Preparation method, system and electronic device of magnesium-carbon slide plate
The data of magnesium carbon skateboards are processed through context encoder and convolutional neural networks and the association matrix of graph neural networks is constructed, which solves the problem of component correlation being ignored in traditional methods and achieves the improvement of magnesium carbon material performance.
Patent Information
- Application Number
- CN202111455647.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-02
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2041-12-02
AI Technical Summary
Traditional methods ignore the correlation between components when determining the optimal ratio of each component of a magnesium-carbon skateboard, resulting in the inability to accurately infer the optimal ratio, affecting the material performance.
The physical performance data and microstructure diagram of magnesium carbon skateboards were processed using context encoder model and convolutional neural network. By constructing the combined and dissociating adjacency matrix, the graph neural network was used to obtain the correlation characterization vectors between components to improve the Accuracy of decoding regression.
Accurately determine the optimal proportion of each component of the material made of magnesium carbon skateboard to improve the performance of magnesium carbon materials.
Smart Images

Figure CN114141320B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of intelligent manufacturing design, and more specifically, to a preparation method, system and electronic equipment of a magnesium-carbon skateboard. Background Art
[0002] In the modern steelmaking process, especially with the development of fast and efficient continuous casting and secondary refining technology and process, the sliding nozzle system has become more and more important and has become an indispensable part of smelting. The sliding nozzle system is a molten steel control device in the casting process of the continuous casting machine. It can accurately adjust the molten steel flow from the ladle to the tundish and from the tundish to the crystallizer, so that the inflow and outflow of molten steel can be balanced, making the continuous casting operation easier to control.
[0003] The slide plate is the most important component of the sliding nozzle system. Since it directly controls the flow of molten steel, it needs to withstand the chemical erosion and physical scouring of high-temperature molten steel for a long time and repeatedly under different casting process conditions, as well as the intense and transient thermal shock and mechanical wear. The use conditions are extremely harsh. Improving the performance of the slide plate has always been one of the key contents of refractory research.
[0004] In order to eliminate the blockage of the nozzle caused by aluminum deoxidation products, the molten steel needs to be treated with calcium, but the concentration of Ca in the molten steel and CaO in the slag water increases, and they will react with the aluminum oxide component in the slide to generate low-melting-point substances, which will cause damage to the slide. Due to the excellent high temperature resistance and corrosion resistance of magnesium oxide, it can be used to manufacture slides for casting calcium-treated steel, but this slide is very susceptible to thermal shock damage. Therefore, using carbon and carbides with good thermal conductivity as the binding phase or magnesium-aluminum spinel as the binding phase can improve the poor thermal shock stability of magnesium-carbon slides to a certain extent.
[0005] At present, a large amount of aluminum powder is introduced into the magnesium-carbon slide plate to produce metal bonding, thereby improving its thermal shock resistance and strength. However, the introduction of a large amount of aluminum powder may intensify the aluminothermic reduction reaction of magnesium oxide, thereby generating a large amount of spinel and causing expansion. Therefore, it is necessary to study the changes in the structure and properties of magnesium-carbon materials with high aluminum content after treatment at different temperatures. The specific content is the influence of the addition amount of aluminum powder, magnesium sand, flake graphite, sulfide powder and carbide powder on the structure and properties of magnesium-carbon materials treated at different temperatures in air.
[0006] Traditionally, the best ratio of each component in the formula of magnesium-carbon skateboard is determined based on experiments. However, on the one hand, traditional experiments adopt the control variable method, which ignores the correlation between each component. At the same time, considering the cost of the experiment, the number of experiments is limited, and it can only vaguely draw some rules, and it cannot accurately infer the best ratio of each component. Therefore, in order to determine the best ratio of each component in the material of magnesium-carbon skateboard to improve the performance of magnesium-carbon material, a preparation method of magnesium-carbon skateboard is desired. Summary of the invention
[0007] In order to solve the above technical problems, the present application is proposed. The embodiments of the present application provide a method, system and electronic device for preparing a magnesium-carbon skateboard, which utilizes the physical performance data of the magnesium-carbon skateboard collected after changing a certain component while other components remain unchanged, as well as the microstructure diagram of the magnesium-carbon skateboard, and then processes it based on the context encoder model and the convolutional neural network model to obtain the characteristic representation of the component, and also constructs an adjacency matrix through logical operations between rules to represent the association relationship between the components, and further uses the graph neural network technology to obtain the association representation vector of each node containing the association information, so as to improve the accuracy of decoding regression. In this way, the optimal proportion of each component in the material of the magnesium-carbon skateboard can be determined, so that the performance of the magnesium-carbon material can be improved.
[0008] According to one aspect of the present application, a method for preparing a magnesium-carbon slide plate is provided, comprising:
[0009] Obtaining optimal physical property data and microstructure diagrams corresponding to each component in the material of the magnesium-carbon skateboard, wherein the material of the magnesium-carbon skateboard includes magnesia, flake graphite, carbores p, spherical asphalt, aluminum silicon alloy powder, sulfide powder, carbonized powder, thermosetting phenolic resin and carbores T60, and the physical property data include linear change rate, weight change rate, volume density, apparent porosity and room temperature compressive strength;
[0010] For the optimal physical property data corresponding to each component, each of the physical property data is passed through a word embedding layer to obtain a plurality of parameter input vectors;
[0011] Input the microstructure image corresponding to each component into a first convolutional neural network as a filter to obtain a structure input vector;
[0012] Passing the plurality of parameter input vectors and the structure input vector through a context-encoding-based encoder to obtain a sequence of encoded feature vectors;
[0013] Concatenating all the encoded feature vectors in the sequence of the encoded feature vectors to obtain feature representation vectors corresponding to the respective components;
[0014] Based on the logical operation rules of conjunction and disjunction between the various components in the material of the magnesium-carbon skateboard, a conjunction adjacency matrix and a disjunction adjacency matrix between the feature representation vectors corresponding to the various components are constructed, wherein the conjunction logical operation rule represents the parallel relationship between the rules, and the disjunction logical operation represents the replacement relationship between the rules;
[0015] Passing the conjunctive adjacency matrix and the disjunctive adjacency matrix through a second convolutional neural network to obtain a first feature matrix and a second feature matrix, wherein the first feature matrix and the conjunctive adjacency matrix have the same scale, and the second feature matrix and the disjunctive adjacency matrix have the same scale;
[0016] Calculate the first feature matrix and the second feature matrix to obtain a fused class adjacency matrix according to the position points;
[0017] Perform two-dimensional concatenation of the feature representation vectors corresponding to the components to obtain a component feature matrix;
[0018] Passing the component feature matrix and the class adjacency matrix through a graph neural network to obtain a graph structure feature matrix, wherein each row of the graph structure feature matrix is an association representation vector containing association information of a corresponding component, wherein the graph neural network is suitable for processing graph data in an irregular non-Euclidean space to extract association information of data samples due to feature information and irregular topological structure information; and
[0019] The associated representation vector of each component is passed through a decoder for regression to obtain the optimal ratio of the component.
[0020] In the above-mentioned method for preparing the magnesium-carbon skateboard, the microstructure image corresponding to each component is input into the first convolutional neural network as a filter to obtain a structural input vector, including: each layer of the first convolutional neural network except the last layer performs convolution processing on the microstructure image to obtain a structural feature image; and the last layer of the first convolutional neural network performs global average pooling based on a feature matrix on the structural feature image to obtain the structural input vector.
[0021] In the above-mentioned method for preparing the magnesium-carbon skateboard, the multiple parameter input vectors and the structural input vector are passed through an encoder based on context coding to obtain a sequence of encoded feature vectors, including: passing the multiple parameter input vectors and the structural input vector through a converter model of the encoder to obtain a sequence of encoded feature vectors.
[0022] In the above-mentioned method for preparing the magnesium-carbon skateboard, based on the logical operation rules of conjunction and disjunction between the various components in the material made of the magnesium-carbon skateboard, the conjunctive adjacency matrix and the disjunctive adjacency matrix between the feature representation vectors corresponding to the various components are constructed, including: based on the logical operation rules of conjunction between the various components in the material made of the magnesium-carbon skateboard, the conjunctive adjacency matrix between the feature representation vectors corresponding to the various components is constructed according to the following formula;
[0023] Wherein, the formula is:
[0024]
[0025] in, It is a conjunctive adjacency matrix, which is used to indicate that the corresponding pair of rules form a conjunctive normal form when the matrix position is 1, and a non-conjunctive normal form when the matrix position is 0.
