Underground mine multi-equipment intelligent scheduling system and method based on knowledge graph

Abstract

The invention discloses an underground mine multi-equipment intelligent scheduling system and method based on a knowledge graph. The system comprises a data acquisition module used for acquiring structured data, semi-structured data and non-structured data of each production system of an underground mine; the knowledge graph construction module is used for constructing a knowledge graph database of each system; the dynamic knowledge graph construction module is used for acquiring and fusing a knowledge graph database and a real-time database of each production system, and obtaining a dynamic knowledge graph and a rule base in the underground mine dispatching field; and the multi-agent collaborative construction module is used for taking each production system as an agent to carry out joint training so as to obtain a multi-agent collaborative optimization model. According to the method, the problem of incomplete and inaccurate information during production scheduling decision making is solved by effectively extracting and integrating data, and in addition, by establishing a multi-intelligent collaborative optimization model, the constraint of dependence on artificial experience is avoided, and the scheduling efficiency and the overall productivity of a mine in a complex scene are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of underground mine production scheduling, specifically to a knowledge graph-based intelligent scheduling system and method for multiple pieces of equipment in underground mines. Background Technology

[0002] Underground mine production scheduling is a complex systems engineering project involving multiple major production systems such as mining, transportation, hoisting, ventilation, drainage, power supply, crushing, and air supply. Traditional underground mine production scheduling methods have the following problems: Data integration challenges: The diverse data types across production systems, including structured, semi-structured, and unstructured data, make effective integration and utilization difficult, resulting in incomplete and inaccurate information during production scheduling decisions.

[0003] Lack of coordination mechanism: There is a lack of effective coordination mechanism between various production systems. When abnormal situations or task changes occur, it is difficult to quickly and accurately adjust the scheduling, thereby affecting production efficiency and production safety.

[0004] Reliance on human experience: The scheduling decision-making process relies on the experience and knowledge of schedulers, lacks scientific decision-making basis and intelligent decision support, and is difficult to cope with complex production scheduling scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide a knowledge graph-based intelligent scheduling system and method for multiple pieces of equipment in underground mines. This invention addresses the problems of difficulty in integrating and utilizing multi-source heterogeneous data in underground mines, the lack of coordination among production systems leading to delayed scheduling responses, and the inability to achieve scientific and intelligent decision-making due to over-reliance on human experience. By constructing a knowledge graph, production data is integrated and knowledge is mined; multi-agent reinforcement learning is used to achieve collaborative scheduling decisions among production systems, thereby improving the intelligence level and production efficiency of underground mine production scheduling.

[0006] To achieve this objective, the present invention provides an intelligent scheduling system for multiple pieces of equipment in underground mines based on knowledge graphs, comprising: The data acquisition module is used to acquire structured, semi-structured, and unstructured data from various production systems in the underground mine. The knowledge graph construction module is used to extract entities from the structured data, the semi-structured data, and the unstructured data, respectively, and to analyze the relationships between entities. It also uses the entities and the relationships between entities to construct a knowledge graph database for each production system. The dynamic knowledge graph construction module is used to perform multi-source fusion between the knowledge graph database and the real-time database of the scheduling system to obtain a dynamic knowledge graph in the field of underground mine scheduling. Then, the relationship patterns between entities are extracted from the dynamic knowledge graph, and the relationship patterns are used to form a rule base. The multi-agent collaborative construction module is used to initialize each of the production systems as an agent, obtain the composite state space of each agent using a dynamic knowledge graph, and obtain the action space of each agent using a rule base. With the goal of maximizing the completion of the production plan, an overall reward function is constructed. A multi-agent collaborative model is constructed using the composite state space, action space, and overall reward function of each agent. A multi-agent collaborative optimization model is obtained by training the multi-agent collaborative model.

[0007] The beneficial effects of this invention are as follows: 1. This invention provides a knowledge graph-based intelligent scheduling system for multiple pieces of equipment in underground mines, solving the problems of difficulty in integrating multi-source heterogeneous data in mines, poor coordination among various production systems leading to slow scheduling response, and reliance on human experience making intelligent decision-making difficult. By integrating various types of data and constructing a knowledge graph, combined with a multi-agent collaborative scheduling model, more efficient and intelligent scheduling management is achieved.

[0008] 2. This invention can extract entities and relationships from structured, semi-structured and unstructured data, and uniformly construct them into a knowledge graph, which improves the integration efficiency and accuracy of mine data and provides a high-quality knowledge foundation for subsequent intelligent scheduling.

