A method for screening corn varieties based on a corn quality evaluation system model

Abstract

The application discloses a corn variety screening method based on a corn quality evaluation system model, and relates to the field of corn variety screening and evaluation. The application collects the bulk density and key nutritional index data of corn samples, constructs a global coupling matrix and inversion characteristic parameters, realizes the driving mapping of the bulk density on each nutritional index, and establishes the nonlinear coupling mapping between the bulk density and crude fat, crude protein, lysine and crude starch by using a single index inversion neural network. Meanwhile, a multi-scale inversion offset modulation factor and an inversion mapping sensitivity function are embedded to realize the comprehensive mapping of input changes and output indexes. The network is trained by using an inversion composite loss function, and the comprehensive score is generated by weighting the predicted indexes and the bulk density in combination with a variety screening and evaluation function, so that the corn varieties are scientifically sorted and screened, the correlation of the nutritional indexes is reflected, and the comprehensive evaluation of the corn quality is realized.

Description

Technical Field

[0001] This invention relates to the field of maize variety screening and evaluation, specifically a maize variety screening method based on a maize quality evaluation system model. Background Technology

[0002] In the fields of agricultural breeding and grain quality evaluation, the selection of maize varieties typically relies on the measurement and evaluation of single or a small number of nutritional indicators, such as crude protein, crude fat, lysine, and crude starch content, while also combining physical traits such as test weight for variety grading. Existing technologies mainly include experimentally determined indicator analysis, statistical evaluation methods, and some multi-indicator weighted scoring systems for comprehensive evaluation of maize varieties. These methods are usually based on traditional measurement and analysis techniques, processing nutritional indicators through linear or simple nonlinear statistical models, and considering, to some extent, the correlation and influence weights between various indicators.

[0003] However, existing technologies have certain limitations in practical applications: First, evaluation using single or a small number of indicators can easily overlook the complex coupling relationships between different nutrients, making it difficult to comprehensively reflect the overall characteristics of maize quality; second, traditional statistical models and simple weighting methods cannot accurately characterize the nonlinear mapping relationship between bulk density and various nutrient indicators, resulting in insufficient prediction accuracy; in addition, existing evaluation systems lack effective quantification of the response intensity and local perturbation characteristics of changes in input variables to output indicators, limiting the scientific nature and operability of variety selection and failing to meet the needs of modern agriculture for refined selection and optimization of maize quality. Summary of the Invention

[0004] This invention proposes a corn variety screening method based on a corn quality evaluation system model, aiming to solve the problems of single-index evaluation, insufficient quantification of nonlinear relationships, and difficulty in reflecting the coupling relationship between various nutritional indicators and test weight in the existing corn variety screening process.

[0005] One method for screening maize varieties based on a maize quality evaluation system model includes the following steps:

[0006] S1. Collect sample data of maize varieties, including test weight, crude fat, crude protein, lysine, and crude starch. Through correlation analysis, quantify and model the correlation and coupling relationship between test weight and crude fat, crude protein, lysine, and crude starch to obtain a global coupling matrix. Use test weight as a driving variable to extract inversion feature parameters. Specifically, the above scheme collects sample data of different maize varieties, including test weight, crude fat, crude protein, lysine, and crude starch. Establish a sample matrix for each indicator. Calculate the linear correlation coefficient and nonlinear mutual information between test weight and each individual indicator through correlation analysis. Merge the two types of correlation matrices to generate a global coupling matrix. This matrix characterizes the overall coupling relationship between test weight and output indicators. Then, using test weight as a driving variable, map it to the output indicators. Form a high-dimensional feature matrix through multinomial basis, radial basis function, and frequency domain basis expansion. Use the preliminary global coupling matrix to evaluate the contribution of each output indicator, extract the most effective inversion feature parameters, and generate an inversion feature parameter matrix to describe the quantitative mapping relationship from test weight to each output indicator.

[0007] S2. Based on the inversion feature parameters, a single-index inversion neural network is constructed, using bulk density as the input variable and the predicted values ​​of crude fat, crude protein, lysine, and crude starch as the output. This network models the nonlinear coupling mapping path between bulk density and crude fat, crude protein, lysine, and crude starch. Specifically, based on the inversion feature parameters obtained in S1, the above scheme sets the input-output framework of the single-index inversion neural network. The input layer includes one node to receive the bulk density variable, and the output layer corresponds to four single indices, each with an independent output channel. An independent single-index inversion neural network is established for each output index, mapping it to a single output node through hidden layers to achieve a single-input, single-output nonlinear mapping path from bulk density to the target index. The number of hidden layers and nodes is determined according to the quantization model. The bulk density input is sequentially passed to the hidden layer nodes, and a nonlinear coupling mapping is constructed layer by layer through weighted connections and nonlinear activation functions.

[0008] S3. Define an inversion mapping sensitivity function and embed the global coupling matrix and multi-scale inversion offset modulation factor into the hidden layer structure of the network to perform sensitivity modulation on the response strength and local perturbation characteristics of each output index to changes in tolerance. Specifically, the above scheme defines an inversion mapping sensitivity function to characterize the response strength and local perturbation characteristics of each output index prediction value to changes in input tolerance. Then, the global coupling matrix is ​​embedded into the hidden layer structure of each single index network, so that the network reflects the overall coupling relationship between tolerance and output index during the hidden layer transmission process. Subsequently, a multi-scale inversion offset modulation factor is embedded in the hidden layer structure to adjust the output response offset caused by changes in tolerance at different scales, forming a sensitivity modulation mechanism, so that the single index network maintains a stable nonlinear mapping to input changes during training.

