Converter smelting multi-element knowledge graph construction method and system supporting intelligent recommendation of scrap steel proportion

Abstract

The invention provides a converter smelting multi-element knowledge graph construction method and system supporting intelligent recommendation of a scrap steel proportion, and the method comprises the steps: constructing a converter smelting scrap steel proportion domain ontology, entities, relationships and semantic rules, employing a BERT-BiLSTM-MHA-CRF model and an ERNIE model to extract domain entities and relationships, carrying out the disambiguation through combining a core reference analysis technology, obtaining a unified entity, and carrying out the intelligent recommendation of the scrap steel proportion. Incorporating into a graph database to form a converter smelting multivariate knowledge graph; and optimal configuration and efficient utilization of waste steel resources of different qualities are achieved, and cost reduction and benefit increase of the steel industry are promoted. According to the method, the scrap steel proportion knowledge implied in the historical smelting data is deeply mined, and the association between the historical knowledge and the new proportion decision-making task is established, so that the influence of human factors in the decision-making process is reduced, the method does not depend on manual proportion knowledge and experience reserve, and the historical knowledge is conveniently reused.

Description

Technical Field

[0001] This invention belongs to the field of metallurgical intelligence and energy conservation and environmental protection technology, specifically involving a method and system for constructing a multi-dimensional knowledge graph for converter smelting that supports intelligent recommendation of scrap steel ratio. Background Technology

[0002] As a core resource of green metallurgy, scrap steel is currently the only renewable steelmaking raw material that can replace iron ore on a large scale. Its efficient utilization is crucial for reducing smelting costs in steel plants and promoting the green and low-carbon transformation of the steel industry. Under my country's production model, which is dominated by converter smelting, formulating and implementing a scientific and reasonable scrap steel blending scheme is a key way to optimize the allocation and efficient utilization of scrap steel resources of different qualities and promote cost reduction and efficiency improvement in the steel industry.

[0003] Developing a scrap steel blending scheme for converter steelmaking is a complex decision-making process involving multi-factor coupling analysis. It requires comprehensive consideration of real-time smelting conditions, including the target steel grade, smelting quality requirements, and raw materials. Due to the complex physicochemical reactions involved in converter smelting, there are non-linear coupling relationships between the blending of different types and qualities of scrap steel and smelting conditions and endpoint values ​​(endpoint steel composition, temperature, etc.), which are difficult to quantify precisely using mathematical models. This leads steel plants to primarily rely on experience-based models developed by domain experts to guide batching personnel in making on-site blending decisions. However, these experience-based models are typically developed under ideal assumptions (such as a constant scrap steel yield when producing the same steel grade), which deviate from actual smelting conditions. Furthermore, they usually only consider the impact of some key variables (such as iron composition) on the scrap steel blending, resulting in limited reference value for the generated blending recommendations. Batching personnel still need to manually adjust and optimize the model-recommended schemes based on their personal blending knowledge and experience. Therefore, the economic and environmental benefits of the final blending scheme largely depend on the knowledge level of the batching personnel, making continuous and stable optimization difficult. Furthermore, the lack of effective means to store, organize, and manage the scrap steel proportioning knowledge generated during converter smelting makes it difficult to establish a connection between historical knowledge and new proportioning decision-making tasks, hindering knowledge reuse and further increasing the reliance on manual proportioning knowledge and experience reserves.

[0004] How to overcome the limitations of existing empirical models, deeply mine the scrap steel ratio knowledge hidden in historical smelting data, and establish the connection between historical knowledge and new ratio decision-making tasks, thereby reducing the influence of human factors in the decision-making process, is an urgent problem to be solved. Summary of the Invention

[0005] The technical problem this invention aims to solve is: a method and system for constructing a multivariate knowledge graph for converter smelting that supports intelligent recommendation of scrap steel proportions, used for... The technical solution adopted by this invention to solve the above-mentioned technical problems is: a method for constructing a multivariate knowledge graph for converter smelting that supports intelligent recommendation of scrap steel proportions, comprising the following steps: S1: Construct an ontology for the scrap steel ratio in converter smelting and define entities, relationships, and semantic rules that support scrap steel ratio recommendations; S2: Acquire converter smelting data and preprocess it, then use the BERT-BiLSTM-MHA-CRF model and ERNIE model to extract domain entities and relationships; S3: Based on the counting vector and cosine similarity metric, the semantic similarity of entities or relations is quantified, and cross-source knowledge alignment is achieved through a preset similarity threshold; combined with core reference resolution technology, heterogeneous data conflicts are eliminated to achieve knowledge fusion, and the verified and semantically aligned unified entities and their attributes and relations are incorporated into the graph database to form a multi-dimensional knowledge graph for converter smelting; S4: Utilize graph databases to store knowledge graphs and provide traceable knowledge for matching recommendations through native indexes and query optimizations.

[0006] According to the above scheme, the specific steps in step S1 are as follows: S11: Organize the core concept system in the converter smelting field and construct the domain ontology related to scrap steel proportioning, including major categories of raw materials, major categories of processes, major categories of results and major categories of indicators; Based on ontology and smelting data, entities related to scrap steel proportioning are defined, including raw material entities, process entities, result entities, and indicator entities; S12: When representing the knowledge of scrap steel ratio in converter smelting, use structural relationships to describe the attribute associations within entities, use containment relationships to describe the hierarchical or compositional relationships between entities, and use semantic relationships to describe the interaction and influence relationships between entities across categories, including raw material attributes-smelting behavior, raw material ratio-process control, and process control-outcome indicators. Define semantic rules for entities, attributes, and relationships, and describe their specific meanings, mechanisms of action, and constraints in the physicochemical processes of converter smelting and scrap steel proportioning decisions.

[0007] According to the above scheme, the specific steps in step S2 are as follows: S21: Obtain converter smelting data and preprocess it, use the BERT-BiLSTM-MHA-CRF model to extract entities, and map the input text sequence to the corresponding entity label sequence. S22: Extracting relationships from preprocessed converter smelting data using the ERNIE model.

