Data center operation and maintenance automatic management method based on knowledge graph

Abstract

The invention discloses a data center operation and maintenance automatic management method based on a knowledge graph, and the method comprises the following steps: S1, collecting task, strategy, personnel and equipment information, and constructing a task and strategy set; s2, constructing a knowledge graph based on the set; s3, generating a matching initial matrix and a candidate solution set in combination with the path structure, the edge weight and the semantic similarity; s4, copying the knowledge graph to generate a disturbance graph, and setting a node state and an edge disturbance coefficient; s5, calculating a disturbance factor; s6, loading a disturbance factor and a confidence coefficient weight, and executing an improved Jaya algorithm to update the candidate solution; s7, matching similar historical sub-graphs, and extracting strategy path loading candidate solutions; s8, when the candidate solution is not updated, generating a reverse disturbance vector; and S9, outputting an optimal task-strategy matching vector. According to the invention, adaptive efficient matching of the operation and maintenance task and the strategy of the data center is realized.

Description

Technical Field

[0001] This invention relates to the field of data center operation and maintenance management technology, and in particular to a knowledge graph-based automated data center operation and maintenance management method. Background Technology

[0002] Currently, data center operation and maintenance management generally adopts a combination of rule configuration and manual decision-making for task scheduling and strategy selection. While this approach offers a certain degree of controllability, it suffers from problems such as information silos, slow response, and low matching efficiency when facing complex environments with large task volumes and frequent updates. Various types of data, such as task information, policy rules, personnel roles, and equipment status, are usually scattered across different systems and lack a unified structure. This makes it impossible to systematically model the dependencies and execution order between tasks, and makes it difficult to achieve fine-grained scheduling and dynamic adjustment in the strategy matching process.

[0003] In recent years, some systems have attempted to introduce knowledge graphs to improve the matching ability between tasks and policies. However, most existing knowledge graphs are only used for static display or tag aggregation after construction, and have not been used to drive semantic matching and path planning for operation and maintenance tasks. The relationship between task nodes and policy nodes is only measured by path length or keyword similarity, lacking comprehensive modeling of path structure, edge weight and semantic diversity. As a result, the generated matching results are difficult to reflect the actual dependency constraints and execution priorities.

[0004] Meanwhile, in the process of candidate strategy generation and updating, mainstream methods mostly adopt general optimization algorithms such as genetic algorithms or particle swarm optimization, which lack specific modeling of knowledge graph structure. The state of candidate solutions cannot reflect the perturbation effect and state coupling relationship between nodes in the graph. The update strategy lacks the identification and control of confidence level, and is prone to getting trapped in local optima. In addition, when candidate solutions are not updated, existing methods mostly adopt simple restart mechanisms, ignoring the guiding role of causal links between tasks and lacking the ability to construct reverse intervention paths for stalled tasks. This makes the system have obvious limitations in the process of complex evolution of multiple tasks and cannot effectively adapt to the automated operation and maintenance needs under dynamic, heterogeneous and uncertain conditions.

[0005] Therefore, how to provide a knowledge graph-based method for automated data center operation and maintenance management is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an automated management method for data center operations and maintenance based on knowledge graphs. This invention fully integrates knowledge graph modeling, semantic matching calculation, perturbation factor-driven optimization mechanism and structural similarity recognition strategy. It describes in detail the entire process of constructing an automated matching path between tasks and strategies in multi-source heterogeneous operations and maintenance information, and has the advantages of clear structural modeling, accurate matching calculation and strong dynamic response capability.

[0007] An automated data center operation and maintenance management method based on knowledge graphs according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect task, strategy, personnel and equipment information, perform structured processing, and construct task sets and strategy sets;

[0009] S2. Construct a knowledge graph based on task sets and policy sets;

[0010] S3. Generate an initial task-policy matching matrix based on the path structure, edge weights, and semantic similarity between task nodes and policy nodes, and construct a set of candidate solutions, where each candidate solution is a task-policy matching vector.

[0011] S4. Copy the knowledge graph to generate a perturbation graph layer, set the node state variables and edge perturbation coefficients, and construct the perturbation factor calculation structure;

[0012] S5. During the candidate solution update process, the perturbation factor is calculated based on the adjacency state and edge perturbation coefficient of the relevant nodes in the perturbation map layer.

[0013] S6. Input the perturbation factor and the task-policy matching initial matrix into the improved Jaya algorithm, load the confidence weights generated based on knowledge graph reasoning, perform asymmetric vector update, and generate the updated candidate solution set.

