A rapeseed seed quality evaluation system based on big data

Abstract

The present invention relates to the technical field of seed processing and analysis, and in particular to a rapeseed seed quality assessment system based on big data. The system comprises the following steps: constructing a data pool by aggregating multi-source heterogeneous rapeseed data to provide high-quality input for subsequent cross-domain feature association; constructing a cross-domain feature interaction map and using a graph neural network for learning to intuitively present cross-domain feature interactions and analyze gene-environment interactions; focusing on key features, integrating and optimizing information, generating high-quality feature vectors through a four-branch multi-head and cross-modal attention parsing map, achieving high-order feature nonlinear fusion, and significantly improving the model's ability to represent multi-source heterogeneous data; constructing an assessment model by decoding vectors and extracting high-yield variety data, which is conducive to the efficient screening of rapeseed varieties; and rapidly expanding the assessment capability to new varieties and promoting intelligent breeding by deploying a generative adversarial network and transferring learning adaptation models.

Description

Technical Field

[0001] The present invention relates to the technical field of seed processing and analysis, and in particular to a rapeseed seed quality evaluation system based on big data. Background Art

[0002] Rapeseed seed quality assessment currently faces several challenges, including the following: Rapeseed seed quality assessment involves heterogeneous data from multiple sources, including genes, environments, and phenotypes. These data come from a wide range of sources and come in various formats. There is a lack of effective integration methods, resulting in low data utilization efficiency and difficulty in fully realizing the value of the data; rapeseed seed quality is influenced by both genes and the environment, and the gene-environment interaction is complex. Traditional methods have difficulty in deeply exploring this complex interaction, resulting in a lack of in-depth understanding of the mechanisms underlying rapeseed seed quality formation; when assessing rapeseed seed quality, it is necessary to consider multidimensional features such as genes, environments, and phenotypes, but traditional methods have shortcomings in feature fusion, making it difficult to fully integrate information from features across different dimensions, leading to inaccurate assessment results; and as rapeseed varieties continue to update, traditional quality assessment systems have difficulty quickly adapting to the feature space of new varieties, resulting in inaccurate quality assessments of new varieties. Therefore, the present invention proposes a rapeseed seed quality assessment system based on big data. Summary of the Invention

[0003] The purpose of the present invention is to solve the problems in the background technology and to propose a rapeseed seed quality evaluation system based on big data.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions:

[0005] A rapeseed seed quality evaluation system based on big data, comprising:

[0006] Multi-source heterogeneous data acquisition and preprocessing module: acquires multi-source heterogeneous rapeseed data, processes and analyzes the acquired multi-source heterogeneous rapeseed data, and generates a multi-source heterogeneous rapeseed data pool;

[0007] Cross-domain feature interaction map construction module: Based on a multi-source heterogeneous rapeseed data pool, a cross-domain feature interaction map is constructed. Graph neural networks are used to perform deep feature learning on the constructed cross-domain feature interaction map, analyze gene-environment interactions, and output an updated cross-domain feature interaction map.

[0008] Multimodal feature fusion and optimization module: This module uses a preset four-branch multi-head attention architecture to analyze cross-domain feature interaction graphs, combines the cross-modal attention mechanism to generate dual-weighted and fused features, and then performs dual optimization after multi-head attention screening to generate an optimized comprehensive feature vector that integrates multi-dimensional information.

[0009] Feature evaluation and analysis module: Decodes and analyzes the optimized comprehensive feature vector to build a rapeseed seed quality evaluation model and extract high-yield rapeseed variety data;

[0010] Transfer learning adaptation deployment module: Use the extracted high-yield rapeseed variety data to deploy a generative adversarial network, adapt the quality assessment model to the new variety feature space through transfer learning, and complete the intelligent assessment of the quality of new rapeseed seeds.

[0011] Furthermore, the multi-source heterogeneous data acquisition and preprocessing module acquires multi-source heterogeneous rapeseed data, processes and analyzes the acquired multi-source heterogeneous rapeseed data, and generates a multi-source heterogeneous rapeseed data pool. The process includes:

[0012] Extracting rapeseed genome-wide SNP marker data from a gene database; collecting field phenotypic data through an IoT sensor network; connecting to a weather station to monitor data streams in real time and obtain field environmental parameters; and using a drone equipped with a multispectral camera to obtain dynamic image data of the rapeseed growth cycle.

[0013] All collected data sets were uniformly stored in an HBase distributed database and preprocessed. Specifically, data cleaning was performed to remove duplicate sites and low-quality sites in the rapeseed genome-wide SNP marker data, retaining the valid SNP site set to obtain a genotype data set; time series records of field environmental parameters were analyzed to establish an environmental parameter time series index; drone aerial image data was used, and computer vision technology was used to extract features from the dynamic image data. Image timestamps were annotated according to growth cycle segments, and vegetation indices were extracted as preliminary phenotypic features; at the same time, the collected field phenotypic data and the extracted preliminary phenotypic features were spatiotemporally aligned to eliminate sampling time deviations, thereby obtaining synergistic phenotypic features;

[0014] Based on the genotype dataset, a SNP-chromosome association matrix was constructed as the genetic dimension framework for the rapeseed raw data pool. Environmental parameter time series indexes and synergistic phenotypic characteristics were mapped to the genetic dimension framework by timestamp to establish cross-domain data associations. The HBase distributed database's column family storage model was used to divide data column families by genotype, environment, and phenotype, and composite encoding rules for row keys were set. Data integrity checks were performed, verifying cross-domain data consistency using hash checksums, completing the construction of the rapeseed raw data pool.

[0015] The rapeseed original data pool is standardized to obtain the rapeseed standardized data pool.

