A knowledge graph-based auxiliary method and system for mine production scheduling

Abstract

This application relates to the field of mine production scheduling and discloses a knowledge graph-based auxiliary method and system for mine production scheduling. This method guides a large model to construct cleaned, structured data by building prompt word templates, thus improving the reliability of knowledge graph data construction. It achieves accurate parsing and cleansing of entities, attributes, and time sequences in multi-source data, enhancing the structure and accuracy of knowledge extraction. Furthermore, it achieves feature enhancement and triple optimization through prompt word semantic constraints and contrastive learning mechanisms, significantly improving the real-time performance and reliability of knowledge graph construction in the mining field. This invention maintains strong adaptability and scalability even under limited sample conditions, effectively supporting intelligent decision-making and dynamic optimization in mine production scheduling, and significantly improving the efficiency and application value of knowledge graph construction.

Description

Technical Field

[0001] This invention relates to the field of mine production scheduling, and discloses a mine production scheduling auxiliary method and system based on knowledge graph. Background Technology

[0002] Mine production scheduling is a core component of mining production, aiming to optimize resource allocation and achieve efficient collaboration among mining, transportation, and processing processes while ensuring safety and environmental constraints. With the expansion of mining scale and the increasing complexity of production processes, scheduling exhibits characteristics such as multiple constraints, dynamic uncertainty, fragmented knowledge, and a growing need for intelligent solutions.

[0003] In recent years, large language models have demonstrated powerful capabilities in natural language understanding and knowledge reasoning, providing new ideas for knowledge graph modeling and intelligent decision-making in mine production scheduling. However, existing large-scale model-driven data preprocessing methods rely on unsupervised representation mechanisms of general natural language processing, which face three semantic loss problems in the construction of knowledge graphs in the mining field: First, the process-specific terminology clusters contained in mining data (such as "grinding and flotation system" and "three-water treatment system") lack anchoring in the domain semantic dictionary, leading to semantic bias in entity boundary identification; Second, the data purification and attribute alignment processes do not embed mining domain knowledge, which easily confuses similar attributes such as "mining process processing capacity of 4,000 tons / day" and "grinding and flotation process processing capacity of 5,000 tons / day", causing attribute normalization confusion; Third, time-series data such as "mining process capacity expansion and renovation in a certain year (increasing the daily processing capacity to 4,000 tons)" and "the 'three-water' system was completed in a certain year (adding water treatment capacity transportation demand)" only record the time of event occurrence (timestamp) and the event result separately, and the association between the timestamp and the entity state is missing, resulting in the knowledge graph's failure to capture dynamic semantic chains. Based on the aforementioned problems in semantic processing of domain data, existing prompt word technologies have fundamental defects in adapting to the construction of domain knowledge graphs: on the one hand, prompt word templates adopt a generalized semantic framework without incorporating semantic constraints of domain data, resulting in low recognition accuracy of some entities by the model; on the other hand, prompt words lack numerical semantic guidance capabilities. For numerical attributes containing logical relationships, such as "when the operating rate reaches more than 75%, the processing volume can be increased to 5000 tons / day," the pre-encoding of "threshold-attribute-association rule" cannot be achieved, causing the numerical semantics output by large models to be disconnected from industrial logic.

[0004] In existing technologies, the integration of contrastive learning with large models only focuses on surface feature optimization and fails to construct a collaborative framework adapted to the domain knowledge graph. Firstly, the construction of contrastive sample pairs lacks domain semantic anchoring, misclassifying non-homologous attributes as positive sample pairs, resulting in a high degree of domain semantic deviation in feature representation. Secondly, the contrastive loss function does not introduce a domain weight allocation mechanism, failing to differentiate and strengthen core semantic dimensions, leading to low recall of implicit relationships. Thirdly, the training process of contrastive learning and large models lacks coordination; the preprocessing results guided by prompt words fail to serve as semantic priors for contrastive sample construction, resulting in a disconnect between domain knowledge and model training, and reduced semantic effectiveness of feature optimization.

[0005] In summary, the core issues that need to be addressed are domain semantic loss, failure of prompt word guidance, and insufficient collaboration in contrastive learning. There is an urgent need to construct an integrated solution that can achieve semantic anchoring of prompt words, feature optimization of contrastive learning, and deep coupling of semantic modeling of large models, so as to break through the technical barriers to knowledge graph construction in the mining field and improve the effectiveness of mine production scheduling. Summary of the Invention

[0006] This invention provides a knowledge graph-based auxiliary method and system for mine production scheduling to solve the problems of domain semantic loss, failure of prompt word guidance, and insufficient collaborative comparison learning in existing technologies.