[0026] In the above-mentioned method for preparing the magnesium-carbon skateboard, based on the logical operation rules of the disjunction between the various components in the material of the magnesium-carbon skateboard, the disjunctive adjacency matrix between the feature representation vectors corresponding to the various components is constructed by the following formula;
[0027] Wherein, the formula is:
[0028]
[0029] in, is the disjunctive adjacency matrix, which is used to indicate that the corresponding pair of rules form a disjunctive normal form with a matrix position of 1 and a non-disjunctive normal form with a matrix position of 0.
[0030] In the above-mentioned method for preparing the magnesium-carbon skateboard, the conjunctive adjacency matrix and the disjunctive adjacency matrix are passed through a second convolutional neural network to obtain a first feature matrix and a second feature matrix, including: each layer of the second convolutional neural network except the last layer performs explicit spatial encoding on the conjunctive adjacency matrix and the disjunctive adjacency matrix according to the following formula to obtain a first feature map and a second feature map, wherein the formula is:
[0031] f i =active(N i ×f i-1 +B i )
[0032] Among them, f i-1 is the input of the second convolutional neural network in the i-th layer, f i is the output of the second convolutional neural network in the i-th layer, N i is the filter of the second convolutional neural network in the i-th layer, and B iis the bias matrix of the i-th layer of the second convolutional neural network, active represents a nonlinear activation function; and the last layer of the second convolutional neural network performs global pooling processing on the first feature map and the second feature map along the channel dimension to obtain the first feature matrix and the second feature matrix.
[0033] In the above-mentioned method for preparing the magnesium-carbon skateboard, it further includes: setting other label values for representing the upper and lower limits of the proportion range with the optimal proportion of the certain component as the center; passing the associated characterization vector of each component through the decoder to obtain the probability that the associated feature vector belongs to other label values; and, based on the probability, determining whether other label values are feasible.
[0034] According to another aspect of the present application, a system for preparing a magnesium-carbon skateboard is provided, comprising:
[0035] A data acquisition unit, used to obtain the optimal physical property data and microstructure diagram corresponding to each component in the material of the magnesium-carbon skateboard, wherein the material of the magnesium-carbon skateboard includes magnesia, flake graphite, carbores p, spherical asphalt, aluminum silicon alloy powder, sulfide powder, carbonized powder, thermosetting phenolic resin and carbores T60, and the physical property data includes linear change rate, weight change rate, volume density, apparent porosity and room temperature compressive strength;
[0036] A word embedding unit, for obtaining a plurality of parameter input vectors by passing each of the optimal physical property data corresponding to each component obtained by the data acquisition unit through a word embedding layer;
[0037] A first convolution unit, used for inputting the microstructure image corresponding to each component obtained by the data acquisition unit into a first convolutional neural network as a filter to obtain a structure input vector;
[0038] an encoding unit, configured to pass the plurality of parameter input vectors obtained by the word embedding unit and the structure input vector obtained by the first convolution unit through an encoder based on context encoding to obtain a sequence of encoded feature vectors;
[0039] A cascading unit, configured to cascade all the coded feature vectors in the sequence of the coded feature vectors obtained by the encoding unit to obtain a feature representation vector corresponding to each component;
[0040] An adjacency matrix construction unit, used to construct a conjunction adjacency matrix and a disjunction adjacency matrix between the feature representation vectors corresponding to the components obtained by the cascade unit based on the conjunction and disjunction logic operation rules between the components in the material of the magnesium-carbon slide plate, wherein the conjunction logic operation rule represents the parallel relationship between the rules, and the disjunction logic operation represents the replacement relationship between the rules;
[0041] a second convolution unit, configured to pass the conjunctive adjacency matrix obtained by the adjacency matrix construction unit and the disjunctive adjacency matrix obtained by the adjacency matrix construction unit through a second convolutional neural network to obtain a first feature matrix and a second feature matrix, wherein the first feature matrix and the conjunctive adjacency matrix have the same scale, and the second feature matrix and the disjunctive adjacency matrix have the same scale;
[0042] A fusion unit, used for calculating a fused class adjacency matrix between the first feature matrix obtained by the second convolution unit and the second feature matrix obtained by the second convolution unit according to position points;
[0043] A two-dimensional splicing unit, used for performing two-dimensional splicing on the feature representation vectors corresponding to the components obtained by the cascade unit to obtain a component feature matrix;
[0044] A graph neural network unit, used for obtaining a graph structure feature matrix by passing the component feature matrix obtained by the two-dimensional splicing unit and the class adjacency matrix obtained by the fusion unit through a graph neural network, wherein each row of the graph structure feature matrix is an association representation vector containing association information of a corresponding component, wherein the graph neural network is suitable for processing graph data in an irregular non-Euclidean space to extract association information of data samples due to feature information and irregular topological structure information; and
[0045] A regression unit is used to pass the associated representation vector of each component obtained by the graph neural network unit through a decoder for regression to obtain the optimal proportion of the component.
[0046] In the above-mentioned magnesium-carbon skateboard preparation system, the first convolution unit is further used for: each layer of the first convolutional neural network except the last layer performs convolution processing on the microstructure image to obtain a structural feature image; and the last layer of the first convolutional neural network performs global average pooling based on the feature matrix on the structural feature image to obtain the structural input vector.
[0047] In the above-mentioned magnesium-carbon skateboard preparation system, the encoding unit is further used to: pass the multiple parameter input vectors and the structure input vector through the converter model of the encoder to obtain the sequence of the encoding feature vectors.
[0048] In the above-mentioned magnesium-carbon skateboard preparation system, the adjacency matrix construction unit is further used to: construct a conjunction adjacency matrix between the feature representation vectors corresponding to each component based on the conjunction logic operation rule between the components in the material of the magnesium-carbon skateboard using the following formula;
[0049] Wherein, the formula is:
[0050]
[0051] in, It is a conjunctive adjacency matrix, which is used to indicate that the corresponding pair of rules form a conjunctive normal form when the matrix position is 1, and a non-conjunctive normal form when the matrix position is 0.
[0052] In the preparation system of the magnesium-carbon skateboard, based on the logical operation rules of the disjunction between the components in the material of the magnesium-carbon skateboard, the disjunction adjacency matrix between the feature representation vectors corresponding to the components is constructed by the following formula;
[0053] Wherein, the formula is:
[0054]
[0055] in, is the disjunctive adjacency matrix, which is used to indicate that the corresponding pair of rules form a disjunctive normal form with a matrix position of 1 and a non-disjunctive normal form with a matrix position of 0.
[0056] In the above-mentioned magnesium-carbon skateboard preparation system, the second convolution unit is further used for: each layer of the second convolutional neural network except the last layer performs explicit spatial encoding on the conjunctive adjacency matrix and the disjunctive adjacency matrix according to the following formula to obtain the first feature map and the second feature map, wherein the formula is:
[0057] f i =active(N i ×f i-1 +B i )
[0058] Among them, f i-1 is the input of the second convolutional neural network in the i-th layer, f i is the output of the second convolutional neural network in the i-th layer, N i is the filter of the second convolutional neural network in the i-th layer, and B i is the bias matrix of the i-th layer of the second convolutional neural network, active represents a nonlinear activation function; and the last layer of the second convolutional neural network performs global pooling processing on the first feature map and the second feature map along the channel dimension to obtain the first feature matrix and the second feature matrix.
[0059] In the preparation system of the above-mentioned magnesium-carbon skateboard, it further includes: setting other label values for representing the upper and lower limits of the proportion range with the optimal proportion of the certain component as the center; passing the associated characterization vector of each component through the decoder to obtain the probability that the associated feature vector belongs to other label values; and, based on the probability, determining whether other label values are feasible.
[0060] According to another aspect of the present application, an electronic device is provided, comprising: a processor; and a memory, in which computer program instructions are stored, and when the computer program instructions are executed by the processor, the processor executes the method for preparing the magnesium-carbon skateboard as described above.
