Knowledge graph-based power communication maintenance conflict identification method and device

Abstract

The application discloses a power communication maintenance conflict identification method and device based on a knowledge graph, relates to the field of information technology, and mainly aims at improving the identification efficiency and accuracy of power communication maintenance conflicts. The method comprises the following steps: obtaining power communication maintenance text information; extracting entities and mining entity relationships between the entities in the power communication maintenance text information; constructing a power communication maintenance knowledge graph based on the entities and the entity relationships, wherein each entity serves as a node in the power communication maintenance knowledge graph, and the entity relationships serve as edges between the nodes in the power communication maintenance knowledge graph; and identifying power communication maintenance conflicts based on the power communication maintenance knowledge graph. The application is suitable for the scene of identifying power communication maintenance conflicts.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method and apparatus for identifying power communication maintenance conflicts based on knowledge graphs. Background Technology

[0002] Power communication networks are a major component of the power system, crucial for power transmission and dispatch, and closely related to people's daily lives. With rapid economic development, electricity demand is increasing. To meet daily power transmission and supply needs, power communication networks are becoming increasingly large and complex. Consequently, the difficulty of power communication network maintenance is also increasing. Currently, a major problem in power communication maintenance is the potential for various types of maintenance conflicts. Therefore, to ensure the smooth progress of power maintenance, it is necessary to identify maintenance conflicts in advance.

[0003] Currently, maintenance conflicts are typically identified through manual review of maintenance tickets. However, this manual review method is time-consuming and labor-intensive, resulting in low efficiency in identifying power and communication maintenance conflicts. At the same time, due to staff negligence or varying levels of technical expertise, the accuracy of power and communication maintenance conflict identification can be low. Summary of the Invention

[0004] This invention provides a knowledge graph-based method and apparatus for identifying power communication maintenance conflicts, which mainly improves the efficiency and accuracy of power communication maintenance conflict identification.

[0005] According to a first aspect of the present invention, a knowledge graph-based method for identifying power communication maintenance conflicts is provided, comprising:

[0006] Obtain power communication maintenance text information;

[0007] Entities are extracted from the power communication maintenance text information, and the relationships between the entities are mined. Based on the entities and their relationships, a power communication maintenance knowledge graph is constructed, wherein each entity is a node in the power communication maintenance knowledge graph, and the entity relationships are edges between the nodes in the power communication maintenance knowledge graph.

[0008] Based on the aforementioned power communication maintenance knowledge graph, power communication maintenance conflicts are identified.

[0009] Optionally, entities are extracted from the power communication maintenance text information, including:

[0010] Obtain a preset entity extraction model, wherein the preset entity extraction model includes a text encoder for extracting feature vectors, a deep self-attention network for learning context-dependent vector representations, and a conditional random field for labeling; the preset entity extraction model is pre-trained based on a sample text dataset.

[0011] The power communication maintenance text information is input into the preset entity extraction model. The feature vector of the power communication maintenance text information is extracted by the text encoder to obtain the feature vector. The text context features are extracted by the deep self-attention network to obtain the context features. The context features are then labeled by the conditional random field to obtain the sequence label. The sequence label contains each entity and its corresponding entity category.

[0012] Optionally, mining the entity relationships between the entities includes:

[0013] A preset relationship recognition model is obtained, wherein the preset relationship recognition model includes an input representation layer for information embedding, a low-level feature extraction layer for low-level feature extraction, a dynamic routing capsule network feature clustering layer with word attention mechanism for global relationship feature extraction, and a relationship prediction layer for entity relationship prediction; the preset relationship recognition model is pre-trained based on an entity text dataset.

[0014] The power communication maintenance text information and each entity are input into the preset relationship recognition model. The input representation layer performs word embedding, character embedding, part-of-speech embedding, and word position embedding on the power communication maintenance text information, and then fuses the word embedding, character embedding, part-of-speech embedding, and word position embedding to obtain fused embedding. The low-level feature extraction layer extracts low-level features from the fused embedding to obtain low-level features. The dynamic routing capsule network feature clustering layer of the word attention mechanism extracts features from the low-level features to obtain global relationship features. The relationship prediction layer predicts relationships between the entities by using the global relationship features.