[0007] Firstly, this application provides a knowledge graph-based auxiliary method for mine production scheduling, including:

[0008] S1: Collect raw data of multi-source production scheduling in the mine, and construct prompt word templates based on the raw data of multi-source production scheduling. The prompt word templates include entity hierarchy, attribute type, and temporal association. Guide the domain pre-trained large model to perform entity indexing, attribute association parsing, and data purification through the prompt word templates to obtain a structured dataset.

[0009] S2: Determine the semantic constraints of the prompt words, construct comparison sample pairs based on the structured dataset and the semantic constraints of the prompt words, optimize features based on the comparison sample pairs, and construct a comparison feature library;

[0010] S3: Construct a few-shot dataset, and use the few-shot dataset to drive the large language model to extract initial triples; determine the index key, match the initial triples with the comparison feature library based on the index key, and output the optimized triples based on the similarity of the match;

[0011] S4: Mine production auxiliary scheduling based on optimized triples.

[0012] Secondly, this application provides a knowledge graph-based mine production scheduling auxiliary system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the method described in the first aspect above.

[0013] The present invention has the following beneficial effects:

[0014] This application presents a knowledge graph-based auxiliary method for mine production scheduling. By constructing prompt word templates, it guides a large model to build structured data that has undergone data purification, thereby improving the reliability of knowledge graph data construction. It achieves accurate parsing and purification of entities, attributes, and time sequences in multi-source data, enhancing the structure and accuracy of knowledge extraction. Furthermore, through prompt word semantic constraints and contrastive learning mechanisms, it achieves feature enhancement and triple optimization, significantly improving the real-time performance and reliability of knowledge graph construction in the mining field. This invention maintains strong adaptability and scalability even under limited sample conditions, effectively supporting intelligent decision-making and dynamic optimization in mine production scheduling, and significantly improving the efficiency and application value of knowledge graph construction.

[0015] In addition to the objectives, features and advantages described above, the present invention has other objectives, features and advantages.

[0016] The present invention will now be described in further detail with reference to the figures. Attached Figure Description

[0017] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0018] Figure 1 This is a flowchart of a preferred embodiment of the present invention: a mine production scheduling assistance method based on a knowledge graph. Detailed Implementation

[0019] The technical solution of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms "an" or "a," and similar terms, do not indicate a quantity limitation, but rather indicate the presence of at least one.

[0021] It should be understood that the knowledge graph-based mine production scheduling auxiliary method provided in this application can be applied to mine production scheduling scenarios.

[0022] Please see Figure 1 This application provides a knowledge graph-based auxiliary method for mine production scheduling, including:

[0023] S1: Collect raw production scheduling data from multiple sources in the mine, and construct a prompt word template based on the raw production scheduling data from multiple sources. The prompt word template includes entity hierarchy, attribute type, and temporal association. Guide the domain pre-trained large model to perform entity indexing, attribute association parsing, and data purification through the prompt word template to obtain a structured dataset.

[0024] Specifically, the raw data for multi-source production scheduling in mines includes enterprise documents and external professional documents. The enterprise documents include internal documents of the mining enterprise, such as production scheduling logs, operating procedures, accident reports, equipment information, accident case compilations, production plans, and work manuals. The external professional documents include national safety management regulations and accident determination documents in the mining industry.

[0025] S2: Determine the semantic constraints of the prompt words, construct comparison sample pairs based on the structured dataset and the semantic constraints of the prompt words, optimize features based on the comparison sample pairs, and construct a comparison feature library.

[0026] S3: Construct a few-sample dataset, and use the few-sample dataset to drive the large language model to extract initial triples; determine the index key, match the initial triples with the comparison feature library based on the index key, and output the optimized triples based on the similarity of the matching.

[0027] In this step, the large language model can be the Qwen / Qwen-7B-Chat model; this is just an example and not a limitation.

[0028] S4: Mine production auxiliary scheduling based on optimized triples.

