An intelligent matching system for supply and demand of new material technology based on a knowledge graph

Abstract

The application discloses a new material technology supply and demand intelligent matching system based on a knowledge graph. A document analysis module processes unstructured supply and demand documents, and generates a structured material parameter signal by means of an edge-enhanced graph attention network; a semantic normalization module performs interval number embedding coding on parameter values, and outputs a normalized parameter vector signal; a heterogeneous supply and demand graph reasoning module constructs a heterogeneous information network, completes supply and demand matching scoring through a meta-path aggregation network, and synchronously generates a matching explanation path signal; a space-time filtering module calls a time sequence knowledge graph to calculate the real-time effective performance of materials, and eliminates supply nodes that do not meet the performance standard; and a matching output module outputs results according to scoring and ranking, and is matched with explainable information. The system effectively solves the mis-matching problem caused by new material supply and demand semantic heterogeneity and material performance degradation.

Description

Technical Field

[0001] This invention relates to the field of intelligent matching technology for supply and demand of new materials based on knowledge graphs, and specifically to an intelligent matching system for supply and demand of new materials technology based on knowledge graphs. Background Technology

[0002] In the research and development and commercialization of new materials, technology suppliers and demanders typically publish or obtain information in the form of unstructured documents, such as experimental test reports, material specifications, and project requirement documents. These documents often contain complex two-dimensional tables with irregular structures such as merged cells and multi-level headers. The material performance parameters contained in these tables have multiple forms of expression, and the same technical indicator may be expressed using completely different terms in different documents or institutions, such as "tensile strength" and "tensile strength," "elastic modulus" and "Young's modulus." Numerical values ​​are often given in the form of ranges or limits, such as "greater than or equal to 500 MPa" or "300 to 350 MPa." Existing material supply and demand matching technologies mostly use information retrieval methods based on keywords or rule templates. When faced with the above heterogeneous expressions, these methods are prone to omissions or mismatches due to terminology mismatches and the inability to handle range semantics. For documents containing complex tables, conventional optical character recognition or text extraction processes linearly concatenate cell contents, severing the spatial topological relationship between parameter names, values, and units. This results in misaligned parameters and values ​​in the extraction results, severely impacting matching accuracy.

[0003] Furthermore, the demand for high-end materials is often reflected in functional descriptions rather than precise parameter lists, such as "high-temperature resistant lightweight structural materials." Such demands implicitly require specific material categories, composite systems, and the synergistic fulfillment of multiple performance characteristics. Existing systems based on surface word matching or single-parameter precise matching cannot infer candidate materials such as silicon carbide-based composites from "high-temperature resistant," lacking the ability to reason about the implicit relationships between supply and demand. On the other hand, the performance of many engineering materials is not constant and degrades over time and with the environment, such as the aging of polymer materials and the decline in corrosion resistance of coating materials. Traditional supply-demand matching technologies construct static databases or static knowledge graphs, which neither record the degradation patterns of material performance over time nor calculate their current effective performance during matching. This results in recommendations containing a large amount of supply information that no longer meets usage requirements. Although knowledge graph technology has been used to enhance material data association, existing material knowledge graphs are mostly static triple structures, lacking modeling methods for temporal attributes, and the matching process generally lacks interpretability. Users cannot understand why a particular supply is recommended or ranked highly, making it difficult to rely on them in key material selection decisions. In summary, these shortcomings result in low efficiency and poor reliability in matching supply and demand for new materials technologies. There is an urgent need for a technical solution that can accurately extract parameters from unstructured documents, unify heterogeneous semantics, support implicit relational reasoning and dynamic performance evaluation, and output matching explanations. Summary of the Invention

[0004] In view of the shortcomings of the prior art, the purpose of this invention is to provide a novel intelligent matching system for supply and demand of materials based on knowledge graphs, which solves the problems of semantic heterogeneity and performance degradation mismatch in material supply and demand. This invention receives unstructured supply and demand documents, uses an edge-enhanced graph attention network to perform forward propagation on text nodes and delimiter nodes to form structured material parameter signals, normalizes heterogeneous terms and range values ​​through material domain ontology contrastive learning and interval number embedding encoding to form normalized parameter vector signals, constructs a heterogeneous supply and demand graph, and uses a predefined meta-path execution aggregation network to score the connection probability of supply and demand node pairs to obtain matching interpretation path signals. Simultaneously, it accesses a temporal knowledge graph to calculate the current effective performance value of the material according to a degradation function and prunes ineffective nodes to form spatiotemporally effective candidate signals, finally outputting a ranked matching result with interpretable paths.

[0005] This invention provides a knowledge graph-based intelligent matching system for the supply and demand of new materials technologies, comprising: The document parsing module receives unstructured supply and demand documents and uses an edge-enhanced graph attention network to perform forward propagation on text nodes and delimiter nodes to form structured material parameter signals. The semantic normalization module receives the structured material parameter signal and performs interval number embedding encoding on the parameter values ​​to form a normalized parameter vector signal. The heterogeneous supply and demand graph reasoning module receives normalized parameter vector signals, constructs a heterogeneous information network containing supply nodes, demand nodes, material nodes, and parameter nodes, executes a meta-path aggregation network based on predefined meta-paths, scores the connection probability of supply and demand node pairs, and forms a matching score signal and a matching interpretation path signal. The spatiotemporal filtering module receives the matching scoring signal, accesses the time-series knowledge graph storing material performance degradation curves, calculates the current effective performance value of the material based on the timestamp, removes supply nodes that do not meet the demand threshold, and forms spatiotemporally effective candidate signals. The matching output module receives valid candidate signals in spatiotemporal space, sorts the matching results according to the matching score, and outputs the meta-path instance with the highest activation weight as an interpretability information signal.

[0006] In one embodiment of the present invention, the document parsing module is further configured to parse unstructured supply and demand documents into a layout-aware heterogeneous graph, wherein detected text fragments are constructed as text nodes, visual dividing line segments are constructed as separator nodes, and row edges and column edges are constructed based on the spatial alignment relationship between text nodes, and spatial position edges are constructed based on the spatial adjacency relationship between text nodes and separator nodes. The initial feature representation of the text nodes and the edge feature representation containing two-dimensional coordinate differences and edge type embeddings are input into the edge-enhanced graph attention network. Message passing and feature aggregation are performed along the row edges and column edges through a multi-head attention mechanism to generate a text node embedding vector carrying cell logical coordinate attribution information, and a structured material parameter signal is output accordingly.