[0008] Furthermore, in step S21, the specific steps are as follows: S211: The BERT layer receives the preprocessed text input and uses a pre-trained Transformer encoder to generate word or subword embedding vectors containing contextual semantic information. S212: The BiLSTM layer receives the output sequence of the BERT layer and uses a bidirectional long short-term memory network to capture long-distance dependencies from the forward and reverse directions of the sequence, respectively. S213: The multi-head attention layer acts on the output of the BiLSTM layer, and calculates the association weights between any two words or sub-words in the sequence in parallel through multiple independent attention heads, and learns their interaction relationships in different representation subspaces; S214: The output of the multi-head attention layer is mapped to the label space through a linear layer to obtain the emission score of each entity label corresponding to each word or subword; S215: The CRF layer receives the transmission score and uses the learned tag transition matrix to calculate the entity tag sequence with the most reasonable structure and the highest score based on the global optimal principle of the entire sequence; S216: Identify and extract entities and categories from the text based on the decoded tag sequence.

[0009] Furthermore, in step S22, the specific steps are as follows: S221: The ERNIE model receives raw text input and preprocesses it, performing entity boundary annotation and relation alignment. S222: The preprocessed text is fed into the embedding layer, the pre-trained weights of the ERNIE model are loaded, and the text is converted into a character vector representation composed of word vectors, segment vectors and position vectors. S223: The initial vector is passed sequentially through multiple shared encoder layers to capture the dependencies between words and extract complex features, resulting in a deep text representation containing semantic and contextual information; S224: Integrate entity information identified by the model as prompt information; identify potential relational subjects in the text based on prompt information and context; identify objects related to the subject in the context to form subject-object pairs; clarify the relationship between the subject and object through the relation classification module, and output relation triples (subject, relation, object).

[0010] According to the above scheme, in step S2, It is the word embedding vector of the i-th tag. It is the segment embedding vector of the i-th tag. It is the embedding vector of the i-th label. The formula for the BERT layer in the BERT-BiLSTM-MHA-CRF model is: ; Forward LSTM at time step t Hidden state for: ; Backward LSTM at time step t Hidden state for: ; [·;·] represents vector concatenation. The formula for the BiLSTM layer in the BERT-BiLSTM-MHA-CRF model is: ; set up W Q , W K and W V The projection matrix is ​​the formula for the multi-head attention layer of the BERT-BiLSTM-MHA-CRF model: , , ; It is the first i The emission fraction at each position, From the label Transfer to label The transition score, the formula for the CRF layer of the BERT-BiLSTM-MHA-CRF model is: ; It is the word embedding vector of the i-th tag. It is the segment embedding vector of the i-th tag. Let be the embedding vector at the position of the i-th label. The formula for the embedding layer of the ERNIE model is: .

[0011] According to the above scheme, the specific steps in step S3 are as follows: S31: Use a bag-of-words model-based counting vector generator to convert the preprocessed relevant entity description text into a numerical vector representation; S32: Calculate cosine similarity to quantify the similarity between vectors corresponding to different entity descriptions; S33: Identify entities that point to the same real-world object based on a preset similarity threshold; S34: Group descriptions with similarity higher than the threshold into the same category to eliminate entity ambiguity caused by differences in expression; S35: Define standardized naming, entity attributes, and semantic specifications for each core entity; merge entities whose names conform to standardized naming mapping rules, key attributes are highly consistent, or semantic relationships match based on core reference resolution technology; create a unified graph entity node and integrate attribute information from all its sources; S36: Verify the semantic consistency of the merged entities; S37: Incorporate validated and semantically aligned unified entities, along with their attributes and relationships, into the knowledge graph.

[0012] Furthermore, in step S3, v 1 and v 2 represents the vector representations of two entities. The cosine similarity calculation formula is: ; w 1. w 2、…、 w m It refers to the words in the text, count( w m () represents the number of times a word appears in the text. The formula for the counting vector is: .

[0013] According to the above scheme, the specific steps in step S4 are as follows: S41: Based on the knowledge system of converter smelting, the core entities and the relationships between entities are modeled as nodes and relationships in a knowledge graph; the nodes and relationships are stored using the native graph data structure of the graph database to form the basic structure of the knowledge graph; S42: Create corresponding indexes for key attributes that are frequently queried for node labels and relationship types; S43: Based on the constructed graph model and index, the knowledge graph can be queried for rapid retrieval, deep association analysis and intelligent application of converter smelting knowledge by querying entity association, path analysis and pattern matching.

[0014] A multi-dimensional knowledge graph construction system for converter smelting that supports intelligent recommendation of scrap steel ratios. The construction submodule is used to build the ontology of scrap steel proportioning in converter smelting and to define entities, relationships and semantic rules that support scrap steel proportioning recommendations. The extraction submodule is used to acquire converter smelting data and perform preprocessing, using the BERT-BiLSTM-MHA-CRF model and the ERNIE model to extract domain entities and relationships; The graph submodule is used to quantify the semantic similarity of entities or relations based on counting vectors and cosine similarity measures, and to achieve cross-source knowledge alignment by setting a preset similarity threshold. Combined with core reference resolution technology, it eliminates heterogeneous data conflicts to achieve knowledge fusion, and incorporates the verified and semantically aligned unified entities and their attributes and relations into the graph database to form a multi-dimensional knowledge graph for converter smelting. The query submodule is used to store knowledge graphs using graph databases and provide traceable knowledge for matching recommendations through native indexes and query optimizations.

[0015] The beneficial effects of this invention are as follows: 1. The present invention provides a method and system for constructing a multivariate knowledge graph for converter smelting that supports intelligent recommendation of scrap steel ratios. By constructing an ontology, entities, relationships, and semantic rules for the scrap steel ratio domain in converter smelting, the system uses BERT-BiLSTM-MHA-CRF and ERNIE models to extract domain entities and relationships. Combined with core reference resolution technology, it disambiguates to obtain unified entities, which are then incorporated into a graph database to form a multivariate knowledge graph for converter smelting. This achieves optimized allocation and efficient utilization of scrap steel resources of different qualities, promoting cost reduction and efficiency improvement in the steel industry.