[0061] According to yet another aspect of the present application, a computer-readable medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the processor executes the method for preparing the magnesium-carbon skateboard as described above.
[0062] Compared with the prior art, the preparation method, system and electronic device of the magnesium-carbon skateboard provided in the present application utilize the physical performance data of the magnesium-carbon skateboard collected after changing a certain component while other components remain unchanged, as well as the microstructure diagram of the magnesium-carbon skateboard, and then process it based on the context encoder model and the convolutional neural network model to obtain the characteristic representation of the component, and also construct an adjacency matrix through logical operations between rules to represent the association relationship between the various components, and further use the graph neural network technology to obtain the association representation vector of each node containing the association information to improve the accuracy of decoding regression. In this way, the optimal proportion of each component in the material of the magnesium-carbon skateboard can be determined, so that the performance of the magnesium-carbon material can be improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0063] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.
[0064] Figure 1 is a flow chart of a method for preparing a magnesium-carbon slide plate according to an embodiment of the present application;
[0065] Figure 2 A schematic diagram of the system architecture of a method for preparing a magnesium-carbon skateboard according to an embodiment of the present application;
[0066] Figure 3 is a block diagram of a system for preparing a magnesium-carbon skateboard according to an embodiment of the present application;
[0067] Figure 4 is a block diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION
[0068] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.
[0069] Scenario Overview
[0070] As mentioned above, in order to eliminate the nozzle blockage caused by aluminum deoxidation products, the molten steel needs to be treated with calcium, but the concentration of Ca in the molten steel and CaO in the slag water increases, and they will react with the aluminum oxide component in the slide to generate low-melting-point substances, which will cause damage to the slide. Due to the excellent high temperature resistance and corrosion resistance of magnesium oxide, it can be used to manufacture slides for casting calcium-treated steel, but such slides are easily damaged by thermal shock. Therefore, using carbon and carbides with good thermal conductivity as the binding phase or magnesium-aluminum spinel as the binding phase can improve the poor thermal shock stability of magnesium-carbon slides to a certain extent.
[0071] At present, a large amount of aluminum powder is introduced into the magnesium-carbon slide plate to produce metal bonding, thereby improving its thermal shock resistance and strength. However, the introduction of a large amount of aluminum powder may intensify the aluminothermic reduction reaction of magnesium oxide, thereby generating a large amount of spinel and causing expansion. Therefore, it is necessary to study the changes in the structure and properties of magnesium-carbon materials with high aluminum content after treatment at different temperatures. The specific content is the influence of the addition amount of aluminum powder, magnesium sand, flake graphite, sulfide powder and carbide powder on the structure and properties of magnesium-carbon materials treated at different temperatures in air.
[0072] Accordingly, in the technical solution of the present application, it is obviously impractical to determine the optimal ratio or to make a classification result that the ratio of a certain component should be adjusted by the control variable method, and there is mutual influence between the various components, rather than an independent relationship. Therefore, the ratio adjustment result or the optimal ratio result can be generated based on the neural network model.
[0073] Specifically, in the technical solution of the present application, the experimental data are the physical performance data of the magnesium-carbon skateboard collected after changing a certain component while keeping other components unchanged (including type and weight), that is, the linear change rate, weight change rate, volume density, apparent porosity and room temperature compressive strength, as well as the microstructure diagram of the magnesium-carbon skateboard. However, since the components are not independent but have correlations with each other, how to obtain the correlation information between the components becomes the key.
[0074] Based on this, the applicant of this application considered that the data corresponding to the above-mentioned single component can be used to obtain the characteristic representation of the component through a deep neural network model. Therefore, if an adjacency matrix used to represent the association relationship between the components can be further obtained, the graph neural network technology can be used to obtain the association representation vector of each node containing the association information.
[0075] Here, the components obviously do not have node topology information represented by nodes, so the ordinary adjacency matrix cannot be used. However, considering that each component can actually be expressed in a regular form, for example, the formula ratio of large crystalline magnesia sand 98 with a specification of 3-1mm is 25%-32%, the adjacency matrix can be constructed through logical operations between rules.
[0076] Specifically, the logical operations between rules usually include conjunction and disjunction, which are represented by the symbols ∧ and ∨, respectively, and are used to represent the parallel or alternative relationship between rules, that is, the meaning of "and" and "or". For the components mentioned above, this relationship also exists. For example, "the formula ratio of large crystalline magnesia sand 98 with a specification of 3-1mm is 25%-32%" and "the formula ratio of large crystalline magnesia sand 98 with a specification of 200 mesh is 24%-28%" is an "or" relationship, while "the formula ratio of large crystalline magnesia sand 98 with a specification of 3-1mm is 25%-32%" and "the formula ratio of flake graphite 98 with a specification of -100 mesh is 1%-1.8" is an "and" relationship. Therefore, discrete logical nodes can be constructed for the rules satisfied by each component based on the above method, thereby forming an adjacency matrix, which can be expressed as:
[0077]
[0078]
[0079] here, is a conjunctive adjacency matrix, which is used to indicate that the corresponding pair of rules form a conjunctive normal form when the matrix position is 1, and a non-conjunctive normal form when the matrix position is 0. is the disjunctive adjacency matrix, which is used to indicate that the corresponding pair of rules form a disjunctive normal form with a matrix position of 1 and a non-disjunctive normal form with a matrix position of 0.
[0080] Then, the conjunctive adjacency matrix and the disjunctive adjacency matrix are mapped to high-dimensional space through convolutional neural networks and fused respectively to obtain the class adjacency matrix corresponding to the feature representation of each node.
[0081] Therefore, in the technical solution of the present application, the optimal physical performance data of a single component is first obtained, that is, the optimal physical performance data obtained by changing the proportion of a certain component while keeping other components unchanged (including type and weight), such as linear change rate, weight change rate, volume density, apparent porosity and room temperature compressive strength, and the corresponding microstructure diagram. And in order to be able to express the correlation between these parameters, a context encoder model is used for context encoding, that is, each physical parameter is obtained by word embedding to obtain a parameter input vector, and the microstructure diagram is obtained by a first convolutional neural network used as a filter to obtain a structural input vector, and then the parameter input vector and the structural input vector are input into the encoder together to obtain a sequence of encoded feature vectors, and each encoded feature vector is cascaded to obtain a feature representation vector of a single component.
[0082] Then, according to the above method, the conjunction adjacency matrix and the disjunction adjacency matrix are passed through the second convolutional neural network to obtain the first feature matrix and the second feature matrix, and the dimension of the feature matrix is the same as the original input matrix. Then, the points of the first feature matrix and the second feature matrix are calculated to obtain the fused class adjacency matrix.
[0083] After that, the feature representation vectors of each component are concatenated in two dimensions to obtain a component feature matrix, and then the graph structure feature matrix is obtained together with the class adjacency matrix through the graph neural network. Each row of the graph structure feature matrix is the association representation vector of the corresponding component containing association information. Here, the graph neural network can be used to process graph data in irregular non-Euclidean space, thereby extracting the association information of the data sample due to the feature information and irregular topological structure information. Therefore, the association representation vector of the obtained single component can improve the accuracy of decoding regression compared to the feature representation vector obtained by directly concatenating.
[0084] Finally, the associated characterization vector of each component is passed through the decoder for regression to obtain the optimal ratio of the component. In addition, after obtaining the optimal ratio of the component, other label values used to represent the upper and lower limits of the ratio range can be set with the optimal ratio as the center. For example, if the optimal ratio is 30%, the label values can be set to 27%, 28%, 29%, 31%, 32% and 33%, and then the probability of other label values belonging to the label values is obtained through decoding regression. For example, the label probabilities of 28%, 29%, 31%, and 32% are greater than the predetermined threshold, so that a feasible ratio range of 28%-32% is obtained.
