Fully-mechanized mining equipment large model decision support method and system

Abstract

The invention relates to the technical field of knowledge maps and large models, and discloses a fully-mechanized coal mining equipment large model decision support method and system. The method comprises the following steps: defining entities in data information of fully-mechanized coal mining equipment and a relationship between the entities to form triple data; converting the triple data into question and answer pairs based on the question and answer template, and optimizing the question and answer pairs; establishing a large language model, training the large language model based on question and answer pairs, performing fine tuning on the large language model based on a LoRA technology, and performing quantification processing on the equipment state in combination with fuzzy comprehensive evaluation to obtain a quantification result; the system uses the fine-tuned large language model to carry out entity and intention recognition on a natural language input by a user, a Cypher query statement is constructed to carry out query in a graph database, sub-graph information is obtained, the sub-graph information and the quantization result serve as cue words and are transmitted to the large language model, and retrieval information supporting large model decision making is obtained. The operation and maintenance efficiency of the fully-mechanized coal mining equipment is improved.

Description

Technical Field

[0001] The present application relates to the field of knowledge graphs and large-scale model technologies, and in particular to a large-scale model decision support method and system for comprehensive mining equipment. Background Art

[0002] With the widespread use of fully mechanized mining equipment in the mining industry, monitoring its operating status, predicting and diagnosing faults, and formulating maintenance strategies have become increasingly important. Traditional fully mechanized mining equipment operation and maintenance relies on manual experience and regular inspections, resulting in slow response, insufficient accuracy, and high maintenance costs. With the development of big data, natural language processing, and knowledge graphs, it has become possible to leverage advanced technologies to enhance the intelligence of fully mechanized mining equipment operation and maintenance.

[0003] Existing knowledge graph construction methods, such as the Chinese patent application with publication number CN117851610A, disclose a knowledge graph construction method and device based on a pre-trained large language model, wherein the method includes obtaining data to be processed; processing the unstructured data in the data to be processed by the pre-trained large language model to obtain the model output result; and constructing a knowledge graph based on the model output result. This application realizes the processing of multiple tasks in financial text data with the help of a pre-trained large language model, thereby realizing the construction of a knowledge graph. However, this application may perform well in financial text data, but may require additional adjustments or retraining to adapt to the field of comprehensive mining equipment.

[0004] There is also a Chinese patent application with publication number CN117633252A, which discloses an auxiliary retrieval method that integrates knowledge graphs and large language models. This invention uses a large language model to extract information about target domain problems; then accurately aligns the extracted information with the entities in the pre-built target domain knowledge graph to obtain the hit entity; finally, based on the hit entity, the corresponding answer is searched in the pre-built target domain knowledge graph. However, this application relies on the accurate alignment of the information extracted by the large language model with the entities in the pre-built knowledge graph. If the semantic alignment is inaccurate, it may introduce erroneous or irrelevant information.

[0005] Therefore, a large-scale model decision support method for fully mechanized mining equipment is needed to realize real-time monitoring of the operating status of fully mechanized mining equipment and fault prediction, while improving the operation and maintenance efficiency of fully mechanized mining equipment. Summary of the Invention

[0006] The present application provides a large-scale model decision support method and system for fully mechanized mining equipment, which are used to improve the operating efficiency of fully mechanized mining equipment.

[0007] In a first aspect, the present application provides a large-scale model decision support method for fully mechanized mining equipment, the method comprising:

[0008] Step S1: acquiring data information of a fully mechanized mining equipment, classifying the data information to obtain multiple information categories, defining entities in each information category, and relationship information between the entities;

[0009] Step S2: Identify the data format of each data information, extract entities, relationship information, and attribute values ​​of entities from the data information in different ways based on different data formats, form triple data, and store the triple data in a Neo4j database;

[0010] Step S3: Establish a question-answer template, convert the triple data into a question-answer pair in Alpaca format based on the question-answer template, establish a natural language processing model, and optimize the question-answer pair based on the natural language processing model to obtain an optimized question-answer pair;

[0011] Step S4: Establishing a large language model for the fully mechanized mining equipment. The large language model uses the Transformer model as its basic architecture, and uses the optimized question-answer pairs as a training set to train the large language model. During the training process, the large language model is fine-tuned based on the LoRA technology, and multiple evaluation indicators for evaluating the fully mechanized mining equipment are defined, and the evaluation indicators are quantified to obtain quantitative results.

[0012] Step S5: Loading the fine-tuned large language model, the user inputs natural language into the fully-mechanized mining equipment decision support system to query fully-mechanized mining equipment information, and the fully-mechanized mining equipment decision support system uses the fine-tuned large language model to perform entity recognition and intent recognition on the natural language, and constructs a Cypher query statement based on the recognized entities and intents;

[0013] Step S6: Use Cypher query statements to query the Neo4j database to determine whether query results are obtained. If so, obtain a local feature subgraph based on the query results. If no query results are obtained, perform a multi-level relational reasoning query to obtain an information redundancy subgraph, construct prompt words based on the local feature subgraph, the information redundancy subgraph, and the quantification results, and pass the prompt words to the large language model to obtain retrieval information that supports the large model's decision.

[0014] In combination with the first aspect, in a first implementation of the first aspect of the present application, triple data is formed, including:

[0015] The data formats include structured data, semi-structured data and unstructured data. If the data information format is structured data, the data information is directly converted into triple data. If the data information format is semi-structured data, the semi-structured data is parsed to obtain knowledge information, and the knowledge information is extracted to obtain entity, relationship information and attribute values ​​of the entity to obtain triple data.

[0016] If the format of the data information is unstructured data, a BiLSTM-CRF model is established. The BiLSTM-CR model includes a bidirectional long short-term memory network model and a conditional random field discriminant model. The context information in the unstructured data is obtained based on the bidirectional long short-term memory network model, and the entity information is identified from the unstructured data based on the context information. The boundary of the entity information is identified based on the conditional random field discriminant model to obtain complete entity information and relationship information between entities, extract the attribute information of the entity, and form triple data based on the entity information, relationship information and attribute information.