[0029] The aforementioned knowledge graph-based mine production scheduling assistance method improves the reliability of knowledge graph data construction by constructing prompt word templates to guide the large model in building cleaned, structured data. It achieves accurate parsing and cleansing of entities, attributes, and time sequences from multi-source data, enhancing the structure and accuracy of knowledge extraction. Furthermore, it improves feature enhancement and triple optimization through prompt word semantic constraints and contrastive learning mechanisms, significantly increasing the real-time performance and reliability of knowledge graph construction in the mining field. This invention maintains strong adaptability and scalability even with limited sample sizes, effectively supporting intelligent decision-making and dynamic optimization in mine production scheduling, and significantly improving the efficiency and application value of knowledge graph construction.

[0030] Furthermore, through multi-dimensional prompt word templates Guide large models to generate structured datasets The multidimensional prompt word templates include:

[0031] Entity-level prompt keyword constraints as follows:

[0032] ;

[0033] L(ε) is an entity The hierarchy label function is used to output the hierarchy to which an entity belongs; L = { , , , A four-level hierarchical set is pre-defined for the mining sector. Indicates mining area level, workshop level, Indicates process level, Indicates equipment / warehouse level, Indicates the parameter level, used to specify the hierarchical positioning of the entity; For entities The parent entity mapping function outputs the direct parent entity of the entity. This is the candidate set of parent entities, used to establish hierarchical relationships between entities; For entities With parent entity The relational function outputs the dependency relationship type between the two.

[0034] In one example, consider a mineral processing workshop in a mine:

[0035] L0 (Workshop Level): Mineral Processing Workshop Empty, no superior entity. Meaningless);

[0036] L1 (Process Level): Crushing process, grinding process, flotation process, thickening process, filter press process, dewatering process (ε) =Mineral processing workshop, =“belongs to”);

[0037] L2 (Equipment Level): PEJ1200×1500 Jaw Crusher, GP200S Cone Crusher, HP500 Cone Crusher, 2YAG-1548AT Circular Vibrating Screen, #7 Belt Scale ( = Crushing process, =“belongs to”); System 1 ball mill, System 2 ball mill ( =Grinding and floating process =“belongs to”);

[0038] L3 (Parameter Level): Daily processing capacity, hourly processing capacity, discharge outlet size, current, power, oil pressure, oil temperature ( For the corresponding equipment, such as "daily processing capacity" =GP200S cone crusher =“representation”).

[0039] Among them, attribute type hint word constraints as follows:

[0040] ;

[0041] For entities The attribute type identifier function is used to output the attribute type; For entities The complete set of attributes; For entities The static property set; For entities A dynamic set of attributes.

[0042] In one example, consider equipment in a mine's ore dressing workshop:

[0043] For the GP200S cone crusher: = Static, =“Model: GP200S, Type: Cone Crusher”; = Dynamic, =“Processing capacity: 210-260 tons / hour”;

[0044] For mineral processing workshops in mines: = Static, =“Initial construction time: 1957, commissioning time: 1984, process type: crushing-grinding-flotation-thickening-pressure filtration”; = Dynamic, =“Daily crushing capacity: 3960-5000 tons, daily grinding and flotation capacity: 5000-5500 tons, daily dewatering tailings transport capacity: 4000-5100 tons / day, equipment operating rate: ≥75%”.

[0045] Temporal correlation keyword constraints as follows:

[0046] ;

[0047] The entity has a temporal attribute. It is the core timing event function of the entity, used to output key events that affect changes in entity attributes; T It is a physical entity The timestamp function is used to output the time when a key event occurs; T is the range of valid timestamps.

[0048] It is a physical entity A set of changing attributes; Before the critical event occurs, the entity The core attribute set; After a critical event occurs, the entity The core attribute set.

[0049] In one example, consider a mineral processing workshop in a mine:

[0050] =“2008 capacity expansion and renovation of crushing process” =2008, =“Medium crushing equipment: Φ1750 cone crusher, fine crushing equipment: Φ2200 cone crusher, daily processing capacity: <4000 tons” =“Medium crushing equipment: GP200S cone crusher, fine crushing equipment: HP500 cone crusher, daily processing capacity: 4000 tons”;

[0051] =“2018 Ball Mill Optimization and Transformation of Grinding and Floating Process” =2018, =“Single system operation, daily processing capacity: <5000 tons” =“One system + two systems coexist, daily processing capacity: 5000-5500 tons”;

[0052] =“The Sanshui system was completed in 2022” =2022, =“The dewatering process only processes tailings; daily conveying capacity: >5100 tons; environmental risk: low” =“The dehydration process involves a three-water system, with a daily conveying capacity of 4,000-5,100 tons and high environmental risks.”