[0009] 3. This invention uses the Transformer model combined with sequence labeling and decoding algorithms to improve the accuracy of entity recognition in text, enhance the system's ability to understand unstructured information, and help build a more complete knowledge graph.

[0010] 4. This invention utilizes the LSTM model to extract the relationship features between entities, thereby improving the ability to identify complex semantic relationships and making the relationship representation in the knowledge graph richer and more accurate.

[0011] 5. This invention integrates a knowledge graph database constructed from information extracted from structured, semi-structured, and unstructured data with real-time data to build a dynamic knowledge graph, and extracts scheduling rules from it, thereby improving the system's real-time perception of mine operation status and the executability of scheduling strategies.

[0012] 6. This invention models each production system as an intelligent agent and designs the state space and action space by combining dynamic knowledge graphs and rule bases, thereby improving the collaborative efficiency and scheduling flexibility among multiple intelligent agents.

[0013] 7. This invention designs an overall reward function that comprehensively considers production progress, system coordination, and anomaly penalties, guiding multi-agent models to make more scientific and reasonable scheduling decisions.

[0014] 8. This invention uses the MAPPO algorithm to train a multi-agent model, enabling it to autonomously learn the optimal scheduling strategy and improve the system's intelligent scheduling capability and adaptability in complex mining environments. Attached Figure Description

[0015] Figure 1 This is a flowchart of a knowledge graph-based intelligent scheduling system for multiple pieces of equipment in underground mines, as described in an embodiment of the present invention. Figure 2 This is a general framework diagram for constructing a dynamic knowledge graph in the field of underground mine scheduling in an embodiment of the present invention; Figure 3 This is a multi-agent design framework diagram in an embodiment of the present invention; Figure 4 This is a schematic diagram of the overall deployment scheme and decision-making process in an embodiment of the present invention; Figure 5 This is a flowchart of a knowledge graph-based intelligent scheduling method for multiple pieces of equipment in underground mines, as described in an embodiment of the present invention. Detailed Implementation

[0016] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: Example 1: like Figures 1-5 The above describes a knowledge graph-based intelligent scheduling system for multiple pieces of equipment in underground mines, comprising: The data acquisition module is used to acquire structured, semi-structured, and unstructured data from various production systems in the underground mine. The knowledge graph construction module is used to extract entities from the structured data, the semi-structured data, and the unstructured data, respectively, and to analyze the relationships between entities. It also uses the entities and the relationships between entities to construct a knowledge graph database for each production system. The dynamic knowledge graph construction module is used to perform multi-source fusion between the knowledge graph database and the real-time database of the scheduling system to obtain a dynamic knowledge graph in the field of underground mine scheduling. Then, the relationship patterns between entities are extracted from the dynamic knowledge graph, and the relationship patterns are used to form a rule base. The multi-agent collaborative construction module is used to initialize each of the production systems as an agent, obtain the composite state space of each agent using a dynamic knowledge graph, and obtain the action space of each agent using a rule base. With the goal of maximizing the completion of the production plan, an overall reward function is constructed. A multi-agent collaborative model is constructed using the composite state space, action space, and overall reward function of each agent. A multi-agent collaborative optimization model is obtained by training the multi-agent collaborative model.

[0017] This invention proposes a knowledge graph-based intelligent scheduling system for multiple pieces of equipment in underground mines. The system acquires multi-source heterogeneous data (structured, semi-structured, and unstructured data) from various production systems within the mine, providing data for knowledge graph construction. The knowledge graph construction module extracts entities and entity relationships from various data types, building knowledge graph databases for each production system and effectively integrating multi-source heterogeneous data. The dynamic knowledge graph construction module further integrates the real-time database of the scheduling system, generating a dynamic knowledge graph and building a rule base, enabling the system to update in real time and adapt to changing scheduling scenarios. The multi-agent collaborative construction module treats each production system as an agent, designing its state space, action space, and reward function to build a multi-agent collaborative model. Through training, an optimized model is obtained, achieving collaborative optimization scheduling among the production systems. This system enables intelligent scheduling management of complex mine production systems, improving scheduling efficiency and decision-making accuracy. By integrating multi-source heterogeneous data and building a unified knowledge system, the system possesses stronger environmental awareness and adaptability, thereby improving the safety and economic benefits of mine operations.

[0018] Structured data refers to data with strict, fixed formats and well-defined patterns.

[0019] Semi-structured data refers to data that has a certain organizational form or label, but does not follow a strict, fixed relational database table structure pattern.