[0016] 2. This invention deeply mines the scrap steel proportioning knowledge hidden in historical smelting data and establishes the correlation between historical knowledge and new proportioning decision-making tasks, thereby reducing the influence of human factors in the decision-making process, not relying on manual proportioning knowledge and experience reserves, and facilitating the reuse of historical knowledge.

[0017] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of an embodiment of the present invention.

[0020] Figure 2 This is a schematic diagram of the construction of a multi-dimensional knowledge graph for converter smelting according to an embodiment of the present invention.

[0021] Figure 3 This is a view of the multi-dimensional knowledge graph of converter smelting according to an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0023] Example 1 See Figure 1 The specific steps for constructing a multi-dimensional knowledge graph for converter smelting that supports intelligent recommendation of scrap steel ratios are as follows: S1: Define the professional fields covered by the knowledge graph, sort out the core concept system, construct the ontology of scrap steel proportioning in converter smelting, and define entities, relationships, and semantic rules that support scrap steel proportioning recommendations; the specific steps are as follows: S11: This section outlines the core conceptual system of converter smelting, constructing a domain ontology related to scrap steel proportioning, encompassing four core categories: raw materials, processes, results, and indicators. The raw materials category includes subcategories such as molten iron, scrap steel, alloys, and auxiliary materials; the processes category includes subcategories such as technology, equipment, and operating rules; the results category includes subcategories such as molten steel, slag, and smelting cycle; and the indicators category includes subcategories such as quality indicators, energy consumption indicators, carbon emission indicators, and efficiency indicators. Based on the above ontology and the provided smelting data, entities closely related to scrap steel proportioning are defined: raw material entities cover scrap steel properties, molten iron properties, alloys, and auxiliary materials; process entities cover converter properties and process operations; result entities cover the final stage of molten steel, slag, and smelting cycle; and indicator entities cover decarburization rate, metal yield, molten steel cleanliness, and oxygen consumption. S12: When representing the scrap ratio in converter smelting, the structural relation "Has" is used to describe the attribute relationships within entities, the containment relation "Contain" is used to describe the hierarchical or compositional relationships between entities, and the semantic relation is used to describe the interactions and influences between entities across categories. This is mainly reflected in three types of relationships: raw material attributes - smelting behavior, raw material ratio - process control, and process control - result indicators. Specifically, this includes scrap type / physical properties - melting rate / endothermic behavior, scrap chemical composition - molten steel composition control, scrap type / scrap ratio - oxygen supply, oxygen supply intensity / oxygen lance position - molten pool stirring / reaction rate, etc. Examples of these multivariate relationships are as follows: molten iron temperature, molten iron Si content, and target endpoint temperature all affect the maximum allowable scrap ratio; the proportion of scrap type A, scrap type B, scrap type C, and molten iron composition all affect the endpoint molten steel composition; oxygen blowing time, oxygen supply, and total scrap amount all affect the smelting cycle. Define clear semantic rules for the above entities, attributes and relationships, and accurately describe their specific meanings, mechanisms of action and constraints in the physicochemical process of converter smelting and scrap steel proportioning decisions; S2: Preprocess the converter smelting data, using the BERT-BiLSTM-MHA-CRF model and the knowledge-enhanced semantic representation model ERNIE to extract domain entities and relations; the specific steps are as follows: S21: Based on the converter smelting preprocessing data, entity extraction is performed using the BERT-BiLSTM-MHA-CRF model, which maps the input text sequence to the corresponding entity label sequence.

[0024] First, the BERT encoding layer receives the preprocessed input from this text and uses a pre-trained Transformer encoder to generate word / subword embedding vectors rich in contextual semantic information. These vectors not only contain the semantics of the words themselves, but also deeply incorporate the contextual relationships of the words within the entire sentence; Secondly, the BiLSTM layer receives the output sequence from BERT and uses a bidirectional long short-term memory network to capture long-distance dependencies from the forward and backward directions of the sequence, respectively. This layer concatenates the forward and backward hidden states of each time step (corresponding to each word / sub-word) to output an enhanced representation that integrates the bidirectional temporal features of the entire sentence.

[0025] Furthermore, the multi-head attention (MHA) layer operates on the output of the BiLSTM, calculating the association weights between any two words / subwords in the sequence in parallel through multiple independent attention heads, and learning their interaction relationships in different representation subspaces. This layer concatenates and merges the attention outputs of each head to generate a sequence representation that focuses on global key contextual information, significantly improving the model's ability to perceive entity boundaries and internal structural features.

[0026] Then, the output of the MHA layer is mapped to the label space through a linear layer to obtain the emission scores of each entity label corresponding to each word / subword. Finally, the CRF layer receives these emission scores and uses its learned label transition matrix to calculate the entity label sequence with the most reasonable structure and the highest score based on the global optimality principle of the entire sequence.

[0027] Finally, based on the decoded tag sequence, entities and their categories in the text can be identified and extracted.

[0028] S22: Based on the converter smelting pretreatment data, perform relation extraction using the ERNIE model.

[0029] First, the ERNIE model receives the raw text input and preprocesses it, performing entity boundary annotation and relation alignment. Secondly, the preprocessed text is fed into the embedding layer, where ERNIE pre-trained weights are loaded, converting the text into a character vector representation composed of word vectors, segment vectors, and position vectors. ERNIE's dynamic masking mechanism enhances the perception of domain entities during this process. The initial vector is then passed sequentially through multiple shared encoder layers, each containing a multi-head self-attention mechanism and a feedforward neural network to capture word-to-word dependencies and extract complex features. Specifically, the multi-head self-attention mechanism calculates word-to-word dependency weights through dot product scaling to capture global contextual information, while the feedforward neural network performs a non-linear transformation on the output of the self-attention layers to extract more complex features. This multi-layer encoding results in a deep text representation rich in semantic and contextual information. Finally, based on this, the entity information identified by the model is integrated as prompt information. Based on the prompt information and context, potential relational subjects in the text are identified, and for each identified subject, the related object in the context is identified to form subject-object pairs. The relationship between the subject and the object is clarified by the relationship classification module, and the relationship triple (subject, relation, object) is output.