[0016] Furthermore, the cross-domain feature interaction map construction module constructs a cross-domain feature interaction map based on a multi-source heterogeneous rapeseed data pool. Graph neural networks are used to perform deep feature learning on the constructed cross-domain feature interaction map, analyze gene-environment interactions, and output an updated cross-domain feature interaction map. The process includes:

[0017] Separate genotype nodes, environmental nodes, and phenotype nodes from the rapeseed standardized data pool, perform one-hot encoding on each node to generate an initial feature vector as the initial feature representation of the graph;

[0018] After completing the initial feature representation of the nodes, the association analysis of gene nodes and environmental nodes is performed: the mutual information value of the gene nodes and environmental nodes is calculated, where the mutual information value is used to quantify the degree of association between the two node types; based on the calculated mutual information value, the mutual information threshold is set to filter out significantly associated node pairs; based on the screening process, a gene-environment binary edge set is constructed;

[0019] Based on the constructed gene-environment binary edge set and the initial feature representation of the node, a static graph is generated. The features of adjacent nodes are aggregated through the graph convolution layer to capture information about the adjacent nodes around the node. This captured information is then integrated into the feature representation of the target node, thereby generating a node embedding representation. The output static graph contains a node embedding matrix and an adjacency matrix. The node embedding matrix is used to record the feature representation of each node after feature aggregation, while the adjacency matrix is used to represent the connection relationships between nodes in the static graph.

[0020] The generated static graph is fed into a three-layer graph attention network, and a gated recurrent unit is used to capture the temporal changes in gene-environment edge weights.

[0021] Define a gene-environment interaction strength index, which is calculated as the edge weight multiplied by the node embedding cosine similarity, where the edge weight represents the strength of the association between the gene node and the environment node, and the node embedding cosine similarity is used to measure the similarity between the two nodes in the feature space. Based on the gene-environment interaction strength index, select the top N strongly associated edges with the index value, where N is a preset positive integer that limits the number of strongly associated edges selected from the numerous gene-environment edges sorted by the interaction strength index value.

[0022] A graph comparative learning framework is introduced to further analyze the top N strongly correlated edges in the index values: a perturbation operation is performed on the static graph, which includes random modifications to node features and edge weights; the rate of change of interaction intensity is calculated by comparing the differences in node representation of the graph before and after the perturbation; a change rate threshold M is set to filter key feature edges whose interaction intensity change rate is greater than or equal to the change rate threshold M; all the filtered key feature edges are integrated into the static graph to form the final dynamic graph, that is, the updated cross-domain feature interaction graph.

[0023] Furthermore, the generated static graph is fed into a three-layer graph attention network, and the gated recurrent unit is used to capture the temporal changes of gene-environment edge weights. The process includes:

[0024] Each layer of the graph attention network uses a multi-head attention mechanism, which learns the attention relationship between nodes from multiple different subspaces. In each layer, the attention coefficient between nodes is calculated, where the attention coefficient is used to reflect the importance and correlation strength between nodes. Based on the calculated attention coefficient, the node representation is updated so that the node integrates the information of its adjacent nodes.

[0025] Using the rate of change of environmental parameters as the gating signal, the gated recurrent unit dynamically modifies the weight value of the gene-environment edge according to the changes in environmental parameters.

[0026] Furthermore, the multimodal feature fusion and optimization module analyzes the cross-domain feature interaction map through a preset four-branch multi-head attention architecture, combines the cross-modal attention mechanism to generate dual weights and fuse features, and then performs dual optimization after multi-head attention screening to generate an optimized comprehensive feature vector that integrates multi-dimensional information. The process includes:

[0027] The constructed cross-domain feature interaction map is input into the preset four-branch multi-head attention submodule: the four-branch multi-head attention submodule is divided into units, specifically including gene focus head, environment focus head, phenotype focus head and interaction focus head, which process gene, environment, phenotype and gene-environment interaction features respectively;

[0028] After completing the unit division, each focal head begins to independently perform internal calculations. At the same time, a cross-modal attention mechanism is introduced between the gene and environment branches: the dot product operation is performed on the gene features and the environmental features to calculate the cross-correlation strength between the two. Based on the calculated cross-correlation strength value, a weighted cross feature is generated.

[0029] The output results of the four branches are spliced together. The resulting feature vector contains comprehensive information about genes, environment, phenotype, and gene-environment interaction. The spliced feature vector is input into a multi-layer perceptron, which further analyzes the input feature vector and ultimately generates a four-dimensional weight matrix. The four-dimensional weight matrix is normalized using the Softmax function to obtain the gene-environment interaction weight and the phenotypic response weight.

[0030] Based on the gene-environment interaction weight and the phenotypic response weight, a weighted sum operation is performed on the gene, environment, phenotype and gene-environment interaction characteristics to form an initial fusion vector;

[0031] Perform a multi-head self-attention operation on the initial fusion vector: The multi-head self-attention mechanism captures the correlation between elements in the initial fusion vector from multiple different subspaces, thereby calculating the global correlation estimate. Set a global correlation threshold K, and by comparing the global correlation estimate with the threshold K, filter out the core feature subsets whose global correlation estimate is greater than or equal to K.

[0032] A residual connection submodule is set up to input the features of the cross-domain feature interaction map together with the screened core feature subset into the convolutional neural network to generate a residual correction term; a non-negative matrix decomposition operation is performed on the residual correction term to extract latent factors as the components of the optimized comprehensive feature vector; the extracted latent factors are spliced with the previously screened core features, and the final optimized comprehensive feature vector is generated by weighting the attention score.