[0015] Optionally, the low-level feature extraction layer includes a bidirectional long short-term memory neural network, which includes a forward network layer and a backward network layer;

[0016] The step of extracting low-level features from the fused embedding through the low-level feature extraction layer to obtain low-level features includes:

[0017] The fused embedding is input into the bidirectional long short-term memory neural network for global sequence feature extraction, to obtain the forward global sequence features corresponding to the forward network layer and the backward global sequence features corresponding to the backward network layer. The forward global sequence features and the backward global sequence features are fused to obtain the low-level features.

[0018] Optionally, the dynamic routing capsule network feature clustering layer of the word attention mechanism includes: a low-level capsule network, a word attention mechanism, and a high-level capsule network;

[0019] The dynamic routing capsule network feature clustering layer using the word attention mechanism extracts features from the low-level features to obtain global relation features, including:

[0020] The low-level features are input into the low-level capsule network for feature extraction to obtain local features. The local features are then input into the word attention mechanism for weight allocation to obtain weighted local features. Finally, the weighted local features are input into the high-level capsule network for feature extraction to obtain global relation features.

[0021] Optionally, the step of identifying power communication maintenance conflicts based on the power communication maintenance knowledge graph includes:

[0022] Based on the power communication maintenance knowledge graph, at least one of the following conflict identification methods is used for power communication maintenance: time conflict identification, resource conflict identification, and logical conflict identification.

[0023] After identifying power communication maintenance conflicts based on the power communication maintenance knowledge graph, the method further includes:

[0024] Early warnings are issued for identified power communication maintenance conflicts.

[0025] Optionally, before identifying power communication maintenance conflicts based on the power communication maintenance knowledge graph, the method further includes:

[0026] The nodes and edges in the power communication maintenance knowledge graph are converted into vector representations. A multi-head attention mechanism is used to calculate the vector representation of the power communication maintenance knowledge graph. In the multi-head attention mechanism, multiple attention heads are set, and each head focuses on and calculates the vectors of the power communication maintenance knowledge graph from different perspectives. The outputs of the multiple attention heads are concatenated and linearly transformed to obtain a power communication maintenance knowledge graph with attention weights.

[0027] According to a second aspect of the present invention, a knowledge graph-based power communication maintenance conflict identification device is provided, comprising:

[0028] The acquisition unit is used to acquire power communication maintenance text information;

[0029] The construction unit is used to extract entities and mine entity relationships between entities from the power communication maintenance text information, and construct a power communication maintenance knowledge graph based on the entities and entity relationships, wherein each entity is a node in the power communication maintenance knowledge graph, and the entity relationship is an edge between the nodes in the power communication maintenance knowledge graph.

[0030] The identification unit is used to identify power communication maintenance conflicts based on the power communication maintenance knowledge graph.

[0031] According to a third aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described knowledge graph-based power communication maintenance conflict identification method.

[0032] According to a fourth aspect of the present invention, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described knowledge graph-based power communication maintenance conflict identification method.

[0033] According to the present invention, a method and apparatus for identifying power communication maintenance conflicts based on a knowledge graph are provided. Compared with the current method of determining maintenance conflicts by manually reviewing maintenance tickets, the present invention obtains power communication maintenance text information, extracts entities from the power communication maintenance text information and mines the entity relationships between the entities, and constructs a power communication maintenance knowledge graph based on the entities and the entity relationships. Each entity is a node in the power communication maintenance knowledge graph, and the entity relationships are edges between the nodes in the power communication maintenance knowledge graph. Finally, power communication maintenance conflicts are identified based on the power communication maintenance knowledge graph. Therefore, by constructing a power communication maintenance knowledge graph using structured text information from power communication, and combining it with an attention mechanism to improve conflict identification efficiency, the power communication maintenance conflict can be identified based on the power communication maintenance knowledge graph without human intervention, i.e., without manually checking each maintenance order. This improves the automation and efficiency of power communication maintenance conflict identification. At the same time, the knowledge graph can accurately identify complex entity relationships, such as dependencies between equipment and overlapping maintenance times. Humans may make mistakes and oversights when identifying these complex relationships. Thus, this invention can improve the accuracy of power communication maintenance conflict identification. Attached Figure Description

[0034] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0035] Figure 1 The flowchart of a knowledge graph-based power communication maintenance conflict identification method provided by an embodiment of the present invention is shown.

[0036] Figure 2 This diagram illustrates the structure of a preset relationship recognition model provided in an embodiment of the present invention.