[0009] S4. By defining an inversion composite loss function, the constructed single-index inversion neural network containing the inversion mapping sensitivity function is trained. Specifically, the above scheme combines the inversion mapping sensitivity function defined in S3 to construct an inversion composite loss function, which includes a prediction error term, a sensitivity constraint term, and an optional global coupling constraint term. The network parameters are optimized by minimizing the loss function. The training samples are input into the single-index inversion neural network containing the sensitivity function, and the network weights and hidden layer parameters are iteratively updated through the backpropagation algorithm until the network converges to a stable nonlinear mapping relationship. Training stops when the loss function converges to a preset threshold or reaches the maximum number of iterations, thus completing the single-index network training.

[0010] S5. Input new maize variety sample data into the trained single-index inversion neural network, outputting predicted values ​​for crude fat, crude protein, lysine, and crude starch. Based on the output predicted values ​​and combined with the set variety screening evaluation function, the predicted nutritional index values ​​and bulk density data are weighted and combined to screen and score each maize variety sample. Specifically, the above scheme inputs new maize variety samples into the trained single-index inversion neural network to obtain the predicted vectors of the four output indicators. The relative deviation of each indicator is calculated according to the set target benchmark, and a deviation vector is formed. The coupling gain is calculated using the global coupling matrix to quantify the synergistic effect between the output indicators. For indicators below the set threshold, a penalty term is calculated for each item and accumulated. The bulk density component, the weighted combination of the deviation vector, the coupling gain, and the penalty term are combined according to the pre-set linear combination rules to form the final comprehensive score value. This score value is then used to sort and screen each maize variety sample, thereby achieving scientific screening and scoring of maize varieties.

[0011] The beneficial effects of the invention are:

[0012] (1) This invention collects data on bulk density and key nutritional indicators of corn samples, and constructs a global coupling matrix and inversion feature parameters to realize the driving mapping of bulk density to each nutritional indicator; it uses a single-index inversion neural network to establish a nonlinear coupling mapping path between bulk density and crude fat, crude protein, lysine and crude starch, and embeds a multi-scale inversion offset modulation factor and an inversion mapping sensitivity function to enable the network to take into account the response characteristics of input changes and the coupling relationship of each output indicator during the training process.

[0013] (2) This invention trains the network by defining an inversion composite loss function to achieve accurate prediction of each single indicator; finally, by combining the variety screening evaluation function, the predicted nutritional index values ​​and bulk density data are weighted and combined to generate a comprehensive score, and the samples of each maize variety are scientifically sorted and screened, thereby improving the quantitative accuracy and operability of maize variety screening, while reflecting the intrinsic correlation between nutritional indicators, and realizing a comprehensive evaluation and optimization selection of maize quality. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of a method for screening maize varieties based on a maize quality evaluation system model, according to an embodiment of the present invention. Detailed Implementation

[0015] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention; that is, the described embodiments are only a part of the embodiments of the invention, and not all of them. The components of the embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0017] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention. It should be noted that relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0018] Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or machine that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or machine. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or machine that includes said element.

[0019] The features and performance of the present invention will be further described in detail below with reference to embodiments.

[0020] Example 1

[0021] Among them, such as Figure 1 A method for screening maize varieties based on a maize quality evaluation system model includes the following steps:

[0022] S1. Collect sample data of maize varieties, including bulk density, crude fat, crude protein, lysine and crude starch. Through correlation analysis, quantitatively model the correlation and coupling relationship between bulk density and crude fat, crude protein, lysine and crude starch to obtain a global coupling matrix. Use bulk density as a driving variable to extract inversion feature parameters.

[0023] S2. Based on the inversion feature parameters, construct a single-index inversion neural network with bulk density as the input variable and the predicted values ​​of crude fat, crude protein, lysine, and crude starch as the output, and model the nonlinear coupling mapping path between bulk density and crude fat, crude protein, lysine, and crude starch.

[0024] S3. Define the inversion mapping sensitivity function, embed the global coupling matrix and multi-scale inversion offset modulation factor into the hidden layer structure of the network, and perform sensitivity modulation on the response strength of each output index and the local perturbation characteristics of the weight change;

[0025] S4. Train the constructed single-index inversion neural network containing the inversion mapping sensitivity function by defining the inversion composite loss function;

[0026] S5. Input new maize variety sample data into the trained single-index inversion neural network, output the predicted values ​​of crude fat, crude protein, lysine and crude starch. Based on the output predicted values, combined with the set variety screening evaluation function, the predicted nutritional index values ​​and test weight data are weighted and combined to screen and score each maize variety sample.

