Industrial chain differentiation multi-dimensional intelligent evaluation method

Abstract

The application provides an industry chain differentiation multi-dimensional intelligent evaluation method, relates to the technical field of intelligent evaluation, and comprises the following steps: collecting multi-dimensional business time sequence data of a plurality of operating entities in a target industry chain, and constructing an attribute heterogeneous graph; direction decoupling is performed on high-dimensional feature vectors of each node in the attribute heterogeneous graph, and intrinsic direction representation and directional response bias of each node are obtained; three primitive state nodes located at different industry chain levels are selected in the attribute heterogeneous graph, intrinsic direction representation and directional response bias of the three nodes are extracted, and the three primitive state nodes form an evaluation triad.The multi-entity multi-dimensional time sequence data of the industry chain is modeled as a heterogeneous graph, and through direction decoupling, triad tensor conduction, evolution on a manifold and triad closed loop aggregation, the differentiated nonlinear conduction law between nodes in the industry chain can be captured, a multi-dimensional intelligent evaluation index can be output, and the objectivity, sensitivity and interpretability of evaluation are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent assessment technology, and in particular to a multi-dimensional intelligent assessment method for industry chain differentiation. Background Technology

[0002] Existing multidimensional differentiation assessment methods for the industrial chain quantify the degree of differentiation across multiple dimensions, serving as a crucial foundation for identifying weak links and supporting precise policy implementation. Currently, commonly used assessment methods largely rely on constructing industrial chain models based on homogeneous graphs or static networks, reflecting the overall differentiation status of the industrial chain by analyzing statistical characteristics such as network density, clustering coefficients, or node centrality. However, these approaches often reveal shortcomings when dealing with the multidimensional and directed business interactions between operating entities within the industrial chain.

[0003] For example, in conducting differentiated assessments of the new energy vehicle power battery industry chain in a specific city to assist in attracting investment and strengthening the chain, it is necessary to comprehensively consider multi-dimensional time-series data on raw material supply, technology licensing, and talent mobility among enterprises to identify collaborative bottlenecks and weak links in the industry chain. Existing technologies typically simplify these heterogeneous interactive information into a single type of undirected relation edge, or aggregate multi-dimensional time-series data into static weights before analysis. This approach may struggle to effectively reflect the differences in the transmission of different business dimensions along the upstream and downstream directions of the industry chain, and it may also overlook the anisotropic transmission effects caused by attribute differences between upstream resource-based enterprises, midstream manufacturing enterprises, and downstream application-based enterprises. Due to the lack of ability to characterize the multi-dimensional information flow paths and closed-loop feedback coupling relationships between key levels of the industry chain, the resulting differentiated assessment results sometimes fail to accurately reflect the inherent multi-dimensional collaborative and constraint structure of the industry chain, thus affecting the accuracy of locating local weaknesses. Summary of the Invention

[0004] This invention provides a multi-dimensional intelligent evaluation method for supply chain differentiation, which improves the ability to model the asymmetric transmission of material flow and information flow between levels.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] Firstly, a multi-dimensional intelligent assessment method for supply chain differentiation, the method including:

[0007] Step 1: Collect multi-dimensional business time-series data of multiple operating entities in the target industry chain and construct an attribute heterogeneous graph; decouple the high-dimensional feature vectors of each node in the attribute heterogeneous graph in terms of direction to obtain the eigenvector representation and directional response bias of each node.

[0008] Step 2: Select three primitive state nodes located at different levels of the industrial chain in the attribute heterogeneous graph, extract the eigenvalue representation and directional response bias of each of the three, and the three primitive state nodes constitute an evaluation triplet.

[0009] Step 3: By evaluating the eigenvector representation and directional response bias of the first two primitive state nodes in the triplet, an anisotropic interconnection tensor is obtained. The interconnection tensor constitutes the information interaction path between the first two primitive state nodes.

[0010] Step 4: On the conduction manifold determined by the information action path, perform walkpoint migration evolution to obtain the characteristic coordinates of the steady-state evaluation stationary point. The initial coordinates of the walkpoint migration evolution are obtained by aggregating the eigenvector representations of the first two primitive state nodes.

[0011] Step 5: Using the characteristic coordinates of the steady-state evaluation stationary point, calculate the guiding coupling degree between the steady-state evaluation stationary point and the first two primitive state nodes respectively. Based on the magnitude of the guiding coupling degree, construct two auxiliary transmission edges between the third primitive state node in the evaluation triplet and the first two primitive state nodes. The two auxiliary transmission edges and the information action path together constitute a triplet closed-loop transmission network.

[0012] Step 6: Extract the coherent information flow intensity of the ternary closed-loop transmission network and aggregate the extraction results into a multi-dimensional intelligent evaluation index that characterizes the degree of differentiation in the industrial chain.

[0013] In a second aspect, a computing device includes:

[0014] One or more processors;

[0015] A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0016] Thirdly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0017] The above-described solution of the present invention has at least the following beneficial effects:

[0018] This invention, by collecting multidimensional business time-series data and constructing an attribute heterogeneous graph, fully preserves the heterogeneous relationships between operating entities under different business dimensions. This avoids the information loss caused by simplifying multidimensional interactions to a single directed edge, as is common in traditional methods, making the evaluation basis more closely reflect the complex real-world operation of the industry chain. Secondly, it decouples the high-dimensional features of nodes by separating the eigendirection representation of their inherent attributes and the directional response bias representing the intensity of directional responses to neighboring information. This allows for a detailed characterization of the differentiated response characteristics of the same entity to transmission in different directions, accurately capturing the objectively existing anisotropic transmission effects between upstream and downstream sectors of the industry chain, and improving the modeling ability for the asymmetric transmission of material and information flows between levels. Using the eigendirection representation and directional response bias of the first two primitive nodes, an anisotropic interconnected transmission tensor is generated. On the transmission manifold determined by this tensor, a wandering point migration and evolution is performed, dynamically searching for evaluation stationary points that reflect the steady state of information interaction between levels. This process simulates the dynamic equilibrium of supply chain transmission in a fluid geometry evolution manner, overcoming the limitation that static network analysis can only reflect instantaneous cross-sectional characteristics, making the evaluation results more inherently trend-oriented and robust.

[0019] Based on the steady-state assessment station, the directional coupling degree between the station and the first two primitive nodes is calculated, and an auxiliary transmission edge connecting the third primitive node is constructed accordingly, forming a ternary closed-loop transmission network. This incorporates forward transmission and closed-loop feedback coupling into a unified assessment framework. This mechanism realistically reflects the reverse constraint and incentive effect of the downstream industry chain on the upstream, enhancing the model's ability to identify collaborative bottlenecks and constraint structures, and is particularly beneficial for discovering local vulnerable links caused by the lack of feedback. The coherent information flow intensity of the ternary closed-loop transmission network is extracted and aggregated into a multi-dimensional intelligent assessment index that includes structural difference degree, collaborative tightness, and transmission efficiency value. This achieves multi-angle quantification of the degree of differentiation in the industry chain, overcoming the shortcomings of traditional indicators that are singular and difficult to accurately target local weaknesses. It provides a refined and interpretable assessment basis for industry chain supplementation and strengthening decisions. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the multi-dimensional intelligent evaluation method for supply chain differentiation provided in an embodiment of the present invention.

[0021] Figure 2 This is a flowchart illustrating step 3 provided in an embodiment of the present invention. Detailed Implementation

[0022] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0023] like Figure 1 As shown, embodiments of the present invention propose a multi-dimensional intelligent evaluation method for supply chain differentiation, the method comprising the following steps:

[0024] Step 1: Collect multi-dimensional business time-series data of multiple operating entities in the target industry chain and construct an attribute heterogeneous graph; decouple the high-dimensional feature vectors of each node in the attribute heterogeneous graph in terms of direction to obtain the eigenvector representation and directional response bias of each node.

[0025] Step 2: Select three primitive state nodes located at different levels of the industrial chain in the attribute heterogeneous graph, extract the eigenvalue representation and directional response bias of each of the three, and the three primitive state nodes constitute an evaluation triplet.

[0026] Step 3: By evaluating the eigenvector representation and directional response bias of the first two primitive state nodes in the triplet, an anisotropic interconnection tensor is obtained. The interconnection tensor constitutes the information interaction path between the first two primitive state nodes.

[0027] Step 4: On the conduction manifold determined by the information action path, perform walkpoint migration evolution to obtain the characteristic coordinates of the steady-state evaluation stationary point. The initial coordinates of the walkpoint migration evolution are obtained by aggregating the eigenvector representations of the first two primitive state nodes.

[0028] Step 5: Using the characteristic coordinates of the steady-state evaluation stationary point, calculate the guiding coupling degree between the steady-state evaluation stationary point and the first two primitive state nodes respectively. Based on the magnitude of the guiding coupling degree, construct two auxiliary transmission edges between the third primitive state node in the evaluation triplet and the first two primitive state nodes. The two auxiliary transmission edges and the information action path together constitute a triplet closed-loop transmission network.

[0029] Step 6: Extract the coherent information flow intensity of the ternary closed-loop transmission network and aggregate the extraction results into a multi-dimensional intelligent evaluation index that characterizes the degree of differentiation in the industrial chain.