[0017] In combination with the first aspect, in a second implementation of the first aspect of the present application, converting the triple data into a question-answer pair in Alpaca format based on the question-answer template includes:

[0018] The natural language processing model is a PEGASUS model, which defines the question in the question-answer pair as the first information, and the answer corresponding to the question as the second information. All second information related to each first information is collected, and the second information is input into the PEGASUS model. The PEGASUS model fuses all the second information to obtain third information. The third information is a coherent answer containing all the second information, and an optimized question-answer pair is constructed based on the first information and the third information.

[0019] In combination with the first aspect, in a third implementation of the first aspect of the present application, fine-tuning the large language model based on LoRA technology includes:

[0020] In the pre-training stage of the large language model, the LoRA technology is used to keep the original weight matrix of each Transformer layer in the Transformer model unchanged, and the low-rank matrix A and the low-rank matrix B are introduced in the training stage, and the parameters of the two low-rank matrices are adjusted. The fine-tuned first weight matrix W' is obtained based on the first formula, and the first formula is: W'=W+A·B, where W is the original weight matrix and W∈R d×k , indicating that the original weight matrix is ​​a real number matrix with d rows and k columns, A∈R d×r , B∈R r×k, r represents the rank of matrix A or matrix B, where r<<min(d,k). The first weight matrix is ​​used to perform training on the optimized question-answer pairs to achieve fine-tuning of the large language model.

[0021] In combination with the first aspect, in a fourth implementation of the first aspect of the present application, quantifying the evaluation indicator includes:

[0022] A factor set is constructed based on the evaluation indicators, the evaluation indicators are graded to obtain a comment set, a membership function is set, and the membership degree μ(x) of each evaluation indicator to the comment is set based on the membership function. The membership function is: x is the real-time parameter of each evaluation indicator of the comprehensive mining equipment, a, b, c, and d are four preset key parameters, the membership degrees are summarized to obtain a single-factor evaluation matrix, a weight coefficient of each evaluation indicator is set, and a fuzzy synthesis operation is performed based on the weight coefficient and the evaluation matrix to obtain a comprehensive evaluation result vector, which is then normalized and used as the quantified result.

[0023] In combination with the first aspect, in a fifth implementation of the first aspect of the present application, a Cypher query statement is constructed based on the identified entity and intent, including:

[0024] The input natural language is segmented, each segmented word is tagged with a part-of-speech tag and input into the fine-tuned large language model. The model outputs words or phrases that are identified as entities or intents. The identified entities are mapped to corresponding nodes in the Neo4j database. The identified intents are matched with the preset intent list in the system to obtain the intent type. If there are multiple intent types, the multiple intents are merged to obtain a comprehensive intent.

[0025] The fully mechanized mining equipment decision support system determines the query relationship between entities based on the identified intention, constructs a Cypher query statement, retrieves the query relationship in the Neo4j database based on the Cypher query statement, and obtains the query result and returns it to the fully mechanized mining equipment decision support system.

[0026] In combination with the first aspect, in a sixth implementation of the first aspect of the present application, performing a multi-level relational reasoning query to obtain an information redundant subgraph includes:

[0027] A query path is constructed based on the identified entities and the relationships between them. The relationships between entities include first-degree relationships, second-degree relationships, and third-degree relationships. The comprehensive mining equipment decision support system queries the local feature subgraph based on the query path of the first-degree relationship. If the local feature subgraph is not obtained, the query path is expanded to the second-degree relationship and the third-degree relationship in sequence to obtain the corresponding subgraph information. The relevance of each subgraph information is scored, and the subgraph information with a high score is defined as an information redundant subgraph for output.

[0028] In combination with the first aspect, in a seventh implementation of the first aspect of the present application, constructing a prompt word based on the local feature subgraph, the information redundancy subgraph, and the quantization result includes:

[0029] The comprehensive mining equipment decision support system extracts key information from the local feature subgraph or the information redundant subgraph, wherein the key information includes entity name, type, attribute and relationship information between entities, converts the key information into structured information, and combines the natural language input by the user, the structured information and the quantitative results to generate prompt words.

[0030] In combination with the first aspect, in an eighth implementation manner of the first aspect of the present application, the information categories include equipment information, risk information, fault information, cause information, solution information and specification information.

[0031] In a second aspect, the present application provides a large-scale model decision support system for fully mechanized mining equipment, the system comprising:

[0032] A definition module is used to obtain data information of fully mechanized mining equipment, classify the data information to obtain multiple information categories, define entities in each information category, and define relationship information between entities;

[0033] A triple data processing module is used to identify the data format of each data information, extract entity, relationship information and entity attribute values ​​from the data information in different ways based on different data formats, form triple data, and store the triple data in the Neo4j database;

[0034] A question-answer pair formation module is used to establish a question-answer template, convert the triple data into a question-answer pair in Alpaca format based on the question-answer template, establish a natural language processing model, and optimize the question-answer pair based on the natural language processing model to obtain an optimized question-answer pair;

[0035] A fuzzy comprehensive evaluation module is used to establish a large language model for fully mechanized mining equipment. The large language model uses the Transformer model as its basic architecture and uses the optimized question-answer pairs as a training set to train the large language model. During the training process, the large language model is fine-tuned based on LoRA technology, and multiple evaluation indicators for the fully mechanized mining equipment are defined and quantified to obtain quantitative results.

[0036] A query statement construction module is used to load the fine-tuned large language model. The user inputs natural language into the fully-mechanized mining equipment decision support system to query fully-mechanized mining equipment information. The fully-mechanized mining equipment decision support system uses the fine-tuned large language model to perform entity recognition and intent recognition on the natural language, and constructs a Cypher query statement based on the recognized entities and intents;

[0037] The query module is used to use Cypher query statements to query the Neo4j database and determine whether a query result is obtained. If so, a local feature subgraph is obtained based on the query result. If no query result is obtained, a multi-level relationship reasoning query is performed to obtain an information redundant subgraph. A prompt word is constructed based on the local feature subgraph or the information redundant subgraph, and the prompt word is passed to the large language model to obtain retrieval information that supports the large model's decision.