[0037] Figure 3 The flowchart of another knowledge graph-based power communication maintenance conflict identification method provided by an embodiment of the present invention is shown.

[0038] Figure 4 A schematic diagram of a knowledge graph-based power communication maintenance conflict identification device provided in an embodiment of the present invention is shown.

[0039] Figure 5 A schematic diagram of another knowledge graph-based power communication maintenance conflict identification device provided in an embodiment of the present invention is shown.

[0040] Figure 6 A schematic diagram of the physical structure of a computer device provided in an embodiment of the present invention is shown. Detailed Implementation

[0041] The present invention will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the present application can be combined with each other.

[0042] Currently, manually reviewing maintenance tickets to determine maintenance conflicts is time-consuming and labor-intensive. Furthermore, due to staff negligence or varying levels of technical expertise, the accuracy of identifying power and communication maintenance conflicts can be low.

[0043] To address the aforementioned problems, embodiments of the present invention provide a knowledge graph-based method for identifying power communication maintenance conflicts, such as... Figure 1 As shown, the method includes:

[0044] 101. Obtain power communication maintenance text information.

[0045] Specifically, images of documents such as power communication maintenance orders can be converted into power communication maintenance text information. These images can be obtained by scanning paper documents and forms related to power maintenance.

[0046] 102. Extract entities and mine the entity relationships between entities from the power communication maintenance text information. Based on the entities and entity relationships, construct a power communication maintenance knowledge graph, where each entity is a node in the power communication maintenance knowledge graph, and entity relationships are edges between nodes in the power communication maintenance knowledge graph.

[0047] In this embodiment of the invention, after obtaining power communication maintenance text information, it is necessary to construct a power communication maintenance knowledge graph based on the text information. Before constructing the power communication maintenance knowledge graph, it is necessary to extract entities from the power communication maintenance text information. Based on this, step 102 specifically includes: obtaining a preset entity extraction model, wherein the preset entity extraction model includes a text encoder for extracting feature vectors, a deep self-attention network for learning context-related vector representations, and a conditional random field for labeling; the preset entity extraction model is pre-trained based on a sample text dataset; the power communication maintenance text information is input into the preset entity extraction model, the feature vectors of the power communication maintenance text information are extracted by the text encoder to obtain feature vectors, the text context features are extracted by the deep self-attention network to obtain context features, and the context features are labeled by the conditional random field to obtain sequence labeling labels, wherein the sequence labeling labels contain each entity and its corresponding entity category.

[0048] The pre-defined entity extraction model consists of a text encoder, a deep self-attention network, and a conditional random field (CRF). To improve its extraction accuracy, the model needs to be trained and constructed. First, an initial model is built by acquiring a dataset of sample power communication maintenance texts containing entity labels. This dataset is divided into training and testing data. The initial model is trained using the training data and then tested using the testing data. The model that meets the testing conditions is the pre-defined entity extraction model. This model can extract key entities and attributes from power communication maintenance texts. Specifically, the deep self-attention network is used to understand the textual context. Trained on a large amount of text data, the model has a large capacity and strong expressive power, performing well in downstream tasks. Within the named entity recognition framework of the text encoder and CRF, the text encoder produces features, the deep self-attention network learns context vectors, and finally, the CRF generates sequence labels. Entities in the power communication maintenance text can be extracted based on these labels. The sequence labels annotate each entity and its corresponding entity category.