[0030] In this embodiment of the invention, by collecting multi-dimensional business time-series data and constructing an attribute heterogeneous graph, the heterogeneous relationships between operating entities under different business dimensions are fully preserved. This avoids the information loss caused by simplifying multi-dimensional interactions to a single directed edge in traditional methods, making the evaluation basis more closely reflect the complex real-world operation of the industry chain. Secondly, the high-dimensional features of nodes are decoupled directionally, separating the eigendirection representation that characterizes its inherent attributes and the directional response bias that characterizes the intensity of directional response to neighborhood information. This allows for a detailed characterization of the differentiated response characteristics of the same entity to transmission in different directions, thereby accurately capturing the anisotropic transmission effect that objectively exists between upstream and downstream of the industry chain, and improving the modeling ability for the asymmetric transmission of material flow and information flow between levels. Using the eigendirection representation and directional response bias of the first two primitive state nodes, an anisotropic interconnection transmission tensor is generated. On the transmission manifold determined by this, a walkpoint migration evolution is performed, which can dynamically search for evaluation stationary points that reflect the steady state of information interaction between levels. This process simulates the dynamic equilibrium of supply chain transmission in a fluid geometry evolution manner, overcoming the limitation that static network analysis can only reflect instantaneous cross-sectional characteristics, making the evaluation results more inherently trend-oriented and robust.

[0031] Based on the steady-state assessment station, the directional coupling degree between the station and the first two primitive nodes is calculated, and an auxiliary transmission edge connecting the third primitive node is constructed accordingly, forming a ternary closed-loop transmission network. This incorporates forward transmission and closed-loop feedback coupling into a unified assessment framework. This mechanism realistically reflects the reverse constraint and incentive effect of the downstream industry chain on the upstream, enhancing the model's ability to identify collaborative bottlenecks and constraint structures, and is particularly beneficial for discovering local vulnerable links caused by the lack of feedback. The coherent information flow intensity of the ternary closed-loop transmission network is extracted and aggregated into a multi-dimensional intelligent assessment index that includes structural difference degree, collaborative tightness, and transmission efficiency value. This achieves multi-angle quantification of the degree of differentiation in the industry chain, overcoming the shortcomings of traditional indicators that are singular and difficult to accurately target local weaknesses. It provides a refined and interpretable assessment basis for industry chain supplementation and strengthening decisions.

[0032] In a preferred embodiment of the present invention, step 1 includes:

[0033] Step 100: Collect multi-dimensional business time-series data of multiple operating entities in the target industry chain, and construct an attribute heterogeneous graph. Nodes in the attribute heterogeneous graph correspond to operating entities, and directed edges correspond to the supply and demand dependencies between entities, with the edges pointing from upstream suppliers to downstream buyers. The initial attributes of the nodes are high-dimensional feature vectors generated by temporal convolutional encoding of multi-dimensional business indicators, specifically including:

[0034] Collect multi-dimensional business time-series data of all operating entities in the target industry chain. The data collected for a single entity covers T consecutive time steps, and each time step records... The actual observed values ​​of each business indicator constitute a length of T and a dimension of The multivariate time series. The preset time window length T is 52 time units (corresponding to a complete annual cycle when the time unit is a week), and the number of business indicators. The criteria are determined based on the actual industrial chain, including factors such as procurement volume, sales revenue, inventory turnover days, frequency of technological cooperation, and the number of talents entering the industry.

[0035] For the multivariate time series of each business entity, a three-layer one-dimensional convolutional neural network is used for temporal convolutional encoding. The first convolutional layer contains 64 kernels of size 7, the second layer contains 128 kernels of size 5, and the third layer contains 256 kernels of size 3. All convolutional layers have a stride of 1 and use the same padding method to maintain the time dimension length. Each convolutional layer is followed by batch normalization and ReLU activation. After the output of the third convolutional layer, global average pooling is performed along the time dimension to compress the time series of each feature channel into a single value, thereby obtaining a 256-dimensional high-dimensional feature vector corresponding to the entity.

[0036] Each business entity is treated as a node, and connections are established based on whether there is an actual supply and demand dependency between entities. These edges are directed, pointing from the upstream supplier to the downstream buyer. If two entities have a direct supply or purchase relationship in any business dimension, an edge is determined to exist. The dependency direction attribute (upstream to downstream) and dependency strength are recorded on the edge. The dependency strength is taken as the normalized total transaction amount between the two entities in the most recent complete period (i.e., the 52 time units corresponding to the collected data). The normalization method is to divide the total transaction amount by the maximum value of the total transaction amount in all direct upstream and downstream relationships of the entity corresponding to that node, obtaining a value between 0 and 1. The high-dimensional feature vectors of all nodes are assigned as initial attributes to the corresponding nodes in the graph.

[0037] Step 101: Using the node connectivity relationships in the heterogeneous attribute graph as topological constraints, construct a feature compatibility matrix between nodes based on high-dimensional feature vectors. This specifically includes:

[0038] Based on the high-dimensional feature vectors of the nodes obtained in step 100 and the edge connection topology of the graph, a feature compatibility matrix is ​​constructed. Specifically, for each edge in the graph, for the edge connecting node i and node j, the cosine similarity of the high-dimensional feature vectors of the two nodes is calculated, and this similarity value is used as the compatibility value corresponding to the edge. For node pairs in the graph that are not directly connected by an edge, their compatibility value is set to 0. All nodes are arranged by index to form a feature compatibility matrix, and the matrix is... Line 1 The column elements represent the pairs of nodes j with each other. The compatibility value.

[0039] The compatibility matrix is ​​row-wise normalized using the L1 norm, so that the nodes... All non-zero values ​​in the corresponding row (nodes) The sum of all connected nodes is 1. The normalized matrix is ​​the eigencompatibility matrix, whose elements can be understood as nodes. For adjacent nodes Information utilization weight.

[0040] Step 102: For each node, perform context-gated decomposition on the corresponding high-dimensional feature vector using the feature compatibility matrix to separate the eigenvector representation that characterizes the node's inherent attributes, and the directional response bias that characterizes the node's directional response strength to the neighborhood information flow. Specifically, this includes:

[0041] For each node in the attribute heterogeneous graph Using the normalized feature compatibility matrix obtained in step 101, context-gated decomposition is performed on its high-dimensional feature vectors. First, the node... The context vector, i.e., finding the node. In the corresponding row of the feature compatibility matrix, extract the neighbor node indices corresponding to all non-zero elements; extract the high-dimensional feature vectors of these neighbor nodes, and sum them according to their respective compatibility weights to obtain the node... A context vector of the same dimension as a high-dimensional feature vector.

[0042] Node The high-dimensional feature vector of the device itself is concatenated with the context vector along the feature dimension to obtain a concatenated vector with a dimension of 512. This concatenated vector is then input into a gating network, which is a two-layer fully connected structure. The first layer is a fully connected layer containing 128 neurons followed by ReLU activation; the second layer is a fully connected layer containing 256 neurons followed by a sigmoid activation function, outputting a 256-dimensional gating coefficient vector. Each dimension of the vector takes values ​​in the interval (0,1).

[0043] Decompose the original high-dimensional feature vector using the gating coefficient vector: ; ;

[0044] in, For nodes The high-dimensional feature vector, This represents the element-wise product of vectors. This refers to the eigenvector representation that characterizes the inherent properties of a node. To characterize the directional response bias of a node to the directional response strength of its neighborhood information flow, the weight parameters of the gated network are randomly initialized using a preset normal distribution with a mean of 0 and a standard deviation of 0.02.

[0045] This embodiment extracts deep temporal features from multidimensional business data through temporal convolutional coding, constructs an attribute heterogeneous graph that preserves heterogeneous relationships between entities, and avoids information loss. Furthermore, it utilizes graph topological constraints to construct a compatibility matrix and performs gated decomposition, decoupling node representations into intrinsic attributes and directional response biases, enabling the differentiation between an entity's own characteristics and its directional sensitivity to neighborhood information.

[0046] In a preferred embodiment of the present invention, step 2 includes:

[0047] Step 200: Based on the material flow hierarchy of the industry chain, select the first primitive state node located in the upstream raw material supply layer, the second primitive state node located in the midstream manufacturing and processing layer, and the third primitive state node located in the downstream terminal distribution layer from the attribute heterogeneous graph. Specifically, this includes:

[0048] Label each node in the heterogeneous graph with its industry chain hierarchy. Based on the material flow of the industry chain, all business entities are divided into three levels: upstream raw material supply (e.g., iron ore mining, crude oil refining, cotton planting), midstream manufacturing and processing (e.g., steel rolling, parts manufacturing, spinning and weaving), and downstream terminal distribution (e.g., automobile sales, clothing retail, catering services). Labeling is based on the entity's main business category in business registration or industry classification, pre-determined by domain knowledge and assigned as a static attribute to each node. From all nodes labeled as upstream, select nodes that have at least one direct dependency edge with a midstream node, forming a candidate upstream node set, denoted as M. Randomly select a node from the candidate set using a uniform probability distribution, i.e., each candidate node has a probability of 1 / M when selected, and use this node as the first primitive state node.