[0030] The BERT layer of the above BERT-BiLSTM-MHA-CRF model is formulated as follows:

[0031] in, It is the word embedding vector of the i-th tag. It is the segment embedding vector of the i-th tag. It is the position embedding vector of the i-th tag.

[0032] The BiLSTM layer of the above BERT-BiLSTM-MHA-CRF model is formulated as follows:

[0033]

[0034]

[0035] in, Is it a forward LSTM at time step t The hidden state, It is a backward LSTM at time step t The hidden state, [·;·] represents vector concatenation.

[0036] The multi-head attention (MHA) layer of the BERT-BiLSTM-MHA-CRF model described above is formulated as follows:

[0037]

[0038]

[0039] in, , It is a projection matrix.

[0040] The CRF layer of the above BERT-BiLSTM-MHA-CRF model is formulated as follows:

[0041] in, It is the first i The emission fraction at each position, From the label Transfer to label The transfer fraction.

[0042] The embedding layer of the above ERNIE model is defined by the following formula:

[0043] in, It is the word embedding vector of the i-th tag. It is the segment embedding vector of the i-th tag. It is the position embedding vector of the i-th tag.

[0044] S3: Based on a counting vectorizer, text entities are converted into digital vectors, and the semantic similarity of entities / relationships is quantified by calculating cosine similarity. Cross-source knowledge alignment is achieved by setting a similarity threshold. Combined with core reference resolution technology, heterogeneous data conflicts are eliminated to achieve knowledge fusion. The verified and semantically aligned unified entities and their attributes and relationships are formally incorporated into the graph database to form a multi-dimensional knowledge graph for converter smelting. The specific steps are as follows: Entity disambiguation is performed on the extracted entities to identify entities from different data sources that point to the same real-world object.

[0045] First, a CountVectorizer is used to convert the preprocessed relevant entity description text into a numerical vector representation. This vectorizer is based on the Bag-of-Words model, which maps words in the text to feature vectors, with each dimension representing the frequency of a specific word in the description.

[0046] Secondly, the cosine similarity formula is used to calculate the similarity between the corresponding vectors describing different entities. Cosine similarity measures the degree of similarity between two vectors in a direction, and its value range is [0, 1]. The higher the value, the closer the vectors are in direction, and the more similar the texts are in terms of word distribution patterns.

[0047] Next, a judgment is made based on a preset similarity threshold. If the cosine similarity of two entity description vectors is higher than the threshold, they are considered to refer to the same entity.

[0048] Then, descriptions with similarity higher than a threshold are grouped into the same category, thereby eliminating entity ambiguity caused by differences in expression (such as synonyms, near-synonyms, and different phrases describing the same thing).

[0049] Finally, standardized naming, entity attributes, and semantic specifications are defined for each core entity. Based on core reference resolution technology, entities whose names conform to standardized naming mapping rules, whose key attributes are highly consistent, or whose semantic relationships match are merged to create a unified graph entity node. Attribute information from all its sources is integrated to ensure the uniqueness and consistency of entities in the knowledge graph. Furthermore, the semantic consistency of the merged entities is verified to ensure that the merged entities, their attributes, and relationships strictly conform to the defined semantic specifications, eliminating conceptual ambiguity and terminological differences. The verified and semantically aligned unified entities, along with their rich attributes and relationships, are then formally incorporated into the knowledge graph.

[0050] The formula for calculating the cosine similarity is as follows:

[0051] in, V 1 and V 2 is a vector representation of two entities.

[0052] The formula for the above counting vector is as follows:

[0053] in, w 1. w 2、…、 w m It refers to the words in the text, count( w m () represents the number of times a word appears in the text.

[0054] S4: Utilize a graph database to store the knowledge graph, leveraging its native indexing mechanism and query optimization features to ensure efficient knowledge storage and convenient retrieval, providing traceable knowledge support for matching recommendations; the specific steps are as follows: A graph database is used to construct and store a multivariate knowledge graph for converter smelting.

[0055] First, based on the knowledge system of converter smelting, the core entities extracted above and the relationships between entities are modeled as nodes and relationships of a knowledge graph. Then, using the native graph data structure of the graph database, these nodes and relationships are efficiently and persistently stored to form the basic structure of the knowledge graph.

[0056] Then, to ensure the efficient query performance of the knowledge graph, corresponding indexes are created using the graph database's indexing mechanism for key attributes that are frequently queried on node labels and relation types, thereby accelerating the data location speed during queries.

[0057] Finally, based on the constructed graph model and index, efficient and intuitive query statements are written using the query language of the graph database to explore complex entity relationships, path analysis, and pattern matching in the knowledge graph. This enables rapid retrieval, deep association analysis, and intelligent application of converter smelting knowledge, providing decision-making references for intelligent recommendation of converter scrap ratio.

[0058] This embodiment constructs an ontology, entities, relationships, and semantic rules for the scrap steel proportioning domain in converter smelting. It uses the BERT-BiLSTM-MHA-CRF model and the ERNIE model to extract domain entities and relationships, and combines core reference resolution technology to disambiguate and obtain unified entities, which are then incorporated into a graph database to form a multivariate knowledge graph for converter smelting. This achieves optimized allocation and efficient utilization of scrap steel resources of different qualities, and promotes cost reduction and efficiency improvement in the steel industry.