[0027] Further, for the above steps, sample data of maize varieties are collected. This sample data includes five indicators: test weight, crude fat, crude protein, lysine, and crude starch. After data collection, correlation analysis is performed between test weight and crude fat, crude protein, lysine, and crude starch. By calculating correlation coefficients and comparing the interaction strength between each indicator, the correlation and coupling relationship between test weight and the other indicators is quantitatively modeled, thereby constructing a global coupling matrix. This matrix describes the overall interaction relationship between multiple indicators. Based on this, test weight is used as a driving variable to extract inversion feature parameters, providing a basis for subsequent neural network modeling. After obtaining the inversion feature parameters, a single-indicator inversion neural network is constructed based on the modeling results. This network uses test weight as the input variable and the predicted values ​​of crude fat, crude protein, lysine, and crude starch as the output. The network's hidden layer maps the nonlinear coupling path between test weight and each nutrient indicator, enabling the model to achieve inversion prediction of multiple nutrient indicators based on test weight. After the network is constructed, an inversion mapping sensitivity function is further defined, and the global coupling matrix obtained in S1 and the multi-scale inversion offset modulation factor are embedded into the neural network. The hidden layer structure of the network is used to achieve sensitivity modulation of the response intensity and local perturbation characteristics of each output index to changes in bulk density, thus reflecting the dual role of global and local at the network structure level. Subsequently, based on this structure, an inversion composite loss function is defined. By including prediction error terms, sensitivity constraint terms, and necessary regularization terms in the loss, the network can ensure both the fitting accuracy to real data and the reasonable sensitivity performance under changes in bulk density during training, thereby obtaining a stable training model through multiple rounds of iterative updates. Finally, after the model training is completed, new maize variety sample data is input into the trained single-index inversion neural network to obtain predicted values ​​of crude fat, crude protein, lysine, and crude starch. These predicted values ​​are then combined with the pre-set variety screening evaluation function and weighted together with the bulk density data. During the weighting process, the nutritional weight vector and bulk density weight allocation ratio are used, while setting a minimum threshold constraint and a penalty mechanism. Indicators below the threshold are penalized or increased in penalty, ultimately forming a comprehensive screening score for each maize variety sample. Based on the score results, the varieties are ranked according to their quality, thus achieving the screening of high-quality varieties.

[0028] Furthermore, step S1 specifically includes the following sub-steps:

[0029] S101. Collect sample data of different corn varieties, including sample data of bulk density, crude fat, crude protein, lysine and crude starch, and normalize the sample data.

[0030] S102. Calculate the Pearson correlation coefficient between bulk density and each nutritional indicator to obtain the linear correlation matrix, and calculate the mutual information between bulk density and each nutritional indicator to obtain the nonlinear correlation matrix. Merge the linear correlation matrix and the nonlinear correlation matrix into a global coupling matrix.

[0031] S103. Using the initial global coupling matrix as a constraint and mapping the tolerance to the inversion features of the output index, construct a tolerance-driven quantization model, and evaluate the most effective feature set for predicting the output index through contribution assessment, and output the final inversion feature parameters.

[0032] Specifically, there are two types of relationships between bulk density and nutritional indicators: one is a linear relationship (bulk density and crude protein content may be positively correlated with variety), and the other is a non-linear relationship (bulk density and lysine exhibit a threshold effect or saturation effect). In this embodiment, the linear relationship is measured by the Pearson correlation coefficient, capturing the monotonic linear dependence; the non-linear relationship is measured by mutual information, measuring the degree to which the uncertainty of one variable on another is reduced, i.e., capturing the non-linear coupling strength. Both have limitations; using only the Pearson correlation coefficient will miss non-linear relationships, and using only mutual information will lose directionality and linear strength. Therefore, by fusing the two, a global coupling matrix containing both linear and non-linear constraints is constructed. Each element of this matrix not only reflects the dependence strength of bulk density on a single indicator but also reflects the overall coupling pattern. For example, the global coupling matrix is ​​obtained by weighted fusion of a linear correlation matrix and a non-linear correlation matrix. In the weighted fusion, two fusion weight coefficients, corresponding to the linear and non-linear correlation matrices respectively, are used to balance the contributions of linear and non-linear information, and the sum of the two fusion weight coefficients is 1. Further, an example of a global coupling matrix is ​​given:

[0033] ;

[0034] Among them, the They represent Coupling weights of crude fat, crude protein, lysine, and crude starch.

[0035] Specifically, after obtaining the global coupling matrix, it is introduced into the quantization model as a constraint, and the weight is used as the input driving variable to invert the predictive features of each nutritional indicator. This process is essentially a feature selection and mapping process: under the guidance of the global coupling matrix, the coupling contribution of different indicators to the weight is compared, and the feature set with high contribution and low redundancy in predicting the target is selected; the contribution is calculated by measuring the feature importance (weight based on entropy, coefficient size based on regression, or marginal gain based on mutual information); the resulting inverted feature parameters have a subspace mapping of the explanatory power of the weight on the output indicators, and also serve as a constraint prior for the network input in subsequent neural network training.

[0036] Furthermore, in step S103, constructing the quantization model driven by bulk density includes the following sub-steps:

[0037] S1031. The standardized weight index is expanded using polynomial basis, radial basis function and frequency domain basis to obtain a high-dimensional feature matrix. The high-dimensional feature matrix is ​​the mathematical mapping quantization model of the weight to output index inversion feature.

[0038] S1032. The preliminary global coupling matrix is ​​fused with the high-dimensional feature matrix to obtain the inverted feature parameter matrix.