[0020] Unstructured data is data that does not have a predefined data model or organizational form.

[0021] The structured, semi-structured, and unstructured data that needs to be collected from various production systems mainly include: (1) Structured data: mainly equipment efficiency of each production system, plan completion rate, real-time location and coordinates of equipment, and equipment maintenance and repair records; (2) Semi-structured data: Time series data of sensors in each production system in JSON format (sampling frequency ≥ 1Hz), and equipment configuration parameters in XML format; (3) Unstructured data: production plans, geological exploration reports, production technology plans and specific design plans of each production system in text format, and three-dimensional laser point clouds of underground mining areas, roadways, chambers and shafts (point density > 100 points / ㎡).

[0022] The various production systems include: mining system, transportation system, hoisting system, ventilation system, drainage system, power supply system, crushing system, and air supply (compressed air) system.

[0023] Once the knowledge graph is completed, the entities and the relationships between them are stored in the Neoj database to obtain the knowledge graph database.

[0024] In the above technical solution, the method by which the knowledge graph construction module obtains the knowledge graph databases of each production system is as follows: Extract entities from the structured data, semi-structured data, and unstructured data respectively, and analyze the relationships between the entities; Specifically, a pre-trained transformer model is used to perform entity recognition on the text data in the unstructured data of each production system to obtain the entities in the text data of each production system; a pre-trained YOLO model is used to extract the spatial features of the point cloud and image data in the unstructured data of each production system to obtain the entities in the point cloud and image data of each production system; the entities in the text data and the entities in the point cloud and image data are aligned; and all aligned entities in the unstructured data are obtained. A pre-trained LSTM model is used to extract entity relationships from unstructured text and sequence data, thereby obtaining the relationships between entities in unstructured data. Data annotation is performed on structural data and semi-structured data to obtain entities in structural data and semi-structured data, and the relationships between entities are analyzed. The entities and relationships between entities in the structured data, semi-structured data and unstructured data are obtained and stored in the knowledge graph database in the form of triples, thus obtaining the knowledge graph database of each production system.

[0025] The triplet is in the form of (entity1, relation, entity2).

[0026] Compared to unstructured data, there are already mature processing methods for structured or semi-structured data. Structured data can be directly converted into entities and relationships between entities through schema mapping, while semi-structured data requires manual parsing and data labeling before being converted into entities and relationships between entities.

[0027] The text-formatted production plans, geological exploration reports, production technology plans, and specific design plans for each production system are text data in unstructured data, while the three-dimensional laser point clouds of underground mining areas, roadways, chambers, and shafts are point cloud and image data in unstructured data.

[0028] By using a pre-trained LSTM model, entity relationships are extracted from unstructured data, including text data (text-formatted production plans, geological exploration reports, production technology schemes, and specific design schemes for each production system) and sequence data (unstructured records generated by underground mining equipment during the production process, arranged according to time or operation sequence), to obtain the relationships between entities in unstructured data.

[0029] The method for obtaining the knowledge graph database in the knowledge graph construction module extracts entities from unstructured text data and point cloud image data using a pre-trained Transformer model and YOLO, respectively, and aligns them. Simultaneously, an LSTM model is used to extract entity relationships from the unstructured data. Structured and semi-structured data are then labeled to obtain entities and relationships from both structured and semi-structured data. Finally, all entities and relationships extracted from the structured, unstructured, and semi-structured data are stored in triplet form. This method effectively integrates multiple data types, improving the completeness and accuracy of the knowledge graph and providing high-quality knowledge support for subsequent scheduling decisions. Utilizing Transformer and YOLO for entity recognition enhances the system's ability to process unstructured information and improves the overall intelligence level.

[0030] The transformer model is used for entity recognition in unstructured textual data. Its bidirectional context encoding capability can deeply understand the semantics of technical terms, equipment names, and status descriptions in mining texts, reducing ambiguity. Combined with BIOES annotation and Viterbi decoding, it can maintain high accuracy in high-noise and long text scenarios, providing reliable nodes for knowledge graphs.

[0031] YOLO transforms object detection into a single regression problem, achieving real-time detection through grid prediction, anchor box matching, and multi-scale feature fusion. YOLO is used to extract spatial features from point clouds and images in unstructured data, quickly locating targets such as underground equipment, tunnel cross-sections, and material piles. YOLO's real-time detection capability meets the low-latency requirements of mine video / LiDAR streams, transforming visual information into computable entities, achieving text-visual cross-modal alignment, and completing the spatial dimensions of knowledge graphs.

