Optimal path generation method and system for data center operation and maintenance robot, terminal and medium

Abstract

The invention relates to the field of robot navigation, and particularly provides a data center operation and maintenance robot optimal path generation method and system, a terminal and a medium, and the method comprises the steps: constructing a multi-level topological graph of a data center, and fusing the azimuth information of equipment, the historical inspection record of the equipment and the route preference information preset by a user in the node attribute of each level; receiving an inspection task instruction, and folding the topological graph through the large language model to obtain a simplified sub-graph; determining an initial range containing a candidate node set through a large language model; determining a target node in an initial range through the graph neural network model and outputting coordinates of the target node; and generating an inspection action sequence according to the target node through the large language model, and generating and executing an optimal path to the target node by using the reinforcement learning model. According to the invention, the accuracy and reliability of generating the optimal path are improved.

Description

Technical Field

[0001] This application relates to the field of robot navigation, specifically to a method, system, terminal, and medium for generating optimal paths for data center operation and maintenance robots. Background Technology

[0002] Operation and maintenance robots have become key equipment for realizing automated operation and maintenance of data centers. The robot's environmental perception and task execution can be achieved through scene graph-based navigation technology. Scene graph-based navigation technology constructs the environment as a multi-dimensional relationship graph composed of nodes and edges, enabling the robot to better understand the scene structure and navigate.

[0003] Traditional navigation methods based on hierarchical scene maps, while organizing scene data to some extent through a "building-floor-room-object" hierarchy, fail to deeply integrate the unique attribute information of the operation and maintenance scenario, such as equipment historical fault records and user path preferences. This results in insufficient semantic expression and support for operation and maintenance tasks. Furthermore, related solutions only use large language models for coarse room-level localization, relying on traditional algorithms for subsequent path planning. This lack of unified consideration and closed-loop optimization of key factors such as equipment fault frequency and path preferences leads to poor accuracy and reliability of the generated optimal paths. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a method, system, terminal, and medium for generating optimal paths for data center operation and maintenance robots, thereby improving the accuracy and reliability of generating optimal paths.

[0005] In a first aspect, the technical solution of the present invention provides a method for generating the optimal path for a data center operation and maintenance robot, comprising the following steps: Construct a multi-level topology map of the data center. The topology map node hierarchy from top to bottom is building, floor, computer room, area and equipment. The node attributes of each level integrate the location information of the equipment, the historical inspection records of the equipment and the user's preset route preference information. Upon receiving the inspection task instruction, the topology graph is folded using a large language model to obtain a simplified subgraph. The simplified subgraph retains the hierarchical nodes related to the target node and represents the folded node data in the form of node attributes. Using a large language model, semantic parsing is performed on the simplified subgraph based on the inspection task instructions, the robot's current position, the historical inspection route, and the equipment's historical inspection records to determine the initial range containing the candidate node set; the initial range includes at least one candidate machine room, area, or equipment node and its corresponding semantic confidence. Using a graph neural network model, within an initial range, a priority score is calculated for each candidate node based on node features, edge features, and a fault frequency weighting factor from historical inspection records. A list of candidate nodes sorted by the scores is then output. The target node is determined from the list of candidate nodes, and its coordinates are output. The large language model generates an inspection action sequence based on the target node, and the reinforcement learning model is used to generate and execute the optimal path to the target node.

[0006] In an optional implementation, a large language model is used to perform semantic parsing on the simplified subgraph based on the inspection task instructions, the robot's current position, the historical inspection routes, and the equipment's historical inspection records, in order to determine the initial range containing the candidate node set, specifically including: Construct semantic parsing prompt word templates; The prompt word template, simplified sub-graph, inspection task instruction, robot current position, inspection history route, and abnormal records in the equipment's historical inspection records are combined to construct a structured input text sequence. The input text sequence is fed into a large language model, and the output is a set of candidate nodes, which contains a unique identifier for each candidate node and a confidence score of the semantic fit between the candidate node and the inspection task instructions. The initial range is formed by selecting candidate nodes from the candidate node set whose semantic fit confidence is not less than the confidence threshold.

[0007] In an optional implementation, a graph neural network model is used to calculate the priority score of each candidate node within an initial range, based on node features, edge features, and a fault frequency weighting factor from historical inspection records. Specifically, this includes: Extract the corresponding subgraph from the initial range. ,in For a set of nodes, Let it be the set of edges; For each node Constructing node feature vectors The eigenvectors of all nodes constitute the node feature matrix. The node characteristics include device type code, real-time operating status parameters, the timestamp of the most recent inspection, and the code of the region to which it belongs; For each edge Constructing edge feature vectors All edge features constitute the edge feature matrix. Edge features include nodes and The geometric distance and connection type encoding between them; Extract each node from the equipment's historical inspection records. Fault frequency within the preset time window frequency of failures The fault frequency weighting factor is mapped through a nonlinear function. All failure frequency weighting factors Constructing the fault weighting factor vector ; Extract the edges from the user's preset route preference information. The relevant path preference types are identified, and each type is converted into a distance correction coefficient based on a preset mapping relationship. ; Get edges original geometric distance According to the distance correction factor The original geometric distance is dynamically adjusted to generate the effective distance. ; Subgraph Node feature matrix Edge feature matrix and fault weighting factor vector The input is fed into a pre-defined graph neural network model, which consists of at least one graph convolutional layer. Each layer performs message passing and aggregation operations using the following formula:

[0008] in, For the first Nodes in a layered graph convolutional layer eigenvectors, For nodes The set of neighboring nodes, For the first The layer is used to transform the trainable weight matrix of the central node's own features. For the first A differentiable function of the layer, For the first Layer computing nodes and Attention coefficient between them It is a non-linear activation function; edge feature vectors Includes effective distance ,function When generating a message, the effective distance will be considered. Encode and learn; After L layers of graph convolution operations, each node is obtained. The final layer output features The final layer output features Input a fully connected layer and pass it through a sigmoid activation function to compute the priority score for each node.

