Metal mine scheduling decision method and system based on large model

Abstract

The present application relates to the technical field of mine early warning and scheduling, and provides a metal mine scheduling decision method and system based on a large model, wherein the method of the present application breaks the bottom-layer data island of static knowledge and dynamic Internet of Things information of a mine by constructing a unified interactive architecture that fuses multiple source heterogeneous data, significantly reduces the time cost of field personnel in obtaining key production information, and uses a large model intention analysis method based on a few-sample thinking chain to accurately separate static procedure queries and dynamic production scheduling demands, thereby improving the distribution and routing accuracy of complex industrial instructions. Through a scheduling auxiliary decision method that fuses real-time Internet of Things status and mixed integer linear programming, optimal production scheduling parameters are solved under the satisfaction of physical constraints such as ventilation and safety distance, thereby realizing a leap from passive knowledge retrieval to active intelligent production scheduling. In summary, the method of the present application greatly improves the efficiency and accuracy of mine scheduling and decision-making.

Description

Technical Field

[0001] This invention relates to the field of mine early warning and scheduling technology, specifically to a metal mine scheduling decision-making method and system based on a large model. Background Technology

[0002] With the continuous advancement of informatization and intelligentization in mining enterprises, numerous business systems and monitoring devices have been introduced into mine production management to improve overall production efficiency and safety management. However, in actual daily production management and scheduling, the existing information system architecture still faces significant limitations. Currently, the IoT equipment operation monitoring system, production scheduling system, and static safety procedures and operation document library within the mine are often independent and fragmented. Due to the multiple and highly dispersed data sources, on-site staff and managers often need to traverse multiple independent terminals and systems to retrieve dynamic equipment operating parameters and static professional procedures when facing specific process scheduling or daily operations. This multi-channel, fragmented information acquisition and manual integration method not only consumes a lot of time but also greatly increases the cognitive load on personnel, leading to low efficiency in production scheduling and management decisions. Furthermore, mine production operations cover extremely complex and highly interdisciplinary professional fields such as mining, tunneling, electromechanical, transportation, and ventilation. Due to the sheer volume and high level of expertise required for safety regulations, operating procedures, and equipment maintenance manuals, managers and on-site dispatchers struggle to fully and accurately grasp all the cross-disciplinary details. In the absence of unified and efficient knowledge integration and analytical tools, decision-makers, when making production scheduling arrangements or handling complex operating conditions, are easily limited by their own knowledge gaps and experience if they rely solely on personal experience. This can lead to biased judgments under complex process requirements and safety constraints, resulting in violations or incorrect scheduling decisions.

[0003] Therefore, how to break down the barriers between the dynamic operational data of the Internet of Things at the bottom level of the mine and the static professional procedures at the top level, and build a system that can effectively integrate multi-source heterogeneous data, reduce the threshold for information access, and thus provide accurate auxiliary decision-making suggestions for staff, is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] The present invention aims to provide a metal mine scheduling decision-making method and system based on a large model that can accurately assist in mine scheduling decisions.

[0005] To achieve the above objectives, the first aspect of the present invention provides a metal mine scheduling decision-making method based on a large model, the method comprising:

[0006] S1. Receive natural language query commands input by mine staff; S2. Construct a pre-trained large language model to extract semantic features and classify intents of the natural language query instructions, thereby obtaining the static procedure query intent and dynamic process scheduling intent of the natural language query instructions. S3. Based on the first route identifier determined by the static procedure query intent, perform text matching and retrieval in the mine-specific knowledge base file, and combine the retrieved procedure clauses and knowledge fragments to reconstruct and generate a regular question-and-answer response in natural language format through the large language model; S4. Based on the second route identifier determined by the dynamic process scheduling intention, extract the corresponding real-time equipment operation data from the underlying database of the Internet of Things system, input the real-time equipment operation data into the pre-built operations research optimization model for solution, and transform the solved optimal process scheduling parameters into comprehensive auxiliary decision-making suggestions in natural language format through the large language model.

[0007] In one embodiment, the construction of a pre-trained large language model to extract semantic features and classify intent for the natural language query instruction includes: The natural language query command is subjected to text denoising processing and then concatenated with a preset few-sample scenario template in the mining field to generate an enhanced prompt sequence; The enhanced prompt sequence is input into the Transformer encoding layer of the large language model, and an inference intermediate state containing action entities and constraints is generated based on the autoregressive mechanism. Calculate the classification probability distribution of the inference intermediate state in the preset intent slot, and extract the target intent with the highest classification probability and higher than the confidence threshold; Based on the stated target intent, the natural language query instruction is determined to be either a static procedure query intent or a dynamic process scheduling intent.