[0033] Furthermore, the feature evaluation and analysis module optimizes the comprehensive feature vector for decoding analysis, thereby constructing a rapeseed seed quality evaluation model and extracting high-yield rapeseed variety data. The process includes:

[0034] A sparse autoencoder is used to decouple feature domains. L1 regularization is used to force the hidden layer activation values to be sparse, separating the following sub-vectors: gene feature vectors, environmental feature vectors, phenotypic feature vectors, and interaction feature vectors. Each sub-vector corresponds to an independent feature domain.

[0035] A corresponding capsule processing unit is constructed for each decoupled feature domain to form a parallel feature domain processing pipeline. That is, the information transmission weight between each sub-vector is determined through an iterative routing algorithm. At the same time, a corresponding processing pipeline is constructed for each feature domain.

[0036] The decoupled feature sub-vectors are weighted and aggregated based on routing weights and passed layer by layer in the capsule network. First, in the local feature fusion stage, capsules from adjacent feature domains are combined in pairs, and local interaction patterns are refined through dynamic routing. Then, in the global feature fusion stage, all feature domain capsules are integrated into a unified high-level representation, forming a progressive feature refinement path from local to global.

[0037] Based on the results of progressive feature refinement, a multi-scale feature fusion mechanism is established, and the final output contains two levels of fusion features: local feature fusion and global feature fusion. Local feature fusion: retains interactive details through weighted combination of adjacent feature domain capsules; global feature fusion: aggregates all feature domain capsules into a unified global representation through top-level routing, forming a holistic description of rapeseed.

[0038] The feature importance evaluation method is used to quantitatively analyze the local and global fusion features obtained under the multi-scale feature fusion mechanism to determine the contribution of each feature domain and its combination to rapeseed variety labeling.

[0039] Based on the quantitative analysis results, feature domains and their combinations that play a key role in the labeling of high-yield rapeseed varieties are screened out as the criteria for subsequent labeling and screening; ultimately, the labeling results for each rapeseed variety are generated.

[0040] Furthermore, the transfer learning adaptation deployment module uses the extracted high-yield rapeseed variety data to deploy a generative adversarial network. Through transfer learning, the quality assessment model is adapted to the feature space of the new variety. The process of completing the intelligent assessment of the quality of the new rapeseed seed includes:

[0041] The extracted high-yield rapeseed variety data is used as source domain data and input into the preset adversarial training sub-module to pre-train the quality assessment model, so that it can learn the common feature representation related to rapeseed seed quality. A transfer learning strategy is adopted to transfer the knowledge of the pre-trained model from the high-yield variety dataset to the new variety rapeseed seed quality assessment task. At the same time, the parameters of the quality assessment model are adjusted and optimized according to the characteristics of the new variety data.

[0042] Compared with the prior art, the present invention has the following beneficial effects: by acquiring multi-source heterogeneous rapeseed data and constructing a data pool, data from different channels and different structures are integrated, data silos are broken, a comprehensive and rich data foundation is provided for subsequent analysis, and more potential information is mined; by constructing a cross-domain feature interaction map and using graph neural network learning, the interactive relationship between genes and environmental characteristics can be intuitively presented, the gene-environment interaction mechanism can be deeply mined, and a more accurate basis for subsequent analysis can be provided. The output updated map also provides strong support for subsequent analysis; by parsing the map through a four-branch multi-head attention architecture and combining it with a cross-modal attention mechanism, the importance of features in different dimensions can be fully focused on, multi-dimensional information can be integrated, and dual optimization can be used to generate optimized comprehensive feature vectors, effectively improving the quality and representativeness of the features and enhancing the model's ability to process complex data; by decoding the optimized feature vectors to construct an evaluation model and extract high-yield variety data, the model can be directly applied to rapeseed seed quality evaluation, while providing an important reference for subsequent new variety research and helping to quickly screen rapeseed varieties with high-yield potential; by deploying a generative adversarial network using high-yield data and adapting the model through transfer learning, the model can be quickly applied to new varieties, realizing intelligent evaluation and improving evaluation efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a module diagram of a rapeseed seed quality evaluation system based on big data proposed in the present invention. DETAILED DESCRIPTION

[0044] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the implementation regulations described are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0045] Reference Figure 1 , a rapeseed seed quality evaluation system based on big data, including:

[0046] Multi-source heterogeneous data acquisition and preprocessing module: acquires multi-source heterogeneous rapeseed data, processes and analyzes the acquired multi-source heterogeneous rapeseed data, and generates a multi-source heterogeneous rapeseed data pool;

[0047] Cross-domain feature interaction map construction module: Based on a multi-source heterogeneous rapeseed data pool, a cross-domain feature interaction map is constructed. Graph neural networks are used to perform deep feature learning on the constructed cross-domain feature interaction map, analyze gene-environment interactions, and output an updated cross-domain feature interaction map.

[0048] Multimodal feature fusion and optimization module: This module uses a preset four-branch multi-head attention architecture to analyze cross-domain feature interaction graphs, combines the cross-modal attention mechanism to generate dual-weighted and fused features, and then performs dual optimization after multi-head attention screening to generate an optimized comprehensive feature vector that integrates multi-dimensional information.

[0049] Feature evaluation and analysis module: Decodes and analyzes the optimized comprehensive feature vector to build a rapeseed seed quality evaluation model and extract high-yield rapeseed variety data;

[0050] Transfer learning adaptation deployment module: Use the extracted high-yield rapeseed variety data to deploy a generative adversarial network, adapt the quality assessment model to the new variety feature space through transfer learning, and complete the intelligent assessment of the quality of new rapeseed seeds.