[0039] Specifically, the relationship between bulk density and crude fat, crude protein, lysine, and crude starch is not a single pattern, but rather a multi-level, multi-form nonlinear coupling. If only raw bulk density data is used for modeling, the model can only learn linear or approximately linear relationships, making it difficult to cover the complex patterns of maize variety quality. Therefore, it is necessary to extend bulk density to a high-dimensional space to mathematically characterize different types of nonlinear mapping features. Furthermore, to ensure dimensionality consistency, a uniform dimensionality threshold D should be set, meaning all expansion methods generate D-dimensional features. For example, the polynomial basis retains the first D terms, the RBF selects D centers, and the Fourier basis takes D / 2 pairs of sine-cosine terms. By expanding in the above manner, the same dimensionality is output.

[0040] in:

[0041] Polynomial basis expansion: There is a power-law dependence between bulk density and nutritional indicators. For example, after bulk density increases to a certain level, the change in protein or starch content follows a quadratic or cubic trend. By using polynomial basis expansion, the model captures the curvilinear global trend, thus avoiding the limitations of simple linear fitting.

[0042] Radial basis function expansion: The distribution of test weight values ​​for different maize varieties is concentrated in several typical intervals. In some intervals, test weight has a particularly significant impact on nutritional indicators, while in other intervals it has almost no effect. Local sensitivity is difficult to express using linear or polynomial methods, but radial basis functions (Gaussian kernels) can weight and amplify local regions, enabling the model to capture the driving effect of test weight on indicators in a specific interval.

[0043] Frequency domain basis expansion: Maize quality data exhibits periodic fluctuations or oscillations within a certain range. For example, changes in bulk density show a periodic response to lysine content (with changes in breeding environment or generational variation). By using Fourier basis or sine / cosine expansion, bulk density can be mapped to the frequency domain, revealing its potential periodic coupling relationship with nutritional indicators.

[0044] By combining the above technical solutions, the mapping relationship between bulk density and nutritional indicators is comprehensively quantified from three dimensions: global, local, and periodic, thereby forming a high-dimensional feature matrix. , wherein Represents a high-dimensional feature matrix, the These represent the polynomial basis vector, radial basis vector, and frequency domain basis vector, respectively, after the bulk density index is expanded using polynomial basis, radial basis function, and frequency domain basis.

[0045] Furthermore, for the center point in the radial basis function, a uniform partitioning method based on statistical distribution can be used for selection. This involves statistically analyzing the samples with the highest and lowest test weights in the corn sample, dividing the interval between the lowest and highest test weights by an average, and using the interval division point as the center point. This center point is the reference test weight value used for radial basis function expansion. Further, for frequency domain basis expansion, a pair of sine / cosine terms can be used for expansion. The sine part characterizes the periodic increasing / decreasing trend between test weight and nutritional indicators, while the cosine part characterizes supplementary information about the periodic phase. The frequency can be defined according to actual needs. Finally, for polynomial bases, the order of the polynomial expansion is generally defined according to requirements.

[0046] Furthermore, while the high-dimensional feature matrix contains rich potential mapping patterns, it lacks filtering and contains a large number of redundant features. The global coupling matrix provides correlation constraint information between weight and nutritional indicators. Therefore, fusing it with the high-dimensional feature matrix can introduce correlation priors into the feature space. Specifically, the global coupling matrix and the high-dimensional feature matrix are fused through matrix operations. First, the high-dimensional feature matrix is ​​linearly mapped so that the dimension of the feature vector of each sample is consistent with the dimension of the output indicator affected by the global coupling matrix, thus achieving dimensional unification. The weights in the global coupling matrix represent the coupling strength between each output indicator under weight-driven conditions, used to measure the statistical correlation between the output indicators. During the fusion process, the weights in the coupling matrix are applied to each feature of the high-dimensional feature matrix. Through element-wise weighting or linear combination, features with weak correlation to the target indicator are suppressed, while features with strong correlation are strengthened, resulting in a feature matrix corrected by coupling constraints. The resulting inversion feature parameter matrix, after the above operations, has each row corresponding to the inversion feature vector of a corn sample, and each column corresponding to a certain dimension value of the fused high-dimensional mapped feature after coupling weight correction. This matrix retains the expressive power of high-dimensional mapping while incorporating the effect of statistical prior constraints, avoiding excessive expansion of the feature space. It provides a compact and effective input representation for single-index inversion neural networks, enabling the network to accurately invert nutritional indicators under the weight-driven approach.

[0047] Furthermore, step S2 specifically includes the following sub-steps:

[0048] S201. Based on the extracted inversion feature parameters, set the input and output framework of the single-index inversion neural network, wherein the input layer includes an input node for receiving the bulky variable, and the output layer corresponds to four single indices to be predicted, including crude fat, crude protein, lysine and crude starch, and each index corresponds to an independent output channel.

[0049] S202. For each output index, establish an independent single-index inversion neural network with tolerance as input, and map it to a single output node through the hidden layer to construct a single-input single-output inversion mapping path between tolerance and the corresponding target output index;

[0050] S203. For each single-index inversion neural network, set the number of hidden layers and node settings according to the quantitative modeling relationship, pass the weight input to the hidden layer nodes in sequence, and establish a mapping layer by layer through weight connection and nonlinear function operation to build a nonlinear coupling path between the weight and the target output index.

