Phosphogypsum element-mineral correlation and elution difference analysis method and device and electronic equipment

Abstract

The invention relates to the technical field of ardealite resource utilization and process optimization, and provides an ardealite element-mineral correlation and elution difference analysis method which comprises the following steps: acquiring element occurrence data of an ardealite sample before and after elution, and constructing a mineral-element incidence matrix before elution and a mineral-element incidence matrix after elution; based on the pre-leaching mineral-element incidence matrix and the post-leaching mineral-element incidence matrix, respectively constructing a pre-leaching incidence diagram and a post-leaching incidence diagram; inputting the pre-elution association graph and the post-elution association graph into a cross-domain graph neural network model, calculating an element-mineral basic correlation matrix based on node embedding features, and calculating an influence coefficient matrix based on cross-domain attention differences; the basic correlation matrix represents the occurrence relation strength between elements and minerals, and the influence coefficient matrix represents the influence degree of the leaching process on the element-mineral correlation strength. Nonlinear correlation can be captured, the influence of the leaching process can be quantified, and process optimization can be effectively guided.

Description

Technical Field

[0001] This invention relates to the field of phosphogypsum resource utilization and process optimization technology, specifically to a method, apparatus and electronic equipment for analyzing the element-mineral correlation and leaching differences of phosphogypsum. Background Technology

[0002] Phosphogypsum is a large-scale industrial solid waste generated during the wet-process phosphoric acid production. While its annual output is high, its comprehensive utilization rate has long been low. Impurities such as phosphorus, fluorine, and organic matter in phosphogypsum severely restrict its resource utilization. Leaching, as a core method for the harmless pretreatment of phosphogypsum, reduces the content of harmful elements by dissolving and removing soluble impurity minerals. However, optimizing the leaching process parameters highly depends on a precise understanding of the correlation between elemental and mineral composition and the mechanisms influencing the leaching process.

[0003] Existing methods for correlation analysis of phosphogypsum components mainly fall into two categories: traditional statistical methods and conventional machine learning methods. Traditional statistical methods, such as Pearson correlation coefficient and Spearman rank correlation coefficient, can only capture linear relationships between variables and cannot effectively characterize the complex nonlinear dependencies between elements and minerals. Existing GAT models are difficult to directly handle dual-temporal / dual-domain data with strong physical correlations before and after leaching, and they are also difficult to quantify the perturbation of microscopic correlations by process parameters in the absence of explicit labels. Furthermore, there is a significant stoichiometric prior relationship between the elemental composition and mineral composition of phosphogypsum, i.e., the molecular formula of the mineral determines its elemental composition. Existing methods fail to fully integrate this chemical prior information, resulting in insufficient physical interpretability of the correlation analysis results.

[0004] In summary, the existing technology has the following clear technical defects: First, it lacks an analytical method that can simultaneously capture nonlinear correlations and quantify cross-domain effects; second, the composition data before and after rinsing are cross-domain data with distribution differences, making it difficult for conventional methods to effectively align and accurately identify correlation trends; third, there is a lack of effective mapping between the analytical results and rinsing process parameters, which cannot provide clear data support for adjusting key process parameters such as temperature, liquid-solid ratio, and stirring rate.

[0005] Therefore, there is an urgent need for a method for analyzing the element-mineral correlation and leaching difference of phosphogypsum that can capture nonlinear correlations, quantify the impact of the leaching process, and effectively guide process optimization. Summary of the Invention

[0006] In view of this, embodiments of this application provide a method, apparatus and electronic device for analyzing element-mineral correlation and leaching differences in phosphogypsum, in order to solve the technical problems that existing methods cannot capture nonlinear correlations, quantify the impact of the leaching process and effectively guide process optimization.

[0007] The first aspect of this application provides a method for analyzing the element-mineral correlation and leaching differences of phosphogypsum, including: Elemental occurrence data of phosphogypsum samples before and after leaching were obtained. Based on the elemental occurrence data, the mass contribution ratio of each mineral to each element was calculated, and mineral-element correlation matrices before and after leaching were constructed respectively. Based on the mineral-element correlation matrix before and after leaching, a correlation graph before and after leaching are constructed respectively; the correlation graph uses elements and minerals as nodes and the quality contribution ratio as edge weights. The pre-rinsing and post-rinsing association graphs are input into a cross-domain graph neural network model, and the node embedding features of the two domains are extracted respectively. The cross-domain attention difference of the corresponding edges of the two domains is calculated. The element-mineral basic correlation matrix is ​​calculated based on the node embedding features, and the influence coefficient matrix is ​​calculated based on the cross-domain attention difference. The basic correlation matrix represents the strength of the occurrence relationship between elements and minerals, and the influence coefficient matrix represents the degree of influence of the leaching process on the correlation strength of each element-mineral.

[0008] In one embodiment, the nodes in the association graph include element nodes and mineral nodes, and the edges are undirected edges connecting mineral nodes and element nodes; The feature vector of the element node is the normalized value of the mass proportion of the element in the sample, and the feature vector of the mineral node is the normalized value of the mass proportion of the mineral in the sample. Wherein, when the quality contribution ratio is less than a preset threshold, the weight of the corresponding edge is set to zero.

[0009] In one embodiment, the cross-domain graph neural network model includes a graph attention layer and a cross-domain attention head; the graph attention layer is used to extract node embedding features of the pre-washing association graph and the post-washing association graph, respectively; the cross-domain attention head is used to calculate the cross-domain attention difference of corresponding edges between the two domains; The method by which the cross-domain attention head calculates cross-domain attention differences is as follows: The cross-domain feature vector is obtained by concatenating the node embedding features before and after rinsing. ,in For node embedding features before rinsing, This refers to the embedding features of nodes after rinsing; Calculate cross-domain attention weights using a fully connected layer: ,in It is the sigmoid activation function. For cross-domain weight matrix, For bias terms; Calculate cross-domain difference values: ,in and These are the attention weights of the corresponding edges in the association graph before and after rinsing, respectively.

[0010] In one embodiment, the underlying correlation matrix The element calculation method is as follows: ; in mineral nodes With element nodes Cosine similarity of embedded features and The node embedding vectors learned by the GAT (Graph Attention Network) before rinsing.

[0011] In one embodiment, the influence coefficient matrix elements Normalized values ​​for cross-domain attention differences: ; in minerals With elements Cross-domain attention differences This is the maximum value among all cross-domain attention differences.

[0012] In one embodiment, the training of the cross-domain graph neural network model employs a joint loss function, which includes a weighted sum of node embedding reconstruction loss and cross-domain difference loss; the node embedding reconstruction loss is used to constrain the representational ability of node embedding features on the original input features, and the cross-domain difference loss is used to constrain the learning of cross-domain difference distribution.

[0013] In one embodiment, the method further includes: identifying element-mineral association pairs with influence coefficients higher than a preset threshold based on the influence coefficient matrix, and generating adjustment suggestions for leaching process parameters for the identified association pairs; The rinsing process parameters include at least one of rinsing temperature, liquid-to-solid ratio, and stirring rate.