[0009] In one optional implementation, a large language model is used to generate an inspection action sequence based on the target node, and a reinforcement learning model is used to generate the optimal path to the target node, specifically including: Construct task planning prompt word templates for large language models; The task planning prompt template, target node coordinates, robot current position, and target node attribute information are combined into an input sequence, which is then input into the large language model. The output is a structured sequence of atomic actions, where each atomic action is a dictionary structure containing the action type and corresponding action parameters. The action type includes "move to". The navigation sub-targets are extracted from the atomic action sequence and a reinforcement learning state space is constructed. The optimal path is then generated using the trained reinforcement learning model.

[0010] In one optional implementation, navigation sub-objectives are extracted based on the atomic action sequence and a reinforcement learning state space is constructed. The optimal path is then generated using the trained reinforcement learning model, specifically including: Extract all actions of type "move to" from the atomic action sequence, and list the coordinates in their parameters as a queue of navigation sub-target points to be reached sequentially. ,in This is the first point that needs to be moved to; Constructing the state space of a reinforcement learning model It includes the robot's current coordinates, current sub-target point, recent obstacle point cloud data perceived by LiDAR, and current region features extracted from path preference information; state space The input is fed into a pre-trained reinforcement learning model, which then outputs the optimal action for the current state based on its policy network. The action includes the direction and speed of movement; the robot performs the action. Afterwards, the environment entered a new state. Repeat this process until the robot successfully reaches the current navigation sub-target point. ; From queue Extract the next sub-target point from the data and reconstruct the state space of the reinforcement learning model. This continues until all sub-target points corresponding to the "move to" action have been visited in sequence, thereby generating and executing the optimal path to the final target node. In one alternative implementation, the reward function of the reinforcement learning model is trained based on historical path data, and its reward items include: reward for successfully reaching a sub-target point, reward for conforming to path preferences, reward for moving away from obstacles, and penalty for deviating from a preset path.

[0011] In one optional implementation, the inspection task instruction is received, and the topology graph is folded using a large language model to obtain a simplified subgraph, specifically including: Receive inspection task instructions from users via voice or management platform; these instructions are in natural language text. Construct prompt word templates to guide large language models in graph folding; The prompt word template, multi-level topology map, and inspection task instructions are combined into an input sequence and then input into the large language model. The large language model generates folding decisions based on the semantic relevance of each node and level in the multi-level topology graph to the current task instruction. Based on these folding decisions, it performs folding operations on the multi-level topology graph to generate simplified subgraphs.

[0012] Secondly, the technical solution of the present invention provides an optimal path generation system for data center operation and maintenance robots, comprising: The topology map construction module is used to build a multi-level topology map of the data center. The topology map node hierarchy from top to bottom is building, floor, computer room, area and equipment. The node attributes of each level integrate the location information of the equipment, the historical inspection records of the equipment and the user's preset route preference information. The simplified subgraph generation module is used to receive inspection task instructions, fold the topology graph through a large language model to obtain a simplified subgraph. The simplified subgraph retains the hierarchical nodes related to the target node and represents the folded node data in the form of node attributes. The initial range determination module is used to perform semantic parsing on the simplified subgraph using a large language model based on the inspection task instructions, the robot's current position, the historical inspection route, and the equipment's historical inspection records, in order to determine the initial range containing the candidate node set; the initial range includes at least one candidate machine room, area, or equipment node and its corresponding semantic confidence. The target node determination module is used to calculate the priority score of each candidate node within an initial range using a graph neural network model, based on node features, edge features, and a fault frequency weighting factor in historical inspection records, and output a list of candidate nodes sorted by the scores; determine the target node from the list of candidate nodes and output its coordinates. The optimal path generation and execution module is used to generate an inspection action sequence based on the target node using a large language model, and to generate and execute the optimal path to the target node using a reinforcement learning model.

[0013] Thirdly, the technical solution of the present invention provides a terminal, including: The memory is used to store the optimal path generation program for the data center operation and maintenance robot. A processor is configured to implement the steps of the optimal path generation method for the data center operation and maintenance robot as described above when executing the optimal path generation program for the data center operation and maintenance robot.

[0014] Fourthly, the present invention provides a computer-readable storage medium storing a data center operation and maintenance robot optimal path generation program, wherein when the data center operation and maintenance robot optimal path generation program is executed by a processor, the data center operation and maintenance robot optimal path generation program implements the steps of the data center operation and maintenance robot optimal path generation method as described in any of the above claims.