[0038] The technical solution provided by this application obtains data information about fully mechanized mining equipment, defines entities in the data information and relationship information between entities, extracts entities, relationship information, and attribute values ​​from the data information in different ways according to different data formats, forms triple data, and stores the triple data in a Neo4j database. A triple data processing module generates a data set for fine-tuning the large model, and the large model fine-tuning module is combined to adapt the model to the needs of decision support in the field of fully mechanized mining equipment. Cypher query statements are then used to query the input for entities and intents, obtain knowledge graph subgraphs associated with the input, and provide retrieval-enhanced decision support. The present invention achieves real-time monitoring of the operating status of fully mechanized mining equipment and fault prediction, while improving the operation and maintenance efficiency of fully mechanized mining equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0040] Figure 1 This is a schematic diagram of an embodiment of a large-scale model decision support method for fully mechanized mining equipment in an embodiment of the present application;

[0041] Figure 2 Construct a process diagram for the knowledge graph of fully mechanized mining equipment in this application;

[0042] Figure 3 This is a diagram of the LoRA fine-tuning process in this application;

[0043] Figure 4 This is an example diagram of the accident tree structure of the fully mechanized mining equipment in this application;

[0044] Figure 5 This is a principle flow chart of the large model module of the fully mechanized mining equipment in this application;

[0045] Figure 6 Displays a graph of the loss rate for fine-tuning training in this application;

[0046] Figure 7 This is the principle flow chart of the comprehensive fuzzy evaluation expert module in this application;

[0047] Figure 8 This is a schematic diagram of an embodiment of a large-scale model decision support system for fully mechanized mining equipment in an embodiment of the present application. DETAILED DESCRIPTION

[0048] The embodiments of the present application provide a method and system for supporting decision making for a large model of a comprehensive mining equipment. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or that are inherent to these processes, methods, products or devices.

[0049] Example 1:

[0050] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiment of the present application, an embodiment of a large-scale model decision support method for fully mechanized mining equipment includes:

[0051] Step S1: Acquire data information of the fully mechanized mining equipment, classify the data information to obtain multiple information categories, define entities in each information category, and define relationship information between entities.

[0052] Specifically, in the process of constructing the knowledge graph, two modes, top-down and bottom-up, are often used. The top-down mode is to construct concepts from the top layer, and then gradually refine the branches downward to ensure the rigor of the hierarchy, and then add the collected entities one by one to the pre-constructed concepts; while the bottom-up mode starts to form concepts from the bottom layer, and then gradually abstracts upward to form upper-level concepts. The present invention selects a top-down approach to construct the knowledge graph of comprehensive mining equipment, and the construction process is as follows: Figure 2 shown.

[0053] First, the knowledge modeling phase is explained: Data related to fully mechanized mining equipment is collected, including its operation logs, sensor data, maintenance records, and user feedback. This data is then divided into six categories: "Equipment," "Risk," "Fault," "Cause," "Solution," and "Standard," with concepts defined within each category to describe entities within the domain in greater detail. Attributes are then defined for the concepts, creating a knowledge system with clear hierarchical relationships. Finally, relationship types between entities are defined to describe their connections and dependencies. For example, risks and faults have a "cause" relationship, while faults and solutions have a "correspond" relationship.

[0054] Step S2: Identify the data format of each data information, extract the entity, relationship information and attribute value of the entity from the data information in different ways based on different data formats, form triple data, and store the triple data in the Neo4j database.

[0055] Specifically, refer to Figure 2 , explain knowledge extraction in detail: Knowledge extraction is the process of extracting effective information from data information and converting it into a structured form that can be used for knowledge graphs. Before knowledge extraction, the source of data information should be clarified. The data sources in the field of comprehensive mining equipment mainly include structured data, semi-structured data and unstructured data. For each source of data information, different methods are used to extract entities in the data information, such as equipment name, fault type, personnel name, etc., to determine the relationship between entities and the attributes of entities, such as equipment model, fault severity, personnel position, etc., and convert the identified entities, relationships and attributes into triples.

[0056] The triple data is stored in the Neo4j database, which is a high-performance, scalable storage engine for graph data with graph database characteristics. It is based on the concept of graphs and stores data in the form of nodes and edges. Each node represents an entity, and each edge represents the relationship between entities. The four core elements of labels, nodes, relationships, and attributes together constitute the basic data model of the Neo4j graph database.

[0057] Step S3: Establish a question-answer template, convert the triple data into question-answer pairs in Alpaca format based on the question-answer template, establish a natural language processing model, optimize the question-answer pairs based on the natural language processing model, and obtain optimized question-answer pairs.

[0058] Specifically, before fine-tuning the large language model, the knowledge in the knowledge graph must be converted into a format compatible with fine-tuning technology, namely the Alpaca format. The Alpaca format contains question instructions and corresponding response pairs. Using predefined question-answer templates, triples are converted into question-answer pairs that conform to the Alpaca format. For example, the triples (S, P, O) stored in the knowledge graph are ("hydraulic support", "used for", "fully mechanized mining working face"), respectively. Based on this triple, a question-answer pair that conforms to the Alpaca format is generated: Question: What are hydraulic supports used for? Answer: Hydraulic supports are used for fully mechanized mining working faces.

[0059] Since a node in a knowledge graph may be connected to many other nodes, the question-answer pairs generated by extracting triple data in this case may have many identical questions with different answers, which will reduce the quality of the dataset and have a significant negative impact on model performance. Therefore, data enhancement is needed during data preprocessing to reduce the negative impact. This paper captures the deep semantic information of the text through the pre-trained model PEGASUS, obtains a high-dimensional text vector representation, and uses natural language generation technology to merge multiple answers into a coherent and natural answer text, that is, the optimized answer pair.

[0060] Step S4: Establish a large language model for comprehensive mining equipment. The large language model uses the Transformer model as its basic architecture, and the optimized question-answer pairs are used as the training set to train the large language model. During the training process, the large language model is fine-tuned based on LoRA technology, and various evaluation indicators for comprehensive mining equipment are defined. The evaluation indicators are quantified to obtain quantitative results.