[0053] In this application, entity labeling employs a confidence-based filtering algorithm to determine valid entities and exclude entities with incorrect confidence levels. The confidence formula is as follows:

[0054]

[0055] in, For entities The set of word vectors for candidate words, where Sim is the cosine similarity function. For entity-level prompt semantic vectors, Represents a semantic word segmentation set A single element in Indicates the entity filtering threshold. Indicates the confidence level.

[0056] Furthermore, based on the structured dataset and semantic constraints of the prompt words, this application constructs hierarchical entities, temporal states, and contrastive sample pairs. On this basis, a contrastive learning loss function is introduced to optimize the discriminative ability of samples in the representation space by maximizing the similarity between positive sample representations and minimizing the similarity between negative sample representations, thereby obtaining enhanced feature representations.

[0057] This invention is based on the construction of a comparative sample pair between a structured dataset and prompt words, and includes the following steps:

[0058] Based on structured datasets With semantic constraints on prompt words, a formalized structure is constructed. The comparison sample pairs, where f is the high-dimensional semantic feature mapping tensor and y is the sample label. s1 and s2 are the weights of the sample pairs, and s1 and s2 are different entities in the structured dataset.

[0059] The comparison sample pairs include hierarchical entity comparison sample pairs and temporal state comparison sample pairs.

[0060] Construct hierarchical entity comparison sample pairs, wherein positive sample pairs of hierarchical entity comparison sample pairs satisfy "same parent entity + same type of entity + core attribute Jaccard similarity ≥ 0.5". ".

[0061]

[0062] Negative sample pairs of hierarchical entity comparison samples satisfy the following conditions: "same parent entity + different entity types + core attribute Jaccard similarity < ".

[0063]

[0064] in, It is the parent entity corresponding to sample s. It is the entity type of sample s. It is a sample and Jaccard similarity of the core attribute set. Taking the equipment under the crushing process (parent entity) as an example, the positive sample pair is: (GP200S cone crusher, HP500 cone crusher), which satisfies... =“Crushing process” ="Cone Crusher" (Similar Equipment), Core Attribute Set =“Type: Cone Crusher, Function: Medium / Fine Crushing, Process: Ore Crushing” =“Type: Cone Crusher, Function: Fine Crushing, Process: Mineral Crushing”, calculate if The positive sample pair meets the criteria. The negative sample pair (GP200S cone crusher, #7 belt scale) meets the criteria. =“Crushing process” Cone crusher, =Electronic scale (non-standard device), core attribute set =“Type: Cone Crusher, Function: Crushing, Process: Ore Crushing, Structure: Crushing Chamber” =“Type: Electronic scale, Function: Measuring, Process: Mineral crushing, Structure: Sensor” It meets the negative sample criteria. It is the threshold that satisfies the construction conditions for positive sample pairs. It is the threshold that satisfies the construction conditions for negative sample pairs.

[0065] Construct temporal state comparison sample pairs. Positive sample pairs are semantically or functionally related, occur consecutively within a short time window, or belong to different representations of the same process / entity. Negative sample pairs are semantically or functionally unrelated, temporally conflicting, or cross-entity categories. Among them, the same device in two adjacent time windows but with similar features can be considered positive samples.

[0066] The feature optimization based on the contrastive sample pairs includes: using structured sample encoding for s1 and s2 in P, and using the InfoNCE loss function as the contrastive learning loss function to guide feature optimization;

[0067] The contrastive learning loss function is defined as follows:

[0068]

[0069] Where P represents the set of positive sample entity pairs, and N represents the set of negative sample entity pairs. For temperature parameters, This represents the similarity value between a positive sample and the head and tail entities (h,t). Indicates negative samples to the head and tail entities ( The similarity value of ) This represents a contrastive learning loss function based on the InfoNCE loss function. This loss function is used to enhance the discriminativeness of "head entity-tail entity" pairs, improve the robustness of triple extraction optimization, and ensure the reliability of production scheduling.

[0070] Furthermore, based on the aforementioned mine production scheduling, a few-sample dataset is manually constructed, and initial triples are extracted using a few-sample prompting engineering-driven large language model. Subsequently, the feature library is matched and compared using "head entity - tail entity" as the key, and the triples are optimized based on similarity output, thereby achieving real-time and accurate optimization of triple extraction in the mining field.