[0014] In one embodiment, the elemental occurrence data is acquired through an automated mineralogy system, which is based on a combination of scanning electron microscopy and energy dispersive spectroscopy to output chemical elemental composition, mineral composition, and the occurrence state of elements in each mineral.

[0015] A second aspect of this application provides a device for analyzing the element-mineral correlation and leaching differences of phosphogypsum, comprising: The data acquisition and matrix construction module is used to acquire the elemental distribution data of phosphogypsum samples before and after leaching, calculate the mass contribution ratio of each mineral to each element based on the elemental distribution data, and construct the mineral-element correlation matrix before leaching and the mineral-element correlation matrix after leaching, respectively. The correlation graph construction module is used to construct a correlation graph before leaching and a correlation graph after leaching based on the mineral-element correlation matrix before leaching and the mineral-element correlation matrix after leaching, respectively; the correlation graph uses elements and minerals as nodes and the quality contribution ratio as edge weights. The feature extraction and difference calculation module is used to input the pre-rinsing association graph and the post-rinsing association graph into the cross-domain graph neural network model, extract the node embedding features of the two domains respectively, and calculate the cross-domain attention difference of the corresponding edges of the two domains. The correlation and influence coefficient output module is used to calculate the element-mineral basic correlation matrix based on the node embedding features and to calculate the influence coefficient matrix based on the cross-domain attention difference. The basic correlation matrix represents the strength of the occurrence relationship between elements and minerals, and the influence coefficient matrix represents the degree of influence of the leaching process on the correlation strength of each element-mineral.

[0016] A third aspect of this application provides an electronic device including a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the electronic device enables the phosphogypsum element-mineral correlation and leaching difference analysis method provided in the first aspect of this application.

[0017] A fourth aspect of this application provides a computer program product including a computer program that, when run, causes the method described in the first aspect of this application to be performed.

[0018] The method for analyzing the element-mineral correlation and leaching difference of phosphogypsum provided in the first aspect of this application achieves accurate capture of the nonlinear occurrence relationship between phosphogypsum elements and minerals by constructing a dual-domain correlation graph and introducing a cross-domain graph neural network model; it quantifies the change in correlation strength before and after leaching through a cross-domain attention mechanism, solving the technical bottleneck that traditional methods cannot assess the impact of the process; the output basic correlation matrix and influence coefficient matrix can directly guide the optimization of leaching process parameters, providing reliable data support for the harmless treatment of phosphogypsum.

[0019] It is understood that the beneficial effects of the second to fourth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic flowchart of a method for analyzing the element-mineral correlation and leaching differences of phosphogypsum according to an embodiment of this application; Figure 2 This is a schematic flowchart of a method for analyzing the element-mineral correlation and leaching differences of phosphogypsum according to another embodiment of this application; Figure 3 This is a structural diagram of a multi-head attention mechanism provided in an embodiment of this application; Figure 4 This is a schematic diagram of the basic network structure of the CD-GAT model provided in one embodiment of this application; Figure 5 This is a schematic diagram illustrating the influence of rinsing on elemental composition according to an embodiment of this application; Figure 6 This is a schematic diagram of the structure of the phosphogypsum element-mineral correlation and leaching difference analysis device provided in the embodiments of this application. Detailed Implementation

[0022] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0023] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0024] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0025] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0026] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0027] like Figure 1 , 2 As shown, the method for analyzing the element-mineral correlation and leaching difference of phosphogypsum provided in this application includes the following steps S101 to S104: Step S101: Obtain the elemental distribution data of the phosphogypsum sample before and after leaching, calculate the mass contribution ratio of each mineral to each element based on the elemental distribution data, and construct the mineral-element correlation matrix before leaching and the mineral-element correlation matrix after leaching, respectively.

[0028] In application, obtaining elemental composition data of phosphogypsum samples before and after leaching refers to collecting chemical composition information of phosphogypsum samples using professional mineralogical testing equipment. This elemental composition data includes the total content of each chemical element in the sample, the content of each mineral phase, and the distribution of each element in different minerals. The sample before leaching refers to the raw phosphogypsum without leaching treatment, while the sample after leaching refers to phosphogypsum treated with set process parameters. Calculating the mass contribution ratio of each mineral to each element based on the elemental composition data means quantifying the contribution of each mineral to the total element content by relating the amount of each element in each mineral to the total amount of that element.

[0029] Specifically, data such as chemical elemental composition, mineral composition, and elemental occurrence states are obtained from the AMICS automated mineralogy system; this is denoted as the element set before leaching. The set of elements after rinsing ,in The number of element types, the mineral aggregate before leaching. The set of elements after rinsing ,in Given the number of mineral types, based on elemental occurrence state data, the mass contribution ratio of each mineral to each element is calculated, and mineral-element mass contribution matrices are constructed before leaching. and the mineral-element mass contribution matrix after leaching .

[0030] Step S102: Based on the mineral-element correlation matrix before leaching and the mineral-element correlation matrix after leaching, construct the correlation graph before leaching and the correlation graph after leaching respectively; the correlation graph uses elements and minerals as nodes and the quality contribution ratio as edge weights.

[0031] In application, constructing the correlation graph is the process of transforming the mineral-element correlation matrix into graph-structured data. The correlation graph is a bipartite graph structure, where nodes are divided into two categories: element nodes and mineral nodes. Element nodes represent the various chemical elements detected in phosphogypsum, and mineral nodes represent the various mineral phases identified in the sample. Edges exist only between element nodes and mineral nodes, indicating an affinity relationship between them. The edge weight is the mass contribution ratio calculated in step S101; a larger weight indicates a more significant contribution of the mineral to the element. The correlation graph before and after leaching corresponds to the element-mineral relationship networks in the two states before and after leaching, respectively.

[0032] Specifically, based on the matrix obtained in step S101, correlation graphs before rinsing are constructed respectively. Correlation diagram after rinsing , where the set of nodes For element nodes, For mineral nodes, the total number of nodes is ; Element node feature vector: That is, the normalized value of the mass percentage of the element (where... (Represents content); mineral node feature vector: That is, the normalized value of the proportion of the mineral content; edge set and Undirected edges are established only between mineral nodes and element nodes; there are no connections between nodes of the same type (element-element, mineral-mineral). Edge weight: and The elements in the table are directly used as the weights of the corresponding edges. When the weight is less than a preset threshold (e.g., ...), the weight is used to determine the weight of the edge. When the element is considered to have no significant correlation with the mineral, the edge weight is set to 0 to reduce noise interference.

[0033] Step S103: Input the pre-rinse association graph and the post-rinse association graph into the cross-domain graph neural network model, extract the node embedding features of the two domains respectively, and calculate the cross-domain attention difference of the corresponding edges of the two domains.

[0034] In applications, such as Figure 3As shown, the cross-domain graph neural network model is a deep learning model capable of simultaneously processing two related graph structure data and capturing their differences. After inputting the pre-leaching and post-leaching association graphs into the model, the model utilizes a graph attention mechanism to learn high-dimensional embedding feature vectors for each node in both domains. These embedding feature vectors encode the node's own attributes and its neighborhood structure information. The calculation of cross-domain attention differences involves comparing the attention weights of corresponding edges in the two domains to quantify the change in the association strength of the same element-mineral association pair before and after leaching.