[0032] The LSTM model is used to extract entity relationships from unstructured text and sequence data. The long-range memory mechanism of LSTM can capture the temporal dependencies in device operation logs and alarm sequences, generating feature vectors with temporal semantics. Combined with path information vectors, it can accurately depict dynamic causality such as "device A failure causes device B to stop", laying a high-quality relational data foundation for subsequent rule mining.

[0033] In the above technical solution, the process of performing entity recognition on text data in the unstructured data of various production systems using a pre-trained transformer model is as follows: The text data is manually segmented into words to form labeled text sequence data; The labeled text sequence data is input into a pre-trained transformer model for encoding to obtain the encoded semantic representation vector of the following text. The encoded context semantic representation vector is sequence labeled using the BIOES format to obtain the label sequence corresponding to each character or word.

[0034] The Viterbi algorithm is used to decode and merge entities based on the labeled sequence to obtain entities after text data recognition in unstructured data.

[0035] By employing the Transformer model combined with BIOES annotation and Viterbi decoding algorithms, the accuracy and robustness of entity recognition are improved. Its beneficial effects lie in enhancing the system's ability to understand unstructured text, ensuring the accurate extraction of entity information from the knowledge graph, and laying a solid foundation for subsequent knowledge graph construction and applications.

[0036] In the above technical solution, the knowledge graph construction module uses a pre-trained LSTM model to extract entity relationships from text data and sequence data. The expression is as follows: ; in, Represented as the first i The entity and the first j The relationship feature vector between entities Indicates the first The contextual semantic vector of an entity, Indicates the first The contextual semantic vector of an entity, This represents a path information vector between two entities. Represents the relational projection weight matrix. It is a relational representation of the bias vector.

[0037] This paper presents an expression for extracting entity relationships from textual and sequence data using a pre-trained LSTM model. This expression is used to extract relational features between entities from textual and sequence data. It can effectively improve the system's ability to understand complex semantic relationships and contribute to building richer knowledge graph structures. By introducing relation projection matrices and context vectors, the model's ability to capture long-distance dependencies is enhanced, further improving the expressive power and inference performance of the knowledge graph.

[0038] In the above technical solution, the dynamic knowledge graph construction module includes the process of constructing a dynamic knowledge graph in the field of underground mine scheduling and constructing a rule base: The process of constructing the dynamic knowledge graph in the field of underground mine scheduling is as follows: It receives input data from the knowledge graph database of various production systems and the real-time database of the scheduling system.

[0039] The input data is fed into a pre-trained graph convolutional network, which uses its graph structure feature extraction capability to perform multi-source data fusion and outputs a set of triple relations containing enhanced semantics to generate a dynamic knowledge graph in the field of underground mine scheduling.

[0040] The knowledge graph database contains entity, relationship, and attribute information in the underground mining production system.

[0041] The real-time database contains dynamic scheduling data on equipment operating status, personnel location, and ore transportation status; the real-time database of the scheduling system is obtained through each mine's own scheduling system.

[0042] The rule base construction process is as follows: Type identification and semantic classification are performed on entity relationships in dynamic knowledge graphs to obtain a set of classified relationships; Based on the association rule growth algorithm with time window, this algorithm extracts relation patterns under different production scheduling scenarios from the classification relation set. It is an extension of the classic FP-Growth algorithm. It first cuts the scheduling log into sliding window transactions according to the timestamp, then runs TW-Growth to mine frequent event-scenario pairs, calculates the confidence on the frequent patterns, and outputs the X→Y rules that meet the threshold to complete the rule mining.

[0043] The mined relational patterns are transformed into condition-outcome rules, with the rule structure represented as: if the triggering condition is met, then execute the action; The triggering conditions include combinations of entity states, the execution actions include scheduling decision instructions, and are associated with priorities and applicable constraints; The generated rules are stored in the rule base in an indexed structure.

[0044] The scheduling decision instruction is a set of actions that meet the feasibility conditions.

[0045] The dynamic knowledge graph construction module involves building a dynamic knowledge graph and rule base. This is achieved by fusing data from various production systems using a pre-trained graph convolutional network to generate a set of triple relations with enhanced semantics, thus constructing the dynamic knowledge graph. Furthermore, a rule mining algorithm extracts relation patterns from the classification relation set, transforming them into conditional outcome rules to form the rule base. This approach enables the system to update the knowledge graph in real time, adapting to changing scheduling scenarios. The rule base provides decision support for agents, improving system flexibility and response speed, and enhancing the intelligence and automation of underground mine scheduling.