[0007] In one embodiment of the present invention, the semantic normalization module incorporates a material domain ontology graph. The material domain ontology graph uses material performance terms as entity nodes and synonym and hierarchical relationships as edges. When fine-tuning the pre-trained language model, terms in the structured material parameter signal are used as anchor points. First-order neighbor entities in the material domain ontology graph are used as positive examples, and randomly sampled entities are used as negative examples. An ontology-guided contrastive learning loss function is constructed. By maximizing the vector similarity between the anchor point and the positive example and minimizing the vector similarity between the anchor point and the negative example, the same material property expressed heterogeneously is mapped to a neighboring position in the semantic vector space. During the inference stage, the terms in the structured material parameter signal are converted into semantic vectors and concatenated with the normalized interval number embedding vector to form a normalized parameter vector signal.

[0008] In one embodiment of the present invention, when the semantic normalization module performs interval number embedding encoding on the parameter values, it represents parameter values ​​with range limitations and parameter values ​​with lower or upper limits as interval number vectors, respectively. The interval number vector includes the interval midpoint vector and the interval radius vector. When calculating the overlap between two interval numbers, the degree of intersection is determined based on the distance between the interval midpoints and the interval radius. The degree of intersection is used as the basis for calculating the numerical dimension similarity in subsequent matching scores. The formula for calculating the numerical dimension similarity is as follows: in, The value is the midpoint of the required parameter range. The value is the midpoint of the supply parameter range. The value is the radius of the required parameter range. To provide the radius value of the supply parameter range, Let be the length of the intersection of the two parameter intervals. The absolute distance is the midpoint of the interval. This is the similarity normalization coefficient. The similarity is based on the numerical dimension of the parameter.

[0009] In one embodiment of the present invention, the predefined meta-path of the heterogeneous supply and demand graph inference module includes at least an exact matching meta-path and an inference matching meta-path. The node sequence of the exact matching meta-path is from the demand node through the demand parameter node to the supply parameter node and then to the supply node. The node sequence of the inference matching meta-path is from the demand node through the application scenario node to the material grade node and then to the supply node. When constructing the heterogeneous information network, the heterogeneous supply and demand graph inference module extracts the demand parameter entity and the supply parameter entity from the normalized parameter vector signal, and establishes alignment edges for node pairs in the demand parameter entity and the supply parameter entity whose semantic vector similarity exceeds a preset threshold. At the same time, the application scenario entity is extracted from the demand description and linked to the scenario node, and the material grade entity is extracted from the supply information and linked to the material grade node.

[0010] In one embodiment of the present invention, the process of the heterogeneous supply and demand graph inference module executing the meta-path aggregation network includes: for each pair of supply nodes and demand nodes, traversing all predefined meta-paths, aggregating the features of neighboring nodes on each meta-path through path instance sampling and attention mechanism to obtain the supply node embedding representation and demand node embedding representation under the meta-path; concatenating the supply node embedding representation and demand node embedding representation and inputting them into a multilayer perceptron to obtain the matching contribution score corresponding to the meta-path; performing a weighted summation of the matching contribution scores of all meta-paths, with the weights being learnable meta-path importance parameters, and mapping the weighted summation result to a connection probability score through a nonlinear activation function, the calculation formula of the connection probability score being as follows: Where P is the connection probability score of the supply and demand node pair. Here, K is the total number of predefined meta-paths in the system, and k is the meta-path index. Let k be the learnable importance weights corresponding to the kth meta-path. The matching contribution score of the k-th meta-path output by the multilayer perceptron. Similarity in terms of parameter numerical dimensions. The node embedding feature balance coefficient, Embed vectors for the features of the demand nodes. Embed the feature vector of the supply node. The L2 distance is the embedding vector between the demand node and the supply node.

[0011] In one embodiment of the present invention, when the heterogeneous supply and demand graph inference module forms a matching interpretation path signal, it records the attention weight and instance path sequence of each meta-path during the execution of the meta-path aggregation network, selects the meta-path with the largest contribution to the connection probability score as the meta-path instance with the highest activation weight, and concatenates the node sequence and node attribute values ​​traversed by the meta-path instance into an interpretation description in natural language form. This interpretation description and the matching score signal are transmitted to the spatiotemporal filtering module together.

[0012] In one embodiment of the present invention, in the time-series knowledge graph accessed by the spatiotemporal filtering module, the material performance degradation curve is stored in the form of time-series attributes of material entity nodes. The time-series attributes include initial performance value, degradation coefficient, and timestamp sequence. The current effective performance value is calculated as the product of the initial performance value and a time decay function, which is a negative exponential function with the degradation coefficient as the exponent. The decay factor is calculated based on the time difference between the current system time and the attribute record timestamp, and the calculated current effective performance value is compared with the demand threshold. When the current effective performance value is lower than the demand threshold, the corresponding supply node is marked as failed and removed from the candidate set. A spatiotemporal effective candidate signal is formed by traversing node by node. The formula for calculating the current effective performance value is as follows: in, This represents the current effective performance value of the material. These are the initial property values ​​of the material. The material property degradation coefficient, This is the current system timestamp. For the initial record timestamp of material properties, This is the time difference. Let C be the performance decay time constant, and C be the time decay correction constant. This is the square of the time difference.

[0013] In one embodiment of the present invention, the spatiotemporal filtering module is also connected to an external sensor data interface or a detection report input interface. When new performance detection data is received, the performance detection data is input as a new observation value of the time series into the gated recurrent unit network. The update gate and reset gate mechanism of the gated recurrent unit network are used to update the degradation state hidden vector of the corresponding material entity. The degradation coefficient is refitted based on the updated hidden vector. The updated degradation coefficient is written back to the corresponding time series attribute in the time series knowledge graph. In the next matching request, the calculation of the current effective performance value and the selection of supply nodes are re-executed based on the updated degradation coefficient.