[0061] Specifically, since the large language model is trained on massive amounts of text, the effect of directly applying the original trained model to the decision support task of comprehensive mining equipment is often not as expected. Therefore, the present invention introduces efficient fine-tuning technology to enhance the performance of the large model in the field of comprehensive mining equipment. The mainstream large language model fine-tuning technology can be divided into three categories: Adapter, Prefix-Tuning and LoRA. Among these three types of fine-tuning technologies, LoRA has unique advantages over Adapter and Prefix-Tuning, making it the first choice in many scenarios. First, LoRA greatly reduces the number of parameters required for fine-tuning and GPU memory requirements by freezing the weights of the pre-trained model and introducing a trainable low-rank decomposition matrix in each layer of the Transformer architecture. The LoRA fine-tuning process is as follows: Figure 3 shown.

[0062] The operating status and fault analysis of fully mechanized mining equipment involve multiple complex factors, which are often ambiguous and uncertain. Traditional decision support methods may not be able to fully consider the impact of these fuzzy factors. Fuzzy comprehensive evaluation method can effectively deal with multi-factor and multi-level complex problems through the theory of fuzzy mathematics, and is suitable for comprehensive status evaluation and fault analysis of fully mechanized mining equipment, such as Figure 4 As shown, this is an example diagram of the accident tree structure of the fully mechanized mining equipment. The present invention sets up a comprehensive fuzzy evaluation expert module, and realizes a comprehensive evaluation of the operating status of the fully mechanized mining equipment according to the comprehensive fuzzy evaluation expert module, and provides a scientific basis for subsequent decision support.

[0063] Step S5: Load the fine-tuned large language model. The user inputs natural language into the comprehensive mining equipment decision support system to query comprehensive mining equipment information. The comprehensive mining equipment decision support system uses the fine-tuned large language model to perform entity recognition and intent recognition on the natural language, and constructs a Cypher query statement based on the recognized entities and intents.

[0064] Specifically, if Figure 5 As shown in the figure, it is the principle flow chart of the large model module of comprehensive mining equipment. First, a fine-tuned large language model is loaded. Based on the pre-training, this model is further trained on data in a specific field. The user enters natural language to query through the interface of the comprehensive mining equipment decision support system. The query content is such as equipment status, fault diagnosis, and maintenance suggestions. The system uses the fine-tuned large language model to perform entity recognition and intent recognition on the natural language input by the user. The intent is the goal that the user hopes to achieve or the type of information obtained through the query, such as querying equipment status, solutions, predicting faults, etc.

[0065] Secondly, determine whether the entity and intent query is successful. If the query is successful, a Cypher query statement is constructed based on the identified entity and intent. Cypher is the query language of the graph database Neo4j, which is used to retrieve and operate data in the graph database. The constructed Cypher query statement is sent to the graph database for execution to retrieve information related to the user query.

[0066] Step S6: Use Cypher query statements to query the Neo4j database to determine whether query results are obtained. If so, obtain a local feature subgraph based on the query results. If no query results are obtained, perform a multi-level relational reasoning query to obtain an information redundancy subgraph. Construct prompt words based on the local feature subgraph, the information redundancy subgraph, and the quantification results, and pass the prompt words to the large language model to obtain retrieval information that supports the large model's decision-making.

[0067] Specifically, after the graph database returns the query results, the system checks whether the Cypher query returned any results. If the query results are returned, it means that data related to the user's query has been found, and the system will build a local feature subgraph based on these results. This subgraph contains entities and relationships directly related to the user's query and is part of the knowledge graph. Useful information such as device status, fault causes, solutions, etc. is then extracted from the local feature subgraph. If the query results are not directly obtained, a multi-level relational reasoning query method is used for the query. This method will return a subgraph containing more entities and relationships. This subgraph contains some not directly related or redundant data, but it provides a wider range of information.

[0068] The system constructs prompt words based on local feature subgraphs or information redundant subgraphs. Prompt words are prompt guide words. Prompt mainly generates an input format that can be passed to the large model of comprehensive mining equipment by combining user input and subgraph information returned by the query. The present invention adopts a hierarchical design, combining the structured information of the knowledge graph with the fuzzy comprehensive evaluation results. By constructing a hierarchical prompt link (task instructions-subgraph parsing-expert rule injection-construction of prompt words), the system can generate text that can be processed by the large model based on the structured information in the graph database and combined with the input, and pass it to the large model. This allows the large model to fully consider the real and powerful information from the graph database when generating decision support information, reduce the "hallucination" phenomenon of the large model, and effectively improve the authenticity and reliability of decision information. It consists of the following steps:

[0069] 1. Task instructions: Define the task objectives of the model, enable the model to understand the context and objectives of the input information, and limit the output range.

[0070] 2. Subgraph Parsing: Subgraph construction includes not only direct relationships between entities but also different types of edges (relationships) and their attributes. This step parses the subgraph information returned by the graph database into structured text to facilitate subsequent prompt word construction.

[0071] 3. Expert rule injection: The equipment status level (such as "abnormal") and key parameter membership (such as "the hydraulic support working pressure membership is 0.8") output by the fuzzy comprehensive evaluation expert module are used as context supplements.

[0072] 4. Constructing prompt words: The constructed prompt words are presented in natural language and contain the text content of the subgraph analysis. By splicing the constructed prompt words with the input, the real and powerful information from the graph database is fully considered to guide the model to generate accurate output.

[0073] By inputting the constructed prompt into the large model of the comprehensive mining equipment, the large model will generate intelligent decisions on the operating status, fault analysis, maintenance strategy, etc. of the comprehensive mining equipment. The retrieval enhancement method enables the system to not only rely on the model's reasoning, but also provide more comprehensive, detailed and realistic decision-making suggestions through the support of historical data and professional knowledge.