[0071] The few-sample dataset is constructed by limiting entity types to processes, equipment, process systems, and parameters, and relation types to processes, processing capacity, and material flow, to ensure that the initial triples better meet production scheduling requirements. Prompt words are embedded into the few-sample dataset, and a large model is used to extract entity-relation triples from the mining text to obtain the initial triples.

[0072] In this application, the construction of the initial triplet includes the following steps:

[0073] The local model adapter is initialized by creating a language model adapter instance dedicated to local deployment. The integrity of the local model configuration parameters is ensured through a verification mechanism, and the local deployment and management of the model are achieved through Transformer.

[0074] Template-based prompt word engineering and semantic injection utilize string formatting mapping for prompt word template filling. It organically integrates prompt words by replacing multi-source heterogeneous information with predefined placeholders. The information sources include, but are not limited to, the original input text sequence, few-sample datasets, entity candidate sets, and relational pattern prompts, forming a structured language model input sequence.

[0075] In one example, a few samples are shown below:

[0076] Example 1:

[0077] Text: Construction of the ore dressing plant began in 1957.

[0078] Three elements: [['Mineral Processing Workshop, 'Start of Infrastructure Construction', '1957']]

[0079] Example 2:

[0080] Text: The ore dressing workshop began production in 1984, mainly engaged in the production of copper from mines.

[0081] Three-part group: [['Mineral Processing Workshop', 'Date of Production', '1984'], ['Mineral Processing Workshop', 'Main Products', 'Mine Copper']]

[0082] Example 3:

[0083] Text: The mineral processing workshop involves processes such as crushing, grinding, flotation, thickening, and pressure filtration.

[0084] Ternary group: [['Mineral processing workshop', 'Involved processes', 'Crushing'], ['Mineral processing workshop', 'Involved processes', 'Grinding'], ['Mineral processing workshop', 'Involved processes', 'Flotation'], ['Mineral processing workshop', 'Involved processes', 'Thinning'], ['Mineral processing workshop', 'Involved processes', 'Filtration']]

[0085] Example 4:

[0086] Text: With the development of green and high-quality development, the ore dressing workshop built a rainwater and sewage separation system and a system for treating acid water and underground water inflow in 2022.

[0087] Three components: [['Rainwater and sewage separation, acid water and underground water treatment system', 'Construction time', '2022'], ['Mineral processing workshop', 'Owns', 'Rainwater and sewage separation system'], ['Mineral processing workshop', 'Owns', 'Acid water treatment system'], ['Mineral processing workshop', 'Owns', 'Underground water treatment system']].

[0088] Then, structured triple parsing is performed, which uses a specialized parsing algorithm to extract structured triples from the natural language description of the original text output generated by the language model.

[0089] After the initial triplet (h,r,t) is extracted, it is matched with the corresponding feature library using (h,t) as the index key according to the "head entity – tail entity" structure. The similarity function is defined as follows:

[0090]

[0091] in, ( () is the enhanced feature vector output by the head entity through comparison with the feature library. It is the enhanced feature vector output by the tail entity through comparison with the feature library.

[0092] During the triplet optimization process, the candidate triplet set obtained after double similarity screening is directly determined based on the threshold to form the final optimized triplet set:

[0093]

[0094] Where S(h,t) represents the cosine similarity of the comparison feature library, and Sim(h,t) represents the semantic similarity. Entity filtering threshold This is the final optimized set of triples. The similarity threshold is used. Entity pairs in the final optimized set are automatically included in the comparison feature library for dynamic updates, achieving adaptive expansion and iterative optimization of the feature library.

[0095] For the dynamic update mechanism of the contrast feature library, for each candidate triplet If the similarity within the database satisfies Then update the corresponding entity representation in the feature library:

[0096]

[0097] in, This is a smoothing coefficient used to balance the weights of historical and new knowledge. The tail entity is enhanced by comparing the feature vector output from the feature library. (t) represents the enhanced features of entities in the current high-confidence triple. ( ) is the enhanced feature vector output by the head entity through comparison with the feature library.