[0051] It should be further explained that, in the specific implementation process, the multi-source heterogeneous data acquisition and preprocessing module obtains multi-source heterogeneous rapeseed data, processes and analyzes the obtained multi-source heterogeneous rapeseed data, and generates a multi-source heterogeneous rapeseed data pool in the following process:

[0052] Extracting rapeseed genome-wide SNP marker data from a gene database; collecting field phenotypic data through an IoT sensor network; connecting to a weather station to monitor real-time data streams and obtain field environmental parameters; and using a drone equipped with a multispectral camera to obtain dynamic image data of the rapeseed growth cycle. It is understood that the phenotypic data collected by field sensors includes branch number, yield, etc., and the field environmental parameters include environmental monitoring data such as soil moisture, temperature, and light intensity.

[0053] All collected data sets were uniformly stored in an HBase distributed database and preprocessed. Specifically, data cleaning was performed to remove duplicate sites and low-quality sites in the rapeseed genome-wide SNP marker data, retaining the valid SNP site set to obtain a genotype data set; time series records of field environmental parameters were analyzed to establish an environmental parameter time series index; drone aerial image data was used, and computer vision technology was used to extract features from the dynamic image data. Image timestamps were annotated according to growth cycle segments, and vegetation indices were extracted as preliminary phenotypic features; at the same time, the collected field phenotypic data and the extracted preliminary phenotypic features were spatiotemporally aligned to eliminate sampling time deviations, thereby obtaining synergistic phenotypic features;

[0054] Based on the genotype dataset, a SNP-chromosome association matrix was constructed as the genetic dimension framework for the rapeseed raw data pool. Environmental parameter time series indexes and synergistic phenotypic characteristics were mapped to the genetic dimension framework by timestamp to establish cross-domain data associations. The HBase distributed database's column family storage model was used to divide data column families by genotype, environment, and phenotype, and composite encoding rules for row keys were set. Data integrity checks were performed, verifying cross-domain data consistency using hash checksums, completing the construction of the rapeseed raw data pool.

[0055] The rapeseed original data pool was standardized to obtain a rapeseed standardized data pool; specifically, cluster analysis was performed on the rapeseed original data pool, and the DBSCAN algorithm was used to identify data outliers caused by extreme climate, and abnormal rapeseed data samples were eliminated based on the local density threshold; for missing measurement values, a time series interpolation algorithm was used to fill the missing measurement values; the filled data were subjected to Z-score standardization to eliminate dimensional differences and generate a standardized data pool that conforms to the normal distribution.

[0056] It should be further explained that in the specific implementation process, the cross-domain feature interaction map construction module constructs a cross-domain feature interaction map based on a multi-source heterogeneous rapeseed data pool, uses a graph neural network to perform deep feature learning on the constructed cross-domain feature interaction map, analyzes the gene-environment interaction relationship, and outputs the updated cross-domain feature interaction map. The process is as follows:

[0057] Separate genotype nodes (SNP site sequences), environmental nodes (time series data such as soil moisture and temperature), and phenotypic nodes (time series features such as NDVI and plant height) from the rapeseed standardized data pool, and perform one-hot encoding on each node to generate an initial feature vector as the initial feature representation of the map; for example, obtain the genotype, environment, and phenotypic data of 1,000 samples from the rapeseed standardized data pool. The genotype data contains information on 50 gene loci, the environmental data includes 10 environmental parameters such as temperature, humidity, and light, and the phenotypic data includes 8 phenotypic indicators such as plant height and yield; the genotype node has 50 possible states, which are encoded as a 50-dimensional vector; the environmental node is 10-dimensional, and the phenotypic node is 8-dimensional, to generate the initial feature vector;

[0058] After completing the initial feature representation of the node, perform association analysis on the gene node and the environment node: calculate the mutual information value between the gene node and the environment node , where the mutual information value is used to quantify the degree of association between the two node types, and the formula is:

[0059] ;

[0060] Where, is a gene node, is the gene node state set, is the environment node, is the state set of the environment nodes, is the joint probability, is the edge probability; based on the calculated mutual information value (assuming it is 0.3), a mutual information threshold (such as 0.5) is set to screen out significantly associated node pairs; based on the screening process, a gene-environment binary edge set is constructed (a total of 200 edges are obtained);

[0061] Based on the constructed gene-environment binary edge set and the initial feature representation of the node, a static graph is generated: the features of the adjacent nodes are aggregated through the graph convolution layer to capture the information of the adjacent nodes around the node, and the captured information is integrated into the feature representation of the target node, thereby generating a node embedding representation; it is understandable that when constructing a static graph, the features of the adjacent nodes are aggregated using the graph convolution layer, and this process involves information interaction between multiple nodes in the graph structure; in each step of the graph convolution operation, a node is selected as the current processing object, and this node is the target node; for example, assuming a gene-environment interaction graph, including gene nodes j1, j2, j3 and environment nodes e1, e2; when the gene When node j1 performs graph convolution, gene node j1 is the target node. If gene node j1 is connected to environment nodes e1 and e2, the graph convolution layer first captures the feature information of environment nodes e1 and e2, and then fuses this feature information into the feature representation of gene node j1. The updated features of gene node j1 contain its own initial features and the feature information of environment nodes e1 and e2, thereby more accurately describing the interaction between gene node j1 and the environment. Finally, the output static graph contains a node embedding matrix and an adjacency matrix. The node embedding matrix is used to record the feature representation of each node after feature aggregation, and the adjacency matrix is used to represent the connection relationship between nodes in the static graph.