[0015] As can be seen from the above technical solutions, this application has the following advantages: 1. This application constructs a multi-level topology map that integrates equipment location, historical inspection records, and user path preferences, enabling the map to carry a semantic network rich in operation and maintenance knowledge. This provides a data foundation for subsequent semantic-based intelligent search and decision-making, allowing the robot to deeply understand the context of operation and maintenance tasks. 2. This application utilizes a large language model to perform real-time semantic parsing of inspection task instructions and dynamically determines the folding and preservation strategy of the graph accordingly, generating task-specific simplified subgraphs. This overcomes the drawbacks of fixed folding strategies, compresses the amount of input data, reduces the computational load of the large language model, and improves the speed and accuracy of subsequent search and reasoning while ensuring that key information is not lost. 3. This application designs a multi-model collaborative architecture of LLM, GNN, and RL. The Large Language Model (LLM) is responsible for semantic parsing and preliminary localization. On this basis, the Graph Neural Network (GNN) performs refined weighted scoring by integrating node features, topological relationships, fault history, and path preferences, thereby achieving precise target localization for operation and maintenance. The Reinforcement Learning (RL) finally generates a motion path that takes into account efficiency, safety, and preferences based on the semantic target, thereby improving the accuracy and reliability of generating the optimal path. 4. This application further incorporates fault frequency as a weighting factor into the message passing process of the GNN and transforms path preference into distance correction coefficients for edge features, making the model's decision-making process closely aligned with actual operation and maintenance needs, thereby improving the system's reliability and practicality in real and complex environments. Attached Figure Description

[0016] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of a method for generating the optimal path for a data center operation and maintenance robot, provided in an embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram of a multi-level topological graph structure.

[0019] Figure 3 This is a schematic block diagram of an optimal path generation system for a data center operation and maintenance robot, provided as an embodiment of the present invention.

[0020] Figure 4 This is a schematic diagram of the structure of a terminal provided in an embodiment of the present invention. Detailed Implementation

[0021] To make the purpose, features, and advantages of this application more apparent and understandable, specific embodiments and accompanying drawings will be used to clearly and completely describe the technical solution protected by this application. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] Unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this application and in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0023] Figure 1 This is a schematic flowchart illustrating a method for generating optimal paths for data center operation and maintenance robots, provided in an embodiment of the present invention. Figure 1 The executing entity can be a data center operations and maintenance robot optimal path generation system. The data center operations and maintenance robot optimal path generation method provided in this embodiment is executed by a computer device; correspondingly, the data center operations and maintenance robot optimal path generation system runs on the computer device. Depending on different requirements, the order of the steps in this flowchart can be changed, and some steps can be omitted.

[0024] like Figure 1 As shown, the method includes the following steps.

[0025] S1 constructs a multi-level topology map of the data center. The topology map node hierarchy from top to bottom is building, floor, computer room, area and equipment. The node attributes of each level integrate the location information of the equipment, the historical inspection records of the equipment and the user's preset route preference information.

[0026] S2 receives the inspection task instruction, folds the topology graph through the large language model to obtain a simplified subgraph, which retains the hierarchical nodes related to the target node and represents the folded node data in the form of node attributes.

[0027] S3 uses a large language model to perform semantic parsing on the simplified subgraph based on the inspection task instructions, the robot's current position, the historical inspection route, and the equipment's historical inspection records, in order to determine the initial range containing the candidate node set; the initial range includes at least one candidate machine room, area, or equipment node and its corresponding semantic confidence.

[0028] S4. Using a graph neural network model, within the initial range, calculate the priority score of each candidate node based on node features, edge features, and fault frequency weighting factors in historical inspection records, and output a list of candidate nodes sorted by the scores; determine the target node from the list of candidate nodes and output its coordinates.

[0029] S5 generates an inspection action sequence based on the target node using a large language model, and uses a reinforcement learning model to generate and execute the optimal path to the target node.

[0030] As a refinement and extension of the specific implementation of the above embodiments, in order to fully explain the specific implementation process of this embodiment, the following will provide possible embodiments to describe the specific implementation of the above steps in a non-limiting manner.

[0031] Figure 2 This is a schematic diagram of a multi-level topology map. During the operation of a data center maintenance robot, the onboard visual camera and IMU perceive scene information from the data center, constructing a map node for each rack in the scene. The relative positions between racks are then used to construct map edges connecting the nodes. Based on the geometric structure of the data center, including its building layout, server room distribution, and equipment arrangement, as well as the relative positions of objects, the data center map is visualized from top to bottom as a multi-level topology structure of "building-floor-server room-area-equipment". The attributes of each level of node not only include the spatial location information of the equipment, such as the second row, row A, and east side, but also incorporate historical inspection records, such as anomaly rates and previous anomaly information, as well as user-preset route preference information, enabling the map to comprehensively support the parsing and execution of various task commands.

[0032] ①Architecture The building layer is located at the top of the data center map, representing the building where the data center operations and maintenance robot is located. Its attributes include user-defined inspection preference routes and historical inspection routes.

[0033] ② Floor The sub-levels of a building floor are called floors, which represent all the floors contained in the current building, and each floor has a unique serial number.

[0034] ③ Computer room The floor level is further divided into sub-levels called computer room levels, which represent all the computer rooms in the current floor. Each computer room node contains the corresponding attributes of the computer room, such as: the computer room number, the computer room name, the computer room area, the location in the floor, and the floor number.

[0035] ④ area The fourth layer of the data center map is the area layer, which divides multiple areas according to the placement of equipment in the computer room, such as the second row, row A, east side, etc. The attributes of the area layer include the number of equipment in the area, the room number to which it belongs, and the failure rate of the equipment.

[0036] ⑤ Equipment The sub-layer below the region layer is the device layer, which is also the bottom layer of the data center map. It contains all the devices in the region and has rich device information, such as: the device's serial number, name, surface attributes, area, length, width and height, the serial number of the region it is located in, and historical inspection records.