[0008] In one embodiment, the first route identifier determined based on the static procedure query intent is used to perform text matching and retrieval in the mine-specific knowledge base file. Combined with the retrieved procedure clauses and knowledge fragments, a regular question-and-answer response in natural language format is generated through reconstruction using the large language model, including: An information extraction model is used to extract entities and relationships from unstructured mine regulations texts and assemble them into a knowledge graph. At the same time, a text embedding model is used to map the mine regulations texts into dense vectors to establish a vector index. Based on the first route identifier, a high-dimensional query vector is extracted, and a local semantic vector retrieval is performed to obtain a local semantic slice set; at the same time, based on the first route identifier, named entities are extracted and used as the starting point to perform a global topology graph retrieval to obtain a global knowledge subgraph set containing logical topology; The local semantic slice set and the global knowledge subgraph set are deduplicated and fused with lightweight features; Based on the local semantic slice set and the global knowledge subgraph set after deduplication and lightweight feature fusion, an enhanced context boundary with graph structure constraints is constructed. The enhanced context boundary with graph structure constraints is input into the large language model, and a regular question-and-answer response that follows mine safety regulations is generated through semantic reconstruction.

[0009] In one embodiment, the local semantic vector retrieval, which yields a set of local semantic slices, includes: Calculate the query vector q and the knowledge slice vector in the HNSW vector index. cosine similarity distance matrix It satisfies the following expression: ; The total number of dimensions of a high-dimensional vector; Let be the feature value of the i-th knowledge slice vector in the j-th dimension; To query the feature value of vector q in the j-th dimension; Recall the set of local semantic slices with the highest similarity scores.

[0010] In one embodiment, the global topology graph retrieval, which yields a global knowledge subgraph containing logical topology, includes: Starting with named entities, a multi-hop topological traversal of a graph attention network is performed in the mine knowledge graph. The attention coefficients of the graph attention network are... Satisfy the following expression: ; in, The eigenvectors of node m; The feature vector representing node n; Let W represent the feature vector of node k, which is used as the iteration variable when traversing the set of neighboring nodes; W is the shared weight matrix of the linear transformation of node features. For the set of neighboring nodes; is the transpose of the weight vector of a single-layer feedforward neural network, used to map the concatenated high-dimensional features to an attention scalar; k is the set of neighbor nodes. Index in; Recall the global knowledge subgraph containing logical topological relationships.

[0011] In one embodiment, S4 includes: Based on the second route identifier, the target device category is parsed to obtain a multi-dimensional operational status matrix captured in real time; The multidimensional operating state matrix is ​​used as a dynamic initial boundary condition input to the scheduling algorithm model; with the objective function of maximizing production scheduling efficiency or minimizing completion time, mixed integer linear programming combined with heuristic rules is used to iteratively solve the optimal process scheduling parameters in the pre-constructed operations research optimization model; The optimal process scheduling parameters are injected into the large language model as prior knowledge, and natural language generation technology is used to transform them into logically coherent comprehensive auxiliary decision-making suggestions and output them.

[0012] In one embodiment, the objective function of the production scheduling process constraint model in the pre-constructed operations research optimization model is: ; Where Z represents the minimum total completion time or total operation cost of the current scheduling scheme; N represents the total number of loaders currently participating in the scheduling; and M represents the total number of mining points currently awaiting operation. , which is a 0-1 decision variable, represents the estimated time for the uth loader to complete the task at the vth mining point. Its constraints include the uniqueness constraint of equipment allocation and the safety ventilation constraint of the roadway. This represents the estimated time for the u-th loader to complete the task at the v-th mining point.

[0013] The second aspect of the present invention provides a metal mine scheduling decision system based on a large model, used to execute the metal mine scheduling decision method based on a large model as described above. The system is deployed on a mine private cloud server and includes mine terminal equipment, a multi-source heterogeneous data access and interaction module, a large model intent parsing and routing module, a knowledge base retrieval question and answer engine, an IoT scheduling auxiliary decision engine, a mine-specific knowledge base, and an IoT underlying database. The multi-source heterogeneous data access and interaction module is used to receive natural language query commands input by mine workers through the mine terminal equipment, and simultaneously capture static procedure data in the mine-specific knowledge base and real-time equipment operation data in the underlying IoT database based on the natural language query commands. The large model intent parsing and routing module has a built-in pre-trained large language model, which is used to extract semantic features and classify intents of the natural language query instructions. Based on the results of semantic feature extraction and intent classification of the natural language query instructions, it triggers the knowledge base retrieval question answering engine or the IoT scheduling auxiliary decision engine. The knowledge base retrieval question-answering engine is used to perform text matching and retrieval in the mine-specific knowledge base file based on the received natural language query command, and combine the retrieved procedural clauses and knowledge fragments to reconstruct and generate a regular question-answering response in natural language format through the large language model, and output it to the terminal device. The IoT scheduling auxiliary decision engine is used to extract corresponding real-time device operation data from the underlying database of the IoT system according to the received natural language query command, input the real-time device operation data into a pre-built operations research optimization model for solving, and convert the solved optimal process scheduling parameters into comprehensive auxiliary decision suggestions in natural language format through the large language model, and output them to the terminal device.