[0049] Extract all direct adjacent edges of the first primitive state node and filter the edges that connect to the midstream level nodes. Compare the dependency strength values ​​recorded on these edges, which are the normalized total transaction amount calculated in step 100, and the value range is between 0 and 1. Select the midstream node with the largest dependency strength value as the second primitive state node. Dependency strength reflects the closeness of transactions between upstream and downstream entities. The greater the strength, the stronger the midstream node's dependence on or absorption capacity of upstream raw materials. It is more likely to play a core processing and conversion role in the industrial chain, and is therefore the most structurally important downstream recipient. If there is more than one midstream node corresponding to the maximum strength, calculate the business data completeness of these nodes. Completeness = 1 - (number of missing values ​​in the multidimensional time series data of the entity ÷ total number of data points), where the total number of data points is the time step T (52) × number of business indicators. Select the node with the lowest proportion of missing values ​​(i.e., the highest completeness) as the second primitive state node, because high data completeness can ensure the reliability and stability of subsequent time series modeling and avoid estimation bias due to too many missing values. If nodes with equal dependency strength still have the same completeness, then the node with the smaller name is selected according to the lexicographical order of the entity name. That is, the node with the smaller name is selected according to the comparison rules of character encoding order (such as UTF-8), ensuring that a unique result is obtained under completely equal conditions.

[0050] Obtain all directly adjacent edges of the second primitive state node, filter those connecting to downstream nodes, and sort them in descending order of dependency strength values ​​recorded on the edges. Select the downstream node with the highest dependency strength as the third primitive state node. If multiple downstream nodes are listed in parallel, compare their business data completeness according to the same rules; if they are still the same, select the node with the smallest lexicographical order of its entity name. At this point, a three-tiered node chain along the industry chain—upstream, midstream, and downstream—is uniquely determined.

[0051] Step 201: Extract the eigenvector representation and directional response bias corresponding to the first, second, and third primitive state nodes to form an evaluation triplet, specifically including:

[0052] From all the node features obtained from decoupling in step 102, the following six feature vectors are extracted, namely the eigenvector representations of the first primitive state node, denoted as... The directional response bias of the first elementary state node is denoted as... The eigenvector representation of the second primitive state node is denoted as... The directional response bias of the second elementary state node is denoted as... The eigenvector representation of the third primitive state node is denoted as... The directional response bias of the third elementary state node is denoted as... These six vectors form a complete evaluation triplet.

[0053] In this embodiment, three primitive nodes with upstream and downstream representative significance are selected according to the material flow hierarchy, and their decoupled representations are extracted to form evaluation triplets. The core structure of the interaction between the industrial chain levels is captured in the form of the smallest evaluation unit. This not only reduces the computational complexity of the whole graph analysis, but also focuses on the key links that best reflect the differentiated transmission, thereby improving the efficiency and accuracy of the evaluation.

[0054] like Figure 2 As shown, in a preferred embodiment of the present invention, step 3 includes:

[0055] Step 300: The eigenvector representations of the first and second primitive state nodes are cross-mixed dimension-by-dimensional in the feature space to generate an initial mixed field; using the directional response biases of the first and second primitive state nodes, directional selective gating coefficients are obtained through nonlinear activation, specifically including:

[0056] The eigenvector representations calculated for the first primitive state node A and the second primitive state node B using step 201. , , , Two structured components representing the two-node interaction basis are constructed, namely the initial mixing field. and direction-selective gating coefficient .Will and Perform pairwise cross-mixing along the feature dimensions, for each dimension of the feature vector. ( ), performing two complementary operations dimension by dimension, one of which is the product operation, i.e., calculating The product vector is obtained. The first operation captures the degree of co-activation of two vectors in the same dimension. When the two vectors have the same sign and a large absolute value, the product magnitude is prominent, reflecting strong interaction. The second is the mean operation, which calculates... The mean vector is obtained. This operation extracts the common trend of two vectors while preserving the central location information of features. The product vectors are then... With mean vector The vector is concatenated along the feature axis to form a concatenated vector with a dimension of 512. .

[0057] Will Input an unbiased linear transformation layer, the weight matrix of which... Size is The parameters are initialized using a uniform distribution from Xavier, specifically from... Uniform sampling within the interval, where This is to ensure the stability of the forward propagation output variance. The linear transformation is calculated as follows: Output This is the initial mixing field, which is a compressed representation of the upstream and downstream nodes' self-representation information after product and mean mixing in two channels.

[0058] Using the bias representation of nodes A and B and To obtain the directional response characteristics of each of the two nodes in the feature space, first, the dimension-wise product of the bias vectors is calculated. The resulting vector Each dimension's value reflects the matching of the two nodes' bias directions along that feature dimension. If the two nodes have the same bias direction (both positive or both negative with a larger absolute value), then... If the values ​​are large positive values, then if the directions are opposite (one positive and one negative), then... It is a negative value; if the bias response of either side is weak in this dimension, then The absolute value is relatively small. For Applying a hyperbolic tangent function to each dimension Map values ​​to The interval is used to obtain the nonlinearly compressed directional interaction signal. The directional selectivity gating coefficients are then adjusted through scaling and translation transformations. The scope of action covers both enhancement and inhibition modes: ;

[0059] In the formula This represents a 256-dimensional constant vector with all elements equal to 1.0, and 0.5 is a scalar scaling factor. This operation makes... Each dimension fall into Interval. When When the two nodes are highly aligned in their bias directions along this dimension, the propagation of the subsequent mixed field along this dimension will be enhanced; when When this occurs, it indicates a divergence in direction or a weak response, and conduction is suppressed; This is equivalent to no modulation.

[0060] Step 301 involves anisotropically modulating each feature dimension of the initial mixing field using direction-selective gating coefficients to obtain an anisotropic modulated field, specifically including:

[0061] The direction-selective gating coefficients generated in step 300 For the initial mixing field Anisotropic scaling is performed dimension-wise to obtain anisotropic modulation fields. For each feature dimension Perform the following operation: This operation is an element-wise multiplication, and the result is still a 256-dimensional vector. The direction-selective gating coefficients independently amplify or attenuate the initial mixing field in each dimension, amplifying dimensions that are highly consistent with the bias directions of upstream and downstream nodes, while compressing divergent dimensions. This non-uniform scaling breaks the isotropy of the initial mixing field, causing the transmitted signal to exhibit direction selectivity and intensity differences in the feature space, thus achieving anisotropic transmission at the feature level.

[0062] Step 302: Based on the dependency distance between the first and second primitive state nodes in the attribute heterogeneous graph, a monotonically non-increasing smooth decay function is constructed to perform spatial attenuation processing on the anisotropic modulation field, resulting in an anisotropic interconnection tensor in the feature space. This interconnection tensor constitutes the information interaction path between the first and second primitive state nodes, specifically including:

[0063] On the attribute heterogeneous graph constructed in step 100, the shortest path length from the first primitive state node A to the second primitive state node B is calculated using the number of edges as the path length, and denoted as the dependency distance. If there is a direct dependency edge between A and B, then If it is necessary to pass through one intermediate node, then And so on. Set a distance limit. If the calculated shortest path length exceeds 5, it is forcibly truncated to 5 to limit the conduction strength of long-distance connections and reduce noise interference. An exponential decay method is used to reduce the distance... Mapped to decay weights The calculation formula is: ;

[0064] Where the attenuation rate coefficient The default value is 0.5. This function has the characteristics of being monotonically non-increasing and smooth. hour, ;when hour, The transmission intensity decreases rapidly with increasing distance. This exponential decay ensures that information exchange between distant nodes is effectively suppressed, highlighting the direct interaction between adjacent levels in the industry chain.

[0065] The anisotropic modulation field obtained in step 301 With decay weight Perform scalar multiplication to obtain the final interconnection transport tensor. , It is a 256-dimensional vector. Interconnection transport tensor Numerically, it simultaneously encodes two pieces of information: one is the gating coefficient. The modulation exhibits anisotropic interactive modes, and secondly, the conduction intensity attenuation is determined by graph distance. Therefore, The information transmission path from upstream primitive node A to midstream primitive node B is defined, which will serve as the core transmission parameter on this link in subsequent graph neural propagation.

[0066] This embodiment generates an initial mixing field through dimensional cross-mixing of eigenvectors and uses directional response biases to generate directional selective gating via nonlinear activation, achieving anisotropic modulation of differences in each dimension. This realistically simulates the non-uniformity of transmission intensity across different business dimensions for upstream and downstream enterprises. Combined with the smooth attenuation of dependency distance, an information interaction path with anisotropic and spatial attenuation characteristics is formed, providing a specific calculation method for accurately measuring the tensor characteristics of inter-level transmission.

[0067] In a preferred embodiment of the present invention, step 4 includes:

[0068] Step 400: The initial characteristic coordinates of the virtual evaluation walkpoint are aggregated from the dimension-wise mean values ​​of the eigenvector representations of the first and second primitive state nodes, and the virtual evaluation walkpoint is initialized on the conduction manifold determined by the interconnection conduction tensor. Specifically, this includes:

[0069] Receive the eigenvector representation of the first primitive state node This vector is a 256-dimensional real vector, where the _i_ is a _i_, _j_ The magnitude of the dimensional component reflects the first primitive state node in the th order. The activation response intensity on each characteristic channel, with a larger value indicating a more significant response in that channel; eigenvector representation of the second primitive state node. Both are 256-dimensional real vectors, but the numerical meanings of each dimension are different. Consistent; Receive interconnected transport tensor It is also a 256-dimensional real vector, and its ______ The numerical magnitude of the dimensional component represents the number of nodes in the two elementary states at the 1st dimensional node. The conduction coupling strength on each characteristic channel is such that the larger the value, the tighter the coupling.