[0037] Based on a large language model, the complete data center map is intelligently folded, preserving the target area and related hierarchical nodes, while compressing non-critical or temporarily unnecessary node data. The folded scene data is represented by node attributes, such as the core inspection area of ​​a floor or a high-priority device set. If historical records show a linkage between non-target area devices and the target, the area is set to a semi-folded state to retain monitoring attributes. This mechanism significantly improves the efficiency of the large language model in processing scene data while ensuring the integrity of critical data center information. Specifically, in some optional implementations, step S2 receives the inspection task instruction and folds the topology map using the large language model to obtain a simplified sub-map, specifically including the following steps.

[0038] S2.1 Receives inspection task instructions input by the user via voice or management platform. These instructions are in natural language text.

[0039] The system receives inspection task instructions input by users via voice or a management platform. The instructions are natural language text used to express the user's inspection needs. The system performs preliminary semantic analysis on the instructions, extracting key entities and operational verbs to provide semantic guidance for subsequent graph folding.

[0040] S2.2, Construct prompt word templates to guide large language models in graph folding.

[0041] The prompt word template is used to define the roles of the large language model, the structure of the input data, the format of the output data, and the goal of the folding task; its structure includes: Role definition: "You are a data center topology management expert, responsible for intelligently folding (compressing) the topology structure according to task instructions to focus on key information." Input structure: "Input data includes: 1) a complete multi-level topology map of the data center; 2) inspection task instructions to be executed." Output format: "Please output a JSON object describing the collapsed graph structure. It must contain a folded_nodes field, the value of which is a list of collapsed node IDs; and a retained_nodes field, the value of which is a list of node IDs that need to be retained." Task objective: "Your task is to analyze the task instructions and determine which levels or nodes in the graph are irrelevant to the current task or have low priority, and then collapse them; at the same time, retain the key nodes that are directly related to the task and may be related to it." S2.3 combines the prompt word template, multi-level topology map, and inspection task instructions into an input sequence, which is then input into the large language model.

[0042] S2.4 The large language model generates a folding decision based on the semantic relevance of each node and level in the multi-level topology graph to the current task instruction. Based on the folding decision, the multi-level topology graph is folded to generate a simplified subgraph.

[0043] Based on its built-in knowledge and reasoning capabilities, the large language model performs deep semantic understanding on the input sequence and comprehensively assesses the semantic relevance of each node and level in the topological graph to the current task instruction. Based on the semantic relevance, a folding decision is generated. This folding decision clearly distinguishes between the set of nodes that need to be folded and compressed and the set of nodes that need to be retained in the simplified view.

[0044] Based on the folding decision, a folding operation is performed on the complete topology graph to generate a simplified subgraph. For any node in the folded node set, it is removed from the explicit structure of the graph, and its key information summary (e.g., node type, number of child nodes, key attribute statistics) is extracted into a new attribute field and appended to its parent node. For the retained node set, its original node characteristics and edge relationships are maintained. Finally, the simplified subgraph is output, which is a task-specific, data-compressed view of the original topology graph for subsequent processing.

[0045] The large language model performs reasoning and parsing on the constructed simplified subgraph to locate the equipment, areas, or rooms to be inspected. Simultaneously, it retrieves historical inspection records to filter data such as fault frequency and last inspection time to determine priorities. Subsequently, it standardizes the data center map, current location, task instructions, inspection records, and path preferences, transforming natural language path preferences into computable parameters, such as congested area coordinates and path weight coefficients. Specifically, the node search phase employs a hybrid "LLM+GNN" architecture. During map search, information such as the data center map, the current location of the maintenance robot, and user-issued task instructions are used as input. Based on the large language model, it determines the approximate range of the target object within the data center map. Search data center map data Data center map sub-graph after search completion Defined as:

[0046] Where: Data center map Depend on The structure consists of layers, from top to bottom: building - floor - computer room - area - equipment. Each layer has multiple nodes, represented as follows: ; For user-issued inspection task instructions, such as inspecting equipment A22, inspecting equipment in the second computer room on the third floor, inspecting equipment with a high failure rate, and inspecting according to the third route; The current positioning of data center operation and maintenance robots; The historical inspection routes of the data center maintenance robot in this map; For equipment Historical test records are recorded as follows: 0 for normal and 1 for abnormal. The defined prompt is used to guide the large language model to find the target equipment, area, and computer room that meet the requirements of the current inspection task; This is a function for searching nodes in a large language model.

[0047] The first stage, as in step S3, uses a large language model to perform semantic parsing on the simplified subgraph based on the inspection task instructions, the robot's current position, the historical inspection route, and the equipment's historical inspection records, in order to determine the initial range containing the candidate node set. This specifically includes the following steps.

[0048] S3.1, Construct semantic parsing prompt word templates.

[0049] The prompt word template is used to define the roles of the large language model, the structure of the input data, the format of the output data, and the task objectives; its structure is as follows: Role definition: "You are a data center operations and maintenance expert who can accurately locate the approximate area where the target to be inspected is located based on the provided scene map and task instructions." Input structure: "Input data includes: 1) a simplified scene map describing the data center architecture; 2) the current inspection task instructions; 3) the robot's current floor and machine room location; 4) a summary of historical inspection routes; 5) a summary of historical equipment anomaly records." Output format: "Please output your inference results in strict JSON format, including the following field: candidate_nodes (a list containing the IDs of all possible candidate nodes and their semantic confidence in belonging to the target)." Task objective: "Your task is to analyze the instructions and identify all computer rooms, areas, or equipment nodes that may contain the inspection targets from the map." S3.2 Combine the prompt word template, simplified sub-graph, inspection task instruction, robot current position, inspection history route, and abnormal records in the equipment's historical inspection records to construct a structured input text sequence.