[0014] In one embodiment, the static procedure data includes non-mining safety procedures, blasting operation standards, and mining equipment maintenance manuals; the real-time equipment operation data includes underground ventilation wind speed, real-time load of the hoist, and real-time position of the underground loader.

[0015] A third aspect of the present invention provides a computer system comprising: a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the metal mine scheduling decision method based on a large model as described above.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a large-scale model-based scheduling and decision-making method for metal mines. By constructing a unified interactive architecture that integrates multi-source heterogeneous data, it breaks down the underlying data silos between static mine knowledge and dynamic IoT information, significantly reducing the time cost for on-site personnel to obtain key production information. Based on a small-sample thought chain-based large-scale model intent parsing method, it can accurately separate static procedure queries from dynamic production scheduling requests, improving the distribution and routing accuracy of complex industrial instructions. Through a scheduling-aided decision-making method that integrates real-time IoT status and mixed-integer linear programming, it solves for optimal production scheduling parameters while satisfying physical constraints such as ventilation and safety distances, achieving a leap from passive knowledge retrieval to proactive intelligent production scheduling. In summary, this invention greatly improves the efficiency and accuracy of mine scheduling and decision-making. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a metal mine scheduling decision-making method based on a large model, provided as an embodiment of the present invention; Figure 2 A flowchart of a metal mine scheduling decision system based on a large model, provided for embodiments of the present invention; Figure 3 This is a flowchart illustrating the semantic feature extraction and intent classification performed by the large model intent parsing and routing module in this embodiment of the invention. Figure 4 This is a flowchart of step S3 in which the knowledge base retrieval question-answering engine responds to the first route identifier in an embodiment of the present invention. Detailed Implementation

[0019] To facilitate understanding of the present invention, the present invention will be described more fully and in detail below with reference to the accompanying drawings and preferred embodiments, but the scope of protection of the present invention is not limited to the following specific embodiments.

[0020] Unless otherwise defined, all technical terms used herein have the same meaning as commonly understood by those skilled in the art. The technical terms used herein are for the purpose of describing particular embodiments only and are not intended to limit the scope of the invention.

[0021] Unless otherwise specified, all raw materials, reagents, instruments and equipment used in this invention can be purchased from the market or prepared by existing methods.

[0022] like Figure 1 As shown, an embodiment of a metal mine scheduling decision-making method based on a large model includes the following steps: S1: Receive natural language query commands input by mine staff.

[0023] S2. Construct a pre-trained large language model to extract semantic features and classify intents of the natural language query instructions, thereby obtaining the static procedure query intent and dynamic process scheduling intent of the natural language query instructions.

[0024] In some embodiments, the construction of a pre-trained large language model to extract semantic features and classify intent for the natural language query instruction specifically includes: The natural language query command is denoised and then concatenated with a preset template of a few-sample scenario in the mining field to generate an enhanced prompt sequence; The enhanced prompt sequence is input into the Transformer encoding layer of a pre-trained large language model, and an inference intermediate state containing action entities and constraints is generated based on the autoregressive mechanism. Calculate the classification probability distribution of the inference intermediate state in the preset intent slot, and extract the target intent with the highest classification probability and higher than the confidence threshold; Based on the stated target intent, the natural language query command is determined to be either a static procedure query intent or a dynamic process scheduling intent.

[0025] This step, based on the Few-ShotCoT (Few-ShotCoT) large model intent parsing method, can accurately separate static procedure queries from dynamic production scheduling requests, improving the distribution and routing accuracy of complex industrial instructions.

[0026] S3. Based on the first route identifier, perform text matching and retrieval in the mine-specific knowledge base file, and combine the retrieved procedural clauses and knowledge fragments to reconstruct and generate a regular question-and-answer response in natural language format through the large language model.