[0085] Based on this, the present application proposes a method for preparing a magnesium-carbon skateboard, which includes: obtaining the optimal physical property data and microstructure diagram corresponding to each component in the material of the magnesium-carbon skateboard, wherein the material of the magnesium-carbon skateboard includes magnesia, flake graphite, carboresp, spherical asphalt, aluminum-silicon alloy powder, sulfide powder, carbonized powder, thermosetting phenolic resin and carboresT60, and the physical property data includes linear change rate, weight change rate, volume density, apparent porosity and room temperature compressive strength; for the optimal physical property data corresponding to each component, each of the physical property data is passed through a word embedding layer to obtain a plurality of parameter input vectors; the microstructure diagram corresponding to each component is input into a first convolutional neural network as a filter to obtain a structural input vector; the plurality of parameter input vectors and the structural input vector are passed through an encoder based on context encoding to obtain a sequence of encoded feature vectors; all the encoded feature vectors in the sequence of the encoded feature vectors are cascaded to obtain the feature representation vector corresponding to each component; the feature representation vector corresponding to each component is constructed based on the logical operation rules of conjunction and disjunction between the components in the material of the magnesium-carbon skateboard. A conjunctive adjacency matrix and a disjunctive adjacency matrix between vectors, wherein the conjunctive logic operation rule represents the parallel relationship between the rules, and the disjunctive logic operation represents the replacement relationship between the rules; the conjunctive adjacency matrix and the disjunctive adjacency matrix are passed through a second convolutional neural network to obtain a first feature matrix and a second feature matrix, wherein the first feature matrix has the same scale as the conjunctive adjacency matrix, and the second feature matrix has the same scale as the disjunctive adjacency matrix; the first feature matrix and the second feature matrix are calculated by position points to obtain a fused class adjacency matrix; the feature representation vectors corresponding to the components are two-dimensionally spliced to obtain a component feature matrix; the component feature matrix and the class adjacency matrix are passed through a graph neural network to obtain a graph structure feature matrix, wherein each row of the graph structure feature matrix is an association representation vector containing association information of the corresponding component, wherein the graph neural network is suitable for processing graph data in an irregular non-Euclidean space to extract association information of the data sample due to feature information and irregular topological structure information; and the association representation vector of each component is passed through a decoder for regression to obtain the optimal proportion of the component.
[0086] Figure 1 The figure shows an application scenario diagram of the method for preparing a magnesium-carbon slide plate according to an embodiment of the present application. Figure 1 As shown, in this application scenario, first, the material of the magnesium-carbon slide plate (for example, Figure 1 The optimal physical property data corresponding to each component can be obtained by using an electron microscope camera (e.g., Figure 1C) to obtain its corresponding microstructure diagram, wherein the materials made of the magnesium-carbon skateboard include magnesia, flake graphite, carbores p, spherical asphalt, aluminum silicon alloy powder, sulfide powder, carbonized powder, thermosetting phenolic resin and carbores T60, and the physical property data include linear change rate, weight change rate, volume density, apparent porosity and room temperature compressive strength. Then, the obtained physical property data and the microstructure diagram are input into a server deployed with a preparation algorithm for the magnesium-carbon skateboard (for example, Figure 1 S) as shown in the figure, wherein the server can process the physical property data and the microstructure diagram with the preparation algorithm of the magnesium-carbon skateboard to obtain the optimal ratio of the components. Furthermore, other label values for indicating the upper and lower limits of the ratio range can be set with the optimal ratio as the center. For example, the label values of 27%, 28%, 29%, 31%, 32% and 33% can be set for the optimal ratio of 30%, and then the probability of the other label values being attributed is obtained by decoding regression. For example, the probability of the labels of 28%, 29%, 31%, and 32% is greater than the predetermined threshold, thus obtaining a feasible ratio range of 28%-32%. In this way, the performance of the magnesium-carbon material can be improved by adding the components in proportion according to the obtained optimal ratio range.
[0087] After introducing the basic principles of the present application, various non-limiting embodiments of the present application will be described in detail with reference to the accompanying drawings.
[0088] Exemplary Methods
[0089] Figure 1 The figure shows a flow chart of the method for preparing a magnesium-carbon slide plate. Figure 1As shown, the preparation method of the magnesium-carbon skateboard according to the embodiment of the present application includes: S110, obtaining the optimal physical property data and microstructure diagram corresponding to each component in the material of the magnesium-carbon skateboard, wherein the material of the magnesium-carbon skateboard includes magnesia, flake graphite, carbores p, spherical asphalt, aluminum silicon alloy powder, sulfide powder, carbonized powder, thermosetting phenolic resin and carbores T60, the physical property data include linear change rate, weight change rate, volume density, apparent porosity and room temperature compressive strength; S120, for the optimal physical property data corresponding to each component, each of the physical property data is passed through a word embedding layer to obtain a plurality of parameter input vectors; S130, the microstructure diagram corresponding to each component is input into a first convolutional neural network as a filter to obtain a structure input vector; S140, the plurality of parameter input vectors and the structure input vector are passed through an encoder based on context encoding to obtain a sequence of encoded feature vectors; S150, all the encoded feature vectors in the sequence of encoded feature vectors are cascaded to obtain the feature representation vectors corresponding to each component; S160, based on the logical operation rules of conjunction and disjunction between the components in the material made of the magnesium-carbon skateboard, the conjunctive adjacency matrix and the disjunctive adjacency matrix between the feature representation vectors corresponding to the components are constructed, wherein the conjunctive logical operation rule represents the parallel relationship between the rules, and the disjunctive logical operation represents the replacement relationship between the rules; S170 , passing the conjunctive adjacency matrix and the disjunctive adjacency matrix through a second convolutional neural network to obtain a first feature matrix and a second feature matrix, wherein the first feature matrix has the same scale as the conjunctive adjacency matrix, and the second feature matrix has the same scale as the disjunctive adjacency matrix; S180, calculating the first feature matrix and the second feature matrix according to the position points to obtain a fused class adjacency matrix; S190, performing two-dimensional splicing on the feature representation vectors corresponding to the components to obtain a component feature matrix; S200, passing the component feature matrix and the class adjacency matrix through a graph neural network to obtain a graph structure feature matrix, wherein each row of the graph structure feature matrix is an association representation vector containing association information of the corresponding component, wherein the graph neural network is suitable for processing graph data in an irregular non-Euclidean space to extract association information of the data sample due to feature information and irregular topological structure information; and, S210, passing the association representation vector of each component through a decoder for regression to obtain the optimal proportion of the component.
[0090] Figure 2 The figure shows a schematic diagram of the structure of the method for preparing the magnesium-carbon slide plate according to an embodiment of the present application. Figure 2 As shown, in the network architecture of the preparation method of the magnesium-carbon slide plate, first, the obtained physical property data (for example, Figure 2P1 as shown in ) through a word embedding layer (e.g., Figure 2 WEL as shown in the figure) to obtain multiple parameter input vectors (for example, Figure 2 Then, the microstructure diagram corresponding to each component obtained (for example, Figure 2 P2) as input to the first convolutional neural network (e.g., as shown in Figure 2 ) to obtain a structure input vector (e.g., Figure 2 Then, the plurality of parameter input vectors and the structure input vector are passed through a context-based encoding encoder (e.g., Figure 2 E) to obtain a sequence of encoded feature vectors (e.g., Figure 2 VF1 shown in FIG); then, all the encoded feature vectors in the sequence of the encoded feature vectors are concatenated to obtain the feature representation vectors corresponding to the components (for example, Figure 2 VF2 as shown in the figure); then, based on the logical operation rules of conjunction and disjunction between the components in the material of the magnesium-carbon skateboard, a conjunction adjacency matrix between the feature representation vectors corresponding to the components is constructed (for example, Figure 2 M1) and the disjunctive adjacency matrix (e.g., Figure 2 Then, the conjunctive adjacency matrix and the disjunctive adjacency matrix are passed through a second convolutional neural network (e.g., Figure 2 ) to obtain a first feature matrix (e.g., Figure 2 MF1 as shown in FIG. 1 ) and a second feature matrix (e.g., Figure 2 MF2 shown in FIG); then, the first feature matrix and the second feature matrix are calculated by position point to obtain a fused class adjacency matrix (for example, Figure 2 MF shown in FIG); then, the feature representation vectors corresponding to the components are concatenated in two dimensions to obtain a component feature matrix (for example, Figure 2 Then, the component feature matrix and the class adjacency matrix are connected through a graph neural network (e.g., Figure 2 The GNN shown in the figure) to obtain the graph structure feature matrix (for example, Figure 2 ), each row of the graph structure feature matrix is a correlation representation vector containing correlation information of the corresponding component; and, finally, the correlation representation vector of each component is passed through a decoder for regression (e.g., Figure 2 D) as shown in the figure to obtain the optimal ratio of the components.