[0098] In summary, this invention addresses the problems of domain data gaps, homogeneous attribute normalization confusion, and lack of semantic logic in the construction of mine production scheduling knowledge graphs. It proposes a large-model-driven method for constructing mine production scheduling knowledge graphs based on a deep fusion of prompt word semantic anchoring and contrastive learning. By constructing a multi-dimensional prompt word system, the large model is guided to construct cleaned and structured data, thus improving the reliability of knowledge graph data construction. Accurate parsing and purification of entities, attributes, and time sequences are achieved in multi-source data, improving the structure and accuracy of knowledge extraction. Furthermore, feature enhancement and triple optimization are achieved through prompt word semantic constraints and contrastive learning mechanisms, significantly improving the real-time performance and reliability of mine domain knowledge graph construction. This invention maintains strong adaptability and scalability even under limited sample conditions, effectively supporting intelligent decision-making and dynamic optimization in mine production scheduling, and significantly improving the efficiency and application value of knowledge graph construction.

[0099] This application also provides a knowledge graph-based mine production scheduling assistance system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the aforementioned knowledge graph-based mine production scheduling assistance method. This knowledge graph-based mine production scheduling assistance system can implement various embodiments of the knowledge graph-based mine production scheduling assistance method and achieve the same beneficial effects, which will not be elaborated upon here.

[0100] The above description is merely a specific embodiment of this disclosure, but the scope of protection of this disclosure is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this disclosure should be included within the scope of protection of this disclosure. Therefore, the scope of protection of this disclosure should be determined by the scope of the claims.

Claims

1. A knowledge graph-based auxiliary method for mine production scheduling, characterized in that, include: S1: Collect raw data of multi-source production scheduling in the mine, and construct prompt word templates based on the raw data of multi-source production scheduling. The prompt word templates include entity hierarchy, attribute type and temporal association. Guide the domain pre-trained large model to perform entity indexing, attribute association parsing and data purification through the prompt word templates to obtain a structured dataset. S2: Determine the semantic constraints of the prompt words, construct comparison sample pairs based on the structured dataset and the semantic constraints of the prompt words, optimize features based on the comparison sample pairs, and construct a comparison feature library; S3: Construct a few-shot dataset and use it to drive a large language model to extract initial triples. Determine the index key, match the initial triples with the comparison feature library based on the index key, and output the optimized triples based on the similarity of the matches; S4: Mine production auxiliary scheduling based on optimized triplet; The prompt word template satisfies the following relation: ； In the formula, For entities Satisfying the three-dimensional cue word joint constraint; Constraints on entity-level prompts; Constrain the attribute type hint words; Constraints for time-series related prompt words; Among them, entity-level prompt word constraints as follows: ； L(ε) is an entity The hierarchy label function is used to output the hierarchy to which an entity belongs; L = { , , , A four-level hierarchical set is pre-defined for the mining sector. Indicates mining area level, workshop level, Indicates process level, Indicates equipment / warehouse level, Indicates the parameter level, used to specify the hierarchical positioning of the entity; For entities The parent entity mapping function outputs the direct parent entity of the entity. This is the candidate set of parent entities, used to establish hierarchical relationships between entities; For entities With parent entity The relational function outputs the dependency relationship type between the two. Attribute type hint constraints as follows: ； For entities The attribute type identifier function is used to output the attribute type; For entities The complete set of attributes; For entities The static property set; For entities A dynamic set of attributes; Temporal correlation keyword constraints as follows: ； The entity has a temporal attribute. It is the core timing event function of the entity, used to output key events that affect changes in entity attributes; T It is a physical entity The timestamp function is used to output the time when a key event occurs; T is the range of valid timestamps. It is a physical entity A set of changing attributes; Before the critical event occurs, the entity The core attribute set; After a critical event occurs, the entity The core attribute set; A confidence-based filtering algorithm is used to filter entities to determine valid entities. The formula for calculating the confidence level is as follows: ； in, For entities The set of word vectors for candidate words, where Sim is the cosine similarity function. For entity-level prompt semantic vectors, Represents a semantic word segmentation set A single element in Indicates the entity filtering threshold. Indicates the confidence level.

2. The mine production scheduling assistance method based on knowledge graphs according to claim 1, characterized in that, The multi-source production scheduling raw data includes production scheduling logs, operating procedures, accident reports, equipment information, accident case compilations, production plans, work manuals, safety management regulations, and mining industry accident judgment documents.