[0049] Furthermore, after extracting the entities, in order to construct a knowledge graph, it is also necessary to mine the entity relationships between the entities. Based on this, step 102 includes: obtaining a preset relationship recognition model, wherein the preset relationship recognition model includes an input representation layer for information embedding, a low-level feature extraction layer for low-level feature extraction, a dynamic routing capsule network feature clustering layer with word attention mechanism for global relationship feature extraction, and a relationship prediction layer for entity relationship prediction; the preset relationship recognition model is pre-trained based on an entity text dataset; the power communication maintenance text information and each entity are input into the preset... In the relationship recognition model, the power communication maintenance text information is embedded with words, characters, parts of speech, and word positions through the input representation layer, and these embeddings are fused to obtain a fused embedding. The fused embedding is then extracted using the low-level feature extraction layer to obtain low-level features. The low-level features are then extracted using the dynamic routing capsule network feature clustering layer of the word attention mechanism to obtain global relationship features. Finally, the global relationship features are predicted using the relationship prediction layer to obtain the entity relationships between the entities. The low-level feature extraction layer includes a bidirectional long short-term memory neural network, which comprises a forward network layer and a backward network layer. The method for extracting low-level features from the fused embedding through the low-level feature extraction layer includes: inputting the fused embedding into the bidirectional long short-term memory neural network for global sequence feature extraction, obtaining forward global sequence features corresponding to the forward network layer and backward global sequence features corresponding to the backward network layer; and fusing the forward global sequence features and the backward global sequence features to obtain the low-level features. In another embodiment of the invention, the dynamic routing capsule network feature clustering layer of the word attention mechanism includes: a low-level capsule network, a word attention mechanism, and a high-level capsule network. The method for extracting global relational features from the low-level features through the dynamic routing capsule network feature clustering layer of the word attention mechanism includes: inputting the low-level features into the low-level capsule network for feature extraction, obtaining local features; inputting the local features into the word attention mechanism for weight allocation, obtaining weighted local features; and inputting the weighted local features into the high-level capsule network for feature extraction, obtaining global relational features.