[0014] In one embodiment of the present invention, when the matching output module outputs the matching results according to the matching score, it sorts the supply nodes contained in the spatiotemporal valid candidate signals in descending order of connection probability score, and combines the demand parameter node attributes, supply parameter node attributes and intermediate associated node information traversed by the meta-path instance with the highest activation weight into an interpretable description text, which is output to the interactive interface or data interface along with the sorted supply node identifier.

[0015] This invention provides a novel intelligent matching system for supply and demand in materials technology based on knowledge graphs. The system receives unstructured supply and demand documents, uses an edge-enhanced graph attention network to perform forward propagation on text nodes and delimiter nodes to form structured material parameter signals, and normalizes heterogeneous terms and range values ​​through material domain ontology comparison learning and interval number embedding encoding to form normalized parameter vector signals. It constructs a heterogeneous supply and demand graph and uses a predefined meta-path execution aggregation network to score the connection probability of supply and demand node pairs to obtain matching interpretation path signals. Simultaneously, it accesses a temporal knowledge graph to calculate the current effective performance value of the material according to a degradation function and removes ineffective nodes to form spatiotemporally effective candidate signals. Finally, it outputs a ranked matching result with interpretable paths. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A system architecture diagram of a knowledge graph-based intelligent matching system for the supply and demand of new materials technologies. Figure 2 This is a flowchart illustrating the internal workflow of the semantic normalization module. Figure 3 A flowchart illustrating the reasoning and matching interpretation path for heterogeneous supply and demand graphs. Detailed Implementation

[0018] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0019] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0020] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0021] Please see Figure 1-3 The diagram illustrates a novel intelligent matching system for supply and demand in materials technology based on a knowledge graph, comprising: a document parsing module that receives unstructured supply and demand documents and uses an edge-enhanced graph attention network to perform forward propagation on text nodes and delimiter nodes to form structured material parameter signals; a semantic normalization module that receives the structured material parameter signals and performs interval number embedding encoding on the parameter values ​​to form normalized parameter vector signals; a heterogeneous supply and demand graph reasoning module that receives the normalized parameter vector signals, constructs a heterogeneous information network containing supply nodes, demand nodes, material nodes, and parameter nodes, executes a meta-path aggregation network based on predefined meta-paths, scores the connection probability of supply and demand node pairs, and forms a matching score signal and a matching interpretation path signal; a spatiotemporal filtering module that receives the matching score signal, accesses a temporal knowledge graph storing material performance degradation curves, calculates the current effective performance value of the material based on timestamps, and removes supply nodes that do not meet the demand threshold to form spatiotemporally effective candidate signals; and a matching output module that receives the spatiotemporally effective candidate signals, sorts the matching results according to the matching scores, and outputs the meta-path instance with the highest activation weight as an interpretability information signal.

[0022] Figure 1As shown in the flowchart, this flowchart presents the complete overall operation of a knowledge graph-based intelligent matching system for the supply and demand of new materials technologies. It integrates main business processes, branch filtering, and dynamic parameter updates, with each functional stage seamlessly connected to complete the entire process from raw data input to matching result output. The process begins with receiving unstructured supply and demand documents. This stage is the data entry point for the entire system, responsible for uniformly collecting various forms of new material supply and demand data. These raw documents are generally characterized by messy formats and irregular layouts. This stage summarizes and integrates the raw data, providing a basic data source for subsequent processing. After data integration, the data flows to the document parsing module. This module performs layout parsing and node construction on the unstructured documents, combining graph attention networks to complete feature transfer and aggregation, transforming the scattered and irregular raw documents into structured material parameter signals with standardized formats and clear logic, achieving the initial conversion from unstructured data to structured data. The structured signal is then passed to the semantic normalization module, which performs semantic unification and numerical encoding on material terminology and interval parameter values. This eliminates matching obstacles caused by differences in expression and parameter format between different documents and outputs a standardized normalized parameter vector signal.

[0023] The data then enters the heterogeneous supply and demand graph inference module. This module constructs a heterogeneous information network composed of multiple types of entity nodes, completes feature calculations based on preset inference paths, and simultaneously generates quantified matching score signals and matching interpretation path signals for traceability, intuitively reflecting the matching degree of different supply and demand combinations. The data then enters the current effective performance value determination stage, which is the core branch node in the process. It combines real-time material performance data to conduct compliance verification, and the process splits into two directions here. When the current performance of the material meets the usage standards, the process proceeds to the stage of generating spatiotemporally effective candidate signals. This stage integrates all compliant supply information to form a complete candidate dataset. When the material performance fails to meet the demand standards, it enters the stage of marking and eliminating failed supply nodes, directly removing unqualified supply information from the candidate set and reducing the computational resources occupied by invalid data.

[0024] After filtering, the valid data is fed into the matching output module. This module sorts and integrates the candidate data, ultimately outputting the matching results and interpretable information, displaying the sorted matching list and corresponding matching criteria to the terminal. In addition to the main business logic, the system also has an independent cyclic update loop. External sensor / test report interfaces continuously collect offline measured performance data of materials. New data is sent to the entry control loop unit to update the degradation coefficient. A deep learning network is used to refit the material performance degradation law, correcting the original degradation parameters. Then, a write-back operation to the time-series knowledge graph is performed, storing the updated coefficients. The updated graph data then flows back to the spatiotemporal filtering module, forming a closed loop. This allows the system to continuously optimize the matching rules based on the real-time status of the materials. The entire process considers basic matching, real-time filtering, and dynamic iteration, fully adapting to the actual application scenarios of new material supply and demand matching.