[0046] The specific process of constructing a dynamic knowledge graph in the field of underground mine scheduling: Data from the knowledge graph databases of various production systems and the real-time database of the scheduling system are integrated and used as input to a pre-trained graph convolutional network. The knowledge graph database encompasses entities, relationships, and related attribute information from various production systems in the underground mine, while the real-time database contains various dynamic data during real-time production scheduling, such as the real-time operating status of equipment, the real-time location of personnel, and the real-time transportation status of ore. Leveraging the powerful graph structure data processing and feature extraction capabilities of the pre-trained graph convolutional network, the input data is classified and processed to output more specific triplet relationships. Through its network structure and parameters, the graph convolutional network can learn the representations of different entities and their relationships in the graph structure, thereby fusing knowledge from various production systems with real-time scheduling data.

[0047] The specific process of building the rule base: (1) Relationship Identification and Classification. First, the relationships between entities in the dynamic knowledge graph of underground mine scheduling are identified and classified in detail. By analyzing the type, direction, and semantic information of the relationships, various relationships are divided into different categories. The relationships between equipment can be divided into connection relationships, dependency relationships, collaborative operation relationships, etc.; the relationships between personnel and equipment can be divided into operation relationships, maintenance relationships, etc.; and the relationships between ore and equipment, operation areas, etc. can be divided into transportation relationships, storage relationships, etc.

[0048] (2) Relationship Pattern Mining. Association rule mining algorithms are used to mine patterns of entity relationships in the dynamic knowledge graph. This uncovers frequently occurring, highly correlated, and regular relationship combinations and their changing patterns between entities under different production scheduling scenarios. For example, a relationship pattern might be discovered such as, "When equipment A malfunctions, the operating status of equipment B, which works in conjunction with it, is often affected, and related ore transportation tasks are delayed, requiring the activation of backup equipment C and D to reallocate tasks."

[0049] (3) Rule Generation and Representation. Based on the mined relational patterns, corresponding rules are generated. These rules are represented in an easy-to-understand and processable form, typically using a "if...then..." condition-result structure. For example, for the above-mined relational patterns, the generated rule can be represented as: "If equipment A fails and equipment B is working in conjunction with it, then start backup equipment C and D to redistribute the ore transportation task to ensure the normal operation of production scheduling." During the rule representation process, factors such as rule priority, scope of application, and related constraints are also considered to improve the accuracy and practicality of the rules.

[0050] (4) The generated rules are stored in the rule base according to a certain organizational structure and indexing method, and encoded using PYTHON so that these rules can be retrieved and applied quickly and efficiently during the intelligent scheduling process.

[0051] In the above technical solution, the method by which the multi-agent collaborative construction module obtains the composite state space, action space, and overall reward function of each agent and constructs the multi-agent collaborative model is as follows: Initialize the state space of each agent, obtain the initial state space of each agent, extract the state domain subgraph of the corresponding agent from the dynamic knowledge graph of the underground mine scheduling domain, and obtain the composite state space of each agent based on the initial state space of each agent and the corresponding state domain subgraph. The actions faced by each agent are filtered using a rule base, and actions that conform to the rule base are recommended to the action space to obtain the action space of the agent. With the goal of maximizing the completion of the production plan, an overall reward function is constructed, a multi-agent collaborative model is built, and the strategies of each agent are collaboratively optimized. The state domain subgraph includes the relationships between entities in the production system and the state information of the entities.

[0052] The data contained in each production system in the system state space: The state space of any production system should include equipment or device location information, work requirement information, current work status information, equipment attributes, work scene space attribute data, and production plan completion status.

[0053] The multi-agent collaborative construction module acquires the composite state space, action space, and overall reward function of the agents. Through steps such as initializing the state space, extracting subgraphs of the state domain, filtering the action space, and constructing a multi-agent collaborative model, it achieves collaborative optimization among various production systems. This method abstracts each production system as an agent, enabling distributed collaborative scheduling and improving the system's ability to handle complex tasks. Utilizing subgraphs and rule bases in a dynamic knowledge graph to guide agent behavior enhances the accuracy and flexibility of scheduling decisions, contributing to efficient and safe mine production management.

[0054] Taking a transportation system as an example, the system state space of a transportation system includes: the location of each piece of transportation equipment, the current speed of each piece of transportation equipment, the current job number of each piece of transportation equipment, the estimated job completion time of each piece of transportation equipment, the load weight of each piece of transportation equipment, and the operating status of each piece of transportation equipment.