[0035] The CD-GAT model consists of a pre-wash image attention sublayer, a post-wash image attention sublayer, a cross-domain attention head, and an output layer, with the following specific structure: Graph Attention Sublayer: Employs a multi-head attention mechanism, with each sublayer containing four attention heads to learn the high-dimensional embedding features of the corresponding nodes in the graph. For the graph attention sublayer before rinsing, the nodes... The formula for calculating the embedding features is: ,in To focus on the number of heads, For the image Middle node For nodes Attention weights, through Function normalization yields: In the formula, For attention vectors, For the first The weight matrix of the layer, This represents a vector concatenation operation. For nodes The calculation method for the attention sublayer after rinsing is the same as before rinsing, except that the parameters are independent.

[0036] Cross-domain attention head: This module aims to capture the dynamic changes in node embedding features before and after rinsing, specifically including three stages: feature concatenation, weight calculation, and difference quantization. First, the embedded features of the nodes before and after rinsing are concatenated: Then, cross-domain attention weights are calculated through a fully connected layer to reflect the cross-domain sensitivity of the node's association pair: ,in for Activation function; finally, quantify cross-domain differences: , i.e., diagram and The difference in attention weights between corresponding edges is multiplied by the cross-domain attention weight.

[0037] Objective function: The model is trained using a joint loss function, including node embedding reconstruction loss and cross-domain difference loss. ,in To reconstruct the loss, This is the decoder (fully connected layer), used to reconstruct the node embedded features into the original input features; For cross-domain difference loss, adopt Kullback-Leibler Divergence measures the similarity of the distribution of differences before and after rinsing; The balancing coefficient is used to balance the two losses, and its value ranges from 0.1 to 0.5. In this embodiment, the optimal value is determined through cross-validation.

[0038] For the above model, model training is performed: using... The optimizer updates its parameters, setting the learning rate to 0.001~0.01 and the number of iterations to 500~1000. The optimization proceeds when the loss function converges (the change in loss is less than 0.01 over 50 consecutive iterations). Training stops when the model is fully trained. After training is complete, the model outputs the basic correlation matrix and the cross-domain difference matrix (influence coefficient matrix).

[0039] Step S104: Calculate the element-mineral basic correlation matrix based on the node embedding features, and calculate the influence coefficient matrix based on the cross-domain attention difference; the basic correlation matrix represents the strength of the occurrence relationship between elements and minerals, and the influence coefficient matrix represents the degree of influence of the leaching process on the correlation strength of each element-mineral.

[0040] In application, the basic correlation matrix is ​​calculated based on node embedding features. Cosine similarity and other metrics are used to evaluate the similarity between element nodes and mineral nodes in the embedding space. Higher similarity indicates a closer relationship between the two. The influence coefficient matrix is ​​obtained by normalizing the cross-domain attention difference values. Its element values ​​are between 0 and 1. The closer the value is to 1, the more significant the influence of the leaching process on the element-mineral association pair. This association pair can be used as a key focus for process optimization.

[0041] The basic correlation matrix is ​​constructed by calculating the cosine similarity between the embedded features of element nodes and mineral nodes. , The value range is The closer the value is to 1, the stronger the positive correlation (dominant relationship) between the mineral and the element; the closer the value is to -1, the weaker the correlation or the negative correlation.

[0042] Cross-domain difference matrix (influence coefficient matrix): This matrix represents cross-domain differences. Normalization to Intervals, constructing an influence coefficient matrix ,element A larger value indicates a more significant impact of the leaching process on the correlation between the element and the mineral, suggesting that this correlation pair is a key target for optimizing the leaching process. For example... Figure 5 As shown, the influence of the rinsing process on the elemental composition is [0,1], and the larger the value, the stronger the influence.

[0043] Specifically, in the application scenario of phosphogypsum analysis in this scheme, the phosphogypsum samples before and after leaching are first analyzed using an automated mineralogical system to obtain the content data of major elements such as calcium, sulfur, phosphorus, fluorine, silicon, and aluminum, as well as the composition data of mineral phases such as gypsum, fluorapatite, quartz, and feldspar. A quality contribution matrix is ​​constructed based on the occurrence ratio of elements in each mineral, and then the matrix is ​​transformed into a graph structure and input into a pre-trained cross-domain graph neural network model. The model extracts node features through multi-layer graph attention calculations and captures the difference information before and after leaching through a cross-domain attention head. The final output basic correlation matrix reveals which minerals are the main carriers of specific elements, while the influence coefficient matrix indicates which element-mineral association pairs have changed significantly during the leaching process.

[0044] The elemental distribution data can be acquired using an automated mineralogical system combining scanning electron microscopy and energy dispersive spectroscopy. This system can automatically identify mineral phases and quantitatively analyze elemental distribution. Alternatively, a combined detection scheme using X-ray fluorescence spectrometry and X-ray diffraction can be employed to acquire elemental content and mineral phase information separately before data fusion. The training of the cross-domain graph neural network model can be performed end-to-end using a joint loss function, or a two-stage training strategy of pre-training and fine-tuning can be adopted, first pre-training on large-scale graph data and then fine-tuning on a phosphogypsum-specific dataset.

[0045] This embodiment constructs a dual-domain association graph structure, abstracting the element-mineral occurrence relationships of phosphogypsum into a computable graph data form, laying a data foundation for association analysis using graph neural networks. The introduction of the cross-domain graph neural network model enables the method to capture nonlinear association relationships, overcoming the limitation of traditional statistical methods that can only analyze linear relationships. By outputting the basic correlation matrix and influence coefficient matrix separately, the static occurrence relationship representation and dynamic process influence quantification are decoupled, providing clear data guidance for subsequent process optimization decisions.

[0046] In one embodiment, the nodes in the association graph include element nodes and mineral nodes, and the edges are undirected edges connecting mineral nodes and element nodes; The feature vector of the element node is the normalized value of the mass proportion of the element in the sample, and the feature vector of the mineral node is the normalized value of the mass proportion of the mineral in the sample. Wherein, when the quality contribution ratio is less than a preset threshold, the weight of the corresponding edge is set to zero.

[0047] In application, the element nodes and mineral nodes are the basic building blocks of the correlation graph. Element nodes correspond to various chemical elements detected in the phosphogypsum sample, such as calcium, sulfur, phosphorus, fluorine, silicon, aluminum, and iron; mineral nodes correspond to various mineral phases identified in the sample, such as dihydrate gypsum, fluorapatite, quartz, and potassium feldspar. The nodes are divided based on the essential differences in their physicochemical properties; elements are the basic components of matter, while minerals are solid phases with specific crystal structures and chemical compositions.

[0048] In application, the edges are undirected edges connecting mineral nodes and element nodes. This design reflects the bidirectional nature of the occurrence relationship between elements and minerals. Undirected edges mean that the association strength from element nodes to mineral nodes and from mineral nodes to element nodes is equal, which conforms to the symmetric nature of chemical occurrence relationships. No edges are established between nodes of the same type; that is, there are no direct connections between element nodes and between mineral nodes. This makes the association graph exhibit typical bipartite graph structure characteristics.