[0070] Dimension index Take out each number from 1 to 256 one by one. The Dimensional components and The For each component, calculate its arithmetic mean and use this mean as the initial feature coordinate vector. No. The value of each dimension. After traversing all 256 dimensions, the complete initial feature coordinates are obtained. This coordinate system symmetrically integrates the intrinsic response information of the two primitive state nodes.

[0071] Using interconnected transmission tensors Construct the fundamental kernel matrix of the conduction manifold, with a pre-defined recombination dimension of . Interconnecting the conduction tensor Fill in one line at a time, in line-major order. matrix Calculate matrix product ,in It is a matrix The transpose of the product. It is A symmetric positive semi-definite matrix, hereinafter referred to as the kernel matrix. Based on kernel matrix Define a potential function in a 256-dimensional feature space. For any 256-dimensional coordinate vector... First, reshape it according to the same row priority rule. matrix The potential energy value is given by the following formula: ;

[0072] in, yes transpose, The trace of a matrix is ​​represented by the sum of the elements along its main diagonal. This potential function is expressed at each point in the characteristic space. gradient direction at The gradient matrix is ​​calculated as follows: , and then Expanding it row-majorly into a 256-dimensional vector, the resulting vector is... It points in the direction of the fastest increase in potential energy. The gradient vectors at all points together constitute the gradient vector field in the characteristic space.

[0073] Let constant Take any real value, and all satisfy point This forms a surface of equal height. When When traversing all real numbers, a family of nested contour surfaces is obtained. This family of contour surfaces, together with the gradient vector field, defines the set of transferable paths for virtual walkpoints in the feature space. That is, walkpoints can only move between contour surfaces along the gradient direction and cannot escape the manifold structure defined by this set of paths. This set of paths is the conduction manifold, which is defined by the kernel matrix. The only decision. Set the current feature coordinates of the virtual evaluation walkpoint to... .because It is obtained by fusing the eigenvalues ​​of two nodes, which are located within the domain of the aforementioned conduction manifold, and the initialization is complete.

[0074] Step 401: Perform spectral decomposition on the interconnected conduction tensor to obtain a set of eigenvalues ​​arranged in descending order of amplitude. Extract the largest and second-largest eigenvalues ​​from the eigenvalue set, calculate the ratio of the largest to the second-largest eigenvalue to obtain the principal eigenvalue ratio. Adaptively determine the number of evolution iterations based on the principal eigenvalue ratio, wherein the number of evolution iterations and the principal eigenvalue ratio satisfy a monotonically non-decreasing mapping relationship, specifically including:

[0075] Based on the kernel matrix output in step 400 Perform full eigenvalue decomposition on it and solve the characteristic equation. ,in Let be the non-zero eigenvectors to be solved. This is a scalar eigenvalue. Because... yes Since the matrix is ​​real symmetric, the equation has 16 solutions. Each solution consists of an eigenvalue and its corresponding eigenvector. Solving for these eigenvalues ​​yields all 16 eigenvalues ​​and their corresponding eigenvectors. The 16 eigenvalues ​​are then sorted in descending order of their numerical values. Given a symmetric positive semi-definite matrix with all eigenvalues ​​being non-negative real numbers, the sequence obtained after arranging the eigenvalues ​​is... .

[0076] Extract the maximum eigenvalue from the descending sequence. and the second largest eigenvalue Preset lower limit protection threshold. Protect the second largest eigenvalue. Then force it to be set to ;like If the value remains unchanged, then the original value will remain unchanged.

[0077] Calculate the principal eigenvalue ratio This ratio measures the intensity difference between the dominant and secondary dominant directions in the conduction manifold. The larger the value, the stronger the extension of the manifold along the direction corresponding to the largest eigenvalue is compared to other directions. The more significant the anisotropy, the more iteration steps are needed to fully traverse the strongly coupled structure. The closer the value is to 1, the more evenly the eigenvalues ​​are distributed, the closer the manifold is to isotropy, and the fewer iterations are needed to reach steady state. This is based on the principal eigenvalue ratio. Adaptive determination of the number of evolution iterations Preset lower limit for the number of iterations. upper limit Scaling factor Exponential coefficient Calculate according to the following monotonically non-decreasing mapping formula. : In the formula, Indicates rounding down. Control the number of iterations The overall magnitude of the growth Control the degree of nonlinearity of the growth curve. When hour, , ;along with Increase, Increase Increase synchronously but not exceeding .

[0078] Step 402: In each iteration update, calculate the negative gradient of the potential energy function at the current walk point to obtain the local propagation descent direction; obtain the inertial offset generated in the previous iteration, and perform a weighted synthesis of the local propagation descent direction and the inertial offset generated in the previous iteration to obtain the corrected drift direction, specifically including:

[0079] Based on the initial feature coordinates output in step 400 The current coordinates for the first iteration And based on the number of iterations output in step 401 Set a loop limit. Number of iterations. Start counting from 0, until... Complete all iterations. For the ... In the next iteration, the current feature coordinates of the virtual evaluation walkpoint are obtained. The directional derivative of the interconnected transmission tensor at the current feature coordinates of the virtual evaluation walkpoint is calculated, thereby obtaining the local transmission gradient. This directional derivative is the partial derivative of the potential energy function defined in step 400 with respect to each component of the coordinate vector. Specifically, the 256-dimensional vector is first... The row-major rule has been reshaped to... matrix That is, the first of the matrix Line 1 The column element corresponds to the first element of the original vector. Each component. The interconnect conduction tensor is derived from the kernel matrix output in step 400. Characterize, and substitute into the potential energy function in step 400. Given the gradient expression, calculate the gradient matrix. ;Will gradient matrix Expanding the vector in row-major order yields a 256-dimensional vector, which represents the local propagated gradient. This vector points in the direction in which the potential energy rises the fastest at the current point. Taking its opposite gives the direction of local conduction descent, which points in the direction in which the potential energy decreases the fastest, driving the wandering point to migrate towards the region of potential energy minimum.

[0080] Get the inertial offset generated in the previous iteration Its source rules are: when When there is no previous iteration, then set Given a 256-dimensional vector of all zeros; when hour, This refers to the corrected drift direction stored in step 404 at the end of the previous iteration. This inertial offset carries information about the direction of movement in the previous step, providing a historical trend reference for this direction synthesis. The local gradient is then propagated. Inertial offset compared to the previous round By performing weighted synthesis, the corrected drift direction for the current round is obtained. The synthesis formula is: ;

[0081] in The Euclidean norm of a vector; It is a small constant that prevents division by zero. When the inertial offset is a zero vector, it ensures that the denominator is not zero and maintains numerical stability. This is the inertia weighting coefficient, and its value range is... The first item The normalized inertial offset multiplied by the inertial weight represents the degree to which the historical movement direction is preserved; the second term Normalizing the current negative gradient direction and multiplying it by the complementary weights represents the degree of following the direction of the fastest decrease in local potential energy. This means that the corrected drift direction is composed of 30% of the historical inertial direction and 70% of the current descent direction, so that the wandering can maintain the continuity of the search and avoid oscillations, and can also move towards the potential energy trough area in a timely manner, thereby ensuring that the iteration converges to the steady-state evaluation stationary point.

[0082] Step 403: Perform local eigenvalue decomposition on the interconnection transmission tensor at the current feature coordinates of the virtual evaluation walkpoint to obtain the local curvature tensor; calculate the trace of the local curvature tensor to obtain the local structure coefficients; multiply the reciprocal of the local structure coefficients by the preset reference step size to obtain the compensation step size, specifically including:

[0083] Based on the kernel matrix output in step 400 and current wandering point Based on the characteristics of the manifold, the compensation step size for this iteration is calculated, allowing the walking stride to be adaptively adjusted according to the geometry of the manifold. Since the potential energy function defined in step 400 is a quadratic form... The Hessian matrix (i.e., the local curvature tensor) of this function is constant throughout the entire space, and is related to the current wandering point. The specific location is irrelevant. This Hessian matrix is ​​equal to the identity matrix and the kernel matrix. Kroneck The calculation of its trace can be simplified to the product of the traces of two factor matrices according to the trace property of the Kronecker product. Based on this property, the local structure coefficients... The calculation formula is: ;

[0084] In the formula, It is a kernel matrix traces, that is The sum of the 16 main diagonal elements; the factor 16 comes from the recombination dimension of the feature bundle, i.e., the number of rows or columns of the reconstructed matrix. Local structure coefficients. It quantitatively reflects the overall intensity of fluctuations in the entire conduction manifold. The larger the value, the higher the average curvature of the potential energy surface and the more complex the manifold structure. The smaller the value, the flatter the potential energy surface.

[0085] Preset reference step size This value, once calibrated, provides a suitable single-step displacement amplitude in flat manifold regions, balancing convergence speed and search accuracy. Normalized reference value. Take as the initial time ( ) Calculated Value, that is Zero-prevention small constant This is used to avoid an abnormally large step size due to an excessively small denominator. The formula for calculating the step size is: ;

[0086] In the formula, It is the compensation step size. This is the normalized ratio of the current local structure coefficients relative to the initial reference value, reflecting the relative curvature of the current region. When curvature increases, the denominator increases, and the step size automatically decreases to prevent excessively large steps in undulating regions, which could cross potential energy ridges or cause oscillations. When curvature decreases, the denominator decreases, and the step size increases accordingly, accelerating passage through flat regions. When the manifold structure coefficient... Compared with the initial reference value At that time, The step size is approximately When the local structure coefficient increases and the curvature increases, the denominator becomes larger and the step size automatically decreases to make precise movements, avoiding crossing the potential energy ridge line due to excessive step size in areas of high curvature; when the local structure coefficient decreases and the curvature decreases, the denominator becomes smaller and the step size automatically increases to accelerate crossing flat areas.