[0055] In the above technical solution, the overall reward function is expressed as: Where i′, j′, and k represent the production system, and i′, j′, and k = 1, 2, 3, ..., 8. Indicates the main schedule item, A i′ P represents the actual amount completed by production system i′. i′ w represents the planned completion amount of production system i′. i′ The weights of each production system i′ are represented by α, and the schedule coefficient is represented by α. Denotes the cooperative term, ΔT j′k η represents the job delay from system j′ to k. j′k β represents the collaboration weight, and β represents the collaboration coefficient. F represents the penalty term. i′ Indicates whether the system is faulty, D m μ represents the degree of safety violation of equipment m. i′ With κ m This represents the penalty coefficient.

[0056] The above is the expression for the overall reward function, which comprehensively considers factors such as production plan completion, operational delays, system failures, and safety violations. By setting the main schedule term, coordination term, and penalty term, the performance of the agent can be comprehensively evaluated. This process improves the overall performance of the scheduling strategy by guiding the multi-agent collaborative optimization model towards the global optimum. By quantifying different influencing factors, the system can balance efficiency with safety and stability, ensuring the continuity and reliability of mine production.

[0057] In the above technical solution, the method for obtaining the multi-agent cooperative optimization model is as follows: the MAPPO algorithm is used to train randomly generated scheduling instance data to obtain the multi-agent cooperative optimization model.

[0058] A multi-agent cooperative optimization model is obtained by training randomly generated scheduling instance data using the MAPPO algorithm. The MAPPO algorithm is a multi-agent reinforcement learning algorithm that effectively handles the cooperative and competitive relationships among multiple agents. By training and optimizing the agents' policies, it improves the overall cooperative performance of the system. This method enables agents to autonomously learn optimal strategies in complex underground mine scheduling environments, adapting to different scheduling scenarios and achieving efficient operation and continuous optimization of the intelligent scheduling system for multiple pieces of equipment in underground mines.

[0059] The randomly generated scheduling instance data is generated using a Gaussian distribution random generation method based on the underground mine scheduling scenario. The main information includes the equipment attribute information of each system and the layout of the underground mine mining area and roadways.

[0060] The MAPPO algorithm is a method for training multi-agent collaborative models. Its beneficial effects lie in leveraging advanced reinforcement learning algorithms to enhance the model's learning ability and generalization performance, enabling it to adapt to complex mine scheduling scenarios. Through training on a large amount of scheduling instance data, the model can autonomously learn the optimal scheduling strategy, significantly improving the system's intelligence and automation level, reducing manual intervention, and increasing mine operational efficiency.

[0061] Figure 2 This is represented as the overall framework diagram for constructing a knowledge graph in the field of underground mine scheduling. Here, GCN is short for pre-trained graph convolutional network. The scheduling system obtains the current time scheduling information by parsing the operation records and stores it in the scheduling system database as input to GCN.

[0062] Figure 3 The diagram illustrates a multi-agent design framework, with the specific process described in the figure: By designing a composite state space, the state neighborhood subgraphs of the agents corresponding to each production system are obtained using the initial state spaces of each production system. Then, the composite state space of each agent is obtained using the state neighborhood subgraphs of each agent.

[0063] By designing the action space and reward function, the action space of each agent is obtained by utilizing the set of actions that each agent can take in the rule base. By designing a reward function, the target composition of each agent is obtained. Then, by using a multi-agent collaborative optimization model, the composite actions of each agent are obtained. After the composite actions are mapped into comprehensive scheduling instructions, they are sent to each production system for execution through the scheduling system.

[0064] Figure 4 The diagram above illustrates the overall deployment plan and decision-making process. Real-time scheduling data (i.e., current-time scheduling information) from the scheduling system database of the scheduling system is input into the multi-agent collaborative optimization model to obtain a dynamic knowledge graph in the underground mine scheduling domain. Through this dynamic knowledge graph, the composite state space and rule base of each agent (production system) under its current state are obtained. The rule base is then used to obtain the action space of each agent (production system) under this composite state space. The overall reward function is used to jointly optimize multiple agents to obtain the execution actions of each agent. The execution actions of all agents form a composite action, which is then mapped to obtain a comprehensive scheduling instruction and fed back to the scheduling system.