[0014] S7. Based on the structural position of the task nodes in the knowledge graph in the updated candidate solution set, match the historical subgraphs with similar structures, extract the strategy path and load it into the corresponding candidate solution.

[0015] S8. If the candidate solution set is not updated within a set number of rounds, construct the causal path between task nodes, generate the reverse perturbation vector, and update the candidate solution state.

[0016] S9. When the termination condition is met, output the task-policy matching vector with the highest fitness score.

[0017] Optionally, the improved Jaya algorithm, compared to the original Jaya algorithm, includes:

[0018] A perturbation factor calculated from the perturbation map layer is introduced into the basic update vector to dynamically adjust the update direction and step size;

[0019] The state variables of the candidate solution associated nodes and the edge perturbation coefficients are jointly modeled as perturbation enhancement terms, and these perturbation enhancement terms are embedded during the update process.

[0020] We introduce confidence weights generated by knowledge graph reasoning to perform weighted corrections on perturbation enhancement terms, thereby improving the update magnitude of high-confidence task-policy matching pairs;

[0021] An asymmetric update mechanism is adopted to adjust the update ratios for solutions close to the optimal solution and solutions far from the worst solution, thereby enhancing convergence stability and local search capability.

[0022] Optionally, the knowledge graph includes task nodes, strategy nodes, personnel nodes, and device nodes, with edges representing dependencies, execution order, and resource constraints between nodes.

[0023] Optionally, S3 specifically includes:

[0024] S31. Extract the path structure between task nodes and strategy nodes in the knowledge graph, and count the path length, number of transit nodes and type of connection edges.

[0025] S32. Calculate the path weight score based on the weight of all edges in the path, and set a threshold to filter high-weight paths;

[0026] S33. Based on the semantic tags of task nodes and policy nodes, use an embedding model to generate semantic vectors and calculate the semantic similarity between task nodes and policy nodes.

[0027] S34. Normalize the path weight score and semantic similarity score, then perform a weighted fusion to construct the initial task-policy matching matrix M, where the matching score M... ij The calculation method is as follows:

[0028]

[0029] Where λ1 and λ2 are weight adjustment coefficients, satisfying λ1 + λ2 = 1, |P ij | represents the number of valid paths between task i and strategy j, w k δ represents the weight of path k. ij Indicates the semantic similarity between task nodes and policy nodes;

[0030] S35. Construct a corresponding matching vector for each row in the initial matching matrix to form a candidate solution set. The candidate solution is a set of correspondences between the task set and the policy set, and the dimension is equal to the size of the task set.

[0031] Optionally, S4 specifically includes:

[0032] S41. Copy the constructed knowledge graph, retain the topological connections of task nodes, policy nodes, personnel nodes and device nodes, and construct a perturbation graph layer with the same structure;

[0033] S42. Set a state variable for each node in the disturbance map layer, wherein the state variable represents information about task execution status, strategy matching status or device working status.

[0034] S43. Set a perturbation coefficient for each edge in the perturbation graph. The perturbation coefficient is used to quantify the perturbation sensitivity between the states of connected nodes.

[0035] S44. Construct the perturbation factor calculation structure, and define the perturbation factor of each node as a weighted average of the perturbation weights and state correlations between it and all its neighboring nodes:

[0036]

[0037] Where, θ i d represents the perturbation factor of node i. i This represents the number of adjacent nodes of node i. Let η represent the set of adjacent nodes of node i. ij ρ(s) represents the perturbation coefficient between node i and its neighboring node j. i ,s j ) represents the correlation function between the state variables of node i and node j;

[0038] S45. Bind the perturbation factor calculation structure to the perturbation map layer to form a composite layer containing structure, state and perturbation factors, which is used to support the dynamic update process of the candidate solution population.

[0039] Optionally, S6 specifically includes:

[0040] S61. Jointly encode the perturbation factor matrix and the initial task-policy matching matrix to generate a matching candidate tensor with perturbation structure, which serves as the input representation of the candidate solution population;

[0041] S62. Load the confidence weight vector generated by the knowledge graph reasoning results. The confidence weight vector assigns a confidence scalar to each task-policy pair for use in fitness updates.

[0042] S63. Construct an improved asymmetric update mechanism for the Jaya algorithm, and adjust the update ratios of candidate solutions in directions close to the optimal solution and in directions far from the worst solution respectively.