[0027] In some embodiments, the step of performing text matching and retrieval in a mine-specific knowledge base file based on a first route identifier, and combining the retrieved procedural clauses and knowledge fragments, to reconstruct and generate a regular question-and-answer response in natural language format through the large language model, specifically includes: An information extraction model is used to extract entities and relationships from unstructured mine regulations texts and assemble them into a knowledge graph. At the same time, a text embedding model is used to map the regulations texts into dense vectors to establish a vector index. High-dimensional query vectors are extracted based on the first route identifier, and local semantic vector retrieval is performed to obtain a local semantic slice set. At the same time, named entities are extracted based on the first route identifier and used as the starting point to perform global topology graph retrieval to obtain a global knowledge subgraph set containing logical topology. The local semantic slice set and the global knowledge subgraph set are deduplicated and fused with lightweight features; Based on the local semantic slice set and global knowledge subgraph set that have been deduplicated and feature-lightly fused, an enhanced context boundary with graph structure constraints is constructed. The enhanced context boundary with graph structure constraints is input into the large language model, and a regular question-and-answer response that follows mine safety regulations is generated through semantic reconstruction.

[0028] Local semantic vector retrieval and global topological graph retrieval, through a graph-text dual-path joint retrieval method based on the LightRAG architecture, provide strong physical and logical topological constraints for generating responses for large models, effectively suppressing "knowledge illusion" and avoiding the risk of illegal operations caused by knowledge blind spots.

[0029] In one embodiment, the local semantic vector retrieval includes: Calculate the query vector q and the knowledge slice vector in the HNSW vector index. cosine similarity distance matrix It satisfies the following expression: ; The total number of dimensions of a high-dimensional vector; Let be the feature value of the i-th knowledge slice vector in the j-th dimension; To query the feature value of vector q in the j-th dimension; Recall the set of local semantic slices with the highest similarity scores.

[0030] In one embodiment, the global topology map retrieval includes: Starting with named entities, a multi-hop topological traversal of a graph attention network is performed in the mine knowledge graph. The attention coefficients of the graph attention network are... Satisfy the following expression: ; in, The eigenvectors of node m; The feature vector representing node n; Let W represent the feature vector of node k, which is used as the iteration variable when traversing the set of neighboring nodes; W is the shared weight matrix of the linear transformation of node features. For the set of neighboring nodes; is the transpose of the weight vector of a single-layer feedforward neural network, used to map the concatenated high-dimensional features to an attention scalar; k is the set of neighbor nodes. Index in; Recall the global knowledge subgraph containing logical topological relationships.

[0031] S4. Based on the second route identifier, extract the corresponding real-time device operation data from the underlying database of the IoT system, input the real-time device operation data into the pre-built operations research optimization model for solution, and transform the solved optimal process scheduling parameters into comprehensive auxiliary decision-making suggestions in natural language format through the large language model.

[0032] In one embodiment, S4 specifically includes: Based on the second route identifier, the target device category is parsed to obtain a multi-dimensional operational status matrix captured in real time; The multidimensional operating state matrix is ​​used as a dynamic initial boundary condition input to the scheduling algorithm model; with the objective function of maximizing production scheduling efficiency or minimizing completion time, mixed integer linear programming combined with heuristic rules is used to iteratively solve the optimal process scheduling parameters in the pre-constructed operations research optimization model; The optimal process scheduling parameters are injected into the large language model as prior knowledge, and natural language generation technology is used to transform them into logically coherent comprehensive auxiliary decision-making suggestions and output them.

[0033] This step provides a scheduling-aided decision-making method that integrates real-time IoT status with mixed integer linear programming (MILP). It solves for the optimal scheduling parameters while satisfying physical constraints such as ventilation and safety distance, thus realizing a leap from passive knowledge retrieval to proactive intelligent scheduling.

[0034] In one embodiment, the optimal process scheduling parameters are iteratively solved in a pre-constructed operations research optimization model, wherein the objective function of the production scheduling process constraint model is:

[0035] Where Z represents the minimum total completion time or total operation cost of the current scheduling scheme; N represents the total number of loaders currently participating in the scheduling; and M represents the total number of mining points currently awaiting operation. , which is a 0-1 decision variable, represents the estimated time for the uth loader to complete the task at the vth mining point. Its constraints include the uniqueness constraint of equipment allocation and the safety ventilation constraint of the roadway. This represents the estimated time for the u-th loader to complete the task at the v-th mining point.

[0036] The metal mine scheduling decision-making method based on a large model provided in this embodiment breaks down the underlying data silos between static mine knowledge and dynamic IoT information by constructing a unified interactive architecture that integrates multi-source heterogeneous data. This significantly reduces the time cost for on-site personnel to obtain key production information. It has extremely high industrial applicability and decision-making reliability, and can provide strong intelligent support for the daily process scheduling, safety management, and capacity assessment of mining enterprises.