[0065] Example 2: A knowledge graph-based intelligent scheduling method for multiple pieces of equipment in underground mines includes the following steps: Acquire structured, semi-structured, and unstructured data from various production systems in underground mines; Entities for each data item are extracted from the structured data, the semi-structured data, and the unstructured data, respectively, and the relationships between the entities are analyzed. Knowledge graph databases for each production system are constructed using the entities and the relationships between them. The knowledge graph database is fused with the real-time database of the scheduling system to obtain a dynamic knowledge graph in the field of underground mine scheduling. Then, the relationship patterns between entities are extracted from the dynamic knowledge graph to form a rule base. Each production system is initialized as an agent. A composite state space for each agent is obtained using a dynamic knowledge graph. A rule base is then used to obtain the action space for each agent. An overall reward function is designed with the goal of maximizing the completion of the production plan. A multi-agent collaborative model is constructed using the composite state space, action space, and overall reward function of each agent. A multi-agent collaborative optimization model is obtained by training the multi-agent collaborative model.

[0066] Example 3: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method.

[0067] Example 4: A computer program product includes a computer program that, when executed by a processor, implements the steps of the above-described method.

[0068] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

Claims

1. A knowledge graph-based intelligent scheduling system for multiple pieces of equipment in underground mines, characterized in that, include: The data acquisition module is used to acquire structured, semi-structured, and unstructured data from various production systems in the underground mine. The knowledge graph construction module is used to extract entities from the structured data, the semi-structured data, and the unstructured data, respectively, and to analyze the relationships between entities. It also uses the entities and the relationships between entities to construct a knowledge graph database for each production system. The dynamic knowledge graph construction module is used to perform multi-source fusion of the knowledge graph database and the real-time database of the scheduling system to obtain a dynamic knowledge graph in the field of underground mine scheduling. Then, the relationship patterns between entities are extracted from the dynamic knowledge graph to form a rule base. The multi-agent collaborative construction module is used to initialize each of the production systems as an agent, obtain the composite state space of each agent using a dynamic knowledge graph, and obtain the action space of each agent using a rule base. With the goal of maximizing the completion of the production plan, an overall reward function is constructed. A multi-agent collaborative model is constructed using the composite state space, action space, and overall reward function of each agent. A multi-agent collaborative optimization model is obtained by training the multi-agent collaborative model.

2. The knowledge graph-based intelligent scheduling system for multiple pieces of equipment in underground mines according to claim 1, characterized in that, The method for obtaining the knowledge graph databases of each production system in the knowledge graph construction module is as follows: Extract entities from the structured data, semi-structured data, and unstructured data respectively, and analyze the relationships between the entities; Specifically, a pre-trained transformer model is used to perform entity recognition on the text data in the unstructured data of each production system to obtain the entities in the text data of each production system; a pre-trained YOLO model is used to extract the spatial features of the point cloud and image data in the unstructured data of each production system to obtain the entities in the point cloud and image data of each production system; the entities in the text data and the entities in the point cloud and image data are aligned; and all aligned entities in the unstructured data are obtained. A pre-trained LSTM model is used to extract entity relationships from unstructured text and sequence data, thereby obtaining the relationships between entities in unstructured data. Data annotation is performed on structural data and semi-structured data to obtain entities in structural data and semi-structured data, and the relationships between entities are analyzed. The entities and relationships between entities in the structured data, semi-structured data and unstructured data are obtained and stored in the knowledge graph database in the form of triples, thus obtaining the knowledge graph database of each production system.

3. The knowledge graph-based intelligent scheduling system for multiple pieces of equipment in underground mines according to claim 2, characterized in that, The process of using a pre-trained transformer model to perform entity recognition on text data in unstructured data from various production systems is as follows: The unstructured text data is manually segmented into words to form labeled text sequence data; The labeled text sequence data is input into a pre-trained transformer model for encoding to obtain the encoded semantic representation vector of the following text. The encoded context semantic representation vector is sequence labeled using the BIOES format to obtain the label sequence corresponding to each character or word. The Viterbi algorithm is used to decode and merge entities based on the labeled sequence to obtain entities after text data recognition in unstructured data.

4. The knowledge graph-based intelligent scheduling system for multiple pieces of equipment in underground mines according to claim 2, characterized in that, In the knowledge graph construction module, the expression for extracting entity relationships from text data and sequence data using a pre-trained LSTM model is as follows: v ij =tanh(W r ·[h i ;h j ;p ij ]+b r ); Among them, v ij Let h be the feature vector representing the relationship between the i-th entity and the j-th entity. i h represents the context semantic vector of the i-th entity. j p represents the context semantic vector of the j-th entity. ij W represents the path information vector between two entities. r Let b represent the relational projection weight matrix. r It is a relational representation of the bias vector.