[0049] In application, the feature vector is constructed by normalizing the node attributes. For an element node, its feature vector is the mass percentage of that element in the sample divided by the sum of the mass percentages of all elements; for a mineral node, its feature vector is the mass percentage of that mineral in the sample divided by the sum of the mass percentages of all minerals. The purpose of normalization is to eliminate the influence of differences in the magnitude of content between different elements or minerals, making the subsequent graph neural network learning process more stable.

[0050] In applications, setting the weight of the corresponding edge to zero when the quality contribution ratio is less than a preset threshold is a noise filtering mechanism. The preset threshold is typically set to [value missing]. In terms of magnitude, when the mass contribution of a mineral to a certain element is below a certain threshold, it can be considered that there is no significant relationship between the two, and the corresponding edge connection is regarded as noise and discarded. This processing can reduce the complexity of the graph and reduce the interference of invalid associations on model learning.

[0051] Specifically, in the phosphogypsum analysis scenario, assuming 8 major elements and 6 major minerals are detected in the sample, the association graph contains 14 nodes. Taking the calcium node as an example, its eigenvalue is calculated as: the mass percentage of calcium divided by the sum of the mass percentages of all elements. Taking the gypsum mineral node as an example, its eigenvalue is calculated as: the mass percentage of gypsum divided by the sum of the mass percentages of all minerals. Edge establishment follows a threshold screening principle. For example, if the mass contribution of gypsum to calcium is 0.85, which is much higher than the threshold, then an edge connection is established with a weight of 0.85; if the contribution of a rare mineral to calcium is only 0.0005, which is lower than the threshold, then no edge connection is established.

[0052] The feature vector construction of the element nodes can employ a quality percentage normalization method, an atomic percentage normalization method, or a logarithmic transformation normalization method. The threshold selection of the edge weights can use a fixed threshold method (setting a uniform threshold standard) or an adaptive threshold method (dynamically determining the selection boundary based on the overall distribution characteristics of each element).

[0053] This embodiment establishes a standardized graph representation of element-mineral occurrence relationships by clearly defining the node types and edge connection rules of the association graph. The use of normalized feature vectors eliminates bias caused by differences in magnitude, enabling the model to learn the features of each node fairly. A threshold filtering mechanism effectively removes noise interference from weak associations, improving the signal-to-noise ratio of the graph structure and providing high-quality input data for subsequent graph neural network learning.

[0054] In one embodiment, the cross-domain graph neural network model includes a graph attention layer and a cross-domain attention head; the graph attention layer is used to extract node embedding features of the pre-washing association graph and the post-washing association graph, respectively; the cross-domain attention head is used to calculate the cross-domain attention difference of corresponding edges between the two domains; like Figure 4 As shown, the method by which the cross-domain attention head calculates cross-domain attention differences is as follows: The cross-domain feature vector is obtained by concatenating the node embedding features before and after rinsing. ,in For node embedding features before rinsing, This refers to the embedding features of nodes after rinsing; Calculate cross-domain attention weights using a fully connected layer: ,in It is the sigmoid activation function. For cross-domain weight matrix, For bias terms; Calculate cross-domain difference values: ,in and These are the attention weights of the corresponding edges in the association graph before and after rinsing, respectively.

[0055] In applications, the graph attention layer is one of the core components of cross-domain graph neural network models. Its function is to perform feature learning and information aggregation on the input association graph. The graph attention layer uses an attention mechanism to assign different weights to each edge, enabling each node to distinguish the importance of different neighbors when aggregating neighbor information. For the association graph before and after rinsing, the model is configured with independent graph attention layers to learn their respective node embedding representations.

[0056] In application, the cross-domain attention head is an innovative component of the model, specifically designed to capture changes in the correlation between two domains. The feature concatenation operation connects the embedded feature vectors of the same node before and after rinsing in a dimensional manner, forming a cross-domain feature vector containing information from both domains. A fully connected layer performs a linear transformation on the cross-domain feature vector, and the sigmoid activation function maps the output to the 0-1 interval, obtaining the attention weights that characterize the cross-domain sensitivity of the node.

[0057] In application, the calculation of cross-domain difference values ​​comprehensively considers both the absolute change in edge attention and the cross-domain sensitivity of nodes. First, the absolute difference in the corresponding edge attention weights before and after rinsing is calculated, then multiplied by the cross-domain attention weight of the node for weighting. This design allows association pairs that undergo significant changes at sensitive nodes to obtain higher difference values, which is beneficial for identifying key influencing points in the rinsing process.

[0058] Specifically, assuming the node embedding feature dimension is 64, then the embedding feature of node i before rinsing is a 64-dimensional vector, and the embedding feature of node i after rinsing is also a 64-dimensional vector. After concatenation, a 128-dimensional cross-domain feature vector is obtained. The cross-domain weight matrix has a dimension of 128×1, with the bias term being a scalar. After calculation by a fully connected layer and sigmoid activation, the cross-domain attention weight of the node is obtained, with a value ranging from 0 to 1. If the edge attention weight of the gypsum-calcium association pair is 0.75 before rinsing and becomes 0.60 after rinsing, and the cross-domain attention weight of the node containing this association pair is 0.8, then the cross-domain difference value is |0.75-0.60|×0.8=0.12.

[0059] The graph attention layer can be implemented using a single-head attention mechanism or a multi-head attention mechanism to enhance the model's expressive power. The construction of the cross-domain feature vectors can be achieved through vector concatenation, vector differencing, or element-wise vector multiplication to fuse information from the two domains.

[0060] This embodiment achieves fine-grained quantification of changes in correlation before and after rinsing by designing a dedicated cross-domain attention head. The combination of feature splicing and fully connected layers can adaptively learn cross-domain sensitivity, avoiding the subjectivity brought about by manually setting weights. The weighted calculation strategy for attention differences fully considers the differences in node sensitivity, making the analysis results more targeted and interpretable.

[0061] In one embodiment, the underlying correlation matrix The element calculation method is as follows: ; in mineral nodes With element nodes Cosine similarity of embedded features and The node embedding vectors learned by the GAT (Graph Attention Network) before rinsing.

[0062] In application, the basic correlation matrix is ​​a numerical matrix used to quantify the strength of the occurrence relationship between elements and minerals. The number of rows in the matrix... The number of corresponding mineral nodes and the number of columns The number of corresponding element nodes. Each element value in the matrix represents the correlation strength between the corresponding mineral-element pair, ranging from -1 to 1. The closer the value is to 1, the more similar the two are in the embedding space, i.e., the closer their relationship is; the closer the value is to -1, the less similar the two are; and the closer the value is to 0, the less significant the correlation between them.