[0025] Specifically, the knowledge graph-based intelligent matching system for the supply and demand of new materials technologies begins its operation with unstructured document input. When a materials testing report or requirement specification enters the system as a document, the document parsing module first performs page element detection and graph structure construction. Each text segment in the document is extracted as an independent text node, and table separator lines on the page are identified as separator nodes. Alignment edges are constructed between text nodes based on their two-dimensional coordinates in the same row or column direction. Spatial position edges are established between text nodes and adjacent separator nodes, thus forming a layout-aware heterogeneous graph that preserves the original document's table topology. The edge features of this graph are formed by concatenating the two-dimensional coordinate difference between two nodes with the edge type embedding vector, while the initial features of the text nodes are encoded and output by a pre-trained language model. After the layout-aware heterogeneous graph is constructed, the edge-enhanced graph attention network begins forward propagation. This network employs a multi-head attention mechanism, calculating attention weights for each text node's neighboring nodes. During message passing, not only are the feature vectors of neighboring nodes transmitted, but edge features are also included in the calculation of the attention coefficients, giving nodes in the same table row or column a stronger information aggregation weight. After multiple layers of forward propagation, each text node generates an embedding vector carrying the logical coordinates of its cell. Based on this, the system reorganizes the originally scattered text fragments into structured material parameter signals with material attribute names as keys and parameter values ​​and units as values.

[0026] After receiving the structured material parameter signals, the semantic normalization module processes both the semantic meaning of the terms and the numerical values ​​of the parameters. This module incorporates a material domain ontology graph, where material performance terms serve as entity nodes, and synonymous or hierarchical relationships between terms are connected by edges. Before the system is deployed, the semantic normalization module fine-tunes the pre-trained language model using an ontology-guided contrastive learning loss function. During fine-tuning, terms appearing in the structured material parameters are used as anchor samples. First-order neighbor entities of the anchors are sampled from the ontology graph as positive examples, and irrelevant terms are randomly sampled from the global term set as negative examples. By maximizing the vector similarity between the anchor and positive examples and minimizing the vector similarity between the anchor and negative examples, heterogeneous expressions such as "tensile strength" and "tensile strength" are mapped to adjacent positions in the semantic vector space. The parameter values ​​are processed using interval number embedding encoding. Any parameter value given in range form or as a lower or upper limit is encoded as an interval number vector. This vector consists of a midpoint vector and a radius vector, comprehensively representing the uncertain range of the parameter value rather than a single scalar value. The terminology, after fine-tuning, is converted into a semantic vector, which is then concatenated with the interval number embedding vector to form a normalized parameter vector signal for downstream transmission. The formula for calculating numerical similarity is as follows: In practical operation, the knowledge graph-based intelligent matching system for supply and demand of new materials addresses the industry characteristics of unstructured supply and demand documents and the common use of interval values, single upper limits, or single lower limits for material parameters. It introduces algorithms for interval intersection and numerical dimension similarity in the semantic normalization module to achieve quantitative comparison of parameter values ​​of different forms, completely solving the technical problem that traditional single-point numerical matching cannot adapt to interval-type material parameters. This algorithm includes two sets of correlation calculation formulas. The first formula calculates the interval intersection length of demand and supply parameters, and the second formula calculates the numerical dimension similarity of parameters based on the intersection length and interval distribution characteristics. The value is the midpoint of the required parameter range. The value is the midpoint of the supply parameter range. The value is the radius of the required parameter range. To provide numerical values ​​for the radius of the parameter interval, the four types of basic variables are extracted from the interval number vector by the semantic normalization module, which is the core data source for performing interval operations; The formula represents the actual intersection length of two sets of parameter intervals. It uses two sets of max functions to determine the relative positions of the left and right boundaries of the intervals, and then uses the difference operation to obtain the effective intersection length. When the two sets of parameter intervals have no overlap, the calculation result will automatically return to zero, directly determining that there is no matching basis between the two in the numerical dimension. This operation logic can fully cover the three actual scenarios of complete intersection, partial intersection, and non-intersection of the two intervals, and adapt to the diverse writing forms of new material parameters. The final parameter numerical similarity is composed of the proportion of interval intersection and the midpoint offset penalty term, where σ is the similarity normalization coefficient. Its function is to unify the similarity calculation dimensions of material parameters of different categories and magnitudes, and to eliminate the interference of parameter numerical values ​​on the matching results. The algorithm constructs a midpoint offset penalty mechanism by combining the absolute distance between the midpoints of the supply and demand parameter intervals with an exponential function. The larger the difference between the midpoints of the intervals, the smaller the value of the exponential term, ultimately reducing the overall similarity score. This aligns with the matching principle in the field of new material applications: "prioritizing the overlap of parameter intervals while also considering the deviation of the central index." This algorithm transforms the originally fuzzy interval overlap relationship into a standardized quantitative index in the 0 to 1 range, completing the deep normalization processing of the structured material parameter signals in the numerical dimension. The calculated numerical dimension similarity is not isolated but is transmitted as core feature data to the downstream heterogeneous supply and demand graph inference module, becoming an important input item for subsequent meta-path matching scoring. This enables algorithmic linkage between the semantic normalization module and the heterogeneous graph inference module, allowing the system's parameter matching to simultaneously consider semantic expression and engineering numerical constraints. This ensures the rationality and accuracy of subsequent matching operations from the source and also provides the underlying numerical support for the entire intelligent matching algorithm system.

[0027] like Figure 2As shown in the flowchart, this flowchart illustrates the internal workflow of the semantic normalization module, which is divided into two parallel business branches: semantic processing and numerical encoding. It also incorporates iterative model training logic, with each step clearly defined and working in concert to ultimately achieve full-dimensional standardization of material parameter signals. The process begins with receiving structured material parameter signals. This step specifically receives the structured data output from the upstream document parsing module and serves as the data source foundation for all computations in this module. All subsequent semantic and numerical processing revolves around this signal. From this initial step, the process naturally splits into two parallel processing links. The first link focuses on the semantic unification of material terminology. It begins with extracting material performance terms, where staff accurately identify and extract various professional terms describing material properties and attributes from the structured parameter signals. These terms are the core processing objects for semantic normalization. After terminology extraction, the module proceeds to constructing positive and negative samples based on the ontology graph. The module calls an internally pre-built material domain ontology graph, using the extracted terms as core anchors. Closely related first-order neighbor entities in the ontology graph are selected as positive samples, and unrelated entities are randomly selected as negative samples, thus building a sample set suitable for model training. After the sample construction is completed, the process moves to the stage of constructing an ontology to guide the comparative learning of the loss function. Loss calculation rules are formulated based on the feature differences between positive and negative samples to constrain the model's learning direction and optimization objectives. The process then proceeds to the loss function convergence determination stage, which constitutes the cyclical process of model training. If the loss function does not reach the preset convergence criterion, the process returns to the stage of constructing an ontology to guide the comparative learning of the loss function, continuously iterating and adjusting model parameters to narrow the distance between synonyms and near-synonyms in the vector space. When the loss function converges and the model training achieves the expected results, the process moves to the stage of converting terms into standard semantic vectors. Using the trained model, terms from similar materials with varying expressions are uniformly converted into standard semantic vectors, completing semantic standardization.