5. The knowledge graph-based intelligent scheduling system for multiple pieces of equipment in underground mines according to claim 1, characterized in that, The dynamic knowledge graph construction module includes the process of constructing a dynamic knowledge graph for the underground mine scheduling domain and forming a rule base: The process of constructing the dynamic knowledge graph in the field of underground mine scheduling is as follows: The knowledge graph databases of various production systems and the real-time database of the scheduling system are used as input data and fed into a pre-trained graph convolutional network. The graph structure features in the pre-trained graph convolutional network are used for feature extraction, and the obtained features are fused to obtain a set of triple relations containing enhanced semantics. Using the set of triple relations containing enhanced semantics, a dynamic knowledge graph in the field of underground mine scheduling is generated. The rule base construction process is as follows: Type identification and semantic classification of entity relationships in the dynamic knowledge graph of underground mine scheduling are performed to obtain a set of classification relationships; Based on rule mining algorithms, relationship patterns under different production scheduling scenarios are extracted from a set of classification relationships. The mined relational patterns are transformed into conditional result rules, with the rule structure represented as: if the triggering condition is met, then execute the action; The triggering conditions include combinations of entity states, the execution actions include scheduling decision instructions, and are associated with priorities and applicable constraints; The generated rules are stored in the rule base in an indexed structure.

6. The knowledge graph-based intelligent scheduling system for multiple pieces of equipment in underground mines according to claim 1, characterized in that, In the multi-agent cooperative construction module, the method for obtaining the composite state space, action space, and overall reward function of each agent and constructing the multi-agent cooperative model is as follows: Initialize the state space of each agent, obtain the initial state space of each agent, extract the state domain subgraph of the corresponding agent from the dynamic knowledge graph of the underground mine scheduling domain, and obtain the composite state space of each agent based on the initial state space of each agent and the corresponding state domain subgraph. The actions faced by each agent are filtered using a rule base, and actions that conform to the rule base are recommended to the action space to obtain the action space of the agent. With the goal of maximizing the completion rate of production plans across all production systems, an overall reward function is constructed, a multi-agent collaborative model is built, and the strategies of each agent are collaboratively optimized. The state domain subgraph includes the relationships between entities in the production system and the state information of the entities.

7. The knowledge graph-based intelligent scheduling system for multiple pieces of equipment in underground mines according to claim 6, characterized in that, The overall reward function is expressed as follows: Where i′, j′, and k represent the production system, and i′, j′, and k = 1, 2, 3, ..., 8. Indicates the main schedule item, A i′ P represents the actual amount completed by production system i′. i′ w represents the planned completion amount of production system i′. i′ The weights of each production system i′ are represented by α, and the schedule coefficient is represented by α. Denotes the cooperative term, ΔT j′k η represents the job delay from system j′ to k. j′k β represents the collaboration weight, and β represents the collaboration coefficient. F represents the penalty term. i′ Indicates whether the system is faulty, D m μ represents the degree of safety violation of equipment m. i′ With κ m This represents the penalty coefficient.

8. The knowledge graph-based intelligent scheduling system for multiple pieces of equipment in underground mines according to claim 7, characterized in that, The method for obtaining the multi-agent cooperative optimization model is as follows: the MAPPO algorithm is used to train the model on randomly generated scheduling instance data.

9. A knowledge graph-based intelligent scheduling method for multiple pieces of equipment in underground mines, characterized in that, Includes the following steps: Acquire structured, semi-structured, and unstructured data from various production systems in underground mines; Entities for each data item are extracted from the structured data, the semi-structured data, and the unstructured data, respectively, and the relationships between the entities are analyzed. Knowledge graph databases for each production system are constructed using the entities and the relationships between them. The knowledge graph database is fused with the real-time database of the scheduling system to obtain a dynamic knowledge graph in the field of underground mine scheduling. Then, the relationship patterns between entities are extracted from the dynamic knowledge graph to form a rule base. Each production system is initialized as an agent. A composite state space for each agent is obtained using a dynamic knowledge graph. A rule base is then used to obtain the action space for each agent. An overall reward function is constructed with the goal of maximizing the completion of the production plan. A multi-agent collaborative model is constructed using the composite state space, action space, and overall reward function of each agent. A multi-agent collaborative optimization model is obtained by training the multi-agent collaborative model.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method of claim 9.

Images