[0063] In applications, cosine similarity is a commonly used method for measuring vector similarity. The numerator of the formula is the inner product of two embedded feature vectors, and the denominator is the product of the norms of the two vectors. The inner product reflects the degree of consistency in direction between the two vectors, and the norm normalization eliminates the influence of differences in vector length, ensuring that the final similarity value reflects only the proximity of the vector directions. This measurement method is well-suited for high-dimensional sparse vectors.

[0064] Specifically, assuming the phosphogypsum sample contains 6 major minerals and 8 major elements, the basic correlation matrix has a dimension of 6×8. Taking gypsum minerals and calcium as examples, assuming the embedding feature vector of the gypsum node is [0.8, 0.6, 0.1, ...], and the embedding feature vector of the calcium node is [0.75, 0.55, 0.15, ...]. The inner product of the two vectors is 0.8×0.75+0.6×0.55+..., and the norms of the two vectors are the square roots of the sum of squares of their respective components. The final calculated cosine similarity between gypsum and calcium is approximately 0.92, indicating a strong positive correlation between the two, which is consistent with the chemical fact that calcium is the main component element in the chemical formula of gypsum, CaSO4·2H2O.

[0065] The correlation can be calculated using cosine similarity, negative Euclidean distance mapping, or Pearson correlation coefficient. The embedded feature vector can be the node embedding features before rinsing, or the average or weighted combination of the embedded features before and after rinsing.

[0066] This embodiment establishes a quantitative characterization method for element-mineral occurrence relationships by calculating the basic correlation matrix using cosine similarity. The insensitivity of cosine similarity to vector length allows for correlation comparisons of elements and minerals at different content levels on a uniform scale. Calculating correlations based on pre-leaching embedded features reflects the intrinsic occurrence relationships of the original sample, providing a stable baseline reference for subsequent influence analysis.

[0067] In one embodiment, the influence coefficient matrix elements Normalized values ​​for cross-domain attention differences: ; in minerals With elements Cross-domain attention differences This is the maximum value among all cross-domain attention differences.

[0068] In application, the influence coefficient matrix is ​​a numerical matrix used to quantify the degree of influence of the leaching process on each element-mineral correlation pair. The matrix has the same dimensions as the basic correlation matrix, and each element value represents the strength of the influence of the leaching process on the corresponding correlation pair. Normalization maps all influence coefficient values ​​to the interval between 0 and 1, facilitating horizontal comparisons and threshold screening between different correlation pairs.

[0069] In practice, normalization is achieved by dividing each cross-domain attention difference value by the maximum of all difference values. This maximum normalization method ensures that the most significant association pairs receive a coefficient value of 1, while the coefficient values ​​of other association pairs are scaled proportionally. This process preserves the relative magnitude of the influence between different association pairs while eliminating the influence of absolute numerical magnitudes.

[0070] Specifically, assuming that in the calculated cross-domain attention difference matrix, the difference value for fluorapatite-phosphorus is 0.18, for gypsum-calcium it is 0.05, and for quartz-silicon it is 0.02, with the maximum value being 0.18, then the influence coefficient for fluorapatite-phosphorus is 0.18 / 0.18 = 1.0, for gypsum-calcium it is 0.05 / 0.18 ≈ 0.28, and for quartz-silicon it is 0.02 / 0.18 ≈ 0.11. This indicates that the rinsing process has the most significant impact on the fluorapatite-phosphorus correlation, and this correlation should be the focus of process optimization.

[0071] The normalization process can employ maximum value normalization, minimum-maximum normalization, or Z-score standardization. The threshold determination of the influence coefficient can use a fixed threshold (e.g., 0.5) for binary classification, or a clustering algorithm can be used to automatically divide the group into high-influence and low-influence groups.

[0072] This embodiment constructs an influence coefficient matrix by normalizing the maximum value, achieving a standardized and quantitative characterization of the influence of the rinsing process. The normalized coefficient values ​​have clear physical meanings: the closer the value is to 1, the more significant the influence of the correlation on the rinsing process. The standardized numerical range facilitates the setting of uniform screening thresholds, providing an operable criterion for automatically identifying process-sensitive correlation pairs.

[0073] In one embodiment, the training of the cross-domain graph neural network model employs a joint loss function, which includes a weighted sum of node embedding reconstruction loss and cross-domain difference loss; the node embedding reconstruction loss is used to constrain the representational ability of node embedding features on the original input features, and the cross-domain difference loss is used to constrain the learning of cross-domain difference distribution.

[0074] In application, the joint loss function is the objective function for model training, composed of a weighted combination of multiple loss terms. The design purpose of the joint loss function is to enable the model to optimize multiple objectives simultaneously during the learning process, avoiding overfitting or bias problems caused by a single objective. The weighting coefficients are used to balance the importance of different loss terms, and their optimal values ​​are typically determined through cross-validation.

[0075] In application, the node embedding reconstruction loss is a self-supervised learning objective. This loss maps the node embedding features back to the original input feature space through the decoder and calculates the mean squared error between the reconstructed features and the original features. The role of the reconstruction loss is to ensure that the embedded features retain the key information of the original input and prevent the loss of important node attributes during feature learning.

[0076] In application, the cross-domain difference loss is used to constrain the model to learn meaningful cross-domain difference representations. This loss typically employs a distribution distance metric such as KL divergence to measure the difference between the cross-domain difference distribution output by the model and the expected distribution. The introduction of the cross-domain difference loss enables the model to accurately capture the distribution characteristics of correlation changes before and after rinsing while ensuring reconstruction accuracy.

[0077] Specifically, the mathematical form of the joint loss function is: ,in The reconstruction loss is calculated using MSE (mean squared error). For cross-domain difference loss, KL divergence is used for calculation; λ is the balance coefficient, typically ranging from 0.1 to 0.5. The training process uses the Adam optimizer with a learning rate of 0.001 to 0.01 and 500 to 1000 iterations. The loss function converges when the change in loss over 50 consecutive iterations is less than a certain threshold. Training should be stopped when the time comes.

[0078] The reconstruction loss can be a mean squared error loss function, a cross-entropy loss function, or a contrastive learning loss function. The cross-domain difference loss can be a KL divergence loss, a JS divergence loss, or a Wasserstein distance loss.

[0079] This embodiment achieves synergistic optimization of embedding quality assurance and differential learning capability through joint loss function model training. Reconstruction loss ensures the integrity of embedded features, providing a reliable feature foundation for downstream correlation calculation; cross-domain differential loss enables the model to focus on learning rinsing-related change patterns, improving the accuracy and physical meaning of the influence coefficient matrix.

[0080] In one embodiment, the method further includes: identifying element-mineral association pairs with influence coefficients higher than a preset threshold based on the influence coefficient matrix, and generating adjustment suggestions for leaching process parameters for the identified association pairs; The rinsing process parameters include at least one of rinsing temperature, liquid-to-solid ratio, and stirring rate.

[0081] In applications, identifying high-impact correlation pairs based on the influence coefficient matrix is ​​a crucial step in transforming analysis results into guidance for process optimization. A preset threshold is used as a criterion for screening significantly influential correlation pairs. When the influence coefficient of a correlation pair exceeds this threshold, it indicates that the rinsing process is significantly sensitive to that correlation, and it should be listed as a key focus for process optimization. The threshold is usually determined based on engineering experience or statistical analysis, typically ranging from 0.5 to 0.7.