[0028] The second parallel link is responsible for encoding the parameter values. It first enters the material parameter value analysis stage, decomposing and classifying various numerical parameters in the structured signal, distinguishing between different forms such as range values ​​and single-boundary limits. After numerical analysis, it enters the stage of generating interval midpoint vectors + radius vectors. Following the rules for expressing interval numbers, various range-type parameters are decomposed into midpoint and radius vectors, fully preserving the interval attribute characteristics of the parameters. Then, the interval number embedding encoding stage is executed, encoding and converting the decomposed vectors to generate numerical feature vectors suitable for model operations. The results of the two branch links finally converge at the semantic vector and interval vector concatenation stage, fusing and integrating the semantic feature vectors with the numerical feature vectors to form a combined feature that possesses both semantic and numerical information. Finally, the process enters the output normalized parameter vector signal stage, transmitting the integrated normalized vector signal to downstream modules. The entire internal process relies on a dual-branch parallel mode to achieve synchronous unification of semantics and numerical values, and uses a loop iteration mechanism to ensure the accuracy of semantic mapping. This effectively solves the industry pain points of messy terminology and inconsistent parameter formats in the field of new materials, and provides high-quality data support with unified format and consistent semantics for subsequent heterogeneous graph reasoning and intelligent matching.

[0029] After receiving the normalized parameter vector signal, the heterogeneous supply and demand graph inference module extracts the parameter entity sets of the demand side and the supply side. It also extracts application scenario entities from the demand description and material grade entities from the supply information, using these entities as nodes to construct a heterogeneous information network. The network includes demand nodes, supply nodes, material grade nodes, demand parameter nodes, supply parameter nodes, and application scenario nodes. Edges are established between nodes based on entity co-occurrence relationships and semantic similarity. If the semantic vector similarity between a demand parameter node and a supply parameter node exceeds a preset threshold, an alignment edge is established. This module predefines two types of core meta-paths. The first type is the exact matching meta-path, whose node sequence is from the demand node through the demand parameter node to the supply parameter node and then back to the supply node. This is used to capture scenarios where the demand parameters and supply parameters are directly matched numerically and semantically. The second type is the inference matching meta-path, whose node sequence is from the demand node through the application scenario node to the material grade node and then back to the supply node. This is used to capture the implicit inference chain where the functional description proposed by the demand side points to a specific material category through the application scenario, and then points to the supplier that can provide the material from the material category. For each supply node and demand node to be matched, the system traverses the above meta-paths, samples path instances along each meta-path, and aggregates the features of neighboring nodes on the path through an attention mechanism to obtain the embedded representations of the supply node and demand node under that meta-path. The embedded representations of the two are concatenated and fed into a multilayer perceptron, which outputs the matching contribution score corresponding to the meta-path. The contribution scores of all meta-paths are weighted and summed through learnable meta-path importance weights, and then mapped to the final connection probability score through a nonlinear activation function. During this process, the system synchronously records the attention weight and the sequence of nodes traversed by each meta-path instance. It selects the meta-path that contributes most to the connection probability score, extracts its complete node path and node attribute values, concatenates them into a natural language matching explanation, and outputs it along with the matching score signal. The formula for calculating the connection probability score is as follows: Following the similarity of parameter numerical dimensions output by the semantic normalization module, the heterogeneous supply and demand graph inference module, relying on the constructed heterogeneous information network and multiple predefined meta-paths, employs a supply and demand node pair connection probability scoring algorithm to complete the comprehensive matching degree calculation through multi-feature fusion. This algorithm integrates meta-path inference features, parameter numerical similarity features, and node semantic embedding features, and is the core computational link for the entire system to achieve intelligent matching. In the formula, P is the final connection probability score of the supply and demand node pair, and its value range is constrained between 0 and 1 by the Sigmoid nonlinear activation function σ. The closer the score is to 1, the better the comprehensive matching effect of the supply and demand node pair; the lower the score, the worse the matching feasibility. K is the total number of predefined meta-paths in the system, covering different path types such as precise matching meta-paths and inference matching meta-paths. k is the traversal sequence number of the meta-path, used to calculate the matching contribution result corresponding to each meta-path in turn. The learnable meta-path importance parameter corresponding to the k-th meta-path supports autonomous iterative updates of the model. It can adaptively adjust the weight ratio of each meta-path according to different new material application scenarios. For example, the weight of the precise matching meta-path is higher in the general material matching scenario, and the inference matching meta-path will be given higher weight in the cross-domain new material application scenario, thus overcoming the shortcomings of the traditional fixed weight algorithm in terms of insufficient flexibility. It is the matching contribution score output by a single meta-path after path instance sampling, neighbor node feature aggregation, and multilayer perceptron operation, which individually represents the degree of supply and demand matching under a single path logic; α is the node embedding feature balance coefficient, which is used to dynamically adjust the weight ratio of parameter numerical similarity and node semantic feature similarity, and can focus on hard parameter constraints or semantic logical association according to business needs. and These are the feature embedding vectors for the demand node and the supply node, respectively, generated by the meta-path aggregation network aggregating the features of all neighboring nodes along the path. The L2 distance between two sets of vectors is used to quantify the degree of difference between supply and demand nodes in the overall semantic feature space. During algorithm execution, firstly, for a single meta-path, the parameter numerical similarity and node vector distance are fused and calculated. Then, the calculation results for all meta-paths are weighted and summed using learnable weights. Finally, a standardized connection probability score is output through a nonlinear activation function. This algorithm fully leverages the complementary advantages of multiple meta-paths in heterogeneous information networks, while deeply reusing the calculation results of the preceding interval similarity algorithm. This allows the matching score to consider both hard parameter indicators, path reasoning logic, and node semantic associations. The calculated matching score signal and the matching interpretation path signal are synchronously transmitted to the spatiotemporal filtering module, providing complete pre-matching data for subsequent dynamic screening based on material temporal performance.