[0050] In this embodiment of the invention, to improve the recognition accuracy of the preset relationship recognition model, it is first necessary to train and construct the preset relationship recognition model. Based on this, the method includes: constructing an initial relationship recognition model and obtaining an entity text dataset, wherein the entity text dataset includes sample power communication maintenance text information with relationship labels and each sample entity in the sample power communication maintenance text information; dividing the entity text dataset into training data and test data, using the training data to train the initial relationship recognition model, and using the test data to test the trained initial relationship recognition model, finally determining the initial relationship recognition model that meets the test conditions as the preset relationship recognition model. Further, entity relationship mining refers to extracting the relationship between two entities in text information and ultimately constructing a graph. In this embodiment of the invention, the main identification includes the association between maintenance equipment and maintenance type, and the association between affecting optical paths and interrupting services. The specific implementation method includes: to effectively identify multiple relationships from power communication maintenance text information and solve the problem of keyword sparsity, a capsule network based on a word attention mechanism is proposed to realize knowledge relationship recognition. The structure of the preset relationship recognition model is as follows: Figure 2 As shown, the proposed model is a variation of the original capsule network. The model architecture consists of four layers: an input representation layer, a low-level feature extraction layer, a dynamic routing capsule network feature clustering layer based on word attention mechanism, and a relation prediction layer. Figure 2 In this context, WE and CE represent word embedding and character embedding, respectively; POS represents the part of speech of the word; P represents the position of the word in the sentence; and U1, U2, ..., U... LS1, S2, S3... represent the outputs of the low-level capsule network (features of each entity), respectively. S1, S2, S3... represent the features of each entity after noise reduction, respectively. V1, V2, V3... represent the outputs of the feature clustering layer of the dynamic routing capsule network based on word attention mechanism, respectively. Squashing is a specific function mainly used to perform nonlinear transformation on the capsule outputs. Y1, Y2, Y3... represent the outputs of the relation prediction layer, respectively. Input Representation Layer: The input representation of the model is a combination of four embeddings: character embedding, word embedding, part-of-speech embedding, and position embedding, to obtain the semantic information of each character and word, as well as the context information of the sentence and the word order information of the words in the sentence. Low-Level Feature Extraction Layer: For the word representation of the input layer, global sequence information is obtained through the feedforward and feedback networks of the bidirectional long short-term memory neural network to capture the low-level features of the sentence. For example, if the input sentence describes the maintenance equipment information of the power communication system, this step can better extract the various attributes of this equipment. The dynamic routing capsule network feature clustering layer based on word attention mechanism: Employs a capsule network with dynamic routing and word attention to reduce noise from irrelevant words, enabling the extraction of high-level features and capturing multiple relationships between entities. For example, if the input contains multiple entities and their characteristics, such as maintenance equipment and its attributes in a power communication system, and maintenance types, this step effectively captures the relationships between other maintenance equipment and maintenance types, achieving the capture of multiple relationships between entities. Relationship prediction layer: Utilizes a marginal loss function to predict relationships between entities. Specifically, the method for entity relationship mining using a pre-defined relationship recognition model is as follows: Based on low-level feature extraction: The low-level feature extraction layer can be a bidirectional long short-term memory (LSTM) neural network, which is effective for modeling sequence data in relationship extraction tasks. Therefore, a bidirectional LSTM neural network algorithm containing forward and backward networks is used to obtain global sequence information, thereby acquiring the low-level features of the sentence. In the relationship extraction task, firstly, the embedding combination of the input representation layer is input into the bidirectional LSTM neural network. Then, the forward and backward networks of the bidirectional LSTM neural network are used simultaneously to obtain global sequence features. Finally, the global sequence feature t is used as the output of the bidirectional long short-term memory neural network and as the input of the first layer of the capsule network. Rich feature extraction based on capsule networks: To further obtain the relevance of candidate relations and entities and address the problem of multiple relations in sentences, a capsule network framework for multiple relation extraction is proposed. The capsule network uses neuron vectors trained by a dynamic routing algorithm to train the model. The capsule outputs a vector to richly represent the features extracted by the node. Different dimensions in the vector are used to record different attributes of the same feature, improving text expressiveness. Each directional capsule vector can effectively capture the implicit syntactic, semantic, and word order features of the text. The capsule network is used to extract the relevance features of relations and entities.Capsule networks divide feature extraction into low-level capsules and high-level capsules. Low-level capsules contain local and trivial features, while high-level capsules are aggregated from low-level capsules. Active capsules use a transformation matrix to predict the instantiation parameters of high-level capsules; when multiple predictions are identical, higher-level capsules are activated. Low-level semantic information captured by a bidirectional long short-term memory neural network is input into the first-level capsule, and the current-level capsule uses a non-linear squeezing function to compress the length of each low-level capsule. A dynamic routing capsule network feature clustering layer based on word attention is proposed: using capsule networks, high-level capsules representing relational features can be obtained. High-level capsule networks can express relational features more richly. However, due to the long source sequences in power communication maintenance text information, noise from irrelevant words leads to significant gradient decay, making it difficult to effectively update sequence parameters. A word attention mechanism is proposed to be introduced into the dynamic routing of capsule networks to extract information words related to the relationships between entities in the sentence, and to aggregate the representations of these information words into sentence vectors, thereby solving the sparsity problem of information words in long sentences. The goal of the word attention mechanism is to ignore irrelevant words and focus on information words, with the weight coefficient of information words being greater than that of irrelevant words. Word attention mechanisms assign different weights to different words, with higher weights for informative words and lower weights for irrelevant words, thus increasing the strength of informative words. Utilizing word attention mechanisms, by combining local information at a higher level through local perception, helps reduce useless information in relation extraction. A dynamic routing algorithm based on word attention mechanisms is designed to further capture contextual relevance and important features, reducing the attenuation of useful information in long sentences, and iteratively correcting connection strength. Word attention mechanisms enable the decoder to obtain important input information to address the keyword sparsity problem. In this embodiment of the invention, the power communication maintenance text information and each entity are input into the preset relation recognition model. Through the input representation layer, low-level feature extraction layer, dynamic routing capsule network feature clustering layer based on word attention mechanisms, and feature extraction in the relation prediction layer of the preset relation recognition model, entity relations can be mined. For example, entity relationships include carrying relationships, which describe the relationship between optical paths and optical cables. For instance, optical path A was originally carried by optical cable B, and during maintenance, it was rerouted to a transitional optical cable from point A to point B to point C to point D. Temporal relationships describe the relationship between maintenance tasks and time. For instance, the maintenance task of disconnecting communication on optical cable A was carried out between February 14th and February 18th. Conflict relationships are used to identify the relationships between entities that conflict, and these relationships can be determined and established later according to the judgment rules.

[0051] Furthermore, after extracting the entities from the power communication maintenance text information and mining the entity relationships between them, each entity is used as a node to construct the power communication maintenance knowledge graph, and the entity relationships are used as edges between the corresponding nodes. Finally, the power communication maintenance knowledge graph is constructed based on each node and the edges between them.

[0052] Furthermore, when building a knowledge graph in the context of power communication maintenance, considering specific maintenance time periods, optical cables, optical paths, and services, the amount of information in the graph becomes enormous, and the computational load and complexity of traversal operations also increase accordingly. Therefore, we introduce an attention mechanism to focus on key information to improve efficiency and accuracy. Power communication systems are dynamically changing; new equipment, services, or maintenance tasks are constantly emerging, and the knowledge graph is updated accordingly. The attention mechanism can better adapt to this dynamic change, flexibly adjusting the degree of attention to different information based on the current situation. When a new maintenance task is added, the attention mechanism can quickly identify the existing knowledge graph portion related to that task, avoiding recalculation and analysis of the entire graph. For example, during the maintenance of optical cable A, the status information of the optical transceivers connected to it, and other service optical paths transmitting signals through this optical cable, can have a significant impact on the planning and risk assessment of the current maintenance task.