[0091] In step S110, the optimal physical property data and microstructure diagram corresponding to each component in the material of the magnesium-carbon skateboard are obtained, wherein the material of the magnesium-carbon skateboard includes magnesia, flake graphite, carbores p, spherical asphalt, aluminum silicon alloy powder, sulfide powder, carbonized powder, thermosetting phenolic resin and carbores T60, and the physical property data include linear change rate, weight change rate, volume density, apparent porosity and room temperature compressive strength. As mentioned above, in order to determine the optimal proportion of each component in the material of the magnesium-carbon skateboard to improve the performance of the magnesium-carbon material, in the technical solution of the present application, it is first necessary to obtain the optimal physical property data and microstructure diagram corresponding to each component in the material of the magnesium-carbon skateboard, that is, the optimal physical property data obtained by changing the proportion of a certain component while keeping other components unchanged (including type and weight), for example, including linear change rate, weight change rate, volume density, apparent porosity and room temperature compressive strength, and the corresponding microstructure diagram. Here, the materials used to make the magnesium-carbon skateboard include magnesia sand, flake graphite, carbores p, spherical asphalt, aluminum silicon alloy powder, sulfide powder, carbonized powder, thermosetting phenolic resin and carbores T60.
[0092] In step S120 and step S130, for the optimal physical property data corresponding to each component, each of the physical property data is passed through a word embedding layer to obtain a plurality of parameter input vectors, and the microscopic structure diagram corresponding to each component is input into the first convolutional neural network as a filter to obtain a structural input vector. It should be understood that since the components are not independent but have correlations with each other, in the technical solution of the present application, in order to obtain the correlation information between the components, a context encoder model is used to perform context encoding. That is, first, each of the physical property data is passed through a word embedding layer to obtain a plurality of parameter input vectors; then, the microscopic structure diagram corresponding to each component is input into the first convolutional neural network as a filter for processing to extract the feature distribution representation of the local features of the microscopic structure diagram in a high-dimensional space, thereby obtaining a structural input vector.
[0093] Specifically, in an embodiment of the present application, the process of inputting the microstructure image corresponding to each component into the first convolutional neural network as a filter to obtain a structural input vector includes: first, each layer of the first convolutional neural network except the last layer performs convolution processing on the microstructure image to obtain a structural feature image; then, the last layer of the first convolutional neural network performs global average pooling based on a feature matrix on the structural feature image to obtain the structural input vector. It should be understood that through global average pooling processing, the number of parameters can be reduced and overfitting can be reduced, thereby improving the accuracy of subsequent calculations.
[0094] In step S140 and step S150, the multiple parameter input vectors and the structural input vector are passed through an encoder based on context coding to obtain a sequence of encoded feature vectors, and all the encoded feature vectors in the sequence of the encoded feature vectors are cascaded to obtain the feature representation vectors corresponding to the components. That is, in the technical solution of the present application, after obtaining the multiple parameter input vectors and the structural input vectors, the multiple parameter input vectors and the structural input vectors are further input into an encoder based on context coding for encoding to obtain a sequence of encoded feature vectors. In a specific example, the multiple parameter input vectors and the structural input vectors can be passed through the converter model of the encoder to obtain the sequence of encoded feature vectors. It should be understood that since the converter-based encoder model can encode the input vector based on the context, the sequence of encoded feature vectors obtained can obtain the global association information of each component. Then, all the encoded feature vectors in the sequence of the encoded feature vectors are cascaded to obtain the feature representation vectors corresponding to each component.
[0095] In step S160, based on the logical operation rules of conjunction and disjunction between the components in the material of the magnesium-carbon skateboard, the conjunctive adjacency matrix and the disjunctive adjacency matrix between the feature representation vectors corresponding to the components are constructed, wherein the conjunctive logical operation rules represent the parallel relationship between the rules, and the disjunctive logical operation represents the replacement relationship between the rules. It should be understood that, considering the experimental data of the single component, the feature representation of the component can be obtained through the deep neural network model. Therefore, in the technical solution of the present application, if the adjacency matrix used to represent the association relationship between the components can be further obtained, the graph neural network technology can be used to obtain the association representation vector of each node containing the association information. However, here, the components obviously do not have the node topology structure information represented by the node, so the ordinary adjacency matrix cannot be used, but considering that the components can actually be expressed in the form of rules, for example, the formula ratio of large crystalline magnesia sand 98 with a specification of 3-1mm is 25%-32%. Therefore, in the technical solution of the present application, the adjacency matrix can be constructed by logical operations between the rules.
[0096] Specifically, considering that the logical operations between the rules usually include conjunction and disjunction, they are represented by symbols ∧ and ∨, respectively, to represent the parallel or alternative relationship between the rules, that is, the meaning of "and" and "or". For the components mentioned above, this relationship also exists, for example, "the formula ratio of large crystalline magnesia sand 98 with a specification of 3-1mm is 25%-32%" and "the formula ratio of large crystalline magnesia sand 98 with a specification of 200 mesh is 24%-28%" is an "or" relationship, while "the formula ratio of large crystalline magnesia sand 98 with a specification of 3-1mm is 25%-32%" and "the formula ratio of flake graphite 98 with a specification of -100 mesh is 1%-1.8" is an "and" relationship. Therefore, discrete logical nodes can be constructed for the rules satisfied by each component based on the above method, thereby forming an adjacency matrix, which can be expressed as:
[0097]
[0098]
[0099] here, is a conjunctive adjacency matrix, which is used to indicate that the corresponding pair of rules form a conjunctive normal form when the matrix position is 1, and a non-conjunctive normal form when the matrix position is 0. is the disjunctive adjacency matrix, which is used to indicate that the corresponding pair of rules form a disjunctive normal form with a matrix position of 1 and a non-disjunctive normal form with a matrix position of 0.
[0100] Specifically, in the embodiment of the present application, the process of constructing the conjunctive adjacency matrix and the disjunctive adjacency matrix between the feature representation vectors corresponding to the components based on the logical operation rules of conjunction and disjunction between the components in the material made of the magnesium-carbon skateboard includes: constructing the conjunctive adjacency matrix between the feature representation vectors corresponding to the components based on the logical operation rules of conjunction between the components in the material made of the magnesium-carbon skateboard according to the following formula;
[0101] Wherein, the formula is:
[0102]
[0103] in, It is a conjunctive adjacency matrix, which is used to indicate that the corresponding pair of rules form a conjunctive normal form when the matrix position is 1, and a non-conjunctive normal form when the matrix position is 0.
[0104] Accordingly, in a specific example, based on the logical operation rule of the disjunction between the various components in the material of the magnesium-carbon skateboard, the disjunctive adjacency matrix between the feature representation vectors corresponding to the various components is constructed by the following formula;
[0105] Wherein, the formula is:
[0106]
[0107] in, is the disjunctive adjacency matrix, which is used to indicate that the corresponding pair of rules form a disjunctive normal form with a matrix position of 1 and a non-disjunctive normal form with a matrix position of 0.
[0108] In step S170 and step S180, the conjunctive adjacency matrix and the disjunctive adjacency matrix are passed through a second convolutional neural network to obtain a first feature matrix and a second feature matrix, wherein the first feature matrix has the same scale as the conjunctive adjacency matrix, the second feature matrix has the same scale as the disjunctive adjacency matrix, and the first feature matrix and the second feature matrix are calculated by position points to obtain a fused class adjacency matrix. That is, in the technical solution of the present application, after obtaining the conjunctive adjacency matrix and the disjunctive adjacency matrix, the conjunctive adjacency matrix and the disjunctive adjacency matrix are further processed through a second convolutional neural network respectively to map the adjacency matrix to a high-dimensional space, thereby obtaining a first feature matrix and a second feature matrix, where the dimension of the feature matrix is the same as the original input matrix. Then, the points of the first feature matrix and the second feature matrix are calculated to fuse the two feature matrices, and the class adjacency matrix corresponding to the feature representation of each node can be obtained.