[0062] The generated static graph is input into a three-layer graph attention network, and a gated recurrent unit is used to capture the temporal changes in the weights of gene-environment edges: each layer of the graph attention network adopts a multi-head attention mechanism (such as an 8-head attention mechanism), which learns the attention relationship between nodes from multiple different subspaces, thereby more comprehensively capturing the interaction information between nodes; in each layer, the attention coefficient between nodes is calculated, where the attention coefficient is used to reflect the importance and association strength between nodes; based on the calculated attention coefficient, the node representation is updated so that the node integrates the information of its adjacent nodes; using the rate of change of environmental parameters as the gating signal, the gated recurrent unit dynamically adjusts the weight value of the gene-environment edge according to the changes in environmental parameters, so that the edge weights in the graph are updated in real time as the environment changes, thereby more accurately describing the dynamic interaction relationship between genes and the environment;

[0063] Defining gene-environment interaction strength indices , the indicator is calculated as edge weight Multiply by the node embedding cosine similarity ,Right now , where edge weight represents the strength of the association between gene nodes and environment nodes, and node embedding cosine similarity is used to measure the similarity between two nodes in the feature space. The product of the two comprehensively reflects the strength of gene-environment interaction. Based on the gene-environment interaction strength index, the top N (e.g., 50) strongly associated edges are screened, where N is a preset positive integer. This is used to limit the number of strongly associated edges selected from a large number of gene-environment edges sorted by the interaction strength index value, so as to focus on edges that play a key role in gene-environment interaction and lay the foundation for the quantitative range of subsequent precise screening of key feature edges.

[0064] Introducing a graph contrast learning framework, we further analyze the top N strongly correlated edges in the index values: performing perturbations on the static graph, including random modifications to node features and edge weights; and calculating the rate of change in interaction strength by comparing the differences in node representations before and after the perturbation. :

[0065] ;

[0066] Set a change rate threshold M and filter key feature edges whose interaction strength change rate is greater than or equal to the change rate threshold M; integrate all the filtered key feature edges into the static graph to form the final dynamic graph, that is, the updated cross-domain feature interaction graph;

[0067] For example, the temperature changes by 10% and the humidity changes by 5% within a month. The weight value of the gene-environment edge is dynamically modified according to these change rates; the gene-environment interaction strength index is defined, and the top 50 strongly correlated edges are screened; the static graph is perturbed, such as randomly modifying 5% of the node features and 3% of the edge weights; by comparing the differences in node representation of the graph before and after the perturbation, the interaction strength change rate is calculated, and the change rate threshold M is set to 0.1. 30 key feature edges with a change rate greater than or equal to 0.1 are screened out and integrated into the static graph to form the final dynamic graph.

[0068] It should be further explained that in the specific implementation process, the multimodal feature fusion and optimization module uses a preset four-branch multi-head attention architecture to analyze the cross-domain feature interaction map, combines the cross-modal attention mechanism to generate dual weights and fuse features, and then performs dual optimization after multi-head attention screening to generate an optimized comprehensive feature vector that integrates multi-dimensional information. The process is as follows:

[0069] The constructed cross-domain feature interaction map is input into the preset four-branch multi-head attention submodule: the four-branch multi-head attention submodule is divided into units, specifically including gene focus head, environment focus head, phenotype focus head and interaction focus head, which process gene, environment, phenotype and gene-environment interaction features respectively;

[0070] After completing the unit division, each focus head begins to independently perform internal calculations, including computing query, key, and value matrices. At the same time, a cross-modal attention mechanism is introduced between the gene and environment branches: the dot product operation is performed on the gene features and the environmental features to calculate the cross-correlation strength between the two. Based on the calculated cross-correlation strength value, a weighted cross feature is generated.

[0071] The output results of the four branches (gene focus head, environment focus head, phenotype focus head, interaction focus head) are spliced together. The spliced feature vector contains comprehensive information about genes, environment, phenotype, and gene-environment interaction. The spliced feature vector is input into a multi-layer perceptron, which further analyzes the input feature vector (nonlinear transformation and feature extraction) to ultimately generate a four-dimensional weight matrix. The four-dimensional weight matrix is normalized using the Softmax function to obtain the gene-environment interaction weight and phenotypic response weight. The gene-environment interaction weight is used to reflect the importance of the interaction between genes and the environment, while the phenotypic response weight reflects the sensitivity of the phenotype to changes in genes and the environment.

[0072] Based on the gene-environment interaction weight and the phenotypic response weight, a weighted sum operation is performed on the gene, environment, phenotype and gene-environment interaction characteristics to form an initial fusion vector;

[0073] Perform a multi-head self-attention operation on the initial fusion vector: The multi-head self-attention mechanism captures the correlation between elements in the initial fusion vector from multiple different subspaces, thereby calculating a global correlation estimate, where the global correlation estimate reflects the importance of each element in the vector as a whole. Set a global correlation threshold K (such as 0.8) and compare the global correlation estimate with the threshold K to select a core feature subset with a global correlation estimate greater than or equal to K. The core feature subset contains the most critical and important information in the initial fusion vector.

[0074] A residual connection submodule is set up to input the features of the cross-domain feature interaction map together with the core feature subset obtained by screening into the convolutional neural network to generate a residual correction term; among which, the features of the cross-domain feature interaction map include feature information of genes, environment, phenotype and gene-environment interaction; a non-negative matrix decomposition operation is performed on the residual correction term: the high-dimensional residual correction term is decomposed into a low-dimensional latent factor matrix and a coefficient matrix, and the latent factor is used as the latent feature pattern in the residual correction term; it can be understood that in the gene feature dimension, the latent factor corresponds to the core component of the gene expression pattern in the gene regulatory network; for environmental characteristics, the latent factor reflects the key environmental factor combination affecting rapeseed traits or the potential driving factor of environmental change; in terms of phenotypic characteristics, the latent factor represents the potential source of phenotypic variation or phenotypic classification; for gene-environment interaction characteristics, the latent factor reflects the potential mechanism of gene-environment interaction; the latent factor is extracted as the component unit of the optimized comprehensive feature vector; the extracted latent factor is spliced with the core features screened out previously, and the final optimized comprehensive feature vector is generated by weighting the attention score.