[0087] For the calculated Perform truncation protection. Preset lower limit for step size. Step size limit .like Then Forced to ;like Then Forced to Otherwise, the original value remains unchanged. This protection mechanism constrains the step size within a reasonable range, preventing the search from stalling due to an excessively small step size, and also preventing numerical overflow or skipping of the potential energy extremum region due to an excessively large step size.

[0088] Step 404: Move the virtual evaluation walkpoint along the corrected drift direction to compensate for the step size, update the feature coordinates of the virtual evaluation walkpoint, and record the corrected drift direction generated in this iteration as the inertial offset for the next iteration; after completing the update of the evolution iteration count, use the final feature coordinates of the virtual evaluation walkpoint as the feature coordinates of the steady-state evaluation stationary point, specifically including:

[0089] Based on the corrected drift direction output in step 402 and the compensation step size output in step 403 Perform virtual evaluation of the walkpoint position update and record the inertial offset for use in the next iteration. After all iterations are completed, output the final characteristic coordinates of the steady-state evaluation stationary point. Then, use the compensation step size output in step 403. The corrected drift direction output in step 402 Perform scalar multiplication to obtain the displacement vector. The direction of this displacement vector is determined by the corrected drift direction. The size is determined by the compensation step size. Modulation. Update the feature coordinates of the virtual evaluation walkpoint, i.e., the displacement vector. With current coordinates Adding each element one by one yields new feature coordinates. This will serve as the current position for the next iteration, and the update formula will be: ;

[0090] This update enables the wander point The migration along the corrected drift direction on the potential energy surface is performed using a step size that senses the local curvature. The corrected drift direction generated in this round is then... Record the inertial offset for the next iteration, that is, let This vector will be read and used in step 402 of the next iteration as inertial information reflecting the movement trend of the previous round, participating in the direction synthesis. Since the direction vector itself is recorded directly here (rather than the displacement vector modulated by the step size), the inertial term will be normalized in the next round of synthesis to ensure that its contribution is only reflected in the direction guidance and is not affected by the step size of this round.

[0091] Based on the number of evolution iterations determined in step 401 Make a judgment if the current round count satisfies... If the preset total number of iterations has not been reached, then the iteration rounds will be reduced. Increment by 1, and return to step 402 to continue the next iteration; if That is, all has been completed. The iteration loop terminates after the next update. At the end of the iteration, the final feature coordinates of the wandering point are... These are the characteristic coordinates of the stationary point in the steady-state evaluation. These coordinates are the distance the virtual evaluation walk point travels on the potential energy surface. The steady-state position reached after the evolution driven by wheel inertia and gradient and adaptive step size modulation.

[0092] In this embodiment, the initial walking point is constructed through mean aggregation, naturally inheriting the common principal features of the two nodes and avoiding unidirectional bias. Simultaneously, the transmission manifold is defined using the interconnection transmission tensor, confining the entire walking process within a feature subspace closely related to the information interaction pattern between the two nodes. The ratio of principal eigenvalues ​​of the interconnection transmission tensor is used to adaptively determine the number of iterations, allowing the number of evolutionary steps to be automatically adjusted according to the dominant directional strength of the transmission manifold. More iterations are allocated to more precise convergence for sharper directionality, while fewer iterations are allocated to less diffuse directionality to avoid redundant computation. The drift direction is corrected by weighting the local transmission gradient with historical inertia, ensuring that the walking follows the steepest transmission direction of the current manifold while smoothing the trajectory and suppressing oscillations through the inertia term, enabling the search process to cross local flat regions more quickly and maintain a stable convergence trend. The compensation step size is calculated using the trace of the local curvature tensor. Since the potential energy function is quadratic, its curvature is constant throughout the entire space, thus the compensation step size remains unchanged during each evaluation. However, by normalizing the initial reference values, different evaluation tasks can adaptively set the step size according to the average curvature of the overall manifold structure, achieving adaptive adjustment between tasks and realizing adaptive granular control over complex manifold geometry, thus improving evolutionary stability and stationary point accuracy. Through multiple rounds of compensated step size drift and inertial transfer, the wandering point gradually approaches the structurally stable position on the manifold. The final output steady-state evaluation stationary point incorporates multi-source information such as transmission direction, spatial attenuation, and node offset, providing a low-dimensional and comparable geometric representation for quantifying the effective transmission capability of information pathways between two nodes.

[0093] In a preferred embodiment of the present invention, step 5 includes:

[0094] Step 500: On the conduction manifold, using the interconnection conduction tensor kernel matrix as the metric tensor, calculate the cosine of the manifold inner product angle between the eigenvector representation of the first primitive state node and the characteristic coordinates of the stationary point in steady-state evaluation, as the first guiding coupling degree; calculate the cosine of the manifold inner product angle between the characteristic coordinates of the stationary point in steady-state evaluation and the eigenvector representation of the second primitive state node, as the second guiding coupling degree, specifically including:

[0095] Receive the steady-state stationary point characteristic coordinates output in step 404 and the interconnected conduction tensor kernel matrix generated in step 400 And use the eigenvectors of the first primitive state node to represent Second primitive state node eigendirection representation Calculate the two guiding coupling degrees on the conduction manifold respectively. and This is used to quantify the directional consistency between different nodes and steady-state stationary points. (This is related to the metric matrix.) To align the order of the vectors, first reshape the three 256-dimensional vectors into a 16×16 matrix according to a consistent rule, denoted as . ; It is a matrix representation of the eigenvectors of the first primitive state node in the two-dimensional spanned space, where each row or column corresponds to the contribution of different conduction channels on the manifold to the node features; Similarly, the eigenvector matrix representation of the second primitive state node; It is a matrix representation of the characteristic coordinates of the steady-state stationary point, describing the most stable state mode of the system in the direction of the lowest potential energy.

[0096] In a conduction manifold, the matrix As a metric tensor, define any two matrices The inner product is ,in It is a matrix The transpose of . Based on this inner product, the matrix . The manifold norm is defined as , indicating under manifold metric The length of , where It is a matrix In the kernel matrix The induced inner product is the inner product of itself.

[0097] First guiding coupling degree In essence and The cosine of the geodesic angle under the manifold inner product. The calculation is performed in two steps, with the numerator being the inner product. ,reflect Transformed with longitude gauge The cumulative consistency of their directions means that positive values ​​indicate they are generally moving in the same direction, while negative values ​​indicate they are moving in opposite directions. yes The transpose of the matrix; the denominator is the product of the norms of the two matrix manifolds with an added zero-preserving constant. ,Right now ,in It is the main part of the denominator of the first guiding coupling degree. It is a matrix The transpose of the formula. The zero-prevention constant is only used to avoid the numerical risk of the denominator being zero; its impact on the normal coupling degree value is negligible, ultimately yielding the first guiding coupling degree. Its value range is within .when Approaching When, explain and Under the manifold metric, the orientations are highly consistent, and the first elementary state nodes exhibit a strong inherent tendency to reach this steady state; close to A positive value indicates that the two directions are approximately orthogonal and the coupling is weak; a negative value indicates that the directions are diverging.

[0098] Second guiding coupling degree Calculated using a completely symmetric method, by and In the same metric The following is obtained: ;

[0099] Second guiding coupling degree They also fell into . Reflects the degree of coupling between the intrinsic directions of the second-elemental nodes and the directions of the steady-state stationary points under the manifold metric, and explains the relationship between them and the manifold metric. Consistent It is the direction matrix of the second elementary state. With steady-state stationary point direction matrix In manifold metric The inner product below, It is the main part of the denominator of the second guiding coupling degree.

[0100] Step 501: Construct a first auxiliary transmission edge pointing from the second primitive state node to the third primitive state node, and a second auxiliary transmission edge pointing from the third primitive state node to the first primitive state node; the first and second auxiliary transmission edges are connected with the information interaction path to form a ternary closed-loop transmission network. The first and second guiding coupling degrees are used to characterize the directional consistency between the steady-state evaluation stationary point and the two primitive state nodes. The transmission strength coefficient of the auxiliary transmission edges is calculated based on the interconnection transmission tensor and the coherent information flow intensity, specifically including:

[0101] Receive the first guiding coupling degree obtained in step 500 Coupling with the second guide And utilize the eigenvectors of the third primitive state node to represent Based on the existing information pathway, two auxiliary transmission edges are constructed to form a ternary closed loop. The system already has an information interaction path from the first primitive state node to the second primitive state node, and its transmission strength coefficient is... The calculation method is as follows: take the eigenvector representation of the first node. Eigendirection representation of the second node Reshape it into a matrix of the same order as the metric. , And using the interconnected conduction tensor kernel matrix output in step 400 As a manifold metric, the conduction potential energy from node 1 to node 2 is defined as: ;

[0102] in To prevent a zero constant, this conduction potential energy is essentially... and The cosine of the angle under the manifold metric reflects the degree of alignment of the eigendirections of the two nodes under the metric. The larger the cosine value, the better the semantic structure of node 1 matches that of node 2 along the transmission direction, and the stronger the potential driving force for information flow. This transmission potential energy is used as the intensity coefficient, i.e. Thus Falling The interval, whose size and sign reflect the potential energy flowing from node 1 to node 2. The system already has an information interaction path from the first primitive state node to the second primitive state node; the interconnection conduction tensor of this path has been determined as the kernel matrix in step 400. Its intensity and direction information are encoded by the anisotropic modulation field generated in step 302 through the kernel matrix construction in step 400. The coherent transport scalar of this path... In step 601, the process will be... The slice sequence is obtained by phase space entanglement integration, and there is no need to define the scalar intensity coefficient separately here.