[0043] S64. Define the objective function for asymmetric updates as follows:

[0044]

[0045] in, This represents the matching vector of the i-th candidate solution in round t+1. Let x represent the matching vector of the i-th candidate solution in round t. best Let x represent the optimal candidate solution vector. worstLet r1 and r2 be random variables between 0 and 1, α and β be scaling factors controlling the positive and negative update directions, and θ be the worst candidate solution vector. i Let w represent the perturbation factor of the node corresponding to the i-th candidate solution, δ be the perturbation enhancement coefficient, and w be the perturbation factor of the node corresponding to the i-th candidate solution. i This is the confidence weight vector;

[0046] S65. After performing update calculations for each candidate solution, select individuals with improved fitness scores to generate an updated candidate solution population.

[0047] Optionally, S7 specifically includes:

[0048] S71. Traverse each task node in the candidate solution group, extract the structural position of the task node in the knowledge graph, including the node's connection path, adjacency strategy node identifier and edge type, and construct a set of structural paths.

[0049] S72. Perform graph embedding encoding on each structural path to generate a structural representation vector;

[0050] S73. Calculate path similarity based on the directional consistency between the structural representation vector and the structural vector of the stored subgraph in the historical knowledge graph.

[0051] S74. Extract policy paths from historical subgraphs with similarity higher than a set threshold, and encode the policy paths into completion vectors.

[0052] S75. Load the completion vector into the corresponding task node position in the candidate solution group to generate an updated matching vector containing the policy path.

[0053] Optionally, S8 specifically includes:

[0054] S81. Determine whether the candidate solution group has not undergone structural changes in a consecutive preset number of rounds. If the invariance condition is met, then perform candidate solution state update.

[0055] S82. Construct causal paths between task nodes, extract edge types, connection order, and policy node influence factors between nodes, and form a causal graph structure.

[0056] S83. Constructing the reverse perturbation vector based on the cause-effect graph structure:

[0057]

[0058] Where, r i Let α be the reverse perturbation vector of task node i. ij Let s be the causal weight of task node j with respect to task node i. j With s iare the state vectors of task nodes j and i, respectively, and m is the number of upstream task nodes that have a causal path with node i.

[0059] S84. Apply the reverse perturbation vector to the corresponding position of the task node in the candidate solution state space, update the matching vector and mark it as the state to be updated iteratively.

[0060] The beneficial effects of this invention are:

[0061] This invention breaks down the barriers between multi-source information in data center operations by constructing a knowledge graph that integrates elements such as tasks, strategies, personnel, and equipment, and establishes a matching mechanism between tasks and strategies based on path structure and semantic relationships. In the initial stage of task scheduling, an initial task-strategy matching matrix is ​​generated by combining path connectivity features and semantic label similarity calculations, constraining the matching relationship from both structural and semantic levels, making candidate solutions more reasonable and targeted. Subsequently, a perturbation graph layer is introduced, introducing state variables and perturbation coefficients while retaining the original topology, constructing a node-level perturbation factor model, enabling candidate solution updates to adaptively adjust according to dynamic changes in the graph state.

[0062] During the matching optimization process, an improved Jaya algorithm is employed, enhancing the traditional solution vector update mechanism by integrating perturbation factors and confidence weights generated from knowledge graph inference. An asymmetric update strategy is introduced to differentiate control over different evolution directions, making the search process more robust and efficient. Simultaneously, the system supports retrieving similar subgraphs from the historical graph and loading policy paths based on the structural position of task nodes in the current candidate solution. This is used to complete matching items for structurally incomplete or semantically isolated task nodes, enhancing the expressive power of the matching vector. When the evolution of a candidate solution stagnates, a causal path is constructed and a reverse perturbation vector is generated, effectively breaking the local stable state and guiding a reasonable adjustment of the task-policy matching structure. The entire process achieves dynamic adaptation and closed-loop feedback while ensuring structural consistency and semantic accuracy, addressing the typical challenges of complex structures, dispersed semantics, and variable states in data center operation and maintenance tasks. Attached Figure Description

[0063] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0064] Figure 1 This is a flowchart of a knowledge graph-based automated management method for data center operations and maintenance proposed in this invention;

[0065] Figure 2 This is a schematic diagram of the knowledge graph construction structure for a knowledge graph-based automated management method for data center operations and maintenance proposed in this invention.