[0037] Please see Figure 2 This embodiment also provides a metal mine scheduling decision system based on a large model, used to execute the metal mine scheduling decision method based on a large model as described above. The system is deployed on a mine private cloud server and includes a mine terminal device 100, a multi-source heterogeneous data access and interaction module 210, a large model intent parsing and routing module 220, a knowledge base retrieval question answering engine 230, an IoT scheduling auxiliary decision engine 240, a mine-specific knowledge base 300, and an IoT underlying database 400. The multi-source heterogeneous data access and interaction module 210 is communicatively connected to the mine terminal device 100, the mine-specific knowledge base 300, and the IoT underlying database 400. Specifically, the multi-source heterogeneous data access and interaction module 210 receives natural language query commands input from the mine terminal device 100 and, based on these commands, simultaneously retrieves static procedure data from the mine-specific knowledge base 300 and real-time equipment operation data from the IoT underlying database 400. In this embodiment, the static procedure data includes non-mining safety procedures, blasting operation standards, and mining equipment maintenance manuals in PDF format; the real-time equipment operation data includes underground ventilation wind speed, real-time load of the hoist, and real-time position of the underground loader; the multi-source heterogeneous data access and interaction module 210 constructs a global operation matrix with real-time production context by semantically aligning the above-mentioned static procedure data and real-time equipment operation data.

[0038] The large model intent parsing and routing module has a built-in pre-trained large language model, which is used to extract semantic features and classify intents of the natural language query instructions. Based on the results of semantic feature extraction and intent classification of the natural language query instructions, it triggers the knowledge base retrieval question answering engine or the IoT scheduling auxiliary decision engine. like Figure 3 The flowchart shown is a process for semantic feature extraction and intent classification performed by the large model intent parsing and routing module 220: Step S101: Receive natural language query instructions and perform text denoising preprocessing; Step S102: The denoised instructions are concatenated with a preset mine few-sample scene template to generate an enhanced prompt sequence; Step S103: Input the enhanced cue sequence into the Transformer encoding layer of the pre-trained large language model; Step S104: Generate an intermediate inference state containing action entities and constraints using the autoregressive mechanism of a pre-trained large language model. Step S105: Calculate the classification probability distribution of the inference intermediate state in the preset intent slot; Step S106: Extract the target intent with the highest classification probability and higher than the confidence threshold, and determine the natural language query instruction as a static procedure query intent or a dynamic process scheduling intent based on the target intent.

[0039] Step S107: Determine the target intent. If it is determined to be a static procedure query intent, output the first route identifier to trigger the knowledge base retrieval question and answer engine 230; if it is determined to be a dynamic process scheduling intent, output the second route identifier to trigger the Internet of Things scheduling auxiliary decision engine 240. The knowledge base retrieval question-answering engine is used to perform text matching and retrieval in the mine-specific knowledge base file based on the received natural language query command, and combine the retrieved procedural clauses and knowledge fragments to reconstruct and generate a regular question-answering response in natural language format through the large language model, and output it to the terminal device. Please see Figure 4 The knowledge base retrieval question-answering engine 230 responds to the first route identifier and executes step S3 to achieve the following dual-path retrieval: After extracting the query vector q and the named entity in step S302, the system starts two retrieval branches in parallel.

[0040] The first branch is local semantic vector retrieval. In step S303, the system calculates the query vector q and the knowledge slice vector in the HNSW vector index. The cosine similarity distance matrix is ​​calculated using the following formula: ； Then, in step S305, the set of local semantic slices with the highest similarity scores is recalled.

[0041] The second branch is global topology graph retrieval. In step S304, starting with named entities, a multi-hop topology traversal of the Graph Attention Network (GAT) is performed in the mine knowledge graph. The attention coefficients of the GAT layer are... The calculation is as follows: ; in, Let W represent the node features, and W be the weight matrix. This is the set of neighboring nodes. The algorithm then uses step S306 to recall a global knowledge subgraph containing logical topological relationships.

[0042] Finally, in step S307, the results from both paths are fused, and in step S310, a question-and-answer response that complies with mine safety regulations is generated.

[0043] The IoT scheduling auxiliary decision engine is used to extract corresponding real-time device operation data from the underlying database of the IoT system according to the received natural language query command, input the real-time device operation data into a pre-built operations research optimization model for solution, and convert the solved optimal process scheduling parameters into comprehensive auxiliary decision suggestions in natural language format through the large language model, and output them to the terminal device.