[0074] In a specific embodiment, forming triple data specifically includes the following steps:

[0075] Data formats include structured data, semi-structured data and unstructured data. If the format of the data information is structured data, the data information is directly converted into triple data. If the format of the data information is semi-structured data, the semi-structured data is parsed to obtain knowledge information, and the knowledge information is extracted to obtain entities, relationship information and attribute values ​​of entities to obtain triple data.

[0076] If the format of the data information is unstructured data, a BiLSTM-CRF model is established. The BiLSTM-CR model includes a bidirectional long short-term memory network model and a conditional random field discriminant model. The context information in the unstructured data is obtained based on the bidirectional long short-term memory network model, and the entity information is identified from the unstructured data based on the context information. The boundary of the entity information is identified based on the conditional random field discriminant model to obtain complete entity information and the relationship information between entities, extract the attribute information of the entity, and form triple data based on the entity information, relationship information and attribute information.

[0077] Specifically, structured data is stored in a clearly defined format, usually in tabular form, such as technical specification sheets, equipment maintenance records, etc. Such data can be directly converted into the form of a knowledge graph without extraction, that is, in the form of triples. Semi-structured data has a certain structure, such as web page data, and information can be extracted through HTML parsing and selector technology, and entities, relationships, and attribute information can be identified based on the extracted information.

[0078] The BiLSTM-CRF model is a deep learning model for natural language processing tasks. It integrates two key components: BiLSTM stands for Bidirectional Long Short-Term Memory. Unlike traditional unidirectional recurrent neural networks, BiLSTM includes two recurrent neural networks: one from left to right (forward) and the other from right to left (backward). This allows the model to simultaneously consider contextual information from both the preceding and following text, helping to better capture dependencies and patterns in sequence data. CRF stands for Conditional Random Field, a statistical modeling method for sequence labeling tasks. In the BiLSTM-CRF model, CRF is used to capture dependencies between entities, better capturing their boundaries and structure.

[0079] In a specific embodiment, converting triple data into question-answer pairs in Alpaca format based on the question-answer template specifically includes the following steps:

[0080] The natural language processing model is the PEGASUS model. The question in the question-answer pair is defined as the first information, and the answer corresponding to the question is defined as the second information. All the second information related to each first information is collected and input into the PEGASUS model. The PEGASUS model integrates all the second information to obtain the third information. The third information is a coherent answer containing all the second information. The optimized question-answer pair is formed based on the first and third information.

[0081] Specifically, since an entity may be associated with multiple other entities, answers extracted directly from the knowledge graph may have many variations. To provide more comprehensive and accurate information, these answers need to be integrated. The PEGASUS model is a pre-trained model for text summarization. In the context of data augmentation, the PEGASUS model may be used to generate or optimize answers to make them more coherent and comprehensive.

[0082] For example, when a user asks, "What are the components of a hydraulic support?" there may be multiple correct answers, such as "A hydraulic support includes columns" or "A hydraulic support includes an oil pump." Although these answers have similar structures, they provide different information. It's possible to consolidate these fragmented answers into a more comprehensive answer. For example, "A hydraulic support includes columns" and "A hydraulic support includes an oil pump" could be combined into "A hydraulic support includes columns, an oil pump, boots, panels, etc."

[0083] In a specific embodiment, fine-tuning the large language model based on LoRA technology specifically includes the following steps:

[0084] In the pre-training stage of the large language model, the LoRA technology is used to keep the original weight matrix of each Transformer layer in the Transformer model unchanged, and the low-rank matrix A and low-rank matrix B are introduced in the training stage. The parameters of the two low-rank matrices are adjusted, and the fine-tuned first weight matrix W′ is obtained based on the first formula. The first formula is: W′=W+A·B, where W is the original weight matrix and W∈R d×k , indicating that the original weight matrix is ​​a real number matrix with d rows and k columns, A∈R d×r , B∈R r×k , r represents the rank of matrix A or matrix B, where r<<min(d,k). The first weight matrix is ​​used to train on the optimized question-answer pairs to achieve fine-tuning of the large language model.

[0085] Specifically, in the pre-training stage of the large language model, the model has been trained on a large amount of text data and has learned rich language features and knowledge. The LoRA technology allows the original weight matrix of each layer in the Transformer model to remain unchanged during the fine-tuning process. Instead, the large model is fine-tuned by introducing two low-rank matrices. The rank of the low-rank matrix is ​​much smaller than the dimension of the original weight matrix, which means that they contain fewer parameters. The first weight matrix W is used to train on the optimized question-answer pairs, which means that the model will be further trained on a specific task dataset to improve its performance on the task. In this way, the large language model can adapt to new task requirements while maintaining pre-training knowledge, achieve effective fine-tuning, and reduce the number of parameters that need to be updated during fine-tuning, thereby improving the efficiency of fine-tuning.

[0086] The loss rate of fine-tuning training is as follows Figure 6 As shown in the figure, it can be seen that the training loss tends to be stable after a certain number of iterations, and there is no obvious oscillation or increase. This shows that the model does not overfit during the training process and the training process is relatively stable.

[0087] In a specific embodiment, quantifying the evaluation index specifically includes the following steps:

[0088] Based on the evaluation indicators, a factor set is constructed, and the evaluation indicators are graded to obtain a comment set. The membership function is set, and the membership degree μ(x) of each evaluation indicator to the comment is set based on the membership function. The membership function is: x is the real-time parameter of each evaluation indicator of the fully mechanized mining equipment, a, b, c, and d are four preset key parameters, the membership degrees are summarized to obtain a single-factor evaluation matrix, the weight coefficient of each evaluation indicator is set, and a fuzzy synthesis operation is performed based on the weight coefficient and the evaluation matrix to obtain a comprehensive evaluation result vector, which is then normalized and used as the quantitative result.