[0051] Specifically, in the inversion modeling, bulk density is set as the sole driving factor. Therefore, the input layer contains only one node, but this node is not a simple numerical value. Instead, it represents the inversion feature parameter matrix processed by S1 and S103, implicitly containing global coupling information and nonlinear expansion between bulk density and other nutritional indicators. The output layer is designed with four independent channels, each corresponding to a single indicator: crude fat (F), crude protein (P), lysine (L), and crude starch (S). Each indicator is modeled separately during prediction. This input-output framework serves as the structural basis for building the subsequent single-indicator inversion network, while ensuring that the outputs do not interfere with each other, avoiding the cross-error effects when modeling multiple indicators. Furthermore, the core of the indicator inversion neural network is single input → single output, meaning that an independent neural network is built for each nutritional indicator. Specifically, since the dependencies between different indicators and bulk density vary greatly—for example, bulk density is mainly linearly related to crude protein but nonlinearly related to lysine—different coupling paths are captured through independent modeling. Each network input layer receives bulk density features after high-dimensional expansion, while the hidden layer undertakes the nonlinear mapping process from input to output. The output layer contains only one node, corresponding to a specific target nutrient indicator. This design allows each network to optimize its own inversion relationship between bulk density and indicator independently, without interference from other outputs.

[0052] Furthermore, the hidden layer design is based on the inverted feature parameters and global coupling matrix from the previous step, aiming to provide suitable nonlinear fitting capabilities for each single-index network. Specifically, the bulk input is connected to the hidden layer through input weights. Each hidden layer node weights and corrects the bias of the input, and then completes the mapping through a nonlinear activation function (sigmoid, tanh, or ReLU). The stacking of multiple hidden layers can progressively extract the nonlinear response pattern of bulk to the index, while the number of nodes determines the complexity and fitting capability of the model. Through layer-by-layer mapping, the final output layer nodes can accurately express the nonlinear dependence path of bulk on the target nutrient index.

[0053] Furthermore, step S3 specifically includes the following sub-steps:

[0054] S301. Define an inversion mapping sensitivity function to describe the degree of response and disturbance characteristics of changes in bulk density to the prediction results of various output indicators when it is used as an input variable.

[0055] S302. Introduce the global coupling matrix into the hidden layer structure of the single-index inversion neural network so that the transmission process of the hidden layer reflects the global coupling relationship between the tolerance and each output index.

[0056] S303. Embed a multi-scale inversion migration modulation factor in the hidden layer structure to adjust the response shift caused by changes in bulk density at different scales.

[0057] Specifically, the inversion mapping sensitivity function is essentially a derivative or gain measure used to quantify the response amplitude and direction of the predicted values ​​of each output index when the input bulk weight changes slightly, while also considering the impact of local perturbations on the output. This function is constructed by taking the partial derivative of the single-index network output with respect to the bulk weight input, or by obtaining the response gradient through local perturbation experiments, and combining this with the contribution weights of each index to form a quantitative expression of the sensitivity measure. Through the function definition, the driving strength and potential nonlinear perturbation effect of the bulk weight on nutritional indicators in different numerical ranges can be evaluated. For example, the construction process of the inversion mapping sensitivity function is given. Its core purpose is to quantify the response characteristics and local perturbation effects of the input bulk weight on each output index. Since slight changes in the input will cause varying degrees of fluctuation in the output, partial derivatives are used to describe the local response strength. Furthermore, the response amplitude of the bulk weight to the output varies greatly in different value ranges; therefore, a multi-scale inversion offset modulation factor is added, and the sensitivity is locally enhanced or attenuated using a Gaussian modulation formula. This indicates the degree of deviation between the input weight and the reference value, through... Control the width of the response curve, by By controlling the modulation amplitude, nonlinear sensitivity adjustment of different indicators is achieved within different tolerance ranges, ensuring the inversion mapping is stable and controllable across multiple scales. This allows for dynamic adjustment of sensitivity weights, enabling the network to maintain a reasonable mapping under both small and large changes.

[0058] ;

[0059] ;

[0060] Among them, the This represents the inversion mapping sensitivity function value of the k-th index under the input of the n-th sample weight. This represents the predicted value of the k-th single indicator for the n-th sample. Indicates the index of indicators, the This represents the weight input value of the nth sample. Represents the multi-scale inversion offset modulation factor, the This represents the offset magnitude coefficient, used to control the maximum enhancement of the sensitivity response of the k-th index. This represents a scale control parameter used to adjust the width range of the response of the k-th index, so that the sensitivity exhibits different response curves at different bulk density scales. This represents the reference value for the bulk density scale, used to determine the center position of the input bulk density and control the position of the response peak.

[0061] Furthermore, the global coupling matrix originates from the quantization modeling in S1 and is used to reflect the linear and nonlinear dependence strength between bulk weight and various nutritional indicators. Introducing this matrix into the hidden layer of the neural network is equivalent to introducing additional weighted constraints during node activation propagation, so that when updating weights and calculating hidden layer outputs, the network depends not only on the single-indicator input features but also on the overall coupling relationship. Specifically, the original weight matrix of the hidden layer is matched with the global coupling matrix. The connection weights between the output of each hidden layer node and each output indicator are adjusted according to the corresponding elements in the coupling matrix, or the hidden layer outputs are adjusted through matrix multiplication. This establishes a global collaborative mapping mechanism within the network, explicitly reflecting the overall dependence between bulk weight and each output indicator within the network, enabling the single-indicator network to possess both local nonlinearity and global coupling constraints during inversion.