[0066] Figure 3 This is a schematic diagram of the asymmetric update process of candidate solutions for a knowledge graph-based automated management method for data center operations and maintenance proposed in this invention. Detailed Implementation

[0067] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0068] refer to Figure 1-3 A knowledge graph-based method for automated data center operation and maintenance management includes the following steps:

[0069] S1. Collect task, strategy, personnel and equipment information, perform structured processing, and construct task sets and strategy sets;

[0070] S2. Construct a knowledge graph based on task sets and policy sets;

[0071] S3. Generate an initial task-policy matching matrix based on the path structure, edge weights, and semantic similarity between task nodes and policy nodes, and construct a set of candidate solutions, where each candidate solution is a task-policy matching vector.

[0072] S4. Copy the knowledge graph to generate a perturbation graph layer, set the node state variables and edge perturbation coefficients, and construct the perturbation factor calculation structure;

[0073] S5. During the candidate solution update process, the perturbation factor is calculated based on the adjacency state and edge perturbation coefficient of the relevant nodes in the perturbation map layer.

[0074] S6. Input the perturbation factor and the task-policy matching initial matrix into the improved Jaya algorithm, load the confidence weights generated based on knowledge graph reasoning, perform asymmetric vector update, and generate the updated candidate solution set.

[0075] S7. Based on the structural position of the task nodes in the knowledge graph in the updated candidate solution set, match the historical subgraphs with similar structures, extract the strategy path and load it into the corresponding candidate solution.

[0076] S8. If the candidate solution set is not updated within a set number of rounds, construct the causal path between task nodes, generate the reverse perturbation vector, and update the candidate solution state.

[0077] S9. When the termination condition is met, output the task-policy matching vector with the highest fitness score.

[0078] This invention constructs a knowledge graph by integrating multi-dimensional information such as task information, policy rules, operation and maintenance personnel and equipment configuration, and fully preserves the dependency structure and constraint logic in the task execution process. Combined with path structure analysis and perturbation reasoning mechanism, it realizes accurate modeling of the dynamic and multi-factor relationship between tasks and policies, thereby improving the pertinence and adaptability of task-policy matching and adapting to the complex operation and maintenance scenarios of high concurrency and high coupling in data centers.

[0079] In this embodiment, the improved Jaya algorithm, compared to the original Jaya algorithm, includes:

[0080] A perturbation factor calculated from the perturbation map layer is introduced into the basic update vector to dynamically adjust the update direction and step size;

[0081] The state variables of the candidate solution associated nodes and the edge perturbation coefficients are jointly modeled as perturbation enhancement terms, and these perturbation enhancement terms are embedded during the update process.

[0082] We introduce confidence weights generated by knowledge graph reasoning to perform weighted corrections on perturbation enhancement terms, thereby improving the update magnitude of high-confidence task-policy matching pairs;

[0083] An asymmetric update mechanism is adopted to adjust the update ratios for solutions close to the optimal solution and solutions far from the worst solution, thereby enhancing convergence stability and local search capability.

[0084] The improved Jaya algorithm proposed in this invention constructs an enhanced update mechanism based on the perturbation factor and confidence weight calculated from the perturbation map layer. It not only dynamically adjusts the optimization direction, but also adopts an asymmetric update strategy to distinguish between candidate vectors that approach the optimal solution and those that are far from the worst solution. This significantly improves the search stability and local optimization capability in the task-policy matching process, and maintains high-precision matching efficiency in a multi-task, highly dynamic environment.

[0085] In this embodiment, the knowledge graph includes task nodes, strategy nodes, personnel nodes, and device nodes. The nodes are connected by edges to represent dependencies, execution order, and resource constraints.

[0086] The knowledge graph constructed by this invention covers task nodes, policy nodes, personnel nodes, and equipment nodes. It uses edge relationships to describe task execution constraints, resource allocation, and policy control flow, making the entity and rule information in the operation and maintenance scenario concrete and operable in a unified graph structure. This provides a structural basis for subsequent path analysis, reasoning, and disturbance modeling, and enhances the expressive integrity and computability of the overall modeling system.

[0087] In this embodiment, S3 specifically includes:

[0088] S31. Extract the path structure between task nodes and strategy nodes in the knowledge graph, and count the path length, number of transit nodes and type of connection edges.

[0089] S32. Calculate the path weight score based on the weight of all edges in the path, and set a threshold to filter high-weight paths;

[0090] S33. Based on the semantic tags of task nodes and policy nodes, use an embedding model to generate semantic vectors and calculate the semantic similarity between task nodes and policy nodes.