[0103] To construct a closed loop, a basic coupling factor is defined. This is used to linearly convert the guiding coupling degree output from step 500 into a non-negative strength coefficient of the auxiliary conducting edge. The linear conversion is calculated as follows: ;

[0104] in It is the strength coefficient of the first auxiliary conduction edge. It is the strength coefficient of the second auxiliary conduction edge. The magnitude of reflects the absolute strength of the coupling between the first primitive node and the steady-state stationary point. The stronger the coupling, the more significant the constraint on the node in the steady-state direction. When this situation propagates along the closed loop, it naturally drives the second edge (nodes 2 to 3) to obtain a strength proportional to it. Then control the overall amplitude of this proportional conversion to keep the strength of the auxiliary edge relative to the original path. Near the magnitude of the signal, to prevent overshoot or undershoot. Similarly, Depend on Through the same linear factor The transformation yields a driving effect on the third edge (nodes 3 to 1) reflecting the coupling strength between the second node and the steady state. The resulting directed closed-loop circuit contains three edges, namely edge 1 (nodes 1 to 2), whose interconnection propagation tensor is... Coherent transport scalar In step 601, the intensity modulation coefficient of edge 2 (nodes 2 to 3) is calculated as follows: (First auxiliary edge), its interconnection transport tensor Constructed in step 600; edge 3 (nodes 3 to 1), its intensity modulation coefficient is (Second auxiliary edge), its interconnection transport tensor Constructed in step 600. The value range is [0, 0.85], which is used to modulate the signal in step 600. and The amplitude. This closed-loop structure makes the auxiliary edge strength coefficient able to pass It automatically adjusts to follow the coupling changes between nodes and the steady state, reflecting the network topology's adaptive ability to the coupling situation.

[0105] This embodiment utilizes the metric tensor of the conduction manifold to calculate the cosine of the geodesic angle, quantifying the geometric relationship between primitive nodes and steady-state stagnation points into a continuous guiding coupling degree. This overcomes the limitation of traditional discrete similarity in failing to reflect the tortuous characteristics of the manifold, providing a high-resolution numerical basis for dynamically adjusting the conduction strength. By directly modulating the strength of the two auxiliary conduction edges through the guiding coupling degree, the ternary closed-loop network can adaptively reflect the strength of the situational coupling between nodes, enhancing the network's ability to elastically model information flow paths.

[0106] In a preferred embodiment of the present invention, step 6 includes:

[0107] Step 600: For the first auxiliary transmission edge and the second auxiliary transmission edge, calculate the corresponding interconnection transmission tensor using the eigendirection representation and directional response bias of the two primitive state nodes connected by the corresponding auxiliary transmission edge; slice the interconnection transmission tensor corresponding to each of the three edges in the ternary closed-loop transmission network along the closed-loop information transmission direction to obtain the local transmission slice sequence on each edge, specifically including:

[0108] Obtain the decoupled eigenvector representations and directional response biases of the three primitive state nodes. Based on the output of step 102, the eigenvector representation of the first primitive state node is as follows: , directional response bias is The second primitive state node corresponds to The third primitive state node corresponds to Among them, the intrinsic direction represents Each dimension characterizes the node in the 1st... The activation intensity of inherent attributes on each feature channel reflects the entity's own business characteristics, such as raw material dependence and processing capacity; directional response bias. Each dimension represents the response bias of a node to information flows from different neighborhood directions. A positive value indicates that the dimension has a positive gain tendency for transmission in a certain direction, while a negative value indicates a suppression tendency. The magnitude of the absolute value reflects the sensitivity of the response.

[0109] For the two auxiliary edges in the closed loop other than edge 1, namely edge 2 (nodes 2 to 3) and edge 3 (nodes 3 to 1), an anisotropic fusion strategy is uniformly used to construct their respective interconnection propagation tensors. Taking edge 2 as an example, the relevant vectors are first reshaped into 16×16 matrices to align with the matrix order of the manifold metric: ; ;

[0110] Among them, the reshaped matrix These are the matrix representations of the eigenvectors of the second and third primitive state nodes, respectively, with each row corresponding to an eigenvalue bundle on a local propagation channel; These are matrix representations of the directional response biases of the second and third primitive state nodes, with each row corresponding to the directional response characteristics of that channel. Set the fusion weights. This parameter determines the relative contribution ratio of intrinsic semantic coupling and directional response coupling in the interconnected transmission tensor. This means that matching is primarily based on the inherent properties of the nodes themselves, with directional response information playing a secondary modulation role. The interconnection propagation tensor of edge 2... Defined as: ;

[0111] in This is the first auxiliary conduction edge intensity modulation coefficient calculated in step 501. This coefficient is scaled overall. The amplitude of the auxiliary edge is such that the conduction strength adaptively reflects the coupling strength between the first primitive state node and the steady-state stationary point. yes The transpose of the matrix, yes The transpose of the matrix, the two terms are weighted and added together to make It captures both static attribute coupling and embeds dynamic direction-sensitive matching information; similarly, the interconnection propagation tensor of edge 3 (nodes 3 to 1) for: ;

[0112] in, The second auxiliary conduction edge intensity modulation coefficient is calculated in step 501. , These are the matrix forms of the eigenvector representation and the directional response bias of the first primitive state node, respectively. For edge 1 (nodes 1 to 2)... Since step 302 has generated the interconnection transport tensor from the first primitive state node to the second primitive state node, and converted it into a kernel matrix in step 400, Therefore, This ensures the consistency between manifold metrics and transmission definitions throughout the evaluation framework. The transmission directions of the ternary closed loop are 1 to 2, 2 to 3, and 3 to 1.

[0113] To extract channel-by-channel details along the propagation direction, the interconnect propagation tensor on each edge is... Perform row slicing operation, The A row can be viewed as an independent 16-dimensional vector, denoted as: ;

[0114] in, Representation from the starting node Along the first A local transmission channel leads to the target node. The transmitted feature components; each dimension value of this vector is the result of all feature channels of the starting node via the [missing information]. The coupling projection of each channel onto the corresponding channel of the target node has a modulus that reflects the energy intensity conducted through that channel. The larger the modulus, the stronger the energy conducted through that channel. The relative distribution between the components preserves the cross-coupling pattern between channels. This yields the slice sequence of each edge, including the edge 1 sequence, edge 2 sequence, and edge 3 sequence, all of which have a length of 16.

[0115] Step 601: Perform a coherence integral operation based on phase space entanglement on the local propagation slice sequence on each edge to obtain the coherent propagation scalar corresponding to each edge, specifically including:

[0116] A coherence integral operation is performed on the sequence of each edge in the local propagation slice sequence, compressing the structural information between slices into a coherent propagation scalar, which is used to quantify the degree of order and directional consistency of the propagation process on that edge. Taking the slice sequence of edge 1 as an example, let it be denoted as... Each slice Corresponding to the Conduction profiles of local conduction channels. Calculate the difference vectors between adjacent slices to capture the changing patterns of conduction characteristics as the channel advances: ;

[0117] Each difference vector its first Each component represents the number of channels. arrive The feature gain in that dimension (positive for enhancement, negative for decay). It is the first The conduction profile vectors of each local conduction channel. To reveal the possible ordered structure in the difference sequence, two adjacent difference vectors are concatenated into an extended phase point, and the embedding dimension is taken. Delay step size 2D embedding involves a minimal nonlinear expansion of a 16-bit sequence, effectively capturing the correlation between adjacent change steps while controlling computational cost; a delay step of 1 uses adjacent difference combinations, fully utilizing the natural order of channel indices. Phase points are thus constructed: ;

[0118] in It is an extended phase point. It is the first If the propagation changes of a difference vector exhibit inherent consistency (e.g., a stable monotonic evolution trend between channels), the phase points will tightly cluster along a certain principal direction, forming an approximately one-dimensional entangled structure; if the changes are random and disordered, the phase points tend to be diffusely distributed. When the sequence length is insufficient to form difference pairs ( If a phase point cannot be generated at this time, the coherent scalar 0 is output directly; currently... A total of 14 phase points were obtained. For all Construct the autocorrelation matrix for each phase point and obtain its second-order statistical characteristics: ;

[0119] It is a real symmetric positive semi-definite matrix. It is a phase point The transpose of (a 32-dimensional column vector). Because It is a real symmetric matrix, and there exist orthogonal matrices. and diagonal array Make: ;

[0120] Among them, eigenvalues In descending order Arrangement, corresponding feature vectors Forming an orthonormal basis, the largest eigenvalue This value represents the energy projection of the set of phase points along the most dominant direction. The larger the value, the stronger the condensation of phase points along that direction. It is an orthogonal matrix The transpose of the edge. Define the coherent transitive scalar of this edge as: ;

[0121] because Hengyou Therefore , It is a real matrix The norm of the matrix. When the phase points are arranged almost along a straight line, the matrix is ​​nearly rank one. , A value approaching 1 indicates a highly ordered conduction pattern, with stable and consistent transition modes existing between channels; when phase points are isotropically diffuse, the eigenvalues ​​are uniformly distributed. , Approaching 0 indicates that the change lacks a dominant direction and tends towards disorder. Processing the edge 2 slice sequence and the edge 3 slice sequence using the same procedure yields the coherent propagation scalar. and The final output consists of three scalar values. (Propagation coherence of edge 1). (Propagation coherence of edge 2). (The conduction coherence of edge 3) is used for the next step of closed-loop polymerization.