[0062] Furthermore, changes in tolerance at different scales produce different response amplitudes to the output index. For example, small changes may cause local perturbations, while large changes may cause overall shifts. The multi-scale inversion shift modulation factor introduces scale-dependent weighted adjustment in the hidden layer. By locally amplifying or reducing the input features, the network can adaptively adjust the output response under different tolerance change amplitudes, achieving synchronous control of local perturbations and global trends. This factor can be implemented through higher-order expansion of tolerance or local gradient weighting, and works together in the forward propagation and backward update of the network.

[0063] Furthermore, step S4 specifically includes the following sub-steps:

[0064] S401. Define the inversion composite loss function by combining the constraint terms of the inversion mapping sensitivity function;

[0065] S402. Input the training sample data into a single-index inversion neural network containing a sensitivity function, perform iterative calculations based on the inversion composite loss function, update the network weights and hidden layer parameters through the backpropagation algorithm, and converge to a stable nonlinear mapping relationship;

[0066] S403. Stop training when the loss function converges to the preset threshold or reaches the maximum set number of iterations.

[0067] Specifically, the inversion composite loss function not only considers the error between the predicted value and the target index, but also introduces a sensitivity function constraint to ensure that the output response conforms to the expected nonlinear mapping pattern under different variations in the tolerance input. The loss function consists of two parts: one is the traditional mean squared error or absolute error, used to measure the deviation between the network output and the target value; the other is a sensitivity constraint term, which penalizes situations where the network output is too insensitive or overly sensitive to changes in tolerance, thereby controlling the local and global response characteristics of the hidden layer mapping during training.

[0068] Furthermore, during forward propagation, the input tolerance samples are mapped through hidden layers and adjusted for sensitivity to output predicted values. The composite loss function is then used to calculate the error and sensitivity deviation. During backpropagation, the error is used to update the network weights and hidden layer parameters via gradient descent, while simultaneously updating the modulation parameters related to the sensitivity constraints. This allows the entire network to gradually learn the nonlinear coupling mapping relationship between tolerance and each output index. Throughout the iteration process, the network continuously adjusts its internal mapping weights to optimize the balance between prediction accuracy and sensitivity matching.

[0069] For example, the construction principle of the inversion composite loss function is given: First, it is necessary to measure the deviation between the network output and the target index, so a prediction error term is introduced. By minimizing this term, the network is guaranteed to fit the training samples and accurately predict each single metric value. This represents the predicted value of the k-th single indicator for the n-th sample. Let represent the k-th target benchmark value of the n-th sample; secondly, considering the response characteristics of input tolerance to output index, a sensitivity constraint term is defined for this purpose. By incorporating the inversion mapping sensitivity function into the loss function, the network can maintain reasonable nonlinear response behavior during training, thereby improving target sensitivity. Determined by quantitative modeling or historical data; weighting parameters are introduced to balance prediction accuracy and response constraints. The relative importance of the prediction error term and the sensitivity constraint term can be controlled, and a global coupling matrix can be added if necessary. With coupling weights ,pass The overall synergistic relationship between multiple index outputs is constrained, where y represents four single index sets, thereby ensuring that the network fits the single index data and reflects the global coupling characteristics between the indices. By minimizing the composite loss function containing these three parts, the network weights and hidden layer parameters are optimized, making the inversion mapping stable, interpretable, and consistent with the statistical and physical laws of the input and output.

[0070] Specifically, the inversion composite loss function is expressed as:

[0071] ;

[0072] Among them, the Denotes the inversion composite loss function, the The total number of training samples is represented by the following. Indicates the index of indicators, the These represent crude fat, crude protein, lysine, and crude starch, respectively. Indicates the sample index, the This represents the predicted value of the k-th single indicator for the n-th sample. This represents the k-th target benchmark value of the n-th sample. The sensitivity constraint weight coefficients are represented by the following: This represents the inversion mapping sensitivity function value of the k-th index under the input of the nth sample weight. This represents the target sensitivity value of the k-th indicator. This represents the weight input value of the nth sample. This represents the prediction vector for the nth sample. Represents the global coupling matrix, the This represents the coupling constraint weight coefficient.

[0073] Furthermore, in step S5, the process of setting the variety screening and evaluation function specifically includes the following sub-steps:

[0074] S501. For a new input corn sample, the bulk density and the prediction vector output by the four single-index inversion neural network;

[0075] S502. Based on the set target benchmark, calculate the relative deviation of each indicator and form a deviation vector;

[0076] S503. Calculate the coupling gain based on the global coupling matrix;

[0077] S504. Calculate and accumulate the penalty items generated when the values ​​are below the threshold for each indicator;

[0078] S505. The weighted combination of the tolerance component, the deviation vector, the coupling gain, and the penalty term are combined into the final comprehensive score according to the pre-set linear combination rules.

[0079] Furthermore, the calculation method for the final comprehensive score is specifically expressed as follows:

[0080] ;

[0081] Among them, the This represents the final overall score. Represents a constant bias term, the Indicates density The weighting in the overall score, the The relative value of the bulk density, the This represents the transpose of the weight vector of the four single-indicator sets y in the comprehensive score. Represents the deviation vector, the This represents the weighted sum of the relative deviations of the four nutritional indicators. The scaling factor representing the coupling gain is set through cross-validation. Indicates the coupling gain, the This represents the penalty that is applied when the accumulated value falls below the threshold.

[0082] Furthermore, the relative value of the bulk density is used to represent the relative extent to which the bulk density is higher or lower than the reference value, and its specific calculation process is as follows:

[0083] ;

[0084] Among them, the This represents the reference value for the weight scale, which can be taken as the target weight or the mean of the training samples.