[0050] S3.3 Input the input text sequence into the large language model and output a set of candidate nodes. This set contains a unique identifier for each candidate node and the semantic fit confidence of the candidate node with the inspection task instruction.

[0051] After receiving the input sequence, the Large Language Model (LLM) performs semantic understanding of the inspection task instructions based on its built-in knowledge base and reasoning capabilities. It also comprehensively considers the node relationships in the simplified subgraph, the proximity of the robot's current position, the preference of historical inspection routes, and the high-frequency fault points in the equipment's historical anomaly records to perform comprehensive reasoning.

[0052] The output of the Large Language Model (LLM) is a text sequence conforming to a specified JSON format, which, after parsing, contains a set of candidate nodes. :

[0053] in, Let be the unique identifier of the i-th candidate node, and let be the candidate node. The confidence score for semantic fit with inspection task instructions is a value between 0 and 1. This represents the number of candidate nodes. The confidence output of the large language model is an internally generated probability value. The generation process is based on the large language model's inner product calculation and probability normalization of the embedding vector of the input sequence and the prior concepts in the knowledge base.

[0054] S3.4 Select candidate nodes from the candidate node set whose semantic fit confidence is not less than the confidence threshold to form the initial range.

[0055] Set a semantic confidence threshold θ, and select the candidate node set. All semantic confidence Nodes greater than or equal to the threshold θ constitute the initial range described in this step, which is the search space for subsequent graph neural network models to perform refined priority calculations. The second stage, as in step S4, involves using a graph neural network model to calculate the priority score of each candidate node within the initial range, based on node features, edge features, and a fault frequency weighting factor from historical inspection records. A list of candidate nodes sorted by these scores is then output. The target node is then determined from this list, and its coordinates are output. This process specifically includes the following steps.

[0056] S4.1, Extract the corresponding subgraph from the initial range. ,in For a set of nodes, Let it be the set of edges.

[0057] S4.2, for each node Constructing node feature vectors The eigenvectors of all nodes constitute the node feature matrix. The node characteristics include the device type code, real-time operating status parameters, the timestamp of the most recent inspection, and the code of the region to which it belongs. Real-time operating status parameters include temperature, CPU load, etc.

[0058] S4.3, for each edge Constructing edge feature vectors All edge features constitute the edge feature matrix. Edge features include nodes and The geometric distance and connection type encoding between them.

[0059] Connection type encoding, such as power connection, network connection, physical adjacency.

[0060] S4.4 Extract each node from the equipment's historical inspection records. Fault frequency within the preset time window frequency of failures The fault frequency weighting factor is mapped through a nonlinear function. All failure frequency weighting factors Constructing the fault weighting factor vector .

[0061] Fault frequency weighting factor The weight used to amplify frequently failing nodes during message passing is calculated using the following formula:

[0062] in, This is a preset gain coefficient used to control the strength of the impact of fault history on priority.

[0063] S4.5, extract the relevant information from the user's preset route preference information. The relevant path preference types are identified, and each type is converted into a distance correction coefficient based on a preset mapping relationship. .

[0064] S4.6, Get Edges original geometric distance According to the distance correction factor The original geometric distance is dynamically adjusted to generate the effective distance. . Path preference dynamically adjusts the geometric distance in edge features using distance correction coefficients (such as 0.9 for fire lanes and 1.5 for congested areas) to ensure that the influence of path preference is taken into account when the graph neural network model calculates the node correlation.

[0065] Path preference type, such as fire lanes, congested areas, and regular routes. For example, fire lanes correspond to... =0.9, corresponding to the congested area =1.5, corresponding to the standard channel =1.0.

[0066] The effective distance is used to characterize the ease of passage; the larger the value, the higher the cost for the robot to traverse the path. S4.7, Subgraph Node feature matrix Edge feature matrix and fault weighting factor vector The input is fed into a pre-defined graph neural network model, which consists of at least one graph convolutional layer. Each layer performs message passing and aggregation operations using the following formula:

[0067] in, For the first Nodes in a layered graph convolutional layer eigenvectors, For nodes The set of neighboring nodes, For the first The layer is used to transform the trainable weight matrix of the central node's own features. For the first A differentiable function of the layer, For the first Layer computing nodes and Attention coefficient between them It is a non-linear activation function; edge feature vectors Includes effective distance ,function When generating a message, the effective distance will be considered. To encode and learn. S4.8, after L layers of graph convolution operations, obtain each node. The final layer output features The final layer output features Input a fully connected layer and pass it through a sigmoid activation function to compute the priority score for each node.

[0068] In some optional implementations, a preset strategy for determining the final target node is provided, including a highest-score strategy, a verification-passing strategy, or a manually specified strategy. The final target node is determined from the candidate node list based on the selected strategy. If the highest-score strategy is used, the candidate node with the highest priority score in the list is directly determined as the target node. If the verification-passing strategy is used, candidate nodes in the list are attempted to be confirmed sequentially in descending order of priority score; the confirmation attempt includes: initiating a real-time query of the node's status, or calling a lightweight verification rule for checking; once the first node that passes verification appears, it is immediately determined as the target node. If the manually specified strategy is used, the sorted candidate node list is submitted to the user interface for the user to select; the user's selection instruction is received, and the node specified in the instruction is determined as the target node. In the multi-level topology map, the coordinate attributes of the finally determined target node are queried; the coordinate attributes are spatial pose representations in a preset global coordinate system; a data tuple containing the target node and its coordinates is output to the path planning program.