[0122] Step 602 involves nonlinearly coupling and aggregating the coherent transfer scalars of the three edges according to the closed-loop order to obtain the closed-loop coherent information flow intensity, specifically including:

[0123] Obtain the edge 1 coherent transfer scalar output in step 601. 2. Coherent Transmission Scalar Coherent scalar transfer with edge 3 Obtain the strength coefficient of the first auxiliary conductive edge calculated in step 501. Second auxiliary conduction edge strength coefficient .because and Already through guided coupling and With linear factor =0.85 is obtained by conversion, and its value range is... Since they remain consistent, they can be directly used as modulation factors.

[0124] because and The steps of step 600 have been embedded respectively. and During construction, therefore and The coherent transport scalar has already been calculated under intensity modulation and does not require further modulation. Let = , , It directly enters the closed-loop nonlinear coupling aggregation.

[0125] The coherent transfer scalars of the three modulated edges are nonlinearly coupled and aggregated in closed-loop order (edge ​​1, edge 2, edge 3) to generate a comprehensive closed-loop coherent information flow intensity. This is to reflect the overall synergistic quality of closed-loop transmission. The aggregation formula uses a combination of geometric mean and difference penalty: ;

[0126] smoothness coefficient Geometric mean term Calculate the geometric mean of the coherence of the three edges to reflect their overall average level. The geometric mean is particularly sensitive to low values; if the coherence of any one edge is poor, the product will be significantly lowered, thus truly reflecting the weakest link effect in the transmission chain. If any coherence scalar is zero, the product is zero, in which case we directly set... This indicates that the closed loop cannot form an effective coherent flow.

[0127] The numerator of the exponential penalty term is the sum of the absolute differences between each pair of the coherence values ​​of the three edges: ;

[0128] It reflects the degree of imbalance in coherence across the closed loop. Divided by the smoothing coefficient Then exponential decay This factor penalizes imbalances; the larger the sum of the differences, the much smaller this factor becomes than 1, ultimately reducing the circulation intensity. To control the sensitivity of the penalty, a value of 0.5 means that when the total difference... At that time, the penalty factor is approximately Moderate imbalance produces a significant inhibitory effect, causing the system to prefer a closed-loop state with uniform and coordinated conduction at all sides. The multiplication of these two factors results in… It comprehensively reflects the average coherence level and balance of closed-loop conduction. A value close to 1 indicates that not only are the three edges individually ordered, but the degree of order among them is also highly consistent, and the closed loop as a whole exhibits robust cooperative flow characteristics; a lower value indicates the presence of obvious disorder or imbalance, ultimately affecting the strength of the closed loop's coherent information flow. .

[0129] Step 603: Input the closed-loop coherent information flow intensity into a preset dimension expansion mapper composed of a multi-layer fully connected network to obtain a multi-dimensional intelligent evaluation index. The multi-dimensional intelligent evaluation index includes the structural difference degree, collaboration tightness, and transmission efficiency value of the industrial chain, specifically including:

[0130] Receive the closed-loop coherent flow intensity output in step 602 It is input into a pre-built, fully connected dimensional expansion mapper with fixed parameters to generate structural dissimilarity in one step. Coordination tightness and conduction efficiency value Three semantically defined evaluation indices. The mapper's structure consists of an input layer, two hidden layers, and an output layer connected sequentially. The input layer contains only one neuron, responsible for receiving scalar values. The first hidden layer has 8 neurons, with a corresponding weight matrix. for The vector, where each weight value represents a coefficient that linearly projects the input intensity to the corresponding hidden layer neuron, and after training, the values ​​of these coefficients are distributed in... Within the interval; bias vector It contains 8 components, each providing an independent baseline activation threshold offset for the corresponding neuron, the value of which is in The distribution is small and localized. The second hidden layer contains 16 neurons, with a weight matrix... The size is Each of its elements controls the intensity of the transformation from a feature in the first hidden layer to a neuron in the second hidden layer, and the values ​​of these elements mainly fall within... Interval; bias vector The 16 components then adaptively shift the neurons in the second hidden layer, with an amplitude not exceeding 0.05. The output layer has 3 neurons, and the weight matrix... for Each row of weights is responsible for weighting and combining the 16-dimensional hidden layer features into an exponent, with weight values ​​approximately... Within range; bias vector The three components of the equation are used to finally shift the three output exponents so that their baselines are near zero. Both hidden layers use the ReLU activation function to introduce non-linear feature representation; the output layer uses linear activation so that the three generated exponents can cover the full range of values ​​after semantic calibration.

[0131] All parameters in the mapper are determined and fixed through offline supervised training. Before training, the system first constructs a sample library covering various three-primary-state propagation scenarios. For each scenario in the library, all the aforementioned steps are run to obtain the closed-loop coherent flow intensity. As network input, the complete node and edge information of the scenario is used to calculate the target ground truth in three dimensions. The specific process is as follows: The three primitive state nodes are denoted as nodes 1, 2, and 3, respectively. For each node, the eigenvector representation obtained by decoupling in step 102 is extracted. (256 dimensions), serving as the feature vector of this node. ( The feature vector has been obtained through temporal convolutional encoding in step 100 and context-gated decomposition in step 102. Each dimension takes values ​​within the range [0,1] (guaranteed by the sigmoid activation of the gating coefficients in step 102 and the non-negativity of global average pooling in step 100). Expected structural dissimilarity value. Normalization is defined as the sum of the pairwise Euclidean distances between the three feature vectors, divided by the theoretical maximum possible distance, using the following formula: ;

[0132] in It is the theoretical maximum distance preset according to the value range of the feature components, since each feature dimension is compressed within... Within, the maximum Euclidean distance between two nodes is ,in =256 (the dimension of the high-dimensional feature vector of the node), therefore the maximum Euclidean distance is =16, the sum of the maximum distances between different pairs of nodes is . =48, this value is used to normalize the original distance. interval, making Unaffected by feature dimension and absolute scale, It is the first The feature vectors of each node. The coherence intensities of the three edges in the closed loop are denoted as follows: These are scalar indices calculated in step 602 along each level of the transmission path. Expected value of collaborative tightness. It is obtained by multiplying the harmonic average of the coherence intensities on each side by the equilibrium factor, i.e.: ;

[0133] in , , It is the reciprocal of the coherence strength of the three sides, and their product is... It can simultaneously reflect coherence strength and balance across all sides. Expected value of conduction efficiency. The arithmetic mean conduction ability, orderliness, and imbalance penalty factor are then multiplied together, resulting in the following formula: ;

[0134] arithmetic mean in the formula Represents basic transmission capacity; the imbalance penalty factor also uses the ratio of the minimum to the maximum value, with a heavier penalty for larger differences between edges; orderliness. This reflects the ordered arrangement of coherence intensity along the propagation path in the closed loop. The calculation process is as follows: First, […]. Arranged into a sequence according to the actual transmission order of the closed loop. Then, assign levels to these three values: the smallest value is assigned level 1, the second smallest level 2, and the largest level 3. If equal values ​​are found, assign the average level. This results in a level sequence. An ideal monotonically increasing sequence of ranks is fixed as follows: Calculate the difference between the two sets of rank sequences. Substituting into the Spearman rank correlation coefficient formula: ;

[0135] in The range of values ​​is ,when When increasing sequentially When decreasing sequentially When there is no monotonic relationship Further through reflection Transform the degree of order to The larger the value in the interval, the more ordered the coherence intensity along the conduction path, and therefore the higher the overall flow efficiency. Subsequently, zero-mean unit variance standardization was performed on the three indicators to unify their scales before training. The mean and standard deviation used for standardization were recorded together to solidify the semantic dimensions of each indicator.

[0136] During the training phase, the sample database will be used for... The data is fed into the mapper in batches to obtain the output under the current parameters. and the target truth value Comparison. The loss function used is mean squared error. The optimizer chosen is Adam, whose core parameters include the learning rate and momentum factor. The step size of the control parameter update, in to Select typical values ​​within the range To balance convergence speed and stability; first-order momentum decay factor An exponentially weighted average used to accumulate historical gradients imparts inertia to the update direction; a second-order momentum decay factor. An exponentially weighted average of the accumulated historical gradient squares is used to adaptively adjust the learning rate of each parameter. Training performs a maximum of 100 complete sample iterations, while monitoring the loss on the validation set after each iteration. If the minimum validation loss in the most recent 5 consecutive iterations decreases by less than 1 compared to the previously recorded best historical loss, the training is considered successful. This means that if the loss is no longer decreasing significantly, early termination is triggered to prevent overfitting. After training, All parameters are fixed and used as fixed parameters in the online inference phase.