[0085] Furthermore, the penalty term generated when the accumulated value falls below a threshold. The specific calculation process is as follows:

[0086] ;

[0087] Among them, the Indicates the index of indicators, the These represent crude fat, crude protein, lysine, and crude starch, respectively. This represents the penalty coefficient for indicator k, which takes a positive value and is used to control the severity of the penalty when the value is below a threshold. This represents the minimum acceptable threshold for indicator k; falling below this value will trigger a penalty. This represents the predicted value of a single sample on index k.

[0088] Furthermore, the parameter acquisition process for calculating the final overall score is given below:

[0089] Let the prediction vector be... Target reference vector Minimum threshold vector Nutrition weight vector reference value for bulk density scale Weight Coupling coefficient Penalty coefficient vector Global coupling mapping weight matrix constant bias Among them, the aforementioned These represent the predicted contents of crude fat, crude protein, lysine, and crude starch, respectively; These represent the minimum threshold values ​​for crude fat, crude protein, lysine, and crude starch, respectively, and are set according to requirements; These represent the target values ​​for crude fat, crude protein, lysine, and crude starch, respectively, which are set based on national / industry standards, superior variety databases, or breeding objectives; These represent the weights of crude fat, crude protein, lysine, and crude starch, respectively, set according to nutritional requirements and breeding objectives. These represent the penalty coefficients for crude fat, crude protein, lysine, and crude starch, respectively. They are set according to the importance of the indicators, with key nutrients (such as lysine and protein) generally assigned larger values.

[0090] The relative deviation vector of a single index is calculated using the principal element method:

[0091] ;

[0092] Calculate the coupling gain term :

[0093] .

[0094] Example 2

[0095] Furthermore, as a preferred embodiment of the above examples, an application scenario for a maize variety screening method based on a maize quality evaluation system model is presented. Specifically, at an agricultural academy in a certain province, a research team conducted quality optimization on 50 new maize lines, aiming to screen out superior varieties with crude protein ≥ 9%, lysine ≥ 0.28%, while ensuring that crude starch and crude fat are within reasonable ranges. Because traditional full-index chemical analysis is too costly, the team adopted a single-index inversion neural network method driven by bulk density for prediction and screening.

[0096] Researchers first measured the bulk density of 50 maize variety samples, and simultaneously measured the crude fat, crude protein, lysine, and crude starch content of 15 representative samples for model construction. Correlation analysis revealed a significant positive correlation between bulk density and crude protein and lysine, a certain negative correlation with crude starch, and a moderate correlation with crude fat. Based on this, the research team quantitatively modeled the coupling relationships, forming a global coupling matrix, and extracted inversion feature parameters using bulk density as a driving variable.

[0097] Based on the inversion feature parameters, a single-index inversion neural network was constructed, using bulk density as the input variable and outputting predicted values ​​for crude fat, crude protein, lysine, and crude starch. A global coupling matrix and a multi-scale inversion modulation factor were embedded in the hidden layer of the network, enabling the model to consider the differentiated responses of bulk density changes to each nutrient index during the prediction process. Subsequently, the team designed an inversion composite loss function, which includes not only a prediction error term but also a threshold violation penalty term to ensure that the model output meets the minimum nutrient index standards required for maize variety selection.

[0098] After completing network training, the research team input the test weight data of 50 new varieties into the model, obtaining prediction results for crude fat, crude protein, lysine, and crude starch. Using crude protein ≥ 9% and lysine ≥ 0.28% as minimum threshold conditions, and combining a variety screening evaluation function weighted by test weight and nutritional indicators, all varieties were comprehensively scored. The results showed that varieties A12 and A25 had predicted crude protein exceeding 9.5%, lysine content above 0.30%, and high test weights, achieving comprehensive scores of 92.3 and 94.1 respectively, significantly higher than other samples, and were therefore selected as priority candidate varieties for promotion.

[0099] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for screening maize varieties based on a maize quality evaluation system model, characterized in that, Includes the following steps: S1. Collect sample data of maize varieties, including bulk density, crude fat, crude protein, lysine and crude starch. Through correlation analysis, quantitatively model the correlation and coupling relationship between bulk density and crude fat, crude protein, lysine and crude starch to obtain a global coupling matrix. Use bulk density as a driving variable to extract inversion feature parameters. S2. Based on the inversion feature parameters, construct a single-index inversion neural network with bulk density as the input variable and the predicted values ​​of crude fat, crude protein, lysine, and crude starch as the output, and model the nonlinear coupling mapping path between bulk density and crude fat, crude protein, lysine, and crude starch. S3. Define the inversion mapping sensitivity function, embed the global coupling matrix and multi-scale inversion offset modulation factor into the hidden layer structure of the network, and perform sensitivity modulation on the response strength of each output index and the local perturbation characteristics of the weight change; S4. Train the constructed single-index inversion neural network containing the inversion mapping sensitivity function by defining the inversion composite loss function; S5. Input new maize variety sample data into the trained single-index inversion neural network, output the predicted values ​​of crude fat, crude protein, lysine and crude starch. Based on the output predicted values, combined with the set variety screening evaluation function, the predicted nutritional index values ​​and test weight data are weighted and combined to screen and score each maize variety sample. S2 specifically includes the following sub-steps: S201. Based on the extracted inversion feature parameters, set the input and output framework of the single-index inversion neural network, wherein the input layer includes an input node for receiving the bulky variable, and the output layer corresponds to four single indices to be predicted, including crude fat, crude protein, lysine and crude starch, and each index corresponds to an independent output channel. S202. For each output index, establish an independent single-index inversion neural network with tolerance as input, and map it to a single output node through the hidden layer to construct a single-input single-output inversion mapping path between tolerance and the corresponding target output index; S203. For each single-index inversion neural network, set the number of hidden layers and node settings according to the quantitative modeling relationship, pass the weight input to the hidden layer nodes in sequence, and establish a mapping layer by layer through weight connection and nonlinear function operation to build a nonlinear coupling path between the weight and the target output index. S3 specifically includes the following sub-steps: S301. Define an inversion mapping sensitivity function to describe the degree of response and disturbance characteristics of changes in bulk density to the prediction results of various output indicators when it is used as an input variable. S302. Introduce the global coupling matrix into the hidden layer structure of the single-index inversion neural network so that the transmission process of the hidden layer reflects the global coupling relationship between the tolerance and each output index; S303. Embed a multi-scale inversion migration modulation factor in the hidden layer structure to adjust the response shift caused by changes in bulk density at different scales.