[0053] In another embodiment of the invention, knowledge graph-related models (such as preset entity extraction models, preset relation recognition models, and knowledge graph embedding models) can be jointly trained with neural network models. Through joint optimization, different models can work together better and complement each other's information. For example, during training, the knowledge graph model provides the neural network with more accurate structured knowledge representations, while the feedback from the neural network can help the knowledge graph model further optimize the extraction and representation of entities and relations, thereby improving the performance of the entire conflict analysis method.

[0054] 103. Based on the knowledge graph of power communication maintenance, identify power communication maintenance conflicts.

[0055] In this embodiment of the invention, after constructing the power communication maintenance knowledge graph, power communication maintenance conflicts are finally identified based on the power communication maintenance knowledge graph.

[0056] For example, time conflict determination involves: extracting the time information of maintenance tasks from the knowledge graph, such as obtaining the maintenance time of the A optical cable communication interruption maintenance task as February 14th to February 18th; traversing other maintenance tasks, i.e., traversing all maintenance task entities in the knowledge graph, and obtaining the corresponding time information for each maintenance task; determining time conflicts, i.e., for each traversed maintenance task, checking whether its time overlaps with the time of the A optical cable communication interruption maintenance task. If there is an overlap, establishing a conflict relationship between the two maintenance tasks in the knowledge graph and marking it as a time conflict; for example, if another maintenance task that may cause the interruption of the detour optical cable plan is found to be from February 16th to February 20th, then it has a time conflict with the A optical cable communication interruption maintenance task, and establishing a conflict relationship between them in the knowledge graph.

[0057] Resource conflict determination: (1) Determine the transition optical cable and related optical paths, identify the transition detour optical cable entity from the knowledge graph, and the optical path entity carried on the transition optical cable during maintenance, traverse other optical paths and service adjustment information, and traverse all optical path entities, maintenance tickets, mode sheets, transition schemes and other related information in the knowledge graph. For each optical path entity, check whether there is an adjustment operation (such as adjusting the other port protection service of AB port protection and its optical path to the optical cable route, or adjusting another single-port security control service in the same direction to the optical cable route). (2) If it is found that the adjustment operation of a certain optical path entity will make it carry on the transition detour optical cable, and the optical cable has already carried other optical paths that are not allowed to be adjusted during maintenance, then establish the conflict relationship between the optical path entity and the transition detour optical cable in the knowledge graph and mark it as a resource conflict. For example, if a maintenance ticket is found to be adjusting the optical path of another protection service of the AB port to a transitional detour optical cable, which will cause the optical cable to carry the same AB port channel of the same protection service simultaneously from February 14th to February 18th, exceeding 8 hours under the maintenance mode, posing an operational risk, then a resource conflict is determined to exist, and the corresponding conflict relationship is established in the knowledge graph.

[0058] Logical conflict determination: Logical conflict determination is usually based on rules, time resources, and relationship consistency. Specifically, for example: (1) Rule-based conflict determination: Some logical rules are pre-set in the knowledge graph. For example, "If the maintenance plan clearly states that a certain optical path carries a specific service, then the service module affected by the maintenance ticket should include these services", or "During the same maintenance period, the same optical cable cannot be in both normal operation and open / closed state at the same time", etc. When new information is entered into the knowledge graph, the system will check whether the relationship and attributes between entities are logically consistent according to these rules. If they are not consistent, a logical conflict is determined. For example, in the above example, the maintenance plan indicates that the optical path carries protection and security control services, but the service module affected by the maintenance ticket does not list security control services, which violates the rules and is determined to be a conflict. (2) Time and resource conflict detection: For conflicts involving time and resources, the relationship between time entities and resource entities such as equipment and optical paths in the knowledge graph is analyzed to determine the conflict. For example, check whether different maintenance tasks overlap in time, or whether the same resource (such as a certain optical cable) is occupied by multiple contradictory operations at the same time. If two maintenance tasks are found in the knowledge graph to perform different operations on the same optical cable within the same time period (e.g., one task is to disconnect the optical cable, and the other task is to perform normal communication on the optical cable), then a time and resource conflict is determined. (3) Relationship consistency check: Verify the consistency of relationships between entities in the knowledge graph. For example, check whether the bearer relationship between optical paths and services, and the relationship between maintenance tasks and affected services are reasonable. If it is found that a service claims to be carried by a certain optical path, but there is no corresponding bearer relationship between the optical path and the service in the knowledge graph, or the relationship is incomplete, then a logical conflict is determined.