[0069] After obtaining the data center graph subgraph based on a multimodal collaborative graph search mechanism, the large language model plans a sequence of inspection actions to complete the task instruction based on the current location of the operation and maintenance robot:

[0070] In the formula: These are prompt words used to guide the large language model in generating inspection action sequences; This is the generated sequence of inspection actions; Generate functions for inspection actions of large language models.

[0071] The generated inspection sequence is broken down into individual target points. A deep Q-network reinforcement learning model is used, with robot localization, target coordinates, obstacle distribution, and path preference as the state space, and movement direction and speed as the action space. A reward function is trained based on historical path data (e.g., +10 for paths conforming to preferences, -5 for deviations) to generate the optimal path to each target point. During the execution phase, the robot moves along the planned path and dynamically avoids obstacles using sensors. After completion, the inspection records (equipment status, anomaly information, etc.) are updated, and the execution data is fed back to the graph neural network model and reinforcement learning model for iterative parameter optimization.

[0072] In some optional implementations, step S5 generates an inspection action sequence based on the target node using a large language model, and generates the optimal path to the target node using a reinforcement learning model, specifically including the following steps.

[0073] S5.1, Construct task planning prompt word templates for large language models.

[0074] The prompt word template is used to define the role, input information, output format, and constraints of the large language model in task planning; its structure includes: Role definition: "You are the task planner of a data center operations robot, responsible for breaking down high-level goals into a sequence of executable atomic actions." Input information: "Input includes: 1) the final target node coordinates Ptarget; 2) the robot's current coordinates Probot; 3) the target node's device type and the inspection operation to be performed (e.g., scanning QR codes, infrared temperature measurement, taking photos)." Output format: "Please output a JSON array listing the sequence of atomic actions in order, with each action containing the action_type and action_parameters fields." Constraints: "The sequence of actions must conform to the robot's physical constraints and should begin with navigation to the target node position." S5.2, combine the task planning prompt word template, target node coordinates, robot current position and target node attribute information into an input sequence, input it into the large language model, and output a structured atomic action sequence. Each atomic action is a dictionary structure containing action type and corresponding action parameters; the action type includes "move to".

[0075] In addition to "move to", action types can also include "manipulate robotic arm" and "acquire image", with corresponding action parameters such as: {"coordinates": [x, y, z]}.

[0076] S5.3 Extract navigation sub-targets from atomic action sequences and construct a reinforcement learning state space, then use the trained reinforcement learning model to generate the optimal path.

[0077] S5.31 Extract all actions of type "move to" from the atomic action sequence, and list the coordinates in their parameters as a queue of navigation sub-target points that need to be reached in sequence. ,in This is the first point that needs to be moved to.

[0078] S5.32, Constructing the state space of the reinforcement learning model It includes the robot's current coordinates, current sub-target point, recent obstacle point cloud data perceived by LiDAR, and current area features extracted from path preference information.

[0079] S5.33, state space The input is fed into a pre-trained reinforcement learning model, which then outputs the optimal action for the current state based on its policy network. The action includes the direction and speed of movement; the robot performs the action. Afterwards, the environment entered a new state. Repeat this process until the robot successfully reaches the current navigation sub-target point. .

[0080] S5.34, from the queue Extract the next sub-target point from the data and reconstruct the state space of the reinforcement learning model. This continues until all sub-target points corresponding to the "move to" action have been visited in sequence, thereby generating and executing the optimal path to the final target node.

[0081] From queue Extract the next sub-target point and return to step S5.32 to repeat the process.

[0082] The reward function of the reinforcement learning model is trained based on historical path data, and its reward items include: reward for successfully reaching the sub-target point, reward for conforming to path preferences, reward for moving away from obstacles, and penalty for deviating from the preset path.

[0083] After the robot reaches the target node position through the optimal path, it sequentially executes non-movement operations (such as robotic arm control and data acquisition) in the atomic action sequence. After the task is completed, the execution data of this path (such as actual path points, energy consumption, and time consumption) and inspection result data are fed back to the reinforcement learning model and graph neural network model for iterative optimization of model parameters.

[0084] The above text provides a detailed description of an embodiment of a method for generating optimal paths for data center operation and maintenance robots. Based on the above embodiment, this invention also provides a system for generating optimal paths for data center operation and maintenance robots corresponding to this method.

[0085] Figure 3 This is a schematic block diagram of an optimal path generation system for a data center operation and maintenance robot provided in an embodiment of the present invention. In this embodiment, the optimal path generation system 300 for the data center operation and maintenance robot can be divided into multiple functional modules according to the functions it performs. A module, as referred to in this invention, is a series of computer program segments that can be executed by at least one processor and perform a fixed function, and is stored in memory.

[0086] The topology map construction module 310 is used to construct a multi-level topology map of the data center. The topology map node hierarchy from top to bottom is building, floor, computer room, area and equipment. The node attributes of each level integrate the location information of the equipment, the historical inspection records of the equipment and the user's preset route preference information.

[0087] The simplified subgraph generation module 320 is used to receive inspection task instructions, fold the topology graph through a large language model to obtain a simplified subgraph, which retains the hierarchical nodes related to the target node and represents the folded node data in the form of node attributes.

[0088] The initial range determination module 330 is used to perform semantic parsing on the simplified subgraph using a large language model based on the inspection task instructions, the robot's current position, the historical inspection route, and the historical inspection records of the equipment, in order to determine the initial range containing the candidate node set; the initial range includes at least one candidate machine room, area, or equipment node and its corresponding semantic confidence.

[0089] The target node determination module 340 is used to calculate the priority score of each candidate node within an initial range using a graph neural network model, based on node features, edge features, and fault frequency weighting factors in historical inspection records, and output a list of candidate nodes sorted by the scores; determine the target node from the list of candidate nodes and output its coordinates.