[0030] After receiving the matching score signal and the matching interpretation path signal, the spatiotemporal filtering module does not directly send the scoring result to the output. Instead, it first performs a time-based validity review of the supply nodes. This module maintains a temporal knowledge graph. In addition to regular attributes, each material entity node in the graph carries extended attributes related to time, including initial performance value, degradation coefficient, and a timestamp recording the most recent update time of that attribute. The calculation of the current effective performance value adopts a negative exponential time decay model, that is, the initial performance value is multiplied by a decay factor with the degradation coefficient as the exponent and the difference between the current system time and the recorded timestamp as the independent variable. The system traverses the material entities corresponding to each supply node involved in the matching score signal, calculates the current effective performance value of its key performance parameters one by one, and compares this value with the demand threshold set by the demand side. Any supply node whose current effective performance value is lower than the demand threshold is marked as invalid and removed from the candidate set. Only the supply nodes whose performance still meets the requirements at the current time are retained, forming a spatiotemporally effective candidate signal. When new performance test data is submitted via external sensor data interface or manually input test reports, this module inputs the new observations as time-series data into a gated recurrent unit network. The network's update gate controls the migration degree of historical hidden states to the new state, and the reset gate controls the degree to which historical information is forgotten. The network outputs an updated degradation state hidden vector. Based on this, the system refits the degradation coefficients and writes them back to the time-series attribute fields of the corresponding entities in the time-series knowledge graph, ensuring that subsequent matching requests are always based on the latest material performance degradation patterns for node selection. The formula for calculating the current effective performance value is as follows: After the heterogeneous supply and demand graph reasoning module filters out candidate supply nodes with high matching probability, the spatiotemporal filtering module relies on the temporal knowledge graph of the stored material performance degradation curves and uses an algorithm based on the current effective performance value of the material to conduct a secondary filtering in the temporal dimension. This compensates for the shortcomings of static parameter matching, which ignores the performance degradation of materials during long-term storage and use, making the matching results more consistent with the actual service scenarios of new materials. This algorithm combines a negative exponential decay model, a squared correction term, and a constraint function to accurately simulate the performance degradation law of materials over time, where... The current effective performance value of the material is the core criterion for determining whether the supply node meets the demand threshold. The initial performance values ​​of the material are directly retrieved from the temporal attributes of the material entity nodes in the temporal knowledge graph, representing the original performance indicators of the material when it leaves the factory or is first tested. The degradation coefficient is determined by the material itself and the processing technology. It also supports dynamic updates of the gated loop unit network combined with external sensor data and test reports to ensure the real-time effectiveness of degradation parameters. This is the system's current timestamp. The timestamps corresponding to the initial performance data of the material are used to record the material. The difference between the two timestamps represents the cumulative storage or usage time of the material. C is the performance degradation time constant, used to regulate the overall rate of performance degradation, and can be configured differently according to the life cycle of different materials; C is the time degradation correction constant, specifically calculated and corrected for materials stored for ultra-long periods, to avoid the calculation results deviating from actual operating conditions over long periods. The formula is divided into two main calculation parts, the first part... It is the classic negative exponential decay calculation formula, and also the core expression form of the material performance degradation curve in the time-series knowledge graph, accurately reflecting the performance decline law caused by the natural aging of raw materials; Part Two For the duration correction term, the negative impact of excessively long storage time on material performance is amplified through time difference squared operation. Simultaneously, a maximum value function is used to fix the lower limit of the correction term to 0, preventing calculation results that violate engineering common sense, such as negative performance values. The module compares the calculated current effective performance value with the performance threshold preset by the demand side, marking supply nodes with substandard effective performance values ​​as invalid and removing them from the candidate set, ultimately generating spatiotemporally effective candidate signals. Looking at the entire algorithm system, three sets of calculation formulas form a progressive and closely interconnected complete computational chain: first, basic screening at the parameter numerical level is completed using interval number intersection and similarity algorithms; then, a supply-demand node connection probability scoring algorithm is used to integrate semantic, path, and numerical features to achieve comprehensive matching and ranking; finally, a current effective performance value algorithm is used to complete dynamic filtering at the temporal performance dimension. This process completes the entire screening from three core dimensions: static parameters, logical reasoning, and temporal performance, comprehensively covering various constraints in the supply-demand matching of new material technologies. This ensures that the system's output matching results meet both parameter requirements and the actual usage state of the material, significantly improving the practicality, reliability, and implementation value of the entire intelligent matching system in industrial scenarios.

[0031] like Figure 3As shown, this flowchart illustrates the core operation of the heterogeneous supply and demand graph reasoning module. Each step connects the entire process, including network construction, multi-path reasoning, feature calculation, and result interpretation. The flowchart incorporates multiple parallel branches and loop traversal logic, simultaneously completing the two core functions of matching score calculation and interpretable path generation. The initial step involves receiving the normalized parameter vector signal. This step receives standardized vector data output from the upstream semantic normalization module, upon which all network construction, path reasoning, and score calculation are based. After data reception, the process moves to constructing the heterogeneous information network. Based on the normalized parameter signal, various entity information such as demand, supply, materials, and parameters is extracted to build a heterogeneous information network containing multiple types of nodes. Simultaneously, alignment edges and association edges are constructed based on the semantic and parameter relationships between entities, completing the construction of the network carrier required for reasoning operations. After the network is set up, the system enters the meta-path traversal phase. It sequentially calls various pre-defined inference paths. Here, the process splits into two parallel meta-path branches: the precise matching meta-path and the inference matching meta-path. These two paths correspond to different matching logics. The precise matching meta-path focuses on direct parameter matching, while the inference matching meta-path relies on application scenarios and material grades to conduct correlation inference, uncovering the inherent connections between supply and demand from different dimensions. After the two meta-path branches complete their initial processing, the data is uniformly merged into the path instance sampling + neighbor feature aggregation phase. This phase performs instance sampling on each inference path, aggregating the feature information of all adjacent nodes on the path using an attention mechanism, and extracting the comprehensive features of the demand and supply nodes respectively. This phase also splits the data into two independent subsequent links: one for matching score calculation and the other for interpreting path data records.