[0109] Specifically, in an embodiment of the present application, the process of passing the conjunctive adjacency matrix and the disjunctive adjacency matrix through a second convolutional neural network to obtain a first feature matrix and a second feature matrix includes: first, each layer of the second convolutional neural network except the last layer explicitly spatially encodes the conjunctive adjacency matrix and the disjunctive adjacency matrix according to the following formula to obtain a first feature map and a second feature map, wherein the formula is:
[0110] f i =active(N i ×f i-1 +B i )
[0111] Among them, f i-1 is the input of the second convolutional neural network in the i-th layer, f i is the output of the second convolutional neural network in the i-th layer, N i is the filter of the second convolutional neural network in the i-th layer, and B iis the bias matrix of the second convolutional neural network of the i-th layer, and active represents a nonlinear activation function. Then, the last layer of the second convolutional neural network performs global pooling processing on the first feature map and the second feature map along the channel dimension to obtain the first feature matrix and the second feature matrix. It should be understood that through global pooling processing, the number of parameters can be reduced, thereby reducing overfitting and improving the accuracy of subsequent regression.
[0112] In step S190 and step S200, the feature representation vectors corresponding to the components are two-dimensionally spliced to obtain a component feature matrix, and the component feature matrix and the class adjacency matrix are passed through a graph neural network to obtain a graph structure feature matrix, each row of the graph structure feature matrix is an association representation vector containing association information of the corresponding component, wherein the graph neural network is suitable for processing graph data in an irregular non-Euclidean space to extract association information of the data sample due to feature information and irregular topological structure information. That is, after obtaining the class adjacency matrix, the feature representation vectors of each component are two-dimensionally spliced to obtain a component feature matrix, and then passed through a graph neural network together with the class adjacency matrix to obtain a graph structure feature matrix, then each row of the graph structure feature matrix is an association representation vector containing association information of the corresponding component. It should be understood that here, the graph neural network can be used to process graph data in an irregular non-Euclidean space, so as to extract the association information of the data samples due to the feature information and irregular topological structure information. Therefore, the obtained association representation vector of the single component can improve the accuracy of decoding regression compared with the feature representation vector obtained by directly splicing.
[0113] In step S210, the associated characterization vector of each component is passed through a decoder for regression to obtain the optimal proportion of the component. That is, in the technical solution of the present application, after obtaining the graph structure feature matrix through the graph neural network, the associated characterization vector of each component is further passed through a decoder for regression to obtain the optimal proportion of the component. In addition, after obtaining the optimal proportion of the component, other label values for representing the upper and lower limits of the proportion range can be set with the optimal proportion as the center. For example, if the optimal proportion is 30%, label values of 27%, 28%, 29%, 31%, 32% and 33% can be set, and then the probability of the other label values being attributed is obtained through decoding regression. For example, the label probabilities of 28%, 29%, 31%, and 32% are greater than the predetermined threshold, so that a feasible proportion range of 28%-32% is obtained.
[0114] That is, in a specific example, after obtaining the optimal proportion of the component, other label values used to represent the upper and lower limits of the proportion range are first set with the optimal proportion of the component as the center; then, the associated characterization vector of each component is passed through the decoder to obtain the probability that the associated feature vector belongs to other label values; finally, based on the probability, determine whether other label values are feasible.
[0115] In summary, the preparation method of the magnesium-carbon skateboard of the embodiment of the present application is explained, which utilizes the physical performance data of the magnesium-carbon skateboard collected after changing a certain component while other components remain unchanged, and the microstructure diagram of the magnesium-carbon skateboard, and then processes it based on the context encoder model and the convolutional neural network model to obtain the characteristic representation of the component, and also constructs an adjacency matrix through logical operations between rules to represent the association relationship between the various components, and further uses the graph neural network technology to obtain the association representation vector of each node containing the association information to improve the accuracy of decoding regression. In this way, the optimal proportion of each component in the material of the magnesium-carbon skateboard can be determined, so that the performance of the magnesium-carbon material can be improved.
[0116] Exemplary Systems
[0117] Figure 3 The figure shows a block diagram of a system for preparing a magnesium-carbon slide plate according to an embodiment of the present application. Figure 3As shown, according to the preparation system 400 of the magnesium-carbon skateboard of the embodiment of the present application, it includes: a data acquisition unit 410, which is used to obtain the optimal physical property data and microstructure diagram corresponding to each component in the material of the magnesium-carbon skateboard, wherein the material of the magnesium-carbon skateboard includes magnesia, flake graphite, carbores p, spherical asphalt, aluminum silicon alloy powder, sulfide powder, carbonized powder, thermosetting phenolic resin and carbores T60, the physical property data include linear change rate, weight change rate, volume density, apparent porosity and room temperature compressive strength; a word embedding unit 420, for obtaining multiple parameter input vectors through the word embedding layer for the optimal physical property data corresponding to each component obtained by the data acquisition unit 410; a first convolution unit 430, for inputting the microstructure image corresponding to each component obtained by the data acquisition unit 410 into the first convolution neural network as a filter to obtain a structure input vector; an encoding unit 440, for passing the multiple parameter input vectors obtained by the word embedding unit 420 and the structure input vector obtained by the first convolution unit 430 through an encoder based on context encoding to obtain a sequence of encoded feature vectors; a cascade unit 450, for cascading all the encoded feature vectors in the sequence of the encoded feature vectors obtained by the encoding unit 440 to obtain the feature representation vectors corresponding to each component; an adjacency matrix construction unit 460, for obtaining a feature representation vector corresponding to each component based on the conjunction and The concatenated adjacency matrix and the disjunctive adjacency matrix between the feature representation vectors corresponding to the components obtained by the cascade unit 450 are constructed by using the disjunctive logical operation rule, wherein the concatenated logical operation rule represents the parallel relationship between the rules, and the disjunctive logical operation represents the replacement relationship between the rules; the second convolution unit 470 is used to pass the concatenated adjacency matrix obtained by the adjacency matrix construction unit 460 and the disjunctive adjacency matrix obtained by the adjacency matrix construction unit 460 through a second convolutional neural network to obtain a first feature matrix and a second feature matrix, wherein the first feature matrix has the same scale as the concatenated adjacency matrix, and the second feature matrix has the same scale as the disjunctive adjacency matrix; the fusion unit 480 is used to calculate the first feature matrix obtained by the second convolution unit 470 and the second feature matrix obtained by the second convolution unit 470 according to the position points to obtain a fused class adjacency matrix; the two-dimensional splicing unit 490 is used to perform two-dimensional splicing on the feature representation vectors corresponding to the components obtained by the cascade unit 450 to obtain a component feature matrix;A graph neural network unit 500 is used to obtain a graph structure feature matrix by passing the component feature matrix obtained by the two-dimensional concatenation unit 490 and the class adjacency matrix obtained by the fusion unit 480 through a graph neural network, wherein each row of the graph structure feature matrix is an association representation vector containing association information of a corresponding component, wherein the graph neural network is suitable for processing graph data in an irregular non-Euclidean space to extract association information of data samples due to feature information and irregular topological structure information; and a regression unit 510 is used to pass the association representation vector of each component obtained by the graph neural network unit 500 through a decoder for regression to obtain an optimal ratio of the component. ;
[0118] In one example, in the above-mentioned magnesium-carbon skateboard preparation system 400, the first convolution unit 430 is further used for: each layer of the first convolutional neural network except the last layer performs convolution processing on the microstructure image to obtain a structural feature map; and the last layer of the first convolutional neural network performs global average pooling based on the feature matrix on the structural feature map to obtain the structural input vector.
[0119] In one example, in the above-mentioned magnesium-carbon skateboard preparation system 400, the encoding unit 440 is further used to: pass the multiple parameter input vectors and the structure input vector through the converter model of the encoder to obtain the sequence of the encoding feature vectors.
[0120] In one example, in the above-mentioned magnesium-carbon skateboard preparation system 400, the adjacency matrix construction unit 460 is further used to: construct a conjunction adjacency matrix between the feature representation vectors corresponding to each component based on the conjunction logic operation rule between the components in the material made of the magnesium-carbon skateboard according to the following formula;
[0121] Wherein, the formula is:
[0122]
[0123] in, It is a conjunctive adjacency matrix, which is used to indicate that the corresponding pair of rules form a conjunctive normal form when the matrix position is 1, and a non-conjunctive normal form when the matrix position is 0.