[0044] This embodiment takes the scenario of "segmented loading and ore extraction coordinated scheduling" in underground mining operations as an example. The IoT scheduling auxiliary decision engine 240 responds to the second routing identifier and executes step S4, as follows: When a mine dispatcher issues an instruction via the mine terminal equipment 100, such as "Calculate the optimal ratio of the three loaders and two trucks in the 1050 section and the estimated completion time": In steps S401-S402, the instruction is forwarded to the large model intent parsing and routing module 220 via the interactive terminal interface 211 in the multi-source heterogeneous data access and interaction module 210. In step S403, the large model intent parsing and routing module 220 performs autoregressive inference based on the CoT (Cooperation of Thought) framework, extracts entity elements such as "1050 middle section", "loader", and "truck" through parsing, determines that the instruction involves real-time calculation of production process parameters, and generates a second routing identifier. In steps S404-S407, the IoT data bus 213 in the multi-source heterogeneous data access and interaction module 210 responds to the second routing identifier, calls the IoT underlying database 400, and obtains the current load, remaining fuel, lane coordinates, and truck trajectory data of each loader within the target area. The above dynamic parameters are encapsulated into a multi-dimensional operating state matrix and fed into the operations research and optimization solver 241 in the IoT scheduling auxiliary decision engine 240.

[0045] In step S408, the operations research solver 241 performs an iterative solution based on mixed-integer linear programming (MILP) with the objective function of minimizing the total operation delay and equipment energy efficiency cost. The specific production scheduling process constraint model is shown below: ; Where n represents the total number of loaders currently participating in the scheduling; m represents the total number of ore extraction points currently awaiting operation; and i is the index of the loader number ( j is the index of the mining site number ( ); For 0-1 decision variables, when The time indicates that the i-th loader will be assigned to the j-th ore outlet to perform the loading task; otherwise... ; This represents the estimated time for the i-th loader to complete the task at the j-th ore extraction point. Its constraints include: 1. Uniqueness constraint for equipment allocation: This is used to ensure that each loader is assigned to only one specific ore outlet during the current scheduling cycle; 2. Safety and ventilation constraints in roadways: This is used to ensure that the total required air volume of all active equipment allocated to this section of the level roadway does not exceed the maximum air supply volume of the ventilation system in that area. ,in Let be the air demand coefficient of equipment i.

[0046] In steps S409-S411, the operations research and optimization solver 241 outputs the optimal loader dispatching plan and the expected operating trajectory parameters of each device to the decision translator 242 in the IoT scheduling auxiliary decision engine 240. The decision translator 242, combined with the original query context, reconstructs the complex mathematical scheduling instructions into natural language suggestions (e.g., "It is suggested that loader No. 2 in section 1050 be moved to mine room No. 3, with an estimated completion time shortened by 15 minutes"), and transmits this information back via the interactive terminal interface 211. Finally, in step S412, the comprehensive result, including the scheduling suggestions, device location snapshots, and completion prediction curves, is visualized on the terminal device 100.

[0047] In one embodiment, the multi-source heterogeneous data access and interaction module includes an interactive terminal interface, a knowledge access agent, and an IoT data bus; The interactive terminal interface is used to receive natural language query commands and render and output conventional question-and-answer responses or comprehensive auxiliary decision-making suggestions generated by the large language model. The knowledge access agent is used to call the file server interface to read unstructured mine regulations text and structured equipment ledgers. The IoT data bus is used to connect to the application programming interface of the underlying relational database and capture a multi-dimensional operating status matrix in real time, including device spatial coordinates, instantaneous load and start / stop status.

[0048] In one embodiment, the large-scale model intent parsing and routing module is based on the intent classification algorithm flow of the thought chain, specifically including: The natural language query command is denoised and then concatenated with a preset template of a few-sample scenario in the mining field to generate an enhanced prompt sequence; The enhanced prompt sequence is input into the Transformer encoding layer of the built-in pre-trained large language model, and an inference intermediate state containing action entities and constraints is generated based on the autoregressive mechanism. Calculate the classification probability distribution of the inference intermediate state in the preset intent slot, and extract the target intent with the highest classification probability and higher than the confidence threshold; If the target intent is determined to be a static procedure query, a first route identifier is output to trigger the knowledge base retrieval question and answer engine; if the target intent is determined to be a dynamic process scheduling, a second route identifier is output to trigger the IoT scheduling auxiliary decision engine.