3. The mine production scheduling assistance method based on knowledge graphs according to claim 1, characterized in that, The construction of comparison sample pairs based on the structured dataset and the semantic constraints of the prompt words includes: Based on structured datasets Construct comparison sample pairs with the semantic constraints of the prompt words, satisfying the following relationship: ； In the formula, f is the high-dimensional semantic feature mapping tensor, and y is the sample label. s1 and s2 are the weights of the sample pairs; s1 and s2 are different entities in the structured dataset. For comparison sample pairs; The comparison sample pairs include hierarchical entity comparison sample pairs and temporal state comparison sample pairs; Construct hierarchical entity comparison sample pairs, wherein positive sample pairs of hierarchical entity comparison sample pairs satisfy the following condition: Same parent entity + same type of entity + core attribute Jaccard similarity ≥ The formula is as follows: Negative sample pairs of hierarchical entity comparison sample pairs satisfy the following conditions: Same parent entity + different entity types + core attribute Jaccard similarity The formula is as follows: in, It is the parent entity corresponding to sample s. It is the entity type of sample s. It is a sample and Jaccard similarity of the core attribute set It is the threshold that satisfies the construction conditions for positive sample pairs. It is the threshold that satisfies the construction conditions of negative sample pairs; Construct temporal state comparison sample pairs, wherein positive sample pairs are semantically or functionally related, or occur consecutively within a short time window, or belong to different representations of the same process or the same entity; negative sample pairs are semantically or functionally unrelated, temporally conflicting, or cross entity categories.

4. The mine production scheduling assistance method based on knowledge graphs according to claim 1, characterized in that, The feature optimization based on contrastive sample pairs includes: encoding s1 and s2 in P using structured samples, and guiding feature optimization by using the InfoNCE loss function as the contrastive learning loss function. The contrastive learning loss function is defined as follows: ； Where P represents the set of positive sample entity pairs, and N represents the set of negative sample entity pairs. For temperature parameters, This represents the similarity value between a positive sample and the head and tail entities (h,t). Indicates negative samples to the head and tail entities ( The similarity value of ) This represents the contrastive learning loss function based on the InfoNCE loss function.

5. The mine production scheduling assistance method based on knowledge graphs according to claim 1, characterized in that, In S3, the few-sample dataset is constructed with entity types limited to process, equipment, process system and parameters, and relation types to process, processing capacity and material flow. The few-sample dataset is embedded with prompt word templates, and the large model is called to extract entity-relation triples from the mining text to obtain the initial triples.

6. The mine production scheduling assistance method based on knowledge graphs according to claim 1, characterized in that, The extraction of the initial triplet includes: Initialize the local model adapter; Template prompt word engineering and semantic injection are performed based on the local model adapter. Prompt word template filling is performed based on string formatting mapping. Prompt words are organically integrated by replacing multi-source heterogeneous information with predefined placeholders. The information sources include the original input text sequence, a few-sample dataset, entity candidate set prompts, and relation pattern prompts, forming a structured language model input sequence. The structured triples are parsed, and the original text output generated by the language model is regularized. After the text information is recognized, it is segmented into list information, and the structured initial triples (h,r,t) are extracted from the list information.

7. The mine production scheduling assistance method based on knowledge graphs according to claim 6, characterized in that, The process of comparing the initial triples with the feature library based on the index key and outputting optimized triples based on similarity includes: The index key is determined based on the initial triple (h,r,t): head entity – tail entity; Based on the index key and its corresponding feature library, the similarity function is defined as follows: ； in, ( () is the enhanced feature vector output by the head entity through comparison with the feature library. It is the enhanced feature vector output by the tail entity through comparison with the feature library; Based on similarity filtering, the candidate triplet that is closest to the index key is selected and the optimization steps are performed. During the triplet optimization process, the candidate triplet set obtained after double similarity screening is used to form the final optimized triplet set based on a threshold, as follows: ； Where S(h,t) represents the cosine similarity of the comparison feature library, and Sim(h,t) represents the semantic similarity. Entity filtering threshold This is the final optimized set of triples. The similarity threshold; The entity pairs in the final optimized triplet set are automatically included in the comparison feature library for dynamic updates.

8. The mine production scheduling assistance method based on knowledge graphs according to claim 7, characterized in that, The dynamic updating of the automatically included comparison feature library includes: For each candidate triplet If the similarity within the database satisfies Then the corresponding entity representation in the feature library is updated as follows: ； in, This is a smoothing coefficient used to balance the weights of historical and new knowledge. The tail entity is enhanced by comparing the feature vector output from the feature library. (t) represents the enhanced features of entities in the current high-confidence triple. ( ) is the enhanced feature vector output by the head entity through comparison with the feature library.

9. A knowledge graph-based mine production scheduling auxiliary system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of any of the methods described in claims 1-8.

Images