[0082] In application, generating adjustment suggestions for rinsing process parameters is a process of combining data analysis results with process knowledge. For different types of high-impact correlation pairs, targeted optimization directions for process parameters are given based on their chemical characteristics and dissolution behavior. The generation of adjustment suggestions can be based on a pre-set rule base or on knowledge extraction from historical optimization cases.

[0083] In application, the leaching process parameters include key control variables such as leaching temperature, liquid-to-solid ratio, and stirring rate. Leaching temperature affects the dissolution rate and solubility of impurity minerals; the liquid-to-solid ratio affects the dissolution driving force and mass transfer efficiency; and the stirring rate affects the solid-liquid contact degree and diffusion rate. Different high-impact correlation pairs may correspond to different optimal parameter adjustment strategies.

[0084] Specifically, if the influence coefficient matrix shows that the coefficient of the fluorapatite-phosphorus correlation pair is 0.85, exceeding the preset threshold of 0.5, it is identified as a high-influence correlation pair. Based on the solubility characteristics of fluorapatite, its solubility increases under acidic conditions and the dissolution rate accelerates at high temperatures. The corresponding process adjustment recommendations are: appropriately reduce the stirring speed to 50 to 100 rpm to reduce crystal breakage, and increase the liquid-solid ratio to 2:1 to 3:1 to improve dissolution efficiency. If the coefficient of the potassium feldspar-potassium correlation pair is 0.72, it is recommended to extend the rinsing time to 10 to 15 minutes and increase the stirring speed to 150 to 200 rpm to enhance the mass transfer process.

[0085] The identification of high-impact correlation pairs can employ a fixed threshold screening method, or an adaptive threshold method that automatically determines the screening boundary based on the distribution of influence coefficients. The generation of process adjustment suggestions can utilize a rule-based expert system approach, or a case-based intelligent decision-making method.

[0086] In applications, based on the influence coefficient matrix Identify high-impact association pairs (i.e.) This involves combining prior chemical knowledge to generate process optimization strategies. The specific optimization logic in this embodiment is as follows: Scenario 1: Phosphorus content control, what is the correlation coefficient between "phosphorus-containing gypsum and phosphorus element"? The high concentration indicates that the rinsing process significantly affects the correlation between the two. To reduce the phosphorus content in the product, the stirring speed can be reduced (50~100 r / min) to decrease the shear force on the crystals, or the liquid-solid ratio can be increased (2:1~3:1) to improve the dissolution and diffusion rate of phosphorus-containing minerals such as apatite. Scenario 2: Removal of fluorine and potassium impurities, and the influence coefficient of the "potassium feldspar / fluorite - F / K element" correlation. The high level indicates that the current process is sensitive to impurity removal. To enhance impurity removal, the first-stage rinsing time can be extended (10-15 min), or the stirring rate can be increased (150-200 r / min) to strengthen the mass transfer process and promote the dissolution and dissociation of impurity minerals such as KF.

[0087] This embodiment achieves a closed-loop application from data analysis to process guidance by mapping the influence coefficient matrix to process parameter adjustments. The automatic identification of high-influence correlation pairs avoids the tedious manual investigation and improves analysis efficiency; targeted process recommendations provide clear adjustment directions for rinsing parameter optimization, possessing significant engineering practical value.

[0088] In one embodiment, the elemental occurrence data is acquired through an automated mineralogy system, which is based on a combination of scanning electron microscopy and energy dispersive spectroscopy to output chemical elemental composition, mineral composition, and the occurrence state of elements in each mineral.

[0089] In application, the automated mineralogical system is a highly automated mineral analysis device capable of rapid and comprehensive mineralogical characterization of samples. The system uses a scanning electron microscope (SEM) as the imaging platform to acquire high-resolution micro-area morphological images of the sample; and an energy dispersive spectroscopy (EDS) instrument as the analytical probe to acquire elemental composition information for each micro-area. By combining morphological and elemental data, the system can automatically identify mineral phases and statistically analyze their content distribution.

[0090] In applications, the combined use of scanning electron microscopy (SEM) and energy dispersive spectroscopy (EDS) is the mainstream technique in modern mineralogical analysis. SEM uses an electron beam to scan the sample surface and employs backscattered or secondary electron imaging to resolve submicron-sized mineral particles; EDS detects characteristic X-rays to qualitatively and quantitatively analyze the elemental composition at various points. The combined use of these two techniques allows for simultaneous morphological observation and compositional analysis.

[0091] In application, the output data includes three types of information: chemical elemental composition, mineral composition, and element occurrence state. The chemical elemental composition data provides the total content of each element in the sample; the mineral composition data provides the content ratio of each mineral phase in the sample; and the element occurrence state data provides the distribution of each element in different minerals, which is the direct data source for constructing the quality contribution matrix.

[0092] Specifically, the operational procedure for testing phosphogypsum samples using an automated mineralogical system is as follows: First, the sample is ground and sieved to prepare a resin-embedded sample, which is then polished. Next, the sample is placed on the scanning electron microscope stage, and parameters such as accelerating voltage and beam current are set. The automatic scanning program is started, and the system acquires energy dispersive spectroscopy (EDS) data point-by-point along a preset path. Finally, the software automatically performs mineral identification and statistical analysis, and outputs a test report. Typical output results include: mineral composition data such as 85% dihydrate gypsum, 5% fluorapatite, and 3% quartz; elemental composition data such as 23% calcium, 18% sulfur, and 1.2% phosphorus; and occurrence state data such as the proportion of phosphorus in fluorapatite being 78%.

[0093] The automated mineralogy system can be the AMICS system, which has a high degree of automation and batch analysis capabilities, or it can be the QEMSCAN system or MLA system, both of which have automatic mineral identification and elemental quantitative analysis functions. Sample preparation can employ resin cold embedding or pellet preparation to accommodate different types of sample characteristics.

[0094] The following example illustrates this embodiment using a complete process for obtaining the chemical elemental composition, mineral composition, and occurrence state of major elements in phosphogypsum. The process includes the following steps: Step 1: Dry the phosphogypsum sample at 55℃ for 24 hours to prevent the high temperature from causing the dihydrate gypsum (CaSO4・2H2O) to dehydrate into hemihydrate gypsum (CaSO4・0.5H2O). Grind the dried phosphogypsum sample and pass it through a 200-mesh sieve to obtain the sample to be tested. Step 2: Evenly spread the ground and sieved mineral particles onto a glass slide, and fix them with epoxy resin (EpoThin). Use a low-shrinkage epoxy resin (such as Buehler Epofix) mixed at a resin:hardener ratio of 10:1. Degas under vacuum for 5 minutes (pressure ≤0.05MPa) to eliminate air bubbles. Embed the sample into the resin mold, ensuring the surface is flush with the top of the resin. Cure in a 60℃ oven for 4 hours to avoid mineral peeling caused by room temperature curing shrinkage. After complete curing, polish until the particles are exposed. Use a vibratory polisher (Buehler VibroMet 2) with an amplitude of 20μm, a frequency of 60Hz, and a treatment time of 8 hours to eliminate mechanical stress.