2. The method for screening maize varieties based on a maize quality evaluation system model as described in claim 1, characterized in that, S1 specifically includes the following sub-steps: S101. Collect sample data of different corn varieties, including sample data of bulk density, crude fat, crude protein, lysine and crude starch, and normalize the sample data. S102. Calculate the Pearson correlation coefficient between bulk density and each nutritional indicator to obtain the linear correlation matrix, and calculate the mutual information between bulk density and each nutritional indicator to obtain the nonlinear correlation matrix. Merge the linear correlation matrix and the nonlinear correlation matrix into a global coupling matrix. S103. Using the initial global coupling matrix as a constraint and mapping the tolerance to the inversion features of the output index, construct a tolerance-driven quantization model, and evaluate the most effective feature set for predicting the output index through contribution assessment, and output the final inversion feature parameters.

3. The method for screening maize varieties based on a maize quality evaluation system model as described in claim 2, characterized in that, In step S103, constructing the quantization model driven by bulk density includes the following sub-steps: S1031. The standardized weight index is expanded using polynomial basis, radial basis function and frequency domain basis to obtain a high-dimensional feature matrix. The high-dimensional feature matrix is ​​the mathematical mapping quantization model of the weight to output index inversion feature. S1032. The preliminary global coupling matrix is ​​fused with the high-dimensional feature matrix to obtain the inverted feature parameter matrix.

4. The method for screening maize varieties based on a maize quality evaluation system model as described in claim 1, characterized in that, S4 specifically includes the following sub-steps: S401. Define the inversion composite loss function by combining the constraint terms of the inversion mapping sensitivity function; S402. Input the training sample data into a single-index inversion neural network containing a sensitivity function, perform iterative calculations based on the inversion composite loss function, update the network weights and hidden layer parameters through the backpropagation algorithm, and converge to a stable nonlinear mapping relationship; S403. Stop training when the loss function converges to the preset threshold or reaches the maximum set number of iterations.

5. The method for screening maize varieties based on a maize quality evaluation system model as described in claim 1, characterized in that, In step S5, the process of setting the variety screening and evaluation function specifically includes the following sub-steps: S501. For a new input corn sample, the bulk density and the prediction vector output by the four single-index inversion neural network; S502. Based on the set target benchmark, calculate the relative deviation of each indicator and form a deviation vector; S503. Calculate the coupling gain based on the global coupling matrix; S504. Calculate and accumulate the penalty items generated when the values ​​are below the threshold for each indicator; S505. The weighted combination of the tolerance component, the deviation vector, the coupling gain, and the penalty term are combined into the final comprehensive score according to the pre-set linear combination rules.

6. The method for screening maize varieties based on a maize quality evaluation system model as described in claim 5, characterized in that, The calculation method for the final comprehensive score is specifically expressed as follows: ； Among them, the This represents the final overall score. Represents a constant bias term, the Indicates density The weighting in the overall score, the The relative value of the bulk density, the This represents the transpose of the weight vector of the four single-indicator sets y in the comprehensive score. Represents the deviation vector, the This represents the weighted sum of the relative deviations of the four nutritional indicators. The scaling factor representing the coupling gain, the Indicates the coupling gain, the This represents the penalty that is applied when the accumulated value falls below the threshold.

7. The method for screening maize varieties based on a maize quality evaluation system model as described in claim 6, characterized in that, The relative value of the bulk density is used to represent the relative extent to which the bulk density is higher or lower than the reference value. The specific calculation process is as follows: ； Among them, the This represents the reference value for the weight scale, which can be taken as the target weight or the mean of the training samples.

8. The method for screening maize varieties based on a maize quality evaluation system model as described in claim 6, characterized in that, The penalty term generated when the accumulated value is below the threshold. The specific calculation process is as follows: ； Among them, the Indicates the index of indicators, the These represent crude fat, crude protein, lysine, and crude starch, respectively. This represents the penalty coefficient for indicator k, which takes a positive value and is used to control the severity of the penalty when the value is below a threshold. This represents the minimum acceptable threshold for indicator k; falling below this value will trigger a penalty. This represents the predicted value of a single sample on index k.

Images