[0032] The link for score calculation first enters the stage of calculating the matching contribution score of a single meta-path. The aggregated node features are input into a multilayer perceptron to calculate the matching contribution value corresponding to a single meta-path. The process then proceeds to the stage where all meta-path traversal is completed. This stage forms a loop for meta-path traversal. If there are still meta-paths in the system that have not participated in the calculation, the process will flow back to the stage of starting meta-path traversal to continue executing feature calculations and score calculations for the next path. When all preset meta-paths have completed traversal, the process enters the stage of weighted summation + nonlinear activation. Combining the path weights learned by the model, the contribution scores of all paths are weighted and integrated, and then the score mapping is completed through a nonlinear function. After the calculation is completed, the process enters the stage of outputting the supply and demand node connection probability score, outputting the final quantified comprehensive matching probability score. Another link, branching off from the path instance sampling + neighbor feature aggregation stage, enters the stage of recording attention weights + path instance sequences. Throughout the process, the attention weights of each meta-path and the complete path node sequence are retained, providing the original basis for generating interpretive information. After data recording is complete, the process proceeds to the stage of filtering and activating the meta-path with the highest weight. By comparing the weight values ​​of all paths, the path instance that contributes the most to the final matching result is selected. Next, a natural language explanation description generation stage is initiated, converting the node information and parameter attributes of the path into intuitive textual descriptions. Finally, the matching probability score and the natural language explanation description are aggregated into an output stage for the matching score + matching explanation path signal, and both types of signals are transmitted to the downstream spatiotemporal filtering module. This entire process relies on multi-path branches to achieve multi-dimensional matching reasoning, uses loop traversal to complete full-path calculations, and simultaneously performs score calculations and explanation content generation in parallel. This ensures the accuracy of the matching results while achieving full-link traceability and explainability of the matching process.

[0033] The matching output module receives valid spatiotemporal candidate signals and sorts the retained supply nodes in descending order of connection probability score. For each recommended supply node, the module simultaneously extracts the meta-path instance information with the highest activation weight corresponding to that node, combining the demand parameter node attributes, supply parameter node attributes, and information on possible intermediate related material nodes along the path into an interpretable descriptive text. This text, along with the sorted supply node identifier, is output to the interactive interface or a data interface for downstream systems to call. Users thus not only obtain a recommended supply list sorted by matching degree but can also view the reasoning behind each recommendation, understanding the parameter equivalence relationships or material application scenarios upon which the system derives its recommendations. The entire system forms a complete data flow and signal processing chain, from the structured parsing of documents, the semantic normalization of heterogeneous terms, and the numerical normalization of parameter range values, to the construction of heterogeneous supply and demand graphs and the parallel reasoning and scoring of multiple meta-paths, and then to the filtering and review of time-series degradation information. Each module is decoupled through a clear signal transmission interface and completes its own independent and verifiable computational tasks, thus jointly achieving a systematic solution to the problems of unstructured input, semantic heterogeneity, missing implicit associations, and performance degradation mismatch in the supply and demand matching of materials technology.

[0034] This invention discloses a novel intelligent matching system for supply and demand in materials technology based on knowledge graphs. It receives unstructured supply and demand documents, uses an edge-enhanced graph attention network to perform forward propagation on text nodes and delimiter nodes to form structured material parameter signals, and normalizes heterogeneous terms and range values ​​through material domain ontology comparison learning and interval number embedding encoding to form normalized parameter vector signals. It constructs a heterogeneous supply and demand graph and uses a predefined meta-path execution aggregation network to score the connection probability of supply and demand node pairs to obtain matching interpretation path signals. Simultaneously, it accesses a temporal knowledge graph to calculate the current effective performance value of the material according to a degradation function and prunes ineffective nodes to form spatiotemporally effective candidate signals. Finally, it outputs a ranked matching result with interpretable paths.

[0035] Therefore, the intelligent matching system for supply and demand of new materials technology based on knowledge graphs of the present invention can solve the problems of semantic heterogeneity of material supply and demand and mismatch due to performance degradation.

[0036] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A novel materials technology supply and demand intelligent matching system based on knowledge graphs, characterized in that, include: The document parsing module receives unstructured supply and demand documents and uses an edge-enhanced graph attention network to perform forward propagation on text nodes and delimiter nodes to form structured material parameter signals. The semantic normalization module receives the structured material parameter signal and performs interval number embedding encoding on the parameter values ​​to form a normalized parameter vector signal. The heterogeneous supply and demand graph reasoning module receives the normalized parameter vector signal, constructs a heterogeneous information network including supply nodes, demand nodes, material nodes and parameter nodes, executes a meta-path aggregation network based on predefined meta-paths, scores the connection probability of supply and demand node pairs, and forms a matching score signal and a matching interpretation path signal. The spatiotemporal filtering module receives the matching scoring signal, accesses a time-series knowledge graph storing material performance degradation curves, calculates the current effective performance value of the material based on the timestamp, removes supply nodes that do not meet the demand threshold, and forms spatiotemporal effective candidate signals. The matching output module receives the spatiotemporal valid candidate signals, sorts and outputs the matching results according to the matching score, and attaches the meta-path instance with the highest activation weight as an interpretability information signal.

2. The system according to claim 1, characterized in that, The document parsing module is further configured to parse the unstructured supply and demand document into a layout-aware heterogeneous graph, wherein detected text fragments are constructed as text nodes, visual dividing line segments are constructed as separator nodes, and row edges and column edges are constructed based on the spatial alignment relationship between text nodes, and spatial position edges are constructed based on the spatial adjacency relationship between text nodes and separator nodes. The initial feature representation of the text nodes and the edge feature representation containing two-dimensional coordinate differences and edge type embeddings are input into the edge-enhanced graph attention network. Through a multi-head attention mechanism, message passing and feature aggregation are performed along the direction of row edges and column edges to generate text node embedding vectors carrying cell logical coordinate attribution information, and the structured material parameter signal is output accordingly.