[0091] S34. Normalize the path weight score and semantic similarity score, then perform a weighted fusion to construct the initial task-policy matching matrix M, where the matching score M... ij The calculation method is as follows:

[0092]

[0093] Where λ1 and λ2 are weight adjustment coefficients, satisfying λ1 + λ2 = 1, |P ij | represents the number of valid paths between task i and strategy j, w k δ represents the weight of path k. ij Indicates the semantic similarity between task nodes and policy nodes;

[0094] S35. Construct a corresponding matching vector for each row in the initial matching matrix to form a candidate solution set. The candidate solution is a set of correspondences between the task set and the policy set, and the dimension is equal to the size of the task set.

[0095] This invention combines path structure scoring and semantic similarity scoring mechanisms to construct an initial matrix for task-policy matching. It employs an embedded model to encode the semantic relationship between the task and the policy, integrates structural information such as path length, edge weight, and node type, and constructs the matrix through a normalized weighting strategy. This ensures that the candidate solution set has semantic rationality and structural feasibility, laying the foundation for perturbation updates and path completion.

[0096] In this embodiment, S4 specifically includes:

[0097] S41. Copy the constructed knowledge graph, retain the topological connections of task nodes, policy nodes, personnel nodes and device nodes, and construct a perturbation graph layer with the same structure;

[0098] S42. Set a state variable for each node in the disturbance map layer, wherein the state variable represents information about task execution status, strategy matching status or device working status.

[0099] S43. Set a perturbation coefficient for each edge in the perturbation graph. The perturbation coefficient is used to quantify the perturbation sensitivity between the states of connected nodes.

[0100] S44. Construct the perturbation factor calculation structure, and define the perturbation factor of each node as a weighted average of the perturbation weights and state correlations between it and all its neighboring nodes:

[0101]

[0102] Where, θ i d represents the perturbation factor of node i. i This represents the number of adjacent nodes of node i. Let η represent the set of adjacent nodes of node i. ij ρ(s) represents the perturbation coefficient between node i and its neighboring node j. i ,s j ) represents the correlation function between the state variables of node i and node j;

[0103] S45. Bind the perturbation factor calculation structure to the perturbation map layer to form a composite layer containing structure, state and perturbation factors, which is used to support the dynamic update process of the candidate solution population.

[0104] This invention constructs a perturbation graph layer by replicating the knowledge graph, sets node state variables and edge perturbation coefficients, and designs a perturbation factor calculation structure. It integrates state correlation and structural perturbation sensitivity into a perturbation factor, which drives candidate solutions to make local convergence or jump-diffusion adjustments in the semantic space of the knowledge graph, supporting the dynamic evolution process of the candidate solution group.

[0105] In this embodiment, S6 specifically includes:

[0106] S61. Jointly encode the perturbation factor matrix and the initial task-policy matching matrix to generate a matching candidate tensor with perturbation structure, which serves as the input representation of the candidate solution population;

[0107] S62. Load the confidence weight vector generated by the knowledge graph reasoning results. The confidence weight vector assigns a confidence scalar to each task-policy pair for use in fitness updates.

[0108] S63. Construct an improved asymmetric update mechanism for the Jaya algorithm, and adjust the update ratios of candidate solutions in directions close to the optimal solution and in directions far from the worst solution respectively.

[0109] S64. Define the objective function for asymmetric updates as follows:

[0110]

[0111] in, This represents the matching vector of the i-th candidate solution in round t+1. Let x represent the matching vector of the i-th candidate solution in round t. best Let x represent the optimal candidate solution vector. worst Let r1 and r2 be random variables between 0 and 1, α and β be scaling factors controlling the positive and negative update directions, and θ be the worst candidate solution vector. i Let w represent the perturbation factor of the node corresponding to the i-th candidate solution, δ be the perturbation enhancement coefficient, and w be the perturbation factor of the node corresponding to the i-th candidate solution. i This is the confidence weight vector;

[0112] S65. After performing update calculations for each candidate solution, select individuals with improved fitness scores to generate an updated candidate solution population.

[0113] This invention constructs an input representation tensor with a perturbation structure during the candidate solution update stage, combines it with the confidence weights generated by knowledge graph reasoning, inputs the asymmetric update mechanism of the improved Jaya algorithm, uses the perturbation enhancement coefficient to guide the update direction, designs a matching vector update formula, distinguishes between the optimal and worst solution paths, achieves priority matching of high confidence directions, and enhances the convergence speed and accuracy of candidate solutions.