[0137] During online inference, the closed-loop coherent flow intensity for any new input Calculated sequentially in forward order, the first hidden layer output... Through formula Received, among which scalar Expanded to 8-dimensional features, then combined with bias After summing element by element and truncating with ReLU, a non-negative 8-dimensional representation is formed. The second hidden layer output... Depend on The calculation further maps the 8-dimensional features into a 16-dimensional intermediate representation, and the final output layer performs a linear transformation. This yields a three-dimensional vector, consisting of three evaluation indices. Since the target ground truth has already been standardized during training, the network parameters already embed the corresponding scale and offset. Therefore, the inference output does not require further scaling and can be used directly.

[0138] The structural difference degree reflects the degree of characteristic dispersion of the three primitive nodes on the conduction manifold. The higher the value, the more prominent the structural diversity among the nodes, and the more coordination and adaptation may be needed in the closed loop. The degree of coordination measures the overall coupling level and balance between edges in the information transmission within the closed loop. A high value means that the coherence of each edge is not only strong on its own, but also balanced in strength with each other, indicating a tight and efficient coordination mechanism. The transmission efficiency value integrates transmission strength, path order, and edge balance. A high value indicates that the closed loop completes the flow in a low-loss and highly ordered manner, while a low value indicates that there are blockages, disorder, or severe imbalances.

[0139] This embodiment unifies the construction of interconnected transmission tensors for each edge by integrating eigenvector representations and directional response biases. Combined with closed-loop directional slicing, the transmission characteristics of each edge are transformed into a structured local sequence, preserving complete spatial directional information for subsequent coherent analysis. The coherence integral operation of phase-space entanglement extracts the intrinsic orderliness of the transmission slice sequence through differential embedding and eigenvalue decomposition, effectively suppressing the interference of noise fluctuations on the closed-loop transmission strength assessment and improving the stability and discriminative power of the coherent transmission scalar. Nonlinear coupling aggregation simultaneously considers the consistency of the overall transmission strength and the three edges, penalizing unbalanced closed-loop flow, enabling the closed-loop coherent information flow strength to sensitively reflect the synergy and abnormal interruptions in network transmission. Utilizing a pre-trained fully connected network, the single-dimensional flow strength is expanded into three explicit indices: structural difference, synergy tightness, and transmission efficiency value, achieving a refined assessment of the diversified situation of the industrial chain, with outputs that can directly serve macro-level decision-making.

[0140] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0141] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0142] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A differentiated, multi-dimensional intelligent evaluation method for the industrial chain, characterized in that: The method includes: Step 1: Collect multi-dimensional business time-series data of multiple operating entities in the target industry chain and construct an attribute heterogeneous graph; decouple the high-dimensional feature vectors of each node in the attribute heterogeneous graph in terms of direction to obtain the eigenvector representation of the inherent attributes of each node and the directional response bias of the node to the directional response intensity of the neighborhood information flow. Step 2: Based on the material flow hierarchy of the industrial chain, select the first primitive state node located in the upstream raw material supply layer, the second primitive state node located in the midstream manufacturing and processing layer, and the third primitive state node located in the downstream terminal distribution layer from the attribute heterogeneous graph; extract the eigenvalue representation and directional response bias corresponding to the first primitive state node, the second primitive state node, and the third primitive state node to form an evaluation triplet; Step 3: By evaluating the eigenvector representation and directional response bias of the first two primitive state nodes in the triplet, an anisotropic interconnection tensor is obtained. The interconnection tensor constitutes the information interaction path between the first two primitive state nodes. Step 4: On the conduction manifold determined by the information action path, perform walkpoint migration evolution to obtain the characteristic coordinates of the steady-state evaluation stationary point. The initial coordinates of the walkpoint migration evolution are obtained by aggregating the eigenvector representations of the first two primitive state nodes. Step 5: On the conduction manifold, using the interconnection conduction tensor kernel matrix as the metric tensor, calculate the cosine of the manifold inner product angle between the eigenvector representation of the first primitive state node and the characteristic coordinates of the stationary point in steady-state evaluation, as the first guiding coupling degree; calculate the cosine of the manifold inner product angle between the characteristic coordinates of the stationary point in steady-state evaluation and the eigenvector representation of the second primitive state node, as the second guiding coupling degree. Construct a first auxiliary transmission edge from the second primitive state node to the third primitive state node, and a second auxiliary transmission edge from the third primitive state node to the first primitive state node; the first auxiliary transmission edge and the second auxiliary transmission edge are connected together with the information action path to form a ternary closed-loop transmission network; the first directional coupling degree and the second directional coupling degree are used to characterize the directional consistency between the steady-state evaluation stationary point and the two primitive state nodes. Step 6: For the first auxiliary transmission edge and the second auxiliary transmission edge, calculate the corresponding interconnection transmission tensor using the eigendirection representation and directional response bias of the two primitive state nodes connected by the corresponding auxiliary transmission edge; slice the interconnection transmission tensor corresponding to each of the three edges in the ternary closed-loop transmission network along the closed-loop information transmission direction to obtain the local transmission slice sequence on each edge. Perform a coherence integral operation based on phase space entanglement on the local propagation slice sequence on each edge to obtain the coherent propagation scalar corresponding to each edge; The coherent transfer scalars of the three sides are nonlinearly coupled and aggregated according to the closed-loop order to obtain the closed-loop coherent information flow intensity. The intensity of closed-loop coherent information flow is input into a preset dimension expansion mapper composed of a multi-layer fully connected network to obtain a multi-dimensional intelligent evaluation index, which includes the structural difference, collaboration tightness and transmission efficiency of the industrial chain.

2. The multi-dimensional intelligent evaluation method for supply chain differentiation according to claim 1, characterized in that, The nodes in the attribute heterogeneous graph correspond to business entities, and the directed edges correspond to the supply and demand dependencies between entities. The direction of the edges points from the upstream supplier to the downstream purchaser. The initial attributes of the nodes are 256-dimensional feature vectors generated by temporal convolution encoding of multi-dimensional business indicators.

3. The multi-dimensional intelligent evaluation method for supply chain differentiation according to claim 2, characterized in that, The high-dimensional eigenvectors of each node in the attribute heterogeneous graph are decoupled by direction to obtain the eigenvector representation and directional response bias of each node, including: Using the node connectivity relationships in the attribute heterogeneous graph as topological constraints, a feature compatibility matrix based on high-dimensional feature vectors is constructed between nodes. Specifically, for each edge in the attribute heterogeneous graph, the cosine similarity of the high-dimensional feature vectors of the two nodes is calculated for the edge connecting node i and node j, and this similarity value is used as the compatibility value corresponding to the edge. For node pairs in the attribute heterogeneous graph that do not have any directly connected edges, their compatibility value is set to 0. All nodes are arranged by index to form a feature compatibility matrix. For each node, the context-gated decomposition of the corresponding high-dimensional feature vector is performed using the feature compatibility matrix to separate the eigenvector representation that characterizes the inherent attributes of the node itself, and the directional response bias that characterizes the directional response intensity of the node to the information flow in the neighborhood.

4. The multi-dimensional intelligent evaluation method for supply chain differentiation according to claim 3, characterized in that, Step 3 includes: The eigenvector representations of the first and second primitive state nodes are cross-mixed dimension-wise in the feature space to generate an initial mixing field. The directional response biases of the first and second primitive state nodes are used to obtain the directional selective gating coefficients through nonlinear activation. Anisotropic modulation of each characteristic dimension of the initial mixing field is performed using direction-selective gating coefficients to obtain an anisotropic modulated field; Based on the dependency distance between the first primitive state node and the second primitive state node in the attribute heterogeneous graph, a monotonically non-increasing smooth decay function is constructed to perform spatial decay processing on the anisotropic modulation field, thereby obtaining an anisotropic interconnection tensor in the feature space. The interconnection tensor constitutes the information interaction path between the first primitive state node and the second primitive state node.

5. The intelligent evaluation method for differentiated multi-dimensional industrial chain assessment according to claim 4, characterized in that, Step 4 includes: The initial characteristic coordinates of the virtual evaluation walkpoint are aggregated from the eigenvector representations of the first and second primitive state nodes, and the virtual evaluation walkpoint is initialized on the manifold determined by the interconnection transmission tensor. The interconnected conduction tensor is spectrally decomposed to obtain a set of eigenvalues ​​arranged in descending order of amplitude. The largest and second largest eigenvalues ​​are extracted from the eigenvalue set, and the ratio of the largest to the second largest eigenvalue is calculated to obtain the principal eigenvalue ratio. The number of evolution iterations is adaptively determined based on the principal eigenvalue ratio, wherein the number of evolution iterations and the principal eigenvalue ratio satisfy a monotonically non-decreasing mapping relationship. In each iteration, the negative gradient of the potential function at the current walk point is calculated to obtain the local propagation descent direction; the inertial offset generated in the previous iteration is obtained, and the local propagation descent direction and the inertial offset generated in the previous iteration are weighted and synthesized to obtain the corrected drift direction. The interconnection transmission tensor is decomposed locally at the current feature coordinates of the virtual evaluation walkpoint to obtain the local curvature tensor. The trace of the local curvature tensor is calculated to obtain the local structure coefficients. The reciprocal of the local structure coefficients is multiplied by the preset reference step size to obtain the compensation step size. The virtual evaluation walkpoint is moved along the corrected drift direction to compensate for the step size, and the feature coordinates of the virtual evaluation walkpoint are updated. The corrected drift direction generated in this iteration is recorded as the inertial offset for the next iteration. After the number of evolution iterations is updated, the final feature coordinates of the virtual evaluation walkpoint are used as the feature coordinates of the steady-state evaluation stationary point.

6. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 5.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 5.

Images