[0090] The optimal path generation and execution module 350 is used to generate an inspection action sequence based on the target node using a large language model, and to generate and execute the optimal path to the target node using a reinforcement learning model.

[0091] The optimal path generation system for data center operation and maintenance robots in this embodiment is used to implement the aforementioned optimal path generation method for data center operation and maintenance robots. Therefore, the specific implementation of this system can be found in the embodiment section of the optimal path generation method for data center operation and maintenance robots mentioned above. Thus, the specific implementation can be referred to the description of the corresponding embodiments, and will not be elaborated here.

[0092] Furthermore, since the optimal path generation system for the data center operation and maintenance robot in this embodiment is used to implement the aforementioned optimal path generation method for the data center operation and maintenance robot, its function corresponds to the function of the above method, and will not be repeated here.

[0093] Figure 4 This is a schematic diagram of a terminal 400 provided in an embodiment of the present invention, including: a processor 410, a memory 420, and a communication unit 430. The processor 410 is used to implement the process steps of the above-described embodiment of the optimal path generation method for data center operation and maintenance robots when implementing the optimal path generation program for data center operation and maintenance robots stored in the memory 420.

[0094] The terminal 400 includes a processor 410, a memory 420, and a communication unit 430. These components communicate via one or more buses. Those skilled in the art will understand that the server structure shown in the figure does not constitute a limitation of the present invention. It can be a bus topology or a star topology, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0095] The memory 420 can be used to store the execution instructions of the processor 410. The memory 420 can be implemented by any type of volatile or non-volatile storage terminal or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. When the execution instructions in the memory 420 are executed by the processor 410, the terminal 400 is able to perform some or all of the steps in the above method embodiments.

[0096] The processor 410 serves as the control center of the storage terminal, connecting various parts of the electronic terminal via various interfaces and lines. It executes software programs and / or modules stored in the memory 420, and calls data stored in the memory to perform various functions of the electronic terminal and / or process data. The processor can be composed of integrated circuits (ICs), such as a single packaged IC or multiple packaged ICs with the same or different functions connected together. For example, the processor 410 may consist only of a central processing unit (CPU). In this embodiment of the invention, the CPU may have a single processing core or include multiple processing cores.

[0097] The communication unit 430 is used to establish a communication channel, enabling the storage terminal to communicate with other terminals. It can receive user data sent by other terminals or send user data to other terminals.

[0098] The present invention also provides a computer storage medium, wherein the storage medium may be a magnetic disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0099] The present invention also provides a computer storage medium, wherein the storage medium may be a magnetic disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0100] The computer storage medium stores a data center operation and maintenance robot optimal path generation program. When the data center operation and maintenance robot optimal path generation program is executed by the processor, it implements the process steps of the above-described data center operation and maintenance robot optimal path generation method embodiment.

[0101] Those skilled in the art will clearly understand that the techniques in the embodiments of the present invention can be implemented using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, or any other medium capable of storing program code. It includes several instructions to cause a computer terminal (which may be a personal computer, a server, or a second terminal, a network terminal, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0102] In the embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0103] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0104] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0105] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for generating an optimal path for a data center operation robot, characterized in that, The method comprises the following steps: constructing a multi-level topology graph of the data center, the node levels of the topology graph being building, floor, machine room, area and device from top to bottom, wherein the position information of the device, the historical inspection record of the device and the route preference information preset by the user are fused in the node attribute of each level; receiving an inspection task instruction, folding the topology graph through a large language model to obtain a simplified subgraph, the simplified subgraph retaining the level nodes related to a target node, and the folded node data being represented in the form of node attributes; performing semantic analysis on the simplified subgraph through the large language model according to the inspection task instruction, the current position of the robot, the historical inspection route and the historical inspection record of the device to determine an initial range containing a candidate node set; the initial range containing at least one candidate machine room, area or device node and the corresponding semantic confidence thereof; calculating the priority score of each candidate node in the initial range through a graph neural network model based on the node features, edge features and fault frequency weighting factor in the historical inspection record information, and outputting a candidate node list sorted according to the score; determining the target node from the candidate node list and outputting the coordinates thereof; generating an inspection action sequence according to the target node through the large language model, and generating and executing the optimal path to the target node through a reinforcement learning model.

2. The method of claim 1, wherein, performing semantic analysis on the simplified subgraph through the large language model according to the inspection task instruction, the current position of the robot, the historical inspection route and the historical inspection record of the device to determine an initial range containing a candidate node set, specifically comprising: constructing a semantic analysis prompt word template; combining the prompt word template, the simplified subgraph, the inspection task instruction, the current position of the robot, the historical inspection route and the abnormal record in the historical inspection record of the device into a structured input text sequence; inputting the input text sequence into the large language model to output a candidate node set, the set containing the unique identifier of each candidate node and the semantic fit degree confidence of the candidate node and the inspection task instruction; selecting the candidate nodes with a semantic fit degree confidence not less than a confidence threshold from the candidate node set to form the initial range.