[0075] It should be further explained that, in the specific implementation process, the feature evaluation and analysis module will optimize the comprehensive feature vector for decoding analysis, thereby constructing a rapeseed seed quality evaluation model and extracting high-yield rapeseed variety data as follows:

[0076] The sparse autoencoder is used to decouple the feature domain. Through L1 regularization constraint, the hidden layer activation value is forced to be sparse, and the gene feature sub-vector is separated. , environmental feature vector , phenotypic feature vector and the interaction eigenvectors ; where R represents the real number space (i.e., the eigenvalues are composed of specific numerical values), d represents the dimension of the corresponding eigenvector (i.e., the number of parameters describing the gene, environment, phenotype, or interaction characteristics); each subvector corresponds to an independent feature domain, representing a semantically clear set of features in the original feature space;

[0077] A corresponding capsule processing unit is constructed for each decoupled feature domain (gene / environment / phenotype / interaction) to form a parallel feature domain processing pipeline, that is, the information transfer weight between each sub-vector is determined by an iterative routing algorithm; at the same time, a corresponding processing pipeline is constructed for each feature domain: the initial routing coefficient is set, and the routing coefficient is adjusted according to the prediction consistency between the low-level capsule and the high-level capsule; at the same time, a corresponding processing pipeline is constructed for each feature domain, specifically: the gene pipeline is obtained by initializing the 1D-CNN convolution kernel; the environment pipeline is obtained by loading the pre-trained ResNet-18 weight processing spatiotemporal sequence; the phenotype pipeline is obtained by loading the plant phenotypic prior knowledge by constructing a knowledge distillation submodule; the bilinear fusion matrix is initialized Obtaining a gene-environment interaction pipeline;

[0078] The decoupled feature sub-vectors are weighted and aggregated based on the routing weights and transmitted layer by layer in the capsule network: First, in the local feature fusion stage, the capsules of adjacent feature domains (such as gene-environment, environment-phenotype) are combined in pairs, and the local interaction patterns are refined through dynamic routing; then, in the global feature fusion stage, all feature domain capsules are integrated into a unified high-level representation, forming a progressive feature refinement path from local to global. Specifically, for gene feature refinement, the gene feature sub-vector G is input to 1D-CNN, and the primary gene pattern is output. , perform sequence annotation through the CRF layer, identify disease resistance-related SNP combination patterns, and generate gene quality scores ,in is the activation function, is the weight matrix of gene features, is the bias term of genetic features; for environmental feature refinement, E is input into the spatiotemporal Transformer and decomposed into static attributes Es (soil type, etc.) and dynamic time series Ed (precipitation / temperature series) through the separable attention mechanism to generate the environmental adaptability score QE = ME(Es⊕LE(Ed)), where ME is a multi-layer decision processor (mining the rules of static environmental features through multi-layer neural networks), LE is a time series memory submodule (used to process time-varying environmental dynamic data such as temperature / precipitation), and ⊕ is a feature fusion operation (merging and analyzing static attributes with dynamic time series information); for phenotypic feature refinement, a cross-layer graph structure is constructed, with nodes as phenotypic features and edges as genetic correlations between phenotypes. Information is propagated through the GAT network to update the phenotypic representation P' and generate a phenotypic integrity score. ,in is the attention weight, Represents different phenotypic feature dimensions (such as plant height, number of siliques, 1000-seed weight, and other specific trait indicators); for interactive feature refinement, bilinear pooling is performed: , generating an interaction strength score , where tanh is a bidirectional activation function, represents the matrix transpose, is the interaction feature weight matrix, The baseline correction parameter calculated for the interaction feature (i.e., the interaction feature bias term), is the interaction feature matrix after bilinear pooling (including multi-dimensional interaction information between genes and environment), and F is the norm of the interaction strength score);

[0079] Based on the results of progressive feature refinement, a multi-scale feature fusion mechanism is established, and the final output contains two levels of fusion features: local feature fusion and global feature fusion. Local feature fusion: retains the detailed information of interactions through weighted combination of adjacent feature domain capsules (such as gene-environment fusion capsules); global feature fusion: aggregates all feature domain capsules into a unified global representation through top-level routing, forming a holistic description of rapeseed.

[0080] Using feature importance assessment methods (such as feature importance calculation based on tree models), quantitative analysis is performed on the local and global fusion features obtained under the multi-scale feature fusion mechanism to determine the contribution of each feature domain and its combination to rapeseed variety labeling;

[0081] Based on the results of quantitative analysis, feature domains and their combinations that play a key role in the labeling of high-yield rapeseed varieties (such as certain gene site combinations in the gene feature sub-vector and temperature and humidity combinations in the environmental feature sub-vector) are screened out as standards for subsequent labeling and screening, while features with small contributions or irrelevant features are eliminated. For example, high-yield rapeseed variety data that meet these key characteristics are extracted from the data, and a total of 200 high-yield rapeseed variety data are obtained. Finally, the labeling results for each rapeseed variety are generated.

[0082] It should be further explained that, during the specific implementation process, the transfer learning adaptation deployment module uses the extracted high-yield rapeseed variety data to deploy a generative adversarial network. Through transfer learning, the quality assessment model is adapted to the feature space of the new variety, completing the intelligent assessment of the quality of the new rapeseed seed as follows:

[0083] The extracted high-yield rapeseed variety data is used as source domain data and input into the preset adversarial training sub-module to pre-train the quality assessment model, so that it can learn the common feature representation related to rapeseed seed quality. A transfer learning strategy is adopted to transfer the knowledge of the pre-trained model from the high-yield variety dataset to the new variety rapeseed seed quality assessment task. At the same time, the parameters of the quality assessment model are adjusted and optimized according to the characteristics of the new variety data to ensure that the quality assessment model can accurately evaluate the rapeseed seed quality based on the new variety data.

