tensor respectively to obtain a variability spectrum vector set; performing multi-scale extreme value and sparsity screening processing according to the variability spectrum vector set, calculating extreme points, spectral energy coverage and sparsity centrality of the spectrum vector in the multi-scale variability space, and identifying candidate modeling samples based on the calculation results to obtain a candidate modeling sample set; performing sample screening processing according to the candidate modeling sample set, performing leave-one-out substitution and sensitivity curve analysis in the multi-source variability spectrum space to evaluate the contribution and sensitivity of each candidate sample to the standard ileal amino acid digestibility prediction equation of the layer, and screening modeling samples based on the evaluation results to obtain all modeling samples in the layer feed raw material sample that meet the requirements.

2. The method for screening feedstock samples of layers based on multi-source variability according to claim 1, characterized in that The method comprises the following steps: performing multi-source attribute coding processing according to the multi-source attribute data, embedding and mapping based on the geographical information entropy of the layer feed raw material origin and the periodicity of the processing time to convert the multi-source attribute data into a continuous numerical feature vector to obtain an encoded multi-source attribute feature set; performing multi-modal tensor initialization processing on the encoded multi-source attribute feature set and the conventional component content data, taking each multi-source attribute feature vector as an independent modal dimension of a tensor, and performing dimension alignment and splicing on the conventional component content data as feature components within the modal to obtain an initial multi-modal tensor; performing nonlinear decoupling and compression processing on the initial multi-modal tensor, decoupling the nonlinear coupling between the modes by an adaptive kernel entropy mapping algorithm, and performing dimension reduction compression by a Tucker decomposition method to obtain a multi-source variability tensor representing source heterogeneity and component difference.

3. The multi-source variability based feedstock sample screening method for laying hens according to claim 1, wherein The method comprises the following steps: performing single-modal local rank analysis processing according to the multi-source variability tensor, calculating the local rank distribution of each mode along the origin dimension, the time dimension and the component dimension, and obtaining the variability components of each mode based on a preset dynamic rank variation rate threshold; performing cross-modal coupling analysis processing according to the variability components of each mode, analyzing the synergistic and antagonistic relationship between the variability components in different modes by introducing the layer digestion physiological parameters as a constraint condition to obtain a mode coupling strength matrix; Spectral vector fusion processing is performed according to the inter-modal coupling strength matrix, and each modal variation component and coupling relationship are fused through nonlinear tensor contraction operation to generate a variation degree spectrum vector set representing energy distribution of the sample in a multi-source variation degree space.

4. The method for screening feedstock samples of layers based on multi-source variability according to claim 1, characterized in that Multi-scale extreme value and sparsity screening processing is performed according to the variation degree spectrum vector set, including: Multi-scale extreme value detection processing is performed according to the variation degree spectrum vector set, and an adaptive threshold mechanism based on a preset correlation of standard ileal amino acid digestibility of the laying hen is used to identify extreme value points with significant variation characteristics at different scales in the spectrum space to obtain a multi-scale extreme value sample set; Spectral space coverage analysis processing is performed according to the multi-scale extreme value sample set, and a coverage optimization sample subset is obtained by calculating the coverage contribution of each extreme value sample in the spectrum energy space; Sparsity centrality evaluation processing is performed according to the coverage optimization sample subset, and a sparsity distribution map of the variation degree spectrum space is constructed to identify key sample points that can minimize the number of samples while maintaining the integrity of the spectrum characteristics, and all key sample points are used as candidate modeling samples to construct a candidate modeling sample set.

5. The multi-source variability based feedstock sample screening method for laying hens according to claim 1, wherein Sample screening processing is performed according to the candidate modeling sample set, including: Leave-one-out contribution analysis processing is performed according to the candidate modeling sample set, and the influence of each candidate sample on a preset standard ileal amino acid digestibility prediction equation of the laying hen is calculated by removing the sample one by one in the variation degree spectrum space to obtain a prediction stability perturbation matrix of each sample; Sensitivity optimization processing based on digestion physiological characteristics is performed according to the prediction stability perturbation matrix, and a biological variation coefficient of the standard ileal amino acid digestibility of the laying hen is introduced as a weight factor to weight and reconstruct the perturbation matrix to obtain a sensitivity optimized sample set; Multi-objective screening processing is performed according to the sensitivity optimized sample set, and a Pareto optimal frontier between variation degree coverage and prediction sensitivity is established to minimize the prediction error risk while ensuring the integrity of the spectrum space to obtain all modeling samples meeting the requirements.