[0059] Example 2 The steps in this embodiment are the same as in Embodiment 1, except that each step is applied to a specific historical smelting case of a steel plant. Specifically, it includes the following steps: S1: Define the professional fields covered by the knowledge graph, sort out the core concept system, construct the ontology of scrap steel ratio in converter smelting, and define the entities, relationships and semantic rules that support scrap steel ratio recommendation; Figure 2 The process of constructing a multi-dimensional knowledge graph for converter smelting was demonstrated, including: S11: The converter smelting data related to scrap steel proportions comes from historical smelting cases provided by a steel plant. This data is distributed across heterogeneous data sources, including Excel spreadsheets of steel grade composition, raw material databases, and smelting performance reports. The data specifically covers smelting conditions such as target steel grades, smelting quality requirements, smelting resources, and smelting processes, as well as endpoint values ​​such as molten steel composition and temperature, actual steel output, slag basicity, and metal recovery. By analyzing the core concept system of converter smelting, a domain ontology is constructed, comprising four core categories: raw materials, processes, results, and indicators. The raw materials category includes subcategories such as molten iron, scrap steel, alloys, and auxiliary materials; the processes category includes subcategories such as processes, equipment, and operating rules; the results category includes subcategories such as molten steel, slag, and smelting cycle; and the indicators category includes subcategories such as quality indicators, energy consumption indicators, carbon emission indicators, and efficiency indicators. Based on the above ontology and the provided smelting data, entities closely related to scrap steel proportioning are defined. Raw material entities cover scrap steel properties (including type, chemical composition, physical properties, etc.), molten iron properties (including composition, temperature, etc.), alloys, and auxiliary materials, etc.; process entities cover converter properties (including furnace volume, furnace life, etc.) and process operations (including blowing, slag making, oxygen supply, etc.); result entities cover the final molten steel (including composition, temperature, etc.), slag (including basicity, composition, etc.), and smelting cycle, etc.; indicator entities cover decarburization rate, metal yield, molten steel cleanliness, and oxygen consumption, etc.

[0060] S12: When representing scrap steel proportioning knowledge, the structural relation "Has" is used to describe the attribute associations within entities. For example, entities such as "scrap steel" and "type" and "chemical composition" are connected through the relation "Has". The contain relation "Contain" is used to describe the hierarchical or compositional relationships between entities. For example, entities such as "raw material" and "molten iron" and "scrap steel" are connected through the relation "Contain". Semantic relations are used to describe the interaction and influence relationships between entities across categories. These are mainly reflected in three types of associations: raw material attributes - smelting behavior, raw material proportioning - process control, and process control - result indicators. The raw material properties-smelting behavior section describes how raw material characteristics affect the physicochemical reactions during the smelting process. For example, scrap type / physical properties (such as density and size) affect melting rate / endothermic behavior, and scrap chemical composition affects steel composition control (such as residual element control). The raw material proportioning-process control section describes how the raw material proportioning (especially the scrap proportioning) determines or constrains process operating parameters. For example, scrap type / scrap ratio affects total oxygen supply / oxygen supply intensity / oxygen lance position. The process control-result indicators section describes how process operations affect the final smelting results and performance indicators. For example, total oxygen supply / oxygen supply intensity / oxygen lance position affects molten pool stirring intensity / reaction rate / decarburization rate. Oxygen blowing time / oxygen supply / total scrap volume affects the smelting cycle. Specific multivariate relationships are illustrated below: molten iron temperature, molten iron Si content, and target endpoint temperature all affect the maximum allowable scrap ratio; the proportion of scrap type A, scrap type B, scrap type C, and molten iron composition all affect the endpoint steel composition; oxygen blowing time, oxygen supply, and total scrap volume all affect the smelting cycle. Define explicit semantic rules for the aforementioned entities, attributes, and relationships, precisely describing their specific meanings, mechanisms of action, and constraints in the physicochemical processes of converter smelting and scrap steel proportioning decisions. For example, the rules need to clarify whether the "influence" relationship is positive or negative, and the interactions between different factors (such as synergy or antagonism).

[0061] S2: Preprocess the converter smelting data, and use BERT-BiLSTM-MHA-CRF and ERNIE models to extract domain entities and relations; specifically including: S21: Based on the preprocessed data from the converter smelting process, entity extraction is performed using the BERT-BiLSTM-MHA-CRF model. This model maps the input text sequence to the corresponding entity label sequence; specifically, it includes: First, the BERT encoding layer receives the preprocessed input from this paper and uses a pre-trained Transformer encoder to generate word / subword embedding vectors rich in contextual semantic information. These vectors not only contain the semantics of the words themselves but also deeply integrate the contextual relationships of the words within the entire sentence. Second, the BiLSTM layer receives the output sequence from BERT and uses a bidirectional long short-term memory network to capture long-distance dependencies from the forward and backward directions of the sequence, respectively. This layer concatenates the forward and backward hidden states at each time step (corresponding to each word / subword) to output an enhanced representation that integrates the bidirectional temporal features of the entire sentence. Third, the multi-head attention (MHA) layer operates on the output of the BiLSTM, using multiple independent attention heads to compute the association weights between any two words / subwords in the sequence in parallel and learn their interaction relationships in different representation subspaces. This layer concatenates and fuses the attention outputs of each head to generate a sequence representation focused on global key contextual information, significantly improving the model's ability to perceive entity boundaries and internal structural features. Finally, the output of the MHA layer is mapped to the label space through a linear layer to obtain the emission scores of each word / subword for each entity label. Finally, the CRF layer receives these emission scores and, using its learned label transition matrix, calculates the entity label sequence with the most reasonable structure and the highest score based on the global optimality principle for the entire sequence. Ultimately, entities and their categories in the text can be identified and extracted based on the decoded label sequence.

[0062] The BERT layer of the above BERT-BiLSTM-MHA-CRF model is formulated as follows:

[0063] in, It is the word embedding vector of the i-th tag. It is the segment embedding vector of the i-th tag. It is the position embedding vector of the i-th tag.

[0064] The BiLSTM layer of the above BERT-BiLSTM-MHA-CRF model is formulated as follows:

[0065]

[0066]

[0067] in, Is it a forward LSTM at time step t The hidden state, It is a backward LSTM at time step t The hidden state, [·;·] represents vector concatenation.

[0068] The multi-head attention (MHA) layer of the BERT-BiLSTM-MHA-CRF model described above is formulated as follows:

[0069]

[0070]

[0071] in, W Q , W K and W V It is a projection matrix.

[0072] The CRF layer of the above BERT-BiLSTM-MHA-CRF model is formulated as follows:

[0073] in, It is the first i The emission fraction at each position, From the label Transfer to label The transfer fraction.