[0084] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0085] It should be understood that determining B based on A does not mean determining B based solely on A. B can also be determined based on A and / or other information.

[0086] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0087] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A rapeseed seed quality assessment system based on big data, characterized by: Multi-source heterogeneous data acquisition and preprocessing module: acquires multi-source heterogeneous rapeseed data, processes and analyzes the acquired multi-source heterogeneous rapeseed data, and generates a multi-source heterogeneous rapeseed data pool; Cross-domain feature interaction map construction module: Based on a multi-source heterogeneous rapeseed data pool, a cross-domain feature interaction map is constructed. Graph neural networks are used to perform deep feature learning on the constructed cross-domain feature interaction map, analyze gene-environment interactions, and output an updated cross-domain feature interaction map. Multimodal feature fusion and optimization module: This module uses a preset four-branch multi-head attention architecture to analyze cross-domain feature interaction graphs, combines the cross-modal attention mechanism to generate dual weights and fuse features, and then uses multi-head attention to filter and optimize the generated optimized comprehensive feature vector that integrates multi-dimensional information. Specifically: The constructed cross-domain feature interaction map is input into the preset four-branch multi-head attention submodule: the four-branch multi-head attention submodule is divided into units, specifically including gene focus head, environment focus head, phenotype focus head and interaction focus head, which process gene, environment, phenotype and gene-environment interaction features respectively; After completing the unit division, each focal head begins to independently perform internal calculations. At the same time, a cross-modal attention mechanism is introduced between the gene and environment branches: the dot product operation is performed on the gene features and the environmental features to calculate the cross-correlation strength between the two. Based on the calculated cross-correlation strength value, a weighted cross feature is generated. The output results of the four branches are spliced together. The resulting feature vector contains comprehensive information about genes, environment, phenotype, and gene-environment interaction. The spliced feature vector is input into a multi-layer perceptron, which further analyzes the input feature vector and ultimately generates a four-dimensional weight matrix. The four-dimensional weight matrix is normalized using the Softmax function to obtain the gene-environment interaction weight and the phenotypic response weight. Based on the gene-environment interaction weight and the phenotypic response weight, a weighted sum operation is performed on the gene, environment, phenotype and gene-environment interaction characteristics to form an initial fusion vector; Perform a multi-head self-attention operation on the initial fusion vector: The multi-head self-attention mechanism captures the correlation between elements in the initial fusion vector from multiple different subspaces, thereby calculating the global correlation estimate. Set a global correlation threshold K, and by comparing the global correlation estimate with the threshold K, filter out the core feature subsets whose global correlation estimate is greater than or equal to K. A residual connection submodule is set up to input the features of the cross-domain feature interaction map and the selected core feature subset into the convolutional neural network to generate residual correction terms. Non-negative matrix factorization is performed on the residual correction terms to extract latent factors as the components of the optimized comprehensive feature vector. The extracted latent factors are spliced with the previously selected core features and weighted by the attention score to generate the final optimized comprehensive feature vector. Feature evaluation and analysis module: Decodes and analyzes the optimized comprehensive feature vector to build a rapeseed seed quality evaluation model and extract high-yield rapeseed variety data; Transfer learning adaptation deployment module: Use the extracted high-yield rapeseed variety data to deploy a generative adversarial network, adapt the quality assessment model to the new variety feature space through transfer learning, and complete the intelligent assessment of the quality of new rapeseed seeds.

2. The rapeseed seed quality assessment system based on big data according to claim 1, characterized in that: The multi-source heterogeneous data acquisition and preprocessing module acquires multi-source heterogeneous rapeseed data, processes and analyzes the acquired multi-source heterogeneous rapeseed data, and generates a multi-source heterogeneous rapeseed data pool. The process includes: Extracting rapeseed genome-wide SNP marker data from a gene database; collecting field phenotypic data through an IoT sensor network; connecting to a weather station to monitor data streams in real time and obtain field environmental parameters; and using a drone equipped with a multispectral camera to obtain dynamic image data of the rapeseed growth cycle. All collected data sets were uniformly stored in an HBase distributed database and preprocessed. Specifically, data cleaning was performed to remove duplicate sites and low-quality sites in the rapeseed genome-wide SNP marker data, retaining the valid SNP site set to obtain a genotype data set; time series records of field environmental parameters were analyzed to establish an environmental parameter time series index; drone aerial image data was used, and computer vision technology was used to extract features from the dynamic image data. Image timestamps were annotated according to growth cycle segments, and vegetation indices were extracted as preliminary phenotypic features; at the same time, the collected field phenotypic data and the extracted preliminary phenotypic features were spatiotemporally aligned to eliminate sampling time deviations, thereby obtaining synergistic phenotypic features; Based on the genotype dataset, a SNP-chromosome association matrix was constructed as the genetic dimension framework for the rapeseed raw data pool. Environmental parameter time series indexes and synergistic phenotypic characteristics were mapped to the genetic dimension framework by timestamp to establish cross-domain data associations. The HBase distributed database's column family storage model was used to divide data column families by genotype, environment, and phenotype, and composite encoding rules for row keys were set. Data integrity checks were performed, verifying cross-domain data consistency using hash checksums, completing the construction of the rapeseed raw data pool. The rapeseed original data pool is standardized to obtain the rapeseed standardized data pool.