[0095] Step 3: Acquire SEM images from multiple angles (0°, 45°, 90°) to verify the influence of mineral cleavage planes on EDS signals and ensure that the elemental quantification error is ≤5%. A blank resin block needs to be prepared for each batch of samples as a background control to remove interference from elements such as Si and O in the resin.

[0096] Bruker ESPRIT software was used to perform a full-spectrum comparison between the measured energy dispersive spectroscopy (EDS) and a mineral standard library (containing 2000+ minerals). The matching threshold was set to ≥85%, and minerals with a matching degree <70% required further confirmation using XRD. The ZAF correction algorithm was used to calculate the content of major elements in the minerals, ensuring that the error for major elements (CaO+SO3) was ≤2%, and the error for trace elements (P2O5, F) was ≤2%. - The error is ≤10%. Finally, the chemical element composition, mineral composition, and occurrence state of major elements in phosphogypsum are obtained.

[0097] This embodiment employs an automated mineralogical system to acquire elemental occurrence data, ensuring the accuracy and comprehensiveness of the input data. The combined use of scanning electron microscopy and energy dispersive spectroscopy (EDS) provides both morphological and compositional information, achieving integrated mineral identification and elemental analysis. The automated data acquisition and processing workflow reduces human error and improves the consistency and repeatability of the analytical results.

[0098] The above method will be explained in detail below with a complete example of phosphogypsum analysis: I. Sample Preparation and Data Acquisition Samples of phosphogypsum were collected from a phosphate fertilizer company. 500g of the original sample before leaching and 500g of the sample after standard leaching were taken. The samples were dried at 55℃ for 24 hours, then ground and passed through a 200-mesh sieve to obtain the powder to be tested. The powder was then prepared into resin-embedded samples, polished, and analyzed using an automated mineralogical system.

[0099] The test results showed that the sample before leaching mainly contained minerals such as gypsum dihydrate (86.5%), fluorapatite (4.8%), quartz (3.2%), and potassium feldspar (2.1%); the main elements included calcium (23.2%), sulfur (18.5%), phosphorus (1.15%), fluorine (0.42%), and silicon (1.8%). After leaching, the mineral composition and element content of the sample changed to some extent, with the fluorapatite content decreasing to 3.1% and the phosphorus content decreasing to 0.72%.

[0100] II. Connection Graph Construction and Model Reasoning The mass contribution ratio of each mineral to each element was calculated based on the detection data. For example, gypsum dihydrate contributes 0.89 to calcium, and fluorapatite contributes 0.78 to phosphorus. The contribution ratio matrix was transformed into an association graph structure containing 10 element nodes and 6 mineral nodes. After threshold filtering, 42 valid edges were retained.

[0101] The association graphs of the two domains before and after rinsing are input into a pre-trained cross-domain graph neural network model. The model contains two graph attention layers with an embedding dimension of 64. A joint loss function with a balance coefficient λ=0.3 is used during training. The model outputs the embedding feature vectors of each node and the cross-domain attention difference values ​​of each edge.

[0102] III. Results Analysis and Process Recommendations The basic correlation matrix shows that the correlation coefficients between gypsum dihydrate and calcium and sulfur are 0.95 and 0.92, respectively, which is consistent with its chemical formula characteristics; the correlation coefficients between fluorapatite and phosphorus, fluorine and calcium are 0.88, 0.76 and 0.65, respectively, reflecting its occurrence characteristics as the main carrier mineral of these elements.

[0103] The influence coefficient matrix shows that the influence coefficient of the fluorapatite-phosphorus correlation pair is 0.92, and the influence coefficient of the fluorapatite-fluorine correlation pair is 0.78, indicating that the leaching process has the most significant impact on these two correlation pairs. The influence coefficient of the gypsum-calcium correlation pair is only 0.15, indicating that the leaching process has a relatively small impact on the main minerals.

[0104] For the high-impact correlation pair, the system generates the following process optimization suggestions: To further reduce the phosphorus content in the product, it is recommended to increase the liquid-to-solid ratio from the current 1.5:1 to 2.5:1, extend the rinsing time to 12 minutes, and reduce the stirring rate to 80 rpm. It is expected that the phosphorus content can be further reduced by 25% to 30% after optimization.

[0105] It should be understood that the sequence number of each step in the above embodiments does not imply 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 this application.

[0106] This application also provides a device for analyzing the elemental-mineral correlation and leaching difference of phosphogypsum, used to perform the steps in the above-described embodiments of the method for analyzing the elemental-mineral correlation and leaching difference of phosphogypsum. The device for analyzing the elemental-mineral correlation and leaching difference of phosphogypsum can be a virtual appliance in an electronic device, run by the processor of the electronic device, or it can be the electronic device itself.

[0107] like Figure 6 As shown, the phosphogypsum element-mineral correlation and leaching difference analysis device 100 provided in this application embodiment includes: The data acquisition and matrix construction module 101 is used to acquire the elemental distribution data of the phosphogypsum sample before and after leaching, calculate the mass contribution ratio of each mineral to each element based on the elemental distribution data, and construct the mineral-element correlation matrix before leaching and the mineral-element correlation matrix after leaching, respectively. The association graph construction module 102 is used to construct an association graph before leaching and an association graph after leaching based on the mineral-element association matrix before leaching and the mineral-element association matrix after leaching, respectively; the association graph uses elements and minerals as nodes and the quality contribution ratio as edge weights. The feature extraction and difference calculation module 103 is used to input the pre-rinsing association graph and the post-rinsing association graph into the cross-domain graph neural network model, extract the embedding features of the nodes in the two domains respectively, and calculate the cross-domain attention difference of the corresponding edges in the two domains. The correlation and influence coefficient output module 104 is used to calculate the element-mineral basic correlation matrix based on the node embedding features and to calculate the influence coefficient matrix based on the cross-domain attention difference; the basic correlation matrix represents the strength of the occurrence relationship between elements and minerals, and the influence coefficient matrix represents the degree of influence of the leaching process on the correlation strength of each element-mineral.

[0108] In application, the data acquisition and matrix construction module serves as the data entry point for the entire device, responsible for interfacing with external detection equipment and preprocessing the raw data. This module receives the detection report output by the automated mineralogical system, parses and extracts the elemental content, mineral content, and elemental occurrence state data, and generates a mass contribution ratio matrix according to a predefined calculation formula. The module outputs a standardized matrix data format for subsequent modules to use.

[0109] In the application, the association graph construction module is responsible for converting matrix data into graph-structured data. This module creates nodes based on the row and column indices of the matrix, creates weighted edges based on the matrix element values, and applies a threshold filtering strategy to remove weakly associated edges. The module outputs a graph data object that meets the input requirements of a graph neural network, including a node feature matrix, an adjacency matrix, and edge weight vectors.