3. The method of claim 2, wherein, calculating the priority score of each candidate node in the initial range through a graph neural network model based on the node features, edge features and fault frequency weighting factor in the historical inspection record information, specifically comprising: extracting a corresponding subgraph from the initial range wherein is a set of nodes, is a set of edges; For each node Construct node feature vector The feature vectors of all nodes constitute a node feature matrix The node features include device type code, real-time running state parameters, last inspection timestamp, and area code to which the node belongs For each edge Construct edge feature vector , all edge features constitute edge feature matrix , edge features include the geometric distance between nodes and Connection relationship type code; extracting each node from the device historical inspection records failure frequency in a preset time window failure frequency mapping the failure frequency to a failure frequency weighting factor through a nonlinear function all failure frequency weighting factors constituting a failure weighting factor vector ; Extract the edges from the user's preset route preference information. The relevant path preference types are identified, and each type is converted into a distance correction coefficient based on a preset mapping relationship. ; acquisition edge original geometric distance , dynamically adjust the original geometric distance according to the distance correction coefficient to generate an effective distance ; subgraphs , node feature matrix , edge feature matrix , and a failure weighting factor vector are input into a preset graph neural network model, which is composed of at least one graph convolution layer, and each layer performs a message passing and aggregation operation through the following formula: wherein, is the layer, is the feature vector of a node in the layer, is the set of neighboring nodes of the node , is the layer, is the trainable weight matrix of the layer for transforming the node’s own feature, is a differentiable function of the layer, is the layer, is the attention coefficient between the node and is a nonlinear activation function; the edge feature vector contains the effective distance , the function encodes and learns the effective distance when generating the message. After L-layer graph convolution operation, the final layer output features of each node are obtained The final layer output features of each node are input into a fully connected layer and the priority score of each node is calculated via a sigmoid activation function.

4. The method of claim 1, wherein, generating an inspection action sequence according to the target node through the large language model, and generating and executing the optimal path to the target node through a reinforcement learning model, specifically comprising: constructing a task planning prompt word template for the large language model; combining the task planning prompt word template, the coordinates of the target node, the current position of the robot and the attribute information of the target node into an input sequence, inputting the input sequence into the large language model, and outputting a structured atomic action sequence, each atomic action being a dictionary structure containing an action type and corresponding action parameters; the action type includes "move to"; extracting navigation sub-targets from the atomic action sequence and constructing a reinforcement learning state space, and using a trained reinforcement learning model to generate the optimal path.

5. The method of claim 4, wherein, According to the atomic action sequence, a navigation sub-target is extracted and a reinforcement learning state space is constructed, and an optimal path is generated using a trained reinforcement learning model, specifically including: Extract all the action types of "move to" from the atomic action sequence, and list the coordinate points in the parameters in turn as the navigation sub-target point queue that needs to be reached in turn wherein is the first point that needs to be moved to; State space for building reinforcement learning model , including current coordinates of the robot, current sub-target point, recent obstacle point cloud data perceived by the lidar, and current regional features extracted from path preference information; state space input to a pre-trained reinforcement learning model, the reinforcement learning model outputs an optimal action in the current state based on its policy network , the action including a moving direction and speed; the robot performs the action After that, the environment enters a new state , repeat the process until the robot successfully reaches the current navigation sub-goal point ; from the queue the next sub-goal point is taken out, the state space of the reinforcement learning model is reconstructed , until all the sub-goal points corresponding to the "move to" actions are sequentially accessed, thereby generating and executing the optimal path to the final goal node.

6. The method of claim 5, wherein, The reward function of the reinforcement learning model is trained based on historical path data, and the reward item setting includes: a reward for successfully reaching the sub-target point, a reward for meeting the path preference, a reward for moving away from obstacles, and a penalty for deviating from the preset path.

7. The method of claim 1, wherein, Receiving an inspection task instruction, folding the topology graph through a large language model to obtain a simplified subgraph, specifically including: Receiving an inspection task instruction input by a user through voice or a management platform, the instruction being a natural language text; Building a prompt word template for guiding the folding of the large language model; Combining the prompt word template, the multi-level topology graph, and the inspection task instruction into an input sequence and inputting it into the large language model; The large language model generates a folding decision according to the semantic relevance of each node and each level in the multi-level topology graph to the current task instruction, performs a folding operation on the multi-level topology graph according to the folding decision, and generates a simplified subgraph. 8.A system for generating an optimal path for a data center operation robot, characterized in that, It includes: A topology graph construction module for constructing a multi-level topology graph of a data center, the node levels of the topology graph being building, floor, machine room, area, and device from top to bottom, wherein the node attributes of each level are fused with the position information of the device, the historical inspection records of the device, and the route preference information preset by the user; A simplified subgraph generation module for receiving an inspection task instruction, folding the topology graph through a large language model to obtain a simplified subgraph, the simplified subgraph retaining the level nodes related to the target node and representing the folded node data in the form of node attributes; A node initial range determination module for performing semantic analysis on the simplified subgraph according to the inspection task instruction, the current position of the robot, the historical inspection route, and the historical inspection records of the device through a large language model to determine an initial range containing a candidate node set; the initial range contains at least one candidate machine room, area, or device node and its corresponding semantic confidence; A target node determination module for calculating the priority score of each candidate node based on the node features, edge features, and fault frequency weighting factor in the historical inspection record information within the initial range through a graph neural network model, and outputting a candidate node list sorted by the score; determining the target node from the candidate node list and outputting its coordinates; An optimal path generation and execution module for generating an inspection action sequence according to the target node through a large language model, and generating and executing an optimal path to the target node using a reinforcement learning model.

9. A terminal, characterized by comprising: It includes: A memory for storing a data center operation robot optimal path generation program; A processor for executing the data center operation robot optimal path generation program to implement the steps of the data center operation robot optimal path generation method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The readable storage medium stores a data center operation robot optimal path generation program, and the data center operation robot optimal path generation program is executed by the processor to implement the steps of the data center operation robot optimal path generation method according to any one of claims 1 to 7.

Images