[0114] In this embodiment, S7 specifically includes:

[0115] S71. Traverse each task node in the candidate solution group, extract the structural position of the task node in the knowledge graph, including the node's connection path, adjacency strategy node identifier and edge type, and construct a set of structural paths.

[0116] S72. Perform graph embedding encoding on each structural path to generate a structural representation vector;

[0117] S73. Calculate path similarity based on the directional consistency between the structural representation vector and the structural vector of the stored subgraph in the historical knowledge graph.

[0118] S74. Extract strategy paths from historical subgraphs with similarity higher than a set threshold, and encode the strategy paths into completion vectors.

[0119] S75. Load the completion vector into the corresponding task node position in the candidate solution group to generate an updated matching vector containing the policy path.

[0120] This invention extracts the structural path information of task nodes in the knowledge graph from candidate solutions, encodes it to generate structural representation vectors, calculates the similarity between the vectors and historical knowledge subgraphs, loads and completes the strategy paths of subgraphs with high similarity, realizes the local reconstruction of the matching content of task nodes, makes up for the structural omissions in the initial matching, and enhances the structural consistency and semantic integrity of the matching vectors.

[0121] In this embodiment, S8 specifically includes:

[0122] S81. Determine whether the candidate solution group has not undergone structural changes in a consecutive preset number of rounds. If the invariance condition is met, then perform candidate solution state update.

[0123] S82. Construct causal paths between task nodes, extract edge types, connection order, and policy node influence factors between nodes, and form a causal graph structure.

[0124] S83. Constructing the reverse perturbation vector based on the cause-effect graph structure:

[0125]

[0126] Where, r i Let α be the reverse perturbation vector of task node i. ij Let s be the causal weight of task node j with respect to task node i. j With s i are the state vectors of task nodes j and i, respectively, and m is the number of upstream task nodes that have a causal path with node i.

[0127] S84. Apply the reverse perturbation vector to the corresponding position of the task node in the candidate solution state space, update the matching vector and mark it as the state to be updated iteratively.

[0128] This invention detects structural changes in candidate solutions during multiple updates to determine whether they have entered a stagnant state. When conditions are met, it constructs causal paths between task nodes and builds a reverse perturbation vector based on the causal weights and state changes in the path. It then performs reverse state reasoning and correction operations to guide candidate solutions out of local optimal regions, enabling targeted escape and restoration of iteration potential, thus maintaining the activity of the overall search process.

[0129] Example 1:

[0130] To verify the feasibility of this invention in practice, it was applied to the main data center of a large financial enterprise. This data center has frequent daily operation and maintenance tasks, involving many types of tasks, complex personnel and equipment distribution, and long policy response chains. Traditional scheduling relies on manually formulated rules and static scheduling strategies, which is not only inefficient, but also slow to respond to sudden tasks or personnel changes, which seriously restricts the timeliness and accuracy of operation and maintenance response.

[0131] This method first collects the data center's operation and maintenance task logs, predefined policy entries, on-duty personnel rotation records, and various equipment operation status data for the past three months. All information is then structured to construct task sets and policy sets. Based on this, a knowledge graph containing task nodes, policy nodes, personnel nodes, and equipment nodes is constructed, and the dependencies, resource constraints, and execution order between nodes are defined to ensure that the graph structure is complete and the semantic relationships are clear.

[0132] Subsequently, based on the path structure, path weights, and semantic similarity between task nodes and policy nodes in the graph, we constructed an initial task-policy matching matrix. The matrix size is 326×129, corresponding to 326 tasks to be executed and 129 optional policies. By weighted fusion calculation of path weights and semantic similarity, the system generates an initial matching vector set for subsequent optimization.

[0133] During the matching optimization process, the system replicates the knowledge graph to generate a perturbation graph layer, configures state variables for each node, and sets perturbation coefficients for connecting edges. By calculating the structure through perturbation factors, the system dynamically captures the impact of node state changes on the matching results. Furthermore, the improved Jaya algorithm introduces perturbation enhancement terms and confidence weights to achieve an asymmetric update mechanism, gradually filtering out better matching solutions. During the optimization iteration process, if a candidate solution does not undergo structural changes in ten consecutive rounds, the system automatically activates the reverse perturbation mechanism. Through causal path analysis between task nodes, a reverse perturbation vector is generated to escape local convergence states and explore a better solution space.