[0089] Specifically, if Figure 7 As shown in the figure, it is the principle flow chart of the comprehensive fuzzy evaluation expert module. The fuzzy comprehensive evaluation method is first adopted. Based on the fuzzy set theory, the qualitative evaluation of the fully mechanized mining equipment components is converted into a quantitative evaluation. The membership function is used to describe the degree of membership of each influencing factor to the equipment evaluation level. The fuzzy comprehensive evaluation requires the construction of the factor set U, the comment set V and the weight vector A, which are specifically expressed as follows: U = {u1,u2,…,u n}, V={v1,v2,…,v m}, A=(a1,a2,…,a n ), Among them, the factor set U contains all evaluation indicators, namely: U = {hydraulic support working pressure, support system support initial support force, coal mining machine traction speed, coal mining machine cutting power, scraper conveyor chain speed}, and the comment set V is divided into 4 levels: V = {fault, abnormal, normal, good}. The membership degree is obtained by assigning a membership function. The membership function of the present invention adopts a trapezoidal distribution. The trapezoidal distribution function is defined according to the threshold interval of the equipment parameter. The interval classification in the fault tree structure is used as the boundary value of the trapezoidal distribution. The membership function is used to define the membership degree of a single parameter value to a certain comment;

[0090] According to expert evaluation, the weights of the above factors in the fuzzy evaluation decision are: A = (0.25, 0.25, 0.15, 0.15, 0.20);

[0091] Based on real-time data, the degree of membership of each parameter to each comment is calculated to form a matrix single factor evaluation matrix R. This matrix represents the degree of membership of factors to comments and summarizes the degree of membership of all factors to each comment. It is the premise for comprehensive evaluation. ij ∈[0,1]. The specific evaluation matrix is ​​as follows: Selecting fuzzy synthesis operators Calculate the final comprehensive evaluation result vector B and perform normalization processing on the comprehensive evaluation result vector B to obtain B norm , the formula is as follows: After the above expert rule setting and fuzzy comprehensive evaluation, B is obtained based on the real-time collected data. norm , the expert rule module in this section is based on B norm The results determine the status of specific equipment components and input them into the large language model to provide support for subsequent decision-making assistance functions.

[0092] In a specific embodiment, constructing a Cypher query statement based on the identified entities and intents specifically includes the following steps:

[0093] The input natural language is segmented, and each word is tagged with parts of speech and input into the fine-tuned large language model. The model outputs words or phrases that are identified as entities or intents. The identified entities are mapped to corresponding nodes in the Neo4j database. The identified intents are matched with the pre-set intent list in the system to obtain the intent type. If there are multiple intent types, the multiple intents are merged to obtain a comprehensive intent.

[0094] The fully-mechanized mining equipment decision support system determines the query relationship between entities based on the identified intent, constructs a Cypher query statement, retrieves the query relationship in the Neo4j database based on the Cypher query statement, and returns the query results to the fully-mechanized mining equipment decision support system.

[0095] Specifically, the input natural language is segmented and the part of speech and category of each word are marked. Generally, the jieba library in Python or some Chinese language models are used. The language model is used to identify entities such as device names and fault types in the input. All entities are extracted from the input and mapped to the corresponding nodes in the graph database Neo4j. Entity queries can query the corresponding nodes in the knowledge graph, but it is meaningless for the graph database to only return node information. A subgraph full of nodes and relationships must be returned to provide sufficient information for the large model to support decision-making. Therefore, intention queries are also required to explore the relationships that need to be used.

[0096] The present invention uses a predefined relationship list to efficiently execute intent queries. The following are the steps for intent queries: 1. Word segmentation and labeling: Use the word segmentation and labeling steps in entity query to process intent information; 2. Predefined intent classification: Match the extracted keywords with the predefined intent list in the system to determine the query type, such as "what to do" and "how to deal with it" can be divided into the intent of "find a solution"; 3. Multi-intent processing: For cases where the input contains multiple intents, the system will merge multiple intents and generate corresponding multi-level queries.

[0097] Based on the input entities and relationships, the system constructs corresponding Cypher queries. For example, when the input is "What should I do if there is a hydraulic oil leak?", the system recognizes the two entities "hydraulic oil" and "leakage" and the intent "find a solution," converting it into a Cypher query. The Cypher query statement includes: MATCHpath = (e1)-[:Fault symptom is]->(e2)-[:Solution is]->(e3);

[0098] WHERE e1.name = 'Hydraulic oil' AND e2.name = 'Leak'; RETURN path;

[0099] Through Cypher query statements, the system can query the associated subgraph in the knowledge graph. This subgraph can be called a local feature subgraph, which focuses on the key information in the input.

[0100] In a specific embodiment, performing a multi-level relational reasoning query to obtain an information redundant subgraph specifically includes the following steps:

[0101] A query path is constructed based on the identified entities and the relationships between them. The relationships between entities include first-degree relationships, second-degree relationships, and third-degree relationships. The comprehensive mining equipment decision support system queries the local feature subgraph based on the query path of the first-degree relationship. If the local feature subgraph is not obtained, the query path is expanded to the second-degree relationship and third-degree relationship to obtain the corresponding subgraph information. The relevance of each subgraph information is scored, and the subgraph information with the highest score is defined as an information redundant subgraph for output.

[0102] Specifically, multi-level relational reasoning queries have the following steps: 1. Path construction: Path construction usually starts from known entities in the input and expands along layers of relational paths; 2. Relationship level progression: The system will increase the relationship level of reasoning, from first-degree relationships, that is, relationships directly related to the target entity, to second-degree, third-degree and even higher levels to find possible related information; 3. Relevance scoring: The system performs a relevance score on each acquired subgraph information. The score is based on the relevance of the subgraph content to the user query, and may involve natural language processing technology to evaluate the relevance of information. From all acquired subgraphs, the subgraph with the highest score is selected as the information redundant subgraph. Although this subgraph may contain some redundant information, it provides the most comprehensive data to support decision-making.

[0103] Multi-level relational reasoning queries can reveal redundant subgraphs, which may contain irrelevant or redundant data. While the query results may not be as relevant as the local feature subgraphs obtained through entity and intent queries, they do provide robust data support for decision-making within the comprehensive mining equipment model, improving the system's fault tolerance when entity and intent queries fail.