[0124] In one example, in the above-mentioned magnesium-carbon skateboard preparation system 400, based on the logical operation rule of the disjunction between the various components in the material of the magnesium-carbon skateboard, the disjunctive adjacency matrix between the feature representation vectors corresponding to the various components is constructed by the following formula;
[0125] Wherein, the formula is:
[0126]
[0127] in, is the disjunctive adjacency matrix, which is used to indicate that the corresponding pair of rules form a disjunctive normal form with a matrix position of 1 and a non-disjunctive normal form with a matrix position of 0.
[0128] In one example, in the above-mentioned magnesium-carbon skateboard preparation system 400, the second convolution unit 470 is further used to: each layer of the second convolutional neural network except the last layer explicitly spatially encodes the conjunctive adjacency matrix and the disjunctive adjacency matrix according to the following formula to obtain a first feature map and a second feature map, wherein the formula is:
[0129] f i =active(N i ×f i-1 +B i )
[0130] Among them, f i-1 is the input of the second convolutional neural network in the i-th layer, f i is the output of the second convolutional neural network in the i-th layer, N i is the filter of the second convolutional neural network in the i-th layer, and B i is the bias matrix of the i-th layer of the second convolutional neural network, active represents a nonlinear activation function; and the last layer of the second convolutional neural network performs global pooling processing on the first feature map and the second feature map along the channel dimension to obtain the first feature matrix and the second feature matrix.
[0131] In one example, in the preparation system 400 of the above-mentioned magnesium-carbon skateboard, it further includes: setting other label values for representing the upper and lower limits of the proportion range with the optimal proportion of the certain component as the center; passing the associated characterization vector of each component through the decoder to obtain the probability that the associated feature vector belongs to other label values; and, based on the probability, determining whether other label values are feasible.
[0132] Here, those skilled in the art will appreciate that the specific functions and operations of the various units and modules in the above-mentioned magnesium-carbon skateboard preparation system 400 have been described in the above reference. Figure 1 to Figure 2 The preparation method of the magnesium-carbon skateboard has been described in detail, and therefore, its repeated description will be omitted.
[0133] As described above, the system 400 for preparing a magnesium-carbon skateboard according to an embodiment of the present application can be implemented in various terminal devices, such as a server for preparing an algorithm for a magnesium-carbon skateboard. In one example, the system 400 for preparing a magnesium-carbon skateboard according to an embodiment of the present application can be integrated into a terminal device as a software module and / or a hardware module. For example, the system 400 for preparing a magnesium-carbon skateboard can be a software module in the operating system of the terminal device, or can be an application developed for the terminal device; of course, the system 400 for preparing a magnesium-carbon skateboard can also be one of the many hardware modules of the terminal device.
[0134] Alternatively, in another example, the magnesium-carbon skateboard preparation system 400 and the terminal device may also be separate devices, and the magnesium-carbon skateboard preparation system 400 may be connected to the terminal device via a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.
[0135] Exemplary Electronic Devices
[0136] Below, reference Figure 4 To describe the electronic device according to the embodiment of the present application. Figure 4 As shown, the electronic device 10 includes one or more processors 11 and a memory 12. The processor 11 may be a central processing unit (CPU) or other processing units having data processing capability and / or instruction execution capability, and may control other components in the electronic device 10 to perform desired functions.
[0137] The memory 12 may include one or more computer program products, and the computer program product may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory (cache), etc. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 11 may run the program instructions to implement the functions of the method for preparing the magnesium-carbon skateboard of each embodiment of the present application described above and / or other desired functions. Various contents such as feature representation vectors, class adjacency matrices, etc. may also be stored in the computer-readable storage medium.
[0138] In one example, the electronic device 10 may further include: an input system 13 and an output system 14, and these components are interconnected via a bus system and / or other forms of connection mechanisms (not shown).
[0139] The input system 13 may include, for example, a keyboard, a mouse, and the like.
[0140] The output system 14 can output various information to the outside, including the optimal ratio, regression value, etc. The output system 14 can include, for example, a display, a speaker, a printer, a communication network and a remote output device connected thereto, and the like.
[0141] Of course, to simplify, Figure 4 Only some of the components related to the present application in the electronic device 10 are shown, and components such as a bus, an input / output interface, etc. are omitted. In addition, according to specific application situations, the electronic device 10 may also include any other appropriate components.
[0142] Exemplary computer program products and computer-readable storage media
[0143] In addition to the above-mentioned methods and devices, an embodiment of the present application may also be a computer program product, which includes computer program instructions, which, when executed by a processor, enable the processor to perform the steps in the functions of the method for preparing a magnesium-carbon skateboard according to various embodiments of the present application described in the above-mentioned "Exemplary Method" section of this specification.
[0144] The computer program product may be written in any combination of one or more programming languages to write program codes for performing the operations of the embodiments of the present application, including object-oriented programming languages, such as Java, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, as an independent software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server.
[0145] In addition, an embodiment of the present application may also be a computer-readable storage medium on which computer program instructions are stored. When the computer program instructions are executed by a processor, the processor executes the steps in the method for preparing a magnesium-carbon skateboard described in the above "Exemplary Method" section of this specification.
[0146] The computer readable storage medium can adopt any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium can include, for example, but is not limited to, a system, system or device of electricity, magnetism, light, electromagnetic, infrared, or semiconductor, or any combination of the above. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.
[0147] The basic principles of the present application are described above in conjunction with specific embodiments. However, it should be noted that the advantages, strengths, effects, etc. mentioned in the present application are only examples and not limitations, and it cannot be considered that these advantages, strengths, effects, etc. are required by each embodiment of the present application. In addition, the specific details disclosed above are only for the purpose of illustration and ease of understanding, not for limitation, and the above details do not limit the present application to being implemented by adopting the above specific details.
[0148] The block diagrams of the devices, apparatuses, equipment, and systems involved in this application are only illustrative examples and are not intended to require or imply that they must be connected, arranged, and configured in the manner shown in the block diagram. As will be appreciated by those skilled in the art, these devices, apparatuses, equipment, and systems can be connected, arranged, and configured in any manner. Words such as "including", "comprising", "having", etc. are open words, referring to "including but not limited to", and can be used interchangeably with them. The words "or" and "and" used here refer to the words "and / or" and can be used interchangeably with them, unless the context clearly indicates otherwise. The words "such as" used here refer to the phrase "such as but not limited to", and can be used interchangeably with them.
[0149] It should also be noted that in the apparatus, device and method of the present application, each component or each step can be decomposed and / or recombined. Such decomposition and / or recombination should be regarded as equivalent solutions of the present application.
[0150] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use the present application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of the present application. Therefore, the present application is not intended to be limited to the aspects shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.
[0151] The above description has been given for the purpose of illustration and description. In addition, this description is not intended to limit the embodiments of the present application to the forms disclosed herein. Although multiple example aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, changes, additions and sub-combinations thereof.