[0074] In step S21, the BERT model's embedding layer contains 12 layers, converting the input text into a 768-dimensional high-dimensional vector representation. The BiLSTM layer contains two hidden layers, each with a dimension of 400. The multi-head attention (MHA) layer uses 12 attention heads, each with a dimension of 64. The CRF layer performs sequence labeling based on the features extracted from the above layers, decodes the optimal entity label sequence, determines the entities and their boundaries in the text, and achieves entity extraction. S22: Based on the aforementioned converter smelting preprocessing data, relation extraction is performed using the ERNIE model; specifically including: First, the ERNIE model receives the raw text input and preprocesses it, performing entity boundary annotation and relation alignment. Second, the preprocessed text is fed into the embedding layer, loading ERNIE pre-trained weights to convert the text into a character vector representation composed of word vectors, segment vectors, and position vectors. ERNIE's dynamic masking mechanism enhances the perception of domain entities during this process. Then, the initial vectors are sequentially passed through multiple shared encoder layers, each containing a multi-head self-attention mechanism and a feedforward neural network to capture word-to-word dependencies and extract complex features. Specifically, the multi-head self-attention mechanism calculates the dependency weights between words through dot product scaling, capturing global contextual information, while the feedforward neural network performs a non-linear transformation on the output of the self-attention layer to extract more complex features. After multi-layer encoding, a deep text representation rich in semantic and contextual information is obtained. Finally, based on this, the entity information identified by the model is integrated as cue information. Based on the cue information and context, potential relational subjects in the text are identified, and for each identified subject, related objects are identified in the context, forming subject-object pairs. The relationship between the subject and the object is clarified by the relationship classification module, and the relationship triple (subject, relationship, object) is output. The embedding layer of the above ERNIE model is defined by the following formula:

[0075] in, It is the word embedding vector of the i-th tag. It is the segment embedding vector of the i-th tag. It is the position embedding vector of the i-th tag.

[0076] In step S2, the input layer of the ERNIE model preprocesses the text, including word segmentation and the addition of special markers such as [CLS] and [SEP]. The embedding layer converts the processed text into a numerical vector composed of 768-dimensional word vectors, segment vectors, and position vectors. The text vectors are then passed through 12 shared encoder layers, each containing a 12-head multi-head self-attention mechanism and a feedforward neural network. The hidden layer dimension of the feedforward neural network is 3072. Based on this, the ERNIE model uses the output of the NER subtask as input prompt information. It first extracts the subject to identify potential subjects, then extracts objects for the subjects to form subject-object pairs, and finally, the relation classification module clarifies the relationship between the subjects and objects, outputting relation triples. S3: Based on a counting vectorizer, text entities are converted into digital vectors, and the semantic similarity of entities / relationships is quantified by calculating cosine similarity. Cross-source knowledge alignment is achieved by setting a similarity threshold. Combined with core reference resolution technology, heterogeneous data conflicts are eliminated to achieve knowledge fusion. Verified and semantically aligned unified entities and their attributes and relationships are formally incorporated into the graph database to form a multi-dimensional knowledge graph for converter smelting; specifically including: First, a CountVectorizer is used to convert the preprocessed entity description text into numerical vector representations. This vectorizer is based on the Bag-of-Words model, mapping words in the text to feature vectors, with each dimension representing the frequency of a specific word in the description. Second, the cosine similarity formula is used to calculate the similarity between vectors corresponding to different entity descriptions. Cosine similarity measures the similarity of two vectors in a direction, with a value range of [0, 1]. A higher value indicates that the vectors are closer in direction and the texts are more similar in word distribution patterns. Third, a preset similarity threshold is used for judgment. If the cosine similarity of two entity description vectors is higher than the threshold, they are considered to refer to the same entity. Then, descriptions with similarity higher than the threshold are grouped into the same category, thereby eliminating entity ambiguity caused by differences in expression (such as synonyms, near-synonyms, or different phrases describing the same thing). Finally, standardized naming, entity attributes, and semantic specifications are defined for each core entity. Based on core reference resolution technology, entities whose names conform to standardized naming mapping rules, have highly consistent key attributes, or match semantic relationships are merged to create a unified graph entity node. Attribute information from all its sources is integrated to ensure the uniqueness and consistency of entities in the knowledge graph. Furthermore, the semantic consistency of the merged entities is verified to ensure that the merged entities, their attributes, and relationships strictly conform to the defined semantic specifications, eliminating conceptual ambiguity and terminological differences. The verified and semantically aligned unified entities, along with their rich attributes and relationships, are then formally incorporated into the knowledge graph.

[0077] The formula for calculating the cosine similarity is as follows:

[0078] in, v 1 and v 2 is a vector representation of two entities.

[0079] The formula for the above counting vector is as follows:

[0080] in, w 1. w 2、…、 w mWords in the text, count( w m () represents the number of times a word appears in the text.

[0081] In this embodiment, the "CountVectorizer" function from the "sklearn" library is used to convert the text into a vector representation, and the "cosine_similarity" function is used to calculate the cosine similarity between scrap steel entities. Based on the set similarity threshold, similar descriptions are grouped into the same category. The relevant code for knowledge fusion is as follows:

[0082] S4: Utilize graph databases to store knowledge graphs, leveraging their native indexing mechanisms and query optimization features to ensure efficient knowledge storage and convenient retrieval, providing traceable knowledge support for matching recommendations; specifically including: A graph database is used to construct and store a multi-dimensional knowledge graph for converter smelting. Figure 3 This paper presents a visual representation of the constructed multivariate knowledge graph for converter smelting. First, based on the knowledge system of converter smelting, the extracted core entities and their relationships are modeled as nodes and relations in the knowledge graph. These nodes and relations are then efficiently and persistently stored using the native graph data structure of the graph database, forming the basic structure of the knowledge graph. Next, to ensure efficient query performance, indexes are created for frequently queried key attributes on node labels and relation types, leveraging the powerful indexing mechanism of the graph database to accelerate data location during queries. Finally, based on the constructed graph model and indexes, efficient and intuitive query statements are written using the Cypher query language, a dedicated language for graph databases, to explore complex entity relationships, path analysis, and pattern matching within the knowledge graph. This enables rapid retrieval, deep association analysis, and intelligent application of converter smelting knowledge, providing decision-making references for intelligent recommendation of converter scrap ratios.