[0104] In a specific embodiment, constructing prompt words based on the local feature subgraph, the information redundancy subgraph, and the quantization result specifically includes the following steps:

[0105] The decision support system for comprehensive mining equipment extracts key information from local feature subgraphs or information redundant subgraphs. The key information includes entity names, types, attributes, and relationship information between entities. The key information is converted into structured information, and prompt words are generated by combining the natural language input by the user, structured information, and quantitative results.

[0106] Specifically, the system extracts key information from local feature subgraphs or information-redundant subgraphs. This key information contains entities and relationships related to the user's query, identifying the entity names, types, attributes, and relationship information between entities in the subgraph. For example, if the subgraph contains the entity "hydraulic support," the system extracts its attributes (such as model, manufacturer) and relationships with other entities (such as "connected to" conveyor). The extracted key information is converted into a structured format, such as JSON or XML. The natural language input by the user is combined with the structured information and quantitative results to ensure that the prompt word not only contains information from the graph database, but also reflects the user's original query and the quantitative results of the evaluation indicators. The combined information is used to construct a prompt word, which is a prompt-guided sentence. The constructed prompt word is presented in natural language and contains the text content parsed by the subgraph. The constructed prompt word is spliced ​​with the input, fully considering the real and powerful information from the graph database, guiding the model to generate accurate output.

[0107] Example 2:

[0108] The above describes a large model decision support method for fully mechanized mining equipment in the embodiment of the present application. The following describes a large model decision support system for fully mechanized mining equipment in the embodiment of the present application. Figure 8 In one embodiment of the present application, a large-scale model decision support system for fully mechanized mining equipment includes:

[0109] Definition module, used to obtain data information of fully mechanized mining equipment, classify the data information to obtain multiple information categories, define the entities in each information category, and the relationship information between entities;

[0110] The triple data processing module is used to identify the data format of each data information, extract the entity, relationship information and entity attribute values ​​from the data information in different ways based on the different data formats, form triple data, and store the triple data in the Neo4j database;

[0111] The question-answer pair formation module is used to establish a question-answer template, convert triple data into question-answer pairs in Alpaca format based on the question-answer template, establish a natural language processing model, and optimize the question-answer pairs based on the natural language processing model to obtain optimized question-answer pairs;

[0112] The fuzzy comprehensive evaluation module is used to establish a large language model for fully mechanized mining equipment. The large language model uses the Transformer model as its basic architecture and uses the optimized question-answer pairs as the training set to train the large language model. During the training process, the large language model is fine-tuned based on LoRA technology, and multiple evaluation indicators for fully mechanized mining equipment are defined and quantified to obtain quantitative results.

[0113] The query statement construction module is used to load the fine-tuned large language model. Users input natural language into the fully-mechanized mining equipment decision support system to query fully-mechanized mining equipment information. The fully-mechanized mining equipment decision support system uses the fine-tuned large language model to perform entity recognition and intent recognition on the natural language, and constructs Cypher query statements based on the recognized entities and intents.

[0114] The query module is used to use Cypher query statements to query the Neo4j database and determine whether the query results are obtained. If so, a local feature subgraph is obtained based on the query results. If no query results are obtained, a multi-level relationship reasoning query is performed to obtain an information redundant subgraph. A prompt word is constructed based on the local feature subgraph or the information redundant subgraph, and the prompt word is passed to the large language model to obtain retrieval information that supports the large model's decision-making.

[0115] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0116] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0117] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A large-scale model decision support method for fully mechanized mining equipment, characterized in that: The method comprises: Step S1: Acquire data information of fully mechanized mining equipment, classify the data information to obtain multiple information categories, define entities in each information category, and define relationship information between entities; Step S2: Identify the data format of each data information, extract the entity, relationship information and attribute values ​​of the entity from the data information in different ways based on the different data formats, form triple data, and store the triple data in the Neo4j database; Step S3: Establish a question-answer template, convert the triple data into question-answer pairs in Alpaca format based on the question-answer template, establish a natural language processing model, and optimize the question-answer pairs based on the natural language processing model to obtain optimized question-answer pairs; Step S4: Establish a large language model for fully mechanized mining equipment. The large language model uses the Transformer model as its basic architecture, and uses the optimized question-answer pairs as the training set to train the large language model. During the training process, the large language model is fine-tuned based on LoRA technology, and multiple evaluation indicators for fully mechanized mining equipment are defined. The evaluation indicators are quantified to obtain quantitative results. Step S5: The fine-tuned large language model is loaded, and the user inputs natural language into the fully-mechanized mining equipment decision support system to query fully-mechanized mining equipment information. The fully-mechanized mining equipment decision support system uses the fine-tuned large language model to perform entity recognition and intent recognition on the natural language, and constructs a Cypher query statement based on the recognized entities and intents. Step S6: Use Cypher query statements to query the Neo4j database to determine whether query results are obtained. If so, obtain a local feature subgraph based on the query results. If no query results are obtained, perform a multi-level relational reasoning query to obtain an information redundancy subgraph. Based on the local feature subgraph, the information redundancy subgraph, and the quantification results, construct prompt words, and pass the prompt words to the large language model to obtain retrieval information that supports the large model's decision-making.

2. A large-scale model decision support method for fully mechanized mining equipment according to claim 1, characterized in that: Form triple data, including: The data formats include structured data, semi-structured data and unstructured data. If the data information format is structured data, the data information is directly converted into triple data. If the data information format is semi-structured data, the semi-structured data is parsed to obtain knowledge information, and the knowledge information is extracted to obtain entity, relationship information and attribute values ​​of the entity to obtain triple data. If the format of the data information is unstructured data, a BiLSTM-CRF model is established. The BiLSTM-CR model includes a bidirectional long short-term memory network model and a conditional random field discriminant model. The context information in the unstructured data is obtained based on the bidirectional long short-term memory network model, and the entity information is identified from the unstructured data based on the context information. The boundary of the entity information is identified based on the conditional random field discriminant model to obtain complete entity information and relationship information between entities, extract the attribute information of the entity, and form triple data based on the entity information, relationship information and attribute information.