[0049] In one embodiment, the knowledge base retrieval question-answering engine is based on a lightweight graph retrieval enhanced generation architecture, and its graph-text dual-path joint retrieval and generation algorithm specifically includes: An information extraction model is used to extract entities and relationships from unstructured mine regulations texts and assemble them into a knowledge graph. At the same time, a text embedding model is used to map the regulations texts into dense vectors to establish a vector index. Extract a high-dimensional query vector carrying a first routing identifier instruction, calculate cosine similarity in the vector index to recall a local semantic slice set; simultaneously, starting from the named entity extracted from the instruction, perform a multi-hop topology traversal in the knowledge graph via a graph attention network to recall a global knowledge subgraph set containing logical topology. The deduplicated and fused local semantic slice set and the global knowledge subgraph set are used as context boundaries and input into the large language model. The conventional question-and-answer response that follows the mine safety regulations is generated through semantic reconstruction.

[0050] In one embodiment, the IoT scheduling auxiliary decision engine, combined with the operations research optimization algorithm process based on real-time operational data, specifically includes: Based on the second route identifier, the target device category is parsed to obtain a multi-dimensional operating status matrix captured in real time by the IoT data bus; The multidimensional operating state matrix is ​​used as a dynamic initial boundary condition and input into the built-in scheduling algorithm model. With the objective function of maximizing production scheduling efficiency or minimizing completion time, and under the physical constraints such as ventilation volume lower limit and safety distance, the optimal process scheduling parameters are iteratively solved by using mixed integer linear programming combined with heuristic rules. The process scheduling parameters are injected into the large language model as prior knowledge, and natural language generation technology is used to transform them into logically coherent comprehensive auxiliary decision-making suggestions and output them.

[0051] This embodiment also provides a computer system, including: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the metal mine scheduling decision method based on a large model as described above.

[0052] The computer system can be a computing device such as a mobile phone, desktop computer, laptop, handheld computer, or cloud server. The computer system may include, but is not limited to, a processor and memory. For example, the computer system may also include input / output devices, network access devices, and a bus.

[0053] The above are merely preferred embodiments of the present invention. It should be noted that the present invention is not limited to the above embodiments. For those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A metal mine scheduling decision-making method based on a large model, characterized in that, The method includes: S1. Receive natural language query commands input by mine staff; S2. Construct a pre-trained large language model to extract semantic features and classify intents of the natural language query instructions, thereby obtaining the static procedure query intent and dynamic process scheduling intent of the natural language query instructions. S3. Based on the first route identifier determined by the static procedure query intent, perform text matching and retrieval in the mine-specific knowledge base file, and combine the retrieved procedure clauses and knowledge fragments to reconstruct and generate a regular question-and-answer response in natural language format through the large language model; S4. Based on the second route identifier determined by the dynamic process scheduling intention, extract the corresponding real-time equipment operation data from the underlying database of the Internet of Things system, input the real-time equipment operation data into the pre-built operations research optimization model for solving, and transform the solved optimal process scheduling parameters into comprehensive auxiliary decision-making suggestions in natural language format through the large language model. The construction of a pre-trained large language model for semantic feature extraction and intent classification of the natural language query instructions includes: The natural language query command is subjected to text denoising processing and then concatenated with a preset few-sample scenario template in the mining field to generate an enhanced prompt sequence; The enhanced prompt sequence is input into the Transformer encoding layer of the large language model, and an inference intermediate state containing action entities and constraints is generated based on the autoregressive mechanism. Calculate the classification probability distribution of the inference intermediate state in the preset intent slot, and extract the target intent with the highest classification probability and higher than the confidence threshold; Based on the target intent, the natural language query instruction is determined to be either a static procedure query intent or a dynamic process scheduling intent. The objective function of the production scheduling process constraint model in the pre-constructed operations research optimization model is: ； Where Z represents the minimum total completion time or total operation cost of the current scheduling scheme; N represents the total number of loaders currently participating in the scheduling; and M represents the total number of mining points currently awaiting operation. , which is a 0-1 decision variable, represents the estimated time for the uth loader to complete the task at the vth mining point. Its constraints include the uniqueness constraint of equipment allocation and the safety ventilation constraint of the roadway. This represents the estimated time for the u-th loader to complete the task at the v-th mining point.

2. The metal mine scheduling decision-making method based on a large model according to claim 1, characterized in that, The first route identifier, determined based on the static procedure query intent, performs text matching and retrieval in the mine-specific knowledge base file. Combined with the retrieved procedure clauses and knowledge fragments, it is reconstructed using the large language model to generate a regular question-and-answer response in natural language format, including: An information extraction model is used to extract entities and relationships from unstructured mine regulations texts and assemble them into a knowledge graph. At the same time, a text embedding model is used to map the mine regulations texts into dense vectors to establish a vector index. Based on the first route identifier, a high-dimensional query vector is extracted, and a local semantic vector retrieval is performed to obtain a local semantic slice set; at the same time, based on the first route identifier, named entities are extracted and used as the starting point to perform a global topology graph retrieval to obtain a global knowledge subgraph set containing logical topology; The local semantic slice set and the global knowledge subgraph set are deduplicated and fused with lightweight features; Based on the local semantic slice set and the global knowledge subgraph set after deduplication and lightweight feature fusion, an enhanced context boundary with graph structure constraints is constructed. The enhanced context boundary with graph structure constraints is input into the large language model, and a regular question-and-answer response that follows mine safety regulations is generated through semantic reconstruction.