[0083] This embodiment deeply mines the scrap steel proportioning knowledge hidden in historical smelting data and establishes the correlation between historical knowledge and new proportioning decision-making tasks, thereby reducing the influence of human factors in the decision-making process, not relying on manual proportioning knowledge and experience reserves, and facilitating the reuse of historical knowledge.

[0084] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0085] Example 3 This embodiment is used to implement the principle of the above method embodiment to build a converter smelting multi-dimensional knowledge graph construction system that supports intelligent recommendation of scrap steel ratio, including a construction submodule, an extraction submodule, a graph submodule and a query submodule; The construction submodule is used to build the ontology of scrap steel proportioning in converter smelting and to define entities, relationships and semantic rules that support scrap steel proportioning recommendations. The extraction submodule is used to acquire converter smelting data and perform preprocessing, using the BERT-BiLSTM-MHA-CRF model and the ERNIE model to extract domain entities and relationships; The graph submodule is used to quantify the semantic similarity of entities or relations based on counting vectors and cosine similarity measures, and to achieve cross-source knowledge alignment by setting a preset similarity threshold. Combined with core reference resolution technology, it eliminates heterogeneous data conflicts to achieve knowledge fusion, and incorporates the verified and semantically aligned unified entities and their attributes and relations into the graph database to form a multi-dimensional knowledge graph for converter smelting. The query submodule is used to store knowledge graphs using graph databases and provide traceable knowledge for matching recommendations through native indexes and query optimizations.

[0086] Each submodule is mainly used to implement the various steps of the method implementation, which will not be elaborated here.

[0087] It should be noted that, depending on the implementation needs, the various steps / components described in this application can be broken down into more steps / components, or two or more steps / components or parts of the operation of steps / components can be combined into new steps / components to achieve the purpose of this invention.

[0088] This embodiment also includes a processor, a communication interface, a memory, and a communication bus; wherein the processor, communication interface, and memory communicate with each other through the communication bus; the memory stores a computer program, and when the program is executed by the processor, the processor performs the steps of constructing a multivariate knowledge graph for converter smelting that supports intelligent recommendation of scrap steel ratio.

[0089] This embodiment also provides a computer-readable storage medium storing executable instructions that, when executed by a processor, enable the processor to implement a method for constructing a multivariate knowledge graph for converter smelting that supports intelligent recommendation of scrap steel ratios.

[0090] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.

[0091] Furthermore, this application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0092] This application is described with reference to the flowchart of the method and computer program product according to Embodiment 1 and the block diagram of the device (system) according to Embodiment 3. It should be understood that each step or block in the flowchart or block diagram, as well as combinations of steps or blocks in the flowchart or block diagram, can be implemented by computer program instructions.

[0093] These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which are executable by the processor of the computer or other programmable data processing device, produce instructions for implementing the process. Figure 1 One or more processes or boxes Figure 1 A converter smelting multivariate knowledge graph construction system that supports intelligent recommendation of scrap steel ratios for functions specified in one or more boxes.

[0094] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes or boxes Figure 1 The function specified in one or more boxes.

[0095] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes or boxes Figure 1 The steps for constructing a multivariate knowledge graph for converter smelting, specified in one or more boxes, that supports intelligent recommendation of scrap steel ratios.

[0096] The above embodiments are only used to illustrate the design concept and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, all equivalent changes or modifications made based on the principles and design ideas disclosed in the present invention are within the protection scope of the present invention.

Claims

1. A method for constructing a multivariate knowledge graph for converter smelting that supports intelligent recommendation of scrap steel proportions, characterized by: Includes the following steps: S1: Construct an ontology for the scrap steel ratio in converter smelting and define entities, relationships, and semantic rules that support scrap steel ratio recommendations; S2: Acquire converter smelting data and preprocess it, and use the BERT-BiLSTM-MHA-CRF model and ERNIE model to extract domain entities and relationships; S3: Based on the counting vector and cosine similarity metric, the semantic similarity of entities or relations is quantified, and cross-source knowledge alignment is achieved through a preset similarity threshold; combined with core reference resolution technology, heterogeneous data conflicts are eliminated to achieve knowledge fusion, and the verified and semantically aligned unified entities and their attributes and relations are incorporated into the graph database to form a multi-dimensional knowledge graph for converter smelting; S4: Utilize graph databases to store knowledge graphs and provide traceable knowledge for matching recommendations through native indexes and query optimizations.

2. The method for constructing a multivariate knowledge graph for converter smelting supporting intelligent recommendation of scrap steel ratios according to claim 1, characterized in that: The specific steps in step S1 are as follows: S11: Organize the core concept system in the converter smelting field and construct the domain ontology related to scrap steel proportioning, including major categories of raw materials, major categories of processes, major categories of results, and major categories of indicators; Based on ontology and smelting data, entities related to scrap steel proportioning are defined, including raw material entities, process entities, result entities, and indicator entities; S12: When performing knowledge representation of scrap steel proportioning in converter smelting, structural relationships are used to describe the attribute associations within entities, containment relationships are used to describe the hierarchical or compositional relationships between entities, and semantic relationships are used to describe the interaction and influence relationships between entities across categories, including raw material attributes-smelting behavior, raw material proportioning-process control, and process control-outcome indicators. Define semantic rules for entities, attributes, and relationships, and describe their specific meanings, mechanisms of action, and constraints in the physicochemical processes of converter smelting and scrap steel proportioning decisions.

3. The method for constructing a multivariate knowledge graph for converter smelting supporting intelligent recommendation of scrap steel ratios according to claim 1, characterized in that: The specific steps in step S2 are as follows: S21: Obtain converter smelting data and preprocess it, use the BERT-BiLSTM-MHA-CRF model to extract entities, and map the input text sequence to the corresponding entity label sequence. S22: Extracting relationships from preprocessed converter smelting data using the ERNIE model.