[0110] In application, the feature extraction and difference calculation module is the core computing unit of the device, embedding a pre-trained cross-domain graph neural network model. This module receives two-domain association graph data, calls the model for forward inference, and outputs node embedding features and cross-domain attention difference values. The module's computation process can be accelerated on a GPU to meet the efficiency requirements of large-scale sample analysis.

[0111] In the application, the correlation and influence coefficient output module is responsible for transforming the model output into the final analysis results. This module calculates the cosine similarity matrix based on the embedded features, calculates the normalized influence coefficient matrix based on the cross-domain difference values, and outputs the two matrices in an easy-to-read and use format, such as a heatmap visualization or a data table.

[0112] Specifically, the aforementioned functional modules can be deployed on the same computing platform, forming an end-to-end analysis pipeline. The data acquisition module connects to the AMICS system via a data interface to automatically import detection data; the association graph construction module calls a graph computing library to generate the graph structure; the feature extraction module calls a deep learning framework to perform model inference; and the output module calls a visualization library to generate results for display. Data is transferred between modules using standard data formats, ensuring the system's modularity and scalability.

[0113] The data acquisition module can read the test report using file parsing or directly obtain data from the instrument database using a database interface. The feature extraction module can be implemented using the PyTorch deep learning framework, or using the TensorFlow framework or an ONNX-based inference engine.

[0114] This embodiment achieves an automated workflow for element-mineral correlation and leaching difference analysis of phosphogypsum through a modular device architecture design. Each functional module has clearly defined responsibilities and standardized interfaces, facilitating independent development and maintenance. The end-to-end processing reduces manual intervention and improves analytical efficiency. The modular design also facilitates subsequent functional expansion, such as adding new data source interfaces or replacing / upgrading model versions.

[0115] This application also provides an electronic device, including: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor, wherein the processor executes the computer program to implement the steps in the above-described method embodiments.

[0116] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps described in the various method embodiments above.

[0117] This application provides a computer program product, including a computer program, which, when run on an electronic device, enables the electronic device to perform the steps described in the various method embodiments above.

[0118] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0119] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0120] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0121] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0122] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for analyzing the element-mineral correlation and leaching differences in phosphogypsum, characterized in that, include: Elemental occurrence data of phosphogypsum samples before and after leaching were obtained. Based on the elemental occurrence data, the mass contribution ratio of each mineral to each element was calculated, and mineral-element correlation matrices before and after leaching were constructed respectively. Based on the mineral-element correlation matrix before and after leaching, a correlation graph before and after leaching are constructed respectively; the correlation graph uses elements and minerals as nodes and the quality contribution ratio as edge weights. The pre-rinsing and post-rinsing association graphs are input into a cross-domain graph neural network model, and the node embedding features of the two domains are extracted respectively. The cross-domain attention difference of the corresponding edges of the two domains is calculated. The element-mineral basic correlation matrix is ​​calculated based on the node embedding features, and the influence coefficient matrix is ​​calculated based on the cross-domain attention difference. The basic correlation matrix represents the strength of the occurrence relationship between elements and minerals, and the influence coefficient matrix represents the degree of influence of the leaching process on the correlation strength of each element-mineral.

2. The method as described in claim 1, characterized in that, In the association graph, the nodes include element nodes and mineral nodes, and the edges are undirected edges connecting mineral nodes and element nodes; The feature vector of the element node is the normalized value of the mass proportion of the element in the sample, and the feature vector of the mineral node is the normalized value of the mass proportion of the mineral in the sample. Wherein, when the quality contribution ratio is less than a preset threshold, the weight of the corresponding edge is set to zero.

3. The method as described in claim 1, characterized in that, The cross-domain graph neural network model includes a graph attention layer and a cross-domain attention head; the graph attention layer is used to extract node embedding features of the association graph before and after rinsing, respectively. The cross-domain attention head is used to calculate the cross-domain attention difference between corresponding edges in two domains; The method by which the cross-domain attention head calculates cross-domain attention differences is as follows: The cross-domain feature vector is obtained by concatenating the node embedding features before and after rinsing. ,in For node embedding features before rinsing, This refers to the embedding features of nodes after rinsing; Calculate cross-domain attention weights using a fully connected layer: Where σ is the sigmoid activation function, For cross-domain weight matrix, For bias terms; Calculate cross-domain difference values: ,in and These are the attention weights of the corresponding edges in the association graph before and after rinsing, respectively.

4. The method as described in claim 1, characterized in that, The underlying correlation matrix The element calculation method is as follows: in mineral nodes With element nodes Cosine similarity of embedded features and The node embedding vectors learned by the GAT (Graph Attention Network) before rinsing.

5. The method as described in claim 1, characterized in that, The influence coefficient matrix elements Normalized values ​​for cross-domain attention differences: ； in minerals With elements Cross-domain attention differences This is the maximum value among all cross-domain attention differences.

6. The method as described in claim 3, characterized in that, The training of the cross-domain graph neural network model adopts a joint loss function, which includes a weighted sum of node embedding reconstruction loss and cross-domain difference loss. The node embedding reconstruction loss is used to constrain the representation ability of node embedding features on the original input features, and the cross-domain difference loss is used to constrain the learning of cross-domain difference distribution.

7. The method as described in claim 1, characterized in that, Also includes: Based on the influence coefficient matrix, element-mineral correlation pairs with influence coefficients higher than a preset threshold are identified, and adjustment suggestions for leaching process parameters are generated for the identified correlation pairs. The rinsing process parameters include at least one of rinsing temperature, liquid-to-solid ratio, and stirring rate.

8. The method as described in claim 1, characterized in that, The elemental occurrence data is acquired through an automated mineralogy system, which is based on a combination of scanning electron microscopy and energy dispersive spectroscopy to output chemical elemental composition, mineral composition, and the occurrence state of elements in each mineral.

9. A device for analyzing the element-mineral correlation and leaching differences of phosphogypsum, characterized in that, include: The data acquisition and matrix construction module is used to acquire the elemental distribution data of phosphogypsum samples before and after leaching, calculate the mass contribution ratio of each mineral to each element based on the elemental distribution data, and construct the mineral-element correlation matrix before leaching and the mineral-element correlation matrix after leaching, respectively. The correlation graph construction module is used to construct a correlation graph before leaching and a correlation graph after leaching based on the mineral-element correlation matrix before leaching and the mineral-element correlation matrix after leaching, respectively; the correlation graph uses elements and minerals as nodes and the quality contribution ratio as edge weights. The feature extraction and difference calculation module is used to input the pre-rinsing association graph and the post-rinsing association graph into the cross-domain graph neural network model, extract the node embedding features of the two domains respectively, and calculate the cross-domain attention difference of the corresponding edges of the two domains. The correlation and influence coefficient output module is used to calculate the element-mineral basic correlation matrix based on the node embedding features and to calculate the influence coefficient matrix based on the cross-domain attention difference. The basic correlation matrix characterizes the strength of the occurrence relationship between elements and minerals, and the influence coefficient matrix characterizes the degree of influence of the leaching process on the correlation strength of each element-mineral.

10. An electronic device, characterized in that, The device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the computer program, the electronic device performs the method as described in any one of claims 1-8.

Images