[0134] During the experiment, we selected three task batches and compared the traditional rule matching method with the method described in this invention. Each task batch contained 30 to 50 task items. The test indicators included average matching time, strategy execution success rate, and fault warning response time, etc. The results are shown in Table 1:

[0135] Table 1 Comparison of different methods in data center operation and maintenance tasks

[0136]

[0137] As shown in Table 1, the method of this invention significantly outperforms the traditional rule matching method in all three batches. Taking batch A as an example, the average matching time decreased from 38.7 seconds to 14.2 seconds, a reduction of nearly two-thirds, indicating that the system can complete decision generation faster when handling complex task matching. Regarding the strategy execution success rate, this method improved to 94.6%, a significant advantage over the traditional method's 85.3%, indicating more accurate task-strategy matching and more stable strategy invocation. The fault warning response time was also reduced from 27.6 seconds in the traditional method to 13.4 seconds, almost halved, effectively improving the system's speed in handling sudden anomalies.

[0138] Similar trends were observed in batches B and C. In batch B, the response time decreased from 30.1 seconds to 12.9 seconds, and the success rate increased from 82.7% to 92.1%. In batch C, the execution success rate further improved to 95.8%, and the matching time remained consistently below 15 seconds. These data collectively demonstrate that this method maintains high stability and processing efficiency under different task load conditions, not only improving task execution quality but also significantly reducing response time, adapting to the complex and ever-changing operational needs of data centers.

[0139] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A knowledge graph-based automated management method for data center operation and maintenance, characterized in that, Includes the following steps: S1. Collect task, strategy, personnel and equipment information, perform structured processing, and construct task sets and strategy sets; S2. Construct a knowledge graph based on task sets and policy sets; S3. Generate an initial task-policy matching matrix based on the path structure, edge weights, and semantic similarity between task nodes and policy nodes, and construct a set of candidate solutions, where each candidate solution is a task-policy matching vector. S4. Copy the knowledge graph to generate a perturbation graph layer, set the node state variables and edge perturbation coefficients, and construct the perturbation factor calculation structure; S5. During the candidate solution update process, the perturbation factor is calculated based on the adjacency state and edge perturbation coefficient of the relevant nodes in the perturbation map layer. S6. Input the perturbation factor and the task-policy matching initial matrix into the improved Jaya algorithm, load the confidence weights generated based on knowledge graph reasoning, perform asymmetric vector update, and generate the updated candidate solution set. S7. Based on the structural position of the task nodes in the knowledge graph in the updated candidate solution set, match the historical subgraphs with similar structures, extract the strategy path and load it into the corresponding candidate solution. S8. If the candidate solution set is not updated within a set number of rounds, construct the causal path between task nodes, generate the reverse perturbation vector, and update the candidate solution state. S9. When the termination condition is met, output the task-policy matching vector with the highest fitness score.

2. The data center operation and maintenance automation management method based on knowledge graphs according to claim 1, characterized in that, The improved Jaya algorithm, compared to the original Jaya algorithm, includes the following: A perturbation factor calculated from the perturbation map layer is introduced into the basic update vector to dynamically adjust the update direction and step size; The state variables of the candidate solution associated nodes and the edge perturbation coefficients are jointly modeled as perturbation enhancement terms, and these perturbation enhancement terms are embedded during the update process. We introduce confidence weights generated by knowledge graph reasoning to perform weighted corrections on perturbation enhancement terms, thereby improving the update magnitude of high-confidence task-policy matching pairs; An asymmetric update mechanism is adopted to adjust the update ratios for solutions close to the optimal solution and solutions far from the worst solution, thereby enhancing convergence stability and local search capability.

3. The data center operation and maintenance automation management method based on knowledge graphs according to claim 1, characterized in that, The knowledge graph includes task nodes, strategy nodes, personnel nodes, and device nodes. The nodes are connected by edges to represent dependencies, execution order, and resource constraints.

4. The data center operation and maintenance automation management method based on knowledge graphs according to claim 1, characterized in that, S3 specifically includes: S31. Extract the path structure between task nodes and strategy nodes in the knowledge graph, and count the path length, number of transit nodes and type of connection edges. S32. Calculate the path weight score based on the weight of all edges in the path, and set a threshold to filter high-weight paths; S33. Based on the semantic tags of task nodes and policy nodes, use an embedding model to generate semantic vectors and calculate the semantic similarity between task nodes and policy nodes. S34. Normalize the path weight score and semantic similarity score, then perform a weighted fusion to construct the initial task-policy matching matrix M, where the matching score M... ij The calculation method is as follows: Where λ1 and λ2 are weight adjustment coefficients, satisfying λ1 + λ2 = 1, |P ij | represents the number of valid paths between task i and strategy j, w k δ represents the weight of path k. ij Indicates the semantic similarity between task nodes and policy nodes; S35. Construct a corresponding matching vector for each row in the initial matching matrix to form a candidate solution set. The candidate solution is a set of correspondences between the task set and the policy set, and the dimension is equal to the size of the task set.