3. The system according to claim 1, characterized in that, The semantic normalization module incorporates a material domain ontology graph. This ontology graph uses material performance terms as entity nodes and synonym and hierarchical relationships as edges. When fine-tuning the pre-trained language model, terms in the structured material parameter signal are used as anchor points. First-order neighbor entities in the material domain ontology graph are used as positive examples, and randomly sampled entities are used as negative examples. An ontology-guided contrastive learning loss function is constructed. By maximizing the vector similarity between the anchor point and the positive example and minimizing the vector similarity between the anchor point and the negative example, the same material property expressed heterogeneously is mapped to a neighboring position in the semantic vector space. During the inference stage, the terms in the structured material parameter signal are converted into semantic vectors and concatenated with the normalized interval number embedding vector to form the normalized parameter vector signal.

4. The system according to claim 3, characterized in that, When the semantic normalization module performs interval number embedding encoding on parameter values, it represents parameter values ​​with range limitations and parameter values ​​with lower or upper limits as interval number vectors, respectively. Each interval number vector includes an interval midpoint vector and an interval radius vector. When calculating the overlap between two interval numbers, the degree of intersection is determined based on the distance between the interval midpoints and the interval radius. This degree of intersection is used as the basis for calculating the numerical dimension similarity in subsequent matching scores. The formula for calculating the numerical dimension similarity is as follows: in, The value is the midpoint of the required parameter range. The value is the midpoint of the supply parameter range. The value is the radius of the required parameter range. To provide the radius value of the supply parameter range, Let be the length of the intersection of the two parameter intervals. The absolute distance is the midpoint of the interval. This is the similarity normalization coefficient. The similarity is based on the numerical dimension of the parameters.

5. The system according to claim 1, characterized in that, The heterogeneous supply and demand graph inference module predefined meta-paths include at least precise matching meta-paths and inference matching meta-paths. The node sequence of the precise matching meta-path is from the demand node through the demand parameter node to the supply parameter node and then to the supply node. The node sequence of the inference matching meta-path is from the demand node through the application scenario node to the material grade node and then to the supply node. When constructing the heterogeneous information network, the heterogeneous supply and demand graph inference module extracts demand parameter entities and supply parameter entities from the normalized parameter vector signal, and establishes alignment edges for node pairs in the demand parameter entities and supply parameter entities whose semantic vector similarity exceeds a preset threshold. At the same time, the application scenario entity is extracted from the demand description and linked to the scenario node, and the material grade entity is extracted from the supply information and linked to the material grade node.

6. The system according to claim 5, characterized in that, The heterogeneous supply and demand graph inference module executes the meta-path aggregation network process as follows: For each pair of supply and demand nodes, it traverses all predefined meta-paths; for each meta-path, it aggregates the features of neighboring nodes on the path through path instance sampling and attention mechanism to obtain the supply node embedding representation and demand node embedding representation under that meta-path; it concatenates the supply node embedding representation and demand node embedding representation and inputs it into a multilayer perceptron to obtain the matching contribution score corresponding to that meta-path; it performs a weighted summation of the matching contribution scores of all meta-paths, with the weights being learnable meta-path importance parameters, and maps the weighted summation result to the connection probability score through a nonlinear activation function. The calculation formula for the connection probability score is as follows: Where P is the connection probability score of the supply and demand node pair. Here, K is the total number of predefined meta-paths in the system, and k is the meta-path index. Let k be the learnable importance weights corresponding to the kth meta-path. The matching contribution score of the k-th meta-path output by the multilayer perceptron. Similarity in terms of parameter numerical dimensions. The node embedding feature balance coefficient, Embed vectors for the features of the demand nodes. Embed the feature vector of the supply node. The L2 distance is the embedding vector between the demand node and the supply node.

7. The system according to claim 6, characterized in that, When the heterogeneous supply and demand graph reasoning module forms the matching explanation path signal, it records the attention weight and instance path sequence of each meta-path during the execution of the meta-path aggregation network. It selects the meta-path with the largest contribution to the connection probability score as the meta-path instance with the highest activation weight, and concatenates the node sequence and node attribute values ​​traversed by the meta-path instance into an explanation description in natural language form. This explanation description and the matching score signal are transmitted to the spatiotemporal filtering module together.

8. The system according to claim 1, characterized in that, In the temporal knowledge graph accessed by the spatiotemporal filtering module, the material performance degradation curve is stored as a temporal attribute of the material entity node. This temporal attribute includes an initial performance value, a degradation coefficient, and a timestamp sequence. The current effective performance value is calculated as the product of the initial performance value and a time decay function, which is a negative exponential function with the degradation coefficient as the exponent. The decay factor is calculated based on the time difference between the current system time and the attribute record timestamps. The calculated current effective performance value is then compared with a demand threshold. When the current effective performance value is lower than the demand threshold, the corresponding supply node is marked as failed and removed from the candidate set. The spatiotemporal effective candidate signal is formed by traversing each node. The formula for calculating the current effective performance value is as follows: in, This represents the current effective performance value of the material. These are the initial property values ​​of the material. The material property degradation coefficient, This is the current system timestamp. This is the initial timestamp for recording material properties. This is the time difference. Let C be the performance decay time constant, and C be the time decay correction constant. This is the square of the time difference.

9. The system according to claim 8, characterized in that, The spatiotemporal filtering module is also connected to an external sensor data interface or a detection report input interface. When new performance detection data is received, the performance detection data is input as a new observation value of the time series into the gated recurrent unit network. The update gate and reset gate mechanism of the gated recurrent unit network are used to update the degradation state hidden vector of the corresponding material entity. The degradation coefficient is refitted based on the updated hidden vector. The updated degradation coefficient is written back to the corresponding time series attribute in the time series knowledge graph. In the next matching request, the calculation of the current effective performance value and the supply node screening are re-executed based on the updated degradation coefficient.

10. The system according to claim 1, characterized in that, When the matching output module outputs the matching results according to the matching score, it sorts the supply nodes contained in the spatiotemporal valid candidate signals in descending order of connection probability score, and combines the demand parameter node attributes, supply parameter node attributes and intermediate associated node information traversed by the meta-path instance with the highest activation weight into an interpretable description text, which is output to the interactive interface or data interface along with the sorted supply node identifier.

Images