Claims
1. A method for preparing a magnesium-carbon skateboard, characterized in that: include: Obtaining optimal physical property data and microstructure diagrams corresponding to each component in the material of the magnesium-carbon skateboard, wherein the material of the magnesium-carbon skateboard includes magnesia, flake graphite, carbores p, spherical asphalt, aluminum silicon alloy powder, sulfide powder, carbonized powder, thermosetting phenolic resin and carbores T60, and the physical property data include linear change rate, weight change rate, volume density, apparent porosity and room temperature compressive strength; For the optimal physical property data corresponding to each component, each of the physical property data is passed through a word embedding layer to obtain a plurality of parameter input vectors; Input the microstructure image corresponding to each component into a first convolutional neural network as a filter to obtain a structure input vector; Passing the plurality of parameter input vectors and the structure input vector through a context-encoding-based encoder to obtain a sequence of encoded feature vectors; Concatenating all the encoded feature vectors in the sequence of the encoded feature vectors to obtain feature representation vectors corresponding to the respective components; Based on the logical operation rules of conjunction and disjunction between the various components in the material of the magnesium-carbon skateboard, a conjunction adjacency matrix and a disjunction adjacency matrix between the feature representation vectors corresponding to the various components are constructed, wherein the conjunction logical operation rule represents the parallel relationship between the rules, and the disjunction logical operation represents the replacement relationship between the rules; Passing the conjunctive adjacency matrix and the disjunctive adjacency matrix through a second convolutional neural network to obtain a first feature matrix and a second feature matrix, wherein the first feature matrix and the conjunctive adjacency matrix have the same scale, and the second feature matrix and the disjunctive adjacency matrix have the same scale; Calculate the first feature matrix and the second feature matrix to obtain a fused class adjacency matrix according to the position points; Perform two-dimensional concatenation of the feature representation vectors corresponding to the components to obtain a component feature matrix; Passing the component feature matrix and the class adjacency matrix through a graph neural network to obtain a graph structure feature matrix, wherein each row of the graph structure feature matrix is an association representation vector containing association information of a corresponding component, wherein the graph neural network is suitable for processing graph data in an irregular non-Euclidean space to extract association information of data samples due to feature information and irregular topological structure information; and The associated representation vector of each component is passed through a decoder for regression to obtain the optimal ratio of the component.
2. The method for preparing a magnesium-carbon skateboard according to claim 1, wherein: Inputting the microstructure images corresponding to the components into a first convolutional neural network as a filter to obtain a structure input vector, comprising: Each layer of the first convolutional neural network except the last layer performs convolution processing on the microstructure image to obtain a structural feature image; and The last layer of the first convolutional neural network performs global average pooling based on the feature matrix on the structural feature map to obtain the structural input vector.
3. The method for preparing a magnesium-carbon skateboard according to claim 2, wherein: Passing the plurality of parameter input vectors and the structure input vector through a context-encoded encoder to obtain a sequence of encoded feature vectors, comprising: The plurality of parameter input vectors and the structure input vector are passed through a transformer model of the encoder to obtain a sequence of encoded feature vectors.
4. The method for preparing a magnesium-carbon slide plate according to claim 3, wherein: Based on the logical operation rules of conjunction and disjunction between the components in the material of the magnesium-carbon skateboard, a conjunction adjacency matrix and a disjunction adjacency matrix between the feature representation vectors corresponding to the components are constructed, including: Based on the conjunction logic operation rule between the various components in the material of the magnesium-carbon skateboard, the conjunction adjacency matrix between the feature representation vectors corresponding to the various components is constructed using the following formula; Wherein, the formula is: in, It is a conjunctive adjacency matrix, which is used to indicate that the corresponding pair of rules form a conjunctive normal form when the matrix position is 1, and a non-conjunctive normal form when the matrix position is 0.
5. The method for preparing a magnesium-carbon slide plate according to claim 4, wherein: Based on the logical operation rules of the disjunction between the various components in the material of the magnesium-carbon skateboard, the disjunction adjacency matrix between the feature representation vectors corresponding to the various components is constructed by the following formula; Wherein, the formula is: in, is the disjunctive adjacency matrix, which is used to indicate that the corresponding pair of rules form a disjunctive normal form with a matrix position of 1 and a non-disjunctive normal form with a matrix position of 0.
6. The method for preparing a magnesium-carbon slide plate according to claim 5, wherein: Passing the conjunctive adjacency matrix and the disjunctive adjacency matrix through a second convolutional neural network to obtain a first feature matrix and a second feature matrix, comprising: Each layer of the second convolutional neural network except the last layer explicitly spatially encodes the conjunctive adjacency matrix and the disjunctive adjacency matrix to obtain a first feature map and a second feature map according to the following formula, wherein the formula is: f i =active(N i ×f i-1 +B i ) Among them, f i-1 is the input of the second convolutional neural network in the i-th layer, f i is the output of the second convolutional neural network in the i-th layer, N i is the filter of the second convolutional neural network in the i-th layer, and B i is the bias matrix of the second convolutional neural network in the i-th layer, and active represents the nonlinear activation function; and The last layer of the second convolutional neural network performs global pooling processing on the first feature map and the second feature map along the channel dimension to obtain the first feature matrix and the second feature matrix.
7. The method for preparing the magnesium-carbon slide plate according to claim 6, further comprising: Other label values for indicating upper and lower limits of the ratio range are set with the optimal ratio of the certain component as the center; Passing the associated representation vector of each component through the decoder to obtain the probability that the associated feature vector belongs to other label values; and Based on the probabilities, it is determined whether other label values are feasible.
8. A system for preparing a magnesium-carbon slide plate, characterized in that: include: A data acquisition unit, used to obtain the optimal physical property data and microstructure diagram corresponding to each component in the material of the magnesium-carbon skateboard, wherein the material of the magnesium-carbon skateboard includes magnesia, flake graphite, carbores p, spherical asphalt, aluminum silicon alloy powder, sulfide powder, carbonized powder, thermosetting phenolic resin and carbores T60, and the physical property data includes linear change rate, weight change rate, volume density, apparent porosity and room temperature compressive strength; A word embedding unit, for obtaining a plurality of parameter input vectors by passing each of the optimal physical property data corresponding to each component obtained by the data acquisition unit through a word embedding layer; A first convolution unit, used for inputting the microstructure image corresponding to each component obtained by the data acquisition unit into a first convolutional neural network as a filter to obtain a structure input vector; an encoding unit, configured to pass the plurality of parameter input vectors obtained by the word embedding unit and the structure input vector obtained by the first convolution unit through an encoder based on context encoding to obtain a sequence of encoded feature vectors; A cascading unit, configured to cascade all the coded feature vectors in the sequence of the coded feature vectors obtained by the encoding unit to obtain a feature representation vector corresponding to each component; An adjacency matrix construction unit, used to construct a conjunction adjacency matrix and a disjunction adjacency matrix between the feature representation vectors corresponding to the components obtained by the cascade unit based on the conjunction and disjunction logic operation rules between the components in the material of the magnesium-carbon slide plate, wherein the conjunction logic operation rule represents the parallel relationship between the rules, and the disjunction logic operation represents the replacement relationship between the rules; a second convolution unit, configured to pass the conjunctive adjacency matrix obtained by the adjacency matrix construction unit and the disjunctive adjacency matrix obtained by the adjacency matrix construction unit through a second convolutional neural network to obtain a first feature matrix and a second feature matrix, wherein the first feature matrix and the conjunctive adjacency matrix have the same scale, and the second feature matrix and the disjunctive adjacency matrix have the same scale; A fusion unit, used for calculating a fused class adjacency matrix between the first feature matrix obtained by the second convolution unit and the second feature matrix obtained by the second convolution unit according to position points; A two-dimensional splicing unit, used for performing two-dimensional splicing on the feature representation vectors corresponding to the components obtained by the cascade unit to obtain a component feature matrix; A graph neural network unit, used for obtaining a graph structure feature matrix by passing the component feature matrix obtained by the two-dimensional splicing unit and the class adjacency matrix obtained by the fusion unit through a graph neural network, wherein each row of the graph structure feature matrix is an association representation vector containing association information of a corresponding component, wherein the graph neural network is suitable for processing graph data in an irregular non-Euclidean space to extract association information of data samples due to feature information and irregular topological structure information; and A regression unit is used to pass the associated representation vector of each component obtained by the graph neural network unit through a decoder for regression to obtain the optimal proportion of the component.
9. The system for preparing a magnesium-carbon skateboard according to claim 8, wherein: The first convolution unit is further used for: Each layer of the first convolutional neural network except the last layer performs convolution processing on the microstructure image to obtain a structural feature map; and the last layer of the first convolutional neural network performs global average pooling based on a feature matrix on the structural feature map to obtain the structural input vector.
10. An electronic device, comprising: processor; as well as A memory, in which computer program instructions are stored, and when the computer program instructions are executed by the processor, the processor executes the method for preparing the magnesium-carbon skateboard as described in any one of claims 1-7.
Citation Information
Patent Citations
Slide plate brick for special steel and production method thereof
CN107042300A
Method for preparing magnesium oxide whisker using waste magnesium and carbon brick substrate material as raw material
CN108866634A