5. The data center operation and maintenance automation management method based on knowledge graphs according to claim 1, characterized in that, S4 specifically includes: S41. Copy the constructed knowledge graph, retain the topological connections of task nodes, policy nodes, personnel nodes and device nodes, and construct a perturbation graph layer with the same structure; S42. Set a state variable for each node in the disturbance map layer, wherein the state variable represents information about task execution status, strategy matching status or device working status. S43. Set a perturbation coefficient for each edge in the perturbation graph. The perturbation coefficient is used to quantify the perturbation sensitivity between the states of connected nodes. S44. Construct the perturbation factor calculation structure, and define the perturbation factor of each node as a weighted average of the perturbation weights and state correlations between it and all its neighboring nodes: Where, θ i d represents the perturbation factor of node i. i This represents the number of adjacent nodes of node i. Let η represent the set of adjacent nodes of node i. ij ρ(s) represents the perturbation coefficient between node i and its neighboring node j. i ,s j ) represents the correlation function between the state variables of node i and node j; S45. Bind the perturbation factor calculation structure to the perturbation map layer to form a composite layer containing structure, state and perturbation factors, which is used to support the dynamic update process of the candidate solution population.

6. The data center operation and maintenance automation management method based on knowledge graphs according to claim 1, characterized in that, S6 specifically includes: S61. Jointly encode the perturbation factor matrix and the initial task-policy matching matrix to generate a matching candidate tensor with perturbation structure, which serves as the input representation of the candidate solution population; S62. Load the confidence weight vector generated by the knowledge graph reasoning results. The confidence weight vector assigns a confidence scalar to each task-policy pair for use in fitness updates. S63. Construct an improved asymmetric update mechanism for the Jaya algorithm, and adjust the update ratios of candidate solutions in directions close to the optimal solution and in directions far from the worst solution respectively. S64. Define the objective function for asymmetric updates as follows: in, This represents the matching vector of the i-th candidate solution in round t+1. Let x represent the matching vector of the i-th candidate solution in round t. best Let x represent the optimal candidate solution vector. worst Let r1 and r2 be random variables between 0 and 1, α and β be scaling factors controlling the positive and negative update directions, and θ be the worst candidate solution vector. i Let w represent the perturbation factor of the node corresponding to the i-th candidate solution, δ be the perturbation enhancement coefficient, and w be the perturbation factor of the node corresponding to the i-th candidate solution. i This is the confidence weight vector; S65. After performing update calculations for each candidate solution, select individuals with improved fitness scores to generate an updated candidate solution population.

7. The data center operation and maintenance automation management method based on knowledge graphs according to claim 1, characterized in that, Specifically, S7 includes: S71. Traverse each task node in the candidate solution group, extract the structural position of the task node in the knowledge graph, including the node's connection path, adjacency strategy node identifier and edge type, and construct a set of structural paths. S72. Perform graph embedding encoding on each structural path to generate a structural representation vector; S73. Calculate path similarity based on the directional consistency between the structural representation vector and the structural vector of the stored subgraph in the historical knowledge graph. S74. Extract strategy paths from historical subgraphs with similarity higher than a set threshold, and encode the strategy paths into completion vectors. S75. Load the completion vector into the corresponding task node position in the candidate solution group to generate an updated matching vector containing the policy path.

8. The data center operation and maintenance automation management method based on knowledge graph according to claim 1, characterized in that, S8 specifically includes: S81. Determine whether the candidate solution group has not undergone structural changes in a consecutive preset number of rounds. If the invariance condition is met, then perform candidate solution state update. S82. Construct causal paths between task nodes, extract edge types, connection order, and policy node influence factors between nodes, and form a causal graph structure. S83. Constructing the reverse perturbation vector based on the cause-effect graph structure: Where, r i Let α be the reverse perturbation vector of task node i. ij Let s be the causal weight of task node j with respect to task node i. j With s i are the state vectors of task nodes j and i, respectively, and m is the number of upstream task nodes that have a causal path with node i. S84. Apply the reverse perturbation vector to the corresponding position of the task node in the candidate solution state space, update the matching vector and mark it as the state to be updated iteratively.

Images