4. The method for constructing a multivariate knowledge graph for converter smelting supporting intelligent recommendation of scrap steel ratios according to claim 3, characterized in that: The specific steps in step S21 are as follows: S211: The BERT layer receives the preprocessed text input and uses a pre-trained Transformer encoder to generate word or subword embedding vectors containing contextual semantic information. S212: The BiLSTM layer receives the output sequence of the BERT layer and uses a bidirectional long short-term memory network to capture long-distance dependencies from the forward and reverse directions of the sequence, respectively. S213: The multi-head attention layer acts on the output of the BiLSTM layer, and calculates the association weights between any two words or subwords in the sequence in parallel through multiple independent attention heads, and learns their interaction relationships in different representation subspaces; S214: The output of the multi-head attention layer is mapped to the label space through a linear layer to obtain the emission score of each entity label corresponding to each word or subword; S215: The CRF layer receives the transmission score and uses the learned tag transition matrix to calculate the entity tag sequence with the most reasonable structure and the highest score based on the global optimal principle of the entire sequence; S216: Identify and extract entities and categories from the text based on the decoded tag sequence.

5. The method for constructing a multivariate knowledge graph for converter smelting supporting intelligent recommendation of scrap steel ratios according to claim 3, characterized in that: The specific steps in step S22 are as follows: S221: The ERNIE model receives raw text input and preprocesses it, performing entity boundary annotation and relation alignment. S222: Feed the preprocessed text into the embedding layer, load the pre-trained weights of the ERNIE model, and convert the text into a character vector representation composed of word vectors, segment vectors, and position vectors; S223: The initial vector is passed sequentially through multiple shared encoder layers to capture the dependencies between words and extract complex features, resulting in a deep text representation containing semantic and contextual information; S224: Integrate entity information identified by the model as prompt information; Based on prompts and context, identify potential relational subjects in the text; identify objects related to the subjects in the context to form subject-object pairs; The relationship between the subject and the object is clarified by the relationship classification module, and the relationship triple (subject, relationship, object) is output.

6. The method for constructing a multivariate knowledge graph for converter smelting supporting intelligent recommendation of scrap steel ratios according to claim 1, characterized in that: In step S2, It is the word embedding vector of the i-th tag. It is the segment embedding vector of the i-th tag. It is the embedding vector of the i-th label. The formula for the BERT layer in the BERT-BiLSTM-MHA-CRF model is: ； Forward LSTM at time step t Hidden state for: ； Backward LSTM at time step t Hidden state for: ； [·;·] represents vector concatenation. The formula for the BiLSTM layer in the BERT-BiLSTM-MHA-CRF model is: ； set up W Q , W K and W V The projection matrix is ​​the formula for the multi-head attention layer of the BERT-BiLSTM-MHA-CRF model: ， ， ； It is the first i The emission fraction at each position, From the label Transfer to label The transition score, the formula for the CRF layer of the BERT-BiLSTM-MHA-CRF model is: ； It is the word embedding vector of the i-th tag. It is the segment embedding vector of the i-th tag. Let be the embedding vector at the position of the i-th label. The formula for the embedding layer of the ERNIE model is: 。 7. The method for constructing a multivariate knowledge graph for converter smelting supporting intelligent recommendation of scrap steel ratios according to claim 1, characterized in that: The specific steps in step S3 are as follows: S31: Use a bag-of-words model-based counting vector generator to convert the preprocessed relevant entity description text into a numerical vector representation; S32: Calculate cosine similarity to quantify the similarity between vectors corresponding to different entity descriptions; S33: Identify entities that point to the same real-world object based on a preset similarity threshold; S34: Group descriptions with similarity higher than the threshold into the same category to eliminate entity ambiguity caused by differences in expression; S35: Define standardized naming, entity attributes, and semantic specifications for each core entity; merge entities whose names conform to standardized naming mapping rules, whose key attributes are highly consistent, or whose semantic relationships match based on core reference resolution technology; create a unified graph entity node and integrate attribute information from all its sources; S36: Verify the semantic consistency of the merged entities; S37: Incorporate validated and semantically aligned unified entities, along with their attributes and relationships, into the knowledge graph.

8. The method for constructing a multivariate knowledge graph for converter smelting supporting intelligent recommendation of scrap steel ratios according to claim 7, characterized in that: In step S3, v 1 and v 2 represents the vector representations of two entities. The cosine similarity calculation formula is: ； w 1. w 2、…、 w m It refers to the words in the text, count( w m () represents the number of times a word appears in the text. The formula for the counting vector is: 。 9. The method for constructing a multivariate knowledge graph for converter smelting supporting intelligent recommendation of scrap steel ratios according to claim 1, characterized in that: The specific steps in step S4 are as follows: S41: Based on the knowledge system of converter smelting, the core entities and the relationships between entities are modeled as nodes and relationships in a knowledge graph; the nodes and relationships are stored using the native graph data structure of the graph database to form the basic structure of the knowledge graph; S42: Create corresponding indexes for key attributes that are frequently queried for node labels and relationship types; S43: Based on the constructed graph model and index, the knowledge graph can be queried for rapid retrieval, deep association analysis and intelligent application of converter smelting knowledge by querying entity association, path analysis and pattern matching.

10. A converter smelting multi-dimensional knowledge graph construction system supporting intelligent recommendation of scrap steel ratios, characterized by: The construction submodule is used to build the ontology of scrap steel proportioning in converter smelting and to define entities, relationships and semantic rules that support scrap steel proportioning recommendations. The extraction submodule is used to acquire converter smelting data and perform preprocessing, using the BERT-BiLSTM-MHA-CRF model and the ERNIE model to extract domain entities and relationships; The graph submodule is used to quantify the semantic similarity of entities or relations based on counting vectors and cosine similarity measures, and to achieve cross-source knowledge alignment by setting a preset similarity threshold. Combined with core reference resolution technology, it eliminates heterogeneous data conflicts to achieve knowledge fusion, and incorporates the verified and semantically aligned unified entities and their attributes and relations into the graph database to form a multi-dimensional knowledge graph for converter smelting. The query submodule is used to store knowledge graphs using graph databases and provide traceable knowledge for matching recommendations through native indexes and query optimizations.

Images