3. The metal mine scheduling decision-making method based on a large model according to claim 2, characterized in that, The local semantic vector retrieval yields a set of local semantic slices, including: Calculate the query vector q and the knowledge slice vector in the HNSW vector index. cosine similarity distance matrix It satisfies the following expression: ； Where n is the total number of dimensions of the high-dimensional vector; Let be the feature value of the i-th knowledge slice vector in the j-th dimension; To query the feature value of vector q in the j-th dimension; Recall the set of local semantic slices with the highest similarity scores.

4. The metal mine scheduling decision-making method based on a large model according to claim 2, characterized in that, The global topology graph retrieval yields a global knowledge subgraph set containing logical topology, including: Starting with named entities, a multi-hop topological traversal of a graph attention network is performed in the mine knowledge graph. The attention coefficients of the graph attention network are... Satisfy the following expression: ； in, The eigenvectors of node m; The feature vector representing node n; Let W represent the feature vector of node k, which is used as the iteration variable when traversing the set of neighboring nodes; W is the shared weight matrix of the linear transformation of node features. For the set of neighboring nodes; is the transpose of the weight vector of a single-layer feedforward neural network, used to map the concatenated high-dimensional features to an attention scalar; k is the set of neighbor nodes. Index in; Recall the global knowledge subgraph containing logical topological relationships.

5. The metal mine scheduling decision-making method based on a large model according to claim 1, characterized in that, S4 includes: Based on the second route identifier, the target device category is parsed to obtain a multi-dimensional operational status matrix captured in real time; The multidimensional operating state matrix is ​​used as a dynamic initial boundary condition input to the scheduling algorithm model; with the objective function of maximizing production scheduling efficiency or minimizing completion time, mixed integer linear programming combined with heuristic rules is used to iteratively solve the optimal process scheduling parameters in the pre-constructed operations research optimization model; The optimal process scheduling parameters are injected into the large language model as prior knowledge, and natural language generation technology is used to transform them into logically coherent comprehensive auxiliary decision-making suggestions and output them.

6. A metal mine scheduling decision-making system based on a large model, characterized in that, The system is used to execute the metal mine scheduling decision method based on a large model as described in any one of claims 1-5. The system is deployed on a mine private cloud server and includes mine terminal equipment, a multi-source heterogeneous data access and interaction module, a large model intent parsing and routing module, a knowledge base retrieval question and answer engine, an IoT scheduling auxiliary decision engine, a mine-specific knowledge base, and an IoT underlying database. The multi-source heterogeneous data access and interaction module is used to receive natural language query commands input by mine workers through the mine terminal equipment, and simultaneously capture static procedure data in the mine-specific knowledge base and real-time equipment operation data in the underlying IoT database based on the natural language query commands. The large model intent parsing and routing module has a built-in pre-trained large language model, which is used to extract semantic features and classify intents of the natural language query instructions. Based on the results of semantic feature extraction and intent classification of the natural language query instructions, it triggers the knowledge base retrieval question answering engine or the IoT scheduling auxiliary decision engine. The knowledge base retrieval question-answering engine is used to perform text matching and retrieval in the mine-specific knowledge base file based on the received natural language query command, and combine the retrieved procedural clauses and knowledge fragments to reconstruct and generate a regular question-answering response in natural language format through the large language model, and output it to the terminal device. The IoT scheduling auxiliary decision engine is used to extract corresponding real-time device operation data from the underlying database of the IoT system according to the received natural language query command, input the real-time device operation data into a pre-built operations research optimization model for solving, and convert the solved optimal process scheduling parameters into comprehensive auxiliary decision suggestions in natural language format through the large language model, and output them to the terminal device.

7. The metal mine scheduling decision system based on a large model according to claim 6, characterized in that, The static regulatory data includes non-mining safety regulations, blasting operation standards, and mining equipment maintenance manuals; the real-time equipment operation data includes underground ventilation wind speed, real-time load of the hoist, and real-time position of the underground loader.

8. A computer system, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the metal mine scheduling decision method based on a large model as described in any one of claims 1-5.

Images