3. A large-scale model decision support method for fully mechanized mining equipment according to claim 2, characterized in that: Converting the triple data into question-answer pairs in Alpaca format based on the question-answer template includes: The natural language processing model is a PEGASUS model, which defines the question in the question-answer pair as the first information, and the answer corresponding to the question as the second information. All second information related to each first information is collected, and the second information is input into the PEGASUS model. The PEGASUS model fuses all the second information to obtain third information. The third information is a coherent answer containing all the second information, and an optimized question-answer pair is constructed based on the first information and the third information.

4. A large-scale model decision support method for fully mechanized mining equipment according to claim 1, characterized in that: Fine-tune the large language model based on LoRA technology, including: In the pre-training stage of the large language model, the LoRA technique is used to keep the original weight matrix of each Transformer layer in the Transformer model unchanged, and low-rank matrix A and low-rank matrix B are introduced in the training stage. The parameters of the two low-rank matrices are adjusted, and the fine-tuned first weight matrix W is obtained based on the first formula. ′ , and the first formula is: W ′ = W + A·B, where W is the original weight matrix, and W ∈ R d×k , indicating that the original weight matrix is a real matrix with d rows and k columns, A ∈ R d×r , B ∈ R r×k , r represents the rank of matrix A or matrix B, where r << min(d, k). The large language model is fine-tuned by training with the first weight matrix on the optimized question-and-answer pairs.

5. A large-scale model decision support method for fully mechanized mining equipment according to claim 1, characterized in that: The evaluation indicators are quantified, including: A factor set is constructed based on the evaluation indicators, the evaluation indicators are graded to obtain a comment set, a membership function is set, and the membership degree μ(x) of each evaluation indicator to the comment is set based on the membership function. The membership function is: x is the real-time parameter of each evaluation indicator of the comprehensive mining equipment, a, b, c, and d are four preset key parameters, the membership degrees are summarized to obtain a single-factor evaluation matrix, a weight coefficient of each evaluation indicator is set, and a fuzzy synthesis operation is performed based on the weight coefficient and the evaluation matrix to obtain a comprehensive evaluation result vector, which is then normalized and used as the quantified result.

6. According to the large-scale model decision support method for fully mechanized mining equipment of claim 5, a Cypher query statement is constructed based on the identified entities and intents, including: The input natural language is segmented, each segmented word is tagged with a part-of-speech tag and input into the fine-tuned large language model. The model outputs words or phrases that are identified as entities or intents. The identified entities are mapped to corresponding nodes in the Neo4j database. The identified intents are matched with the preset intent list in the system to obtain the intent type. If there are multiple intent types, the multiple intents are merged to obtain a comprehensive intent. The fully mechanized mining equipment decision support system determines the query relationship between entities based on the identified intention, constructs a Cypher query statement, retrieves the query relationship in the Neo4j database based on the Cypher query statement, and obtains the query result and returns it to the fully mechanized mining equipment decision support system.

7. A large-scale model decision support method for fully mechanized mining equipment according to claim 6, characterized in that: Perform multi-level relational reasoning queries to obtain information redundant subgraphs, including: A query path is constructed based on the identified entities and the relationships between them. The relationships between entities include first-degree relationships, second-degree relationships, and third-degree relationships. The comprehensive mining equipment decision support system queries the local feature subgraph based on the query path of the first-degree relationship. If the local feature subgraph is not obtained, the query path is expanded to the second-degree relationship and the third-degree relationship in sequence to obtain the corresponding subgraph information. The relevance of each subgraph information is scored, and the subgraph information with a high score is defined as an information redundant subgraph for output.

8. A large-scale model decision support method for fully mechanized mining equipment according to claim 7, characterized in that: Constructing a prompt word based on the local feature subgraph, the information redundancy subgraph, and the quantization result includes: The comprehensive mining equipment decision support system extracts key information from the local feature subgraph or the information redundant subgraph, wherein the key information includes entity name, type, attribute and relationship information between entities, converts the key information into structured information, and combines the natural language input by the user, the structured information and the quantitative results to generate prompt words.

9. A large-scale model decision support method for fully mechanized mining equipment according to claim 1, characterized in that: The information categories include equipment information, risk information, fault information, cause information, solution information and specification information.

10. A large-scale model decision support system for fully mechanized mining equipment, used to implement a large-scale model decision support method for fully mechanized mining equipment according to any one of claims 1 to 9, characterized in that: The system comprises: A definition module is used to obtain data information of fully mechanized mining equipment, classify the data information to obtain multiple information categories, define entities in each information category, and define relationship information between entities; A triple data processing module is used to identify the data format of each data information, extract entity, relationship information and entity attribute values ​​from the data information in different ways based on different data formats, form triple data, and store the triple data in the Neo4j database; A question-answer pair formation module is used to establish a question-answer template, convert the triple data into a question-answer pair in Alpaca format based on the question-answer template, establish a natural language processing model, and optimize the question-answer pair based on the natural language processing model to obtain an optimized question-answer pair; A fuzzy comprehensive evaluation module is used to establish a large language model for fully mechanized mining equipment. The large language model uses the Transformer model as its basic architecture and uses the optimized question-answer pairs as a training set to train the large language model. During the training process, the large language model is fine-tuned based on LoRA technology, and multiple evaluation indicators for the fully mechanized mining equipment are defined and quantified to obtain quantitative results. A query statement construction module is used to load the fine-tuned large language model. The user inputs natural language into the fully-mechanized mining equipment decision support system to query fully-mechanized mining equipment information. The fully-mechanized mining equipment decision support system uses the fine-tuned large language model to perform entity recognition and intent recognition on the natural language, and constructs a Cypher query statement based on the recognized entities and intents; The query module is used to use Cypher query statements to query the Neo4j database and determine whether a query result is obtained. If so, a local feature subgraph is obtained based on the query result. If no query result is obtained, a multi-level relationship reasoning query is performed to obtain an information redundant subgraph. A prompt word is constructed based on the local feature subgraph or the information redundant subgraph, and the prompt word is passed to the large language model to obtain retrieval information that supports the large model's decision.

Images