6. A multi-source variability-based layer feed raw material sample screening system, characterized in that, including: An acquisition unit is configured to acquire multi-source attribute data and conventional component content data of a laying hen feed raw material sample, wherein the multi-source attribute data includes origin, time, processing plant and batch information, and the conventional component content data includes content of total energy, crude fat, starch and neutral detergent fiber; A construction unit is configured to perform multi-source variation degree tensor construction processing according to the multi-source attribute data and the conventional component content data, map the multi-source attribute data to independent modal dimensions of the tensor, and decouple and compress the conventional component content data as feature components within the modal to obtain a multi-source variation degree tensor; A decomposition unit is configured to perform variation degree spectrum decomposition processing according to the multi-source variation degree tensor, and perform local rank spectrum decomposition along the origin dimension, time dimension and component dimension of the tensor to obtain a variation degree spectrum vector set; The computing unit is configured to perform multi-scale extreme value and sparsity screening processing according to the set of variation spectrum vectors, calculate extreme points, spectral energy coverage and sparsity centrality of the spectrum vectors in a multi-scale variation space, and identify candidate modeling samples based on the calculation results to obtain a candidate modeling sample set; The screening unit is configured to perform sample screening processing according to the candidate modeling sample set, perform leave-one-out substitution and sensitivity curve analysis in the multi-source variation spectrum space, evaluate the contribution and sensitivity of each candidate sample to the standard ileal amino acid digestibility prediction equation of laying hens, and screen modeling samples based on the evaluation results to obtain all modeling samples in the laying hen feed raw material samples that meet the requirements.

7. The multi-source variability based layer feedstock sample screening system according to claim 6, wherein, The construction unit comprises: The first construction sub-unit is configured to perform multi-source attribute encoding processing according to the multi-source attribute data, convert the multi-source attribute data into a continuous numerical feature vector by embedding and mapping based on the geographical information entropy of the laying hen feed raw material origin and the periodicity of the processing time, and obtain an encoded multi-source attribute feature set; The second construction sub-unit is configured to perform multi-modal tensor initialization processing on the encoded multi-source attribute feature set and the conventional component content data, use each multi-source attribute feature vector as an independent modal dimension of a tensor, and perform dimension alignment and splicing on the conventional component content data as a feature component within the modal to obtain an initial multi-modal tensor; The third construction sub-unit is configured to perform nonlinear decoupling and compression processing on the initial multi-modal tensor, decouple the nonlinear coupling between the modes by using an adaptive kernel entropy mapping algorithm, and perform dimension reduction compression by using a Tucker decomposition method to obtain a multi-source variation tensor representing source heterogeneity and component difference.

8. The multi-source variability based layer feedstock sample screening system according to claim 6, wherein, The decomposition unit comprises: The first decomposition sub-unit is configured to perform single-modal local rank analysis processing according to the multi-source variation tensor, calculate the local rank distribution of each mode along the origin dimension, the time dimension and the component dimension, and obtain the variation components of each mode based on a preset dynamic rank variation rate threshold; The second decomposition sub-unit is configured to perform cross-modal coupling analysis processing according to the variation components of each mode, analyze the synergistic and antagonistic relationships between the variation components of different modes by introducing the laying hen digestion physiological parameters as a constraint condition, and obtain a mode coupling strength matrix; The third decomposition sub-unit is configured to perform spectrum vector fusion processing according to the mode coupling strength matrix, fuse the variation components and coupling relationships of each mode by using a nonlinear tensor contraction operation to generate a set of variation spectrum vectors representing the energy distribution of the sample in the multi-source variation space. 9.The multi-source variability based layer feedstock sample screening system according to claim 6, wherein, The computing unit comprises: The first computing sub-unit is configured to perform multi-scale extreme value detection processing according to the set of variation spectrum vectors, identify extreme points with significant variation characteristics at different scales in the spectrum space based on a preset adaptive threshold mechanism related to the standard ileal amino acid digestibility of laying hens, and obtain a multi-scale extreme sample set; The second computing sub-unit is configured to perform spectrum space coverage analysis processing according to the multi-scale extreme sample set, calculate the coverage contribution of each extreme sample in the spectrum energy space, and obtain a coverage optimized sample subset; The third computing subunit is configured to perform sparsity centrality evaluation processing according to the coverage optimization sample subset, to identify key sample points that can minimize the number of samples while maintaining the integrity of the spectrum characteristics by constructing a sparsity distribution map of the variability spectrum space, and to construct a candidate modeling sample set by taking all the key sample points as candidate modeling samples.

10. The multi-source variability based layer feedstock sample screening system according to claim 6, wherein, The screening unit comprises: The first screening subunit is configured to perform leave-one-out contribution analysis processing according to the candidate modeling sample set, to calculate the influence degree of each candidate sample on a standard ileal amino acid digestibility prediction equation of the laying hen by sequentially removing a single candidate sample in the variability spectrum space, and to obtain a prediction stability perturbation matrix of each sample. The second screening subunit is configured to perform sensitivity optimization processing based on the physiological characteristics of digestion according to the prediction stability perturbation matrix, to obtain a sensitivity optimization sample set by introducing the biological coefficient of variation of the standard ileal amino acid digestibility of the laying hen as a weight factor to weight and reconstruct the perturbation matrix. The third screening subunit is configured to perform multi-objective screening processing according to the sensitivity optimization sample set, to obtain all modeling samples that meet the requirements by establishing a Pareto optimal frontier between variability coverage and prediction sensitivity, and to minimize the prediction error risk while ensuring the integrity of the spectrum space.

Images