3. The rapeseed seed quality assessment system based on big data according to claim 2, characterized in that: The cross-domain feature interaction map construction module constructs a cross-domain feature interaction map based on a multi-source heterogeneous rapeseed data pool. It uses a graph neural network to perform deep feature learning on the constructed cross-domain feature interaction map, analyzes the gene-environment interaction relationship, and outputs an updated cross-domain feature interaction map. The process includes: Separate genotype nodes, environmental nodes, and phenotype nodes from the rapeseed standardized data pool, perform one-hot encoding on each node to generate an initial feature vector as the initial feature representation of the graph; After completing the initial feature representation of the nodes, the association analysis of gene nodes and environmental nodes is performed: the mutual information value of the gene nodes and environmental nodes is calculated, where the mutual information value is used to quantify the degree of association between the two node types; based on the calculated mutual information value, the mutual information threshold is set to filter out significantly associated node pairs; based on the screening process, a gene-environment binary edge set is constructed; Based on the constructed gene-environment binary edge set and the initial feature representation of the node, a static graph is generated. The features of adjacent nodes are aggregated through the graph convolution layer to capture information about the adjacent nodes around the node. This captured information is then integrated into the feature representation of the target node, thereby generating a node embedding representation. The output static graph contains a node embedding matrix and an adjacency matrix. The node embedding matrix is used to record the feature representation of each node after feature aggregation, while the adjacency matrix is used to represent the connection relationships between nodes in the static graph. The generated static graph is fed into a three-layer graph attention network, and a gated recurrent unit is used to capture the temporal changes in gene-environment edge weights. Define a gene-environment interaction strength index, which is calculated as the edge weight multiplied by the node embedding cosine similarity, where the edge weight represents the strength of the association between the gene node and the environment node, and the node embedding cosine similarity is used to measure the similarity between the two nodes in the feature space. Based on the gene-environment interaction strength index, select the top N strongly associated edges with the index value, where N is a preset positive integer that limits the number of strongly associated edges selected from the numerous gene-environment edges sorted by the interaction strength index value. A graph comparative learning framework is introduced to further analyze the top N strongly correlated edges in the index values: a perturbation operation is performed on the static graph, which includes random modifications to node features and edge weights; the rate of change of interaction intensity is calculated by comparing the differences in node representation of the graph before and after the perturbation; a change rate threshold M is set to filter key feature edges whose interaction intensity change rate is greater than or equal to the change rate threshold M; all the filtered key feature edges are integrated into the static graph to form the final dynamic graph, that is, the updated cross-domain feature interaction graph.

4. The rapeseed seed quality assessment system based on big data according to claim 3, characterized in that: The process of inputting the generated static graph into a three-layer graph attention network and using a gated recurrent unit to capture the temporal changes in gene-environment edge weights includes: Each layer of the graph attention network uses a multi-head attention mechanism, which learns the attention relationship between nodes from multiple different subspaces. In each layer, the attention coefficient between nodes is calculated, where the attention coefficient is used to reflect the importance and correlation strength between nodes. Based on the calculated attention coefficient, the node representation is updated so that the node integrates the information of its adjacent nodes. Using the rate of change of environmental parameters as the gating signal, the gated recurrent unit dynamically modifies the weight value of the gene-environment edge according to the changes in environmental parameters.

5. The rapeseed seed quality assessment system based on big data according to claim 1, characterized in that: Feature evaluation The analysis module optimizes the comprehensive feature vector for decoding analysis, thereby constructing a rapeseed seed quality evaluation model and extracting high-yield rapeseed variety data. The process includes: A sparse autoencoder is used to decouple feature domains. L1 regularization is used to force the hidden layer activation values to be sparse, separating the following sub-vectors: gene feature vectors, environmental feature vectors, phenotypic feature vectors, and interaction feature vectors. Each sub-vector corresponds to an independent feature domain. A corresponding capsule processing unit is constructed for each decoupled feature domain to form a parallel feature domain processing pipeline. That is, the information transmission weight between each sub-vector is determined through an iterative routing algorithm. At the same time, a corresponding processing pipeline is constructed for each feature domain. The decoupled feature sub-vectors are weighted and aggregated based on routing weights and passed layer by layer in the capsule network. First, in the local feature fusion stage, capsules from adjacent feature domains are combined in pairs, and local interaction patterns are refined through dynamic routing. Then, in the global feature fusion stage, all feature domain capsules are integrated into a unified high-level representation, forming a progressive feature refinement path from local to global. Based on the results of progressive feature refinement, a multi-scale feature fusion mechanism is established, and the final output contains two levels of fusion features: local feature fusion and global feature fusion. Local feature fusion: retains interactive details through weighted combination of adjacent feature domain capsules; global feature fusion: aggregates all feature domain capsules into a unified global representation through top-level routing, forming a holistic description of rapeseed. The feature importance evaluation method is used to quantitatively analyze the local and global fusion features obtained under the multi-scale feature fusion mechanism to determine the contribution of each feature domain and its combination to rapeseed variety labeling. Based on the quantitative analysis results, feature domains and their combinations that play a key role in the labeling of high-yield rapeseed varieties are screened out as the criteria for subsequent labeling and screening; ultimately, the labeling results for each rapeseed variety are generated.

6. The rapeseed seed quality assessment system based on big data according to claim 1, characterized in that: The transfer learning adaptation deployment module uses the extracted high-yield rapeseed variety data to deploy a generative adversarial network. Through transfer learning, the quality assessment model is adapted to the feature space of the new variety. The process of completing the intelligent assessment of the quality of the new rapeseed seeds includes: The extracted high-yield rapeseed variety data is used as source domain data and input into the preset adversarial training sub-module to pre-train the quality assessment model, so that it can learn the common feature representation related to rapeseed seed quality. A transfer learning strategy is adopted to transfer the knowledge of the pre-trained model from the high-yield variety dataset to the new variety rapeseed seed quality assessment task. At the same time, the parameters of the quality assessment model are adjusted and optimized according to the characteristics of the new variety data.

Images