[0059] In another embodiment of the present invention, in order to improve the accuracy of power communication maintenance conflict identification, it is also necessary to optimize the power communication maintenance knowledge graph. Based on this, the method includes: converting the nodes and edges in the power communication maintenance knowledge graph into vector representations, and using a multi-head attention mechanism to calculate the vector representations of the power communication maintenance knowledge graph; in the multi-head attention mechanism, setting multiple attention heads, each head focusing on and calculating the vectors of the power communication maintenance knowledge graph from different perspectives; and concatenating and linearly transforming the outputs of the multiple attention heads to obtain a power communication maintenance knowledge graph with attention weights.

[0060] Specifically, the multimodal knowledge graph vector initialization employs a Heterogeneous Graph Attention Network (HGAT) extended model to fuse graph structure and semantic features, providing targeted processing for various entities and relationships within the power communication knowledge graph. For instance, in a fiber optic cable breakage and maintenance example, the multimodal knowledge graph vector initialization needs to be more targeted in conflict judgment. For the entity of the fiber optic cable in the transition phase, the vector generation should not only consider its topological connection with surrounding equipment (such as optical transmission equipment at both ends and surrounding fiber optic cables) and the fiber optic cable's own functional description semantics (transmission capacity, type of communication service carried), but also factors that may cause conflicts. For example, analyzing the correlation between the services carried by the fiber optic cable in the transition phase; if the correlation is high, it may indicate that the services carried by the fiber optic cable in the transition phase do not have N-1 resistance during maintenance. For maintenance tasks, when generating vectors by fusing task flow logic (such as the optical path detour stage, the fiber optic cable breakage and maintenance stage, and the optical path restoration stage), potential conflicts between each stage and other work order tasks should be taken into account. For example, during the optical path detour and restoration of the original route phases, it is determined whether the detour will affect the normal communication of surrounding equipment and whether it will conflict with other maintenance work along the detour route. During the optical cable breakage and maintenance phase, it is assessed whether it will conflict with other maintenance tasks or methods, and whether related work will interfere with the ongoing temporary detour services. Simultaneously, location coding is introduced to record node structural location information, such as clarifying the location of a certain line's optical cable in the network topology. Combined with the locations of potentially affected surrounding equipment and optical cables, this further enhances the completeness of vector representation in conflict detection.

[0061] Regarding adaptive attention head allocation, within the multi-head attention mechanism framework, when training the attention head allocator using reinforcement learning algorithms for tasks such as fiber optic cable breakage and maintenance, it's crucial to analyze conflict-related factors more deeply. A detailed study of historical conflict cases related to similar fiber optic cable breakage and maintenance is necessary, focusing not only on the temporal correlation between resources and maintenance tasks and resource allocation, but also on conflicts at each stage of the process. For example, for potential conflicts at different stages such as optical path detours, cable breakage, and restoration of the original optical path, it's necessary to intelligently allocate weights to different attention heads based on conflict assessment, ensuring precise focus on key conflict points.

[0062] Hierarchical attention calculations must be closely centered on conflict resolution. During coarse-grained global attention calculations, attention mechanisms are used to quickly locate key subgraph regions related to the fiber optic cable breakage and maintenance task. This involves not only identifying subgraphs containing potentially conflicting equipment groups but also considering conflict areas at each stage of the ticketing process. For example, during optical path detours and route restoration, attention is paid to areas that may interfere with communication of surrounding equipment; during fiber optic cable breakage and maintenance, attention is paid to areas that may affect the normal operation of detour services. Next, fine-grained local attention calculations are performed within the subgraph regions. Residual connections are introduced to avoid information loss, accurately capturing detailed information and complex relationships within the subgraph, with particular attention to conflict-related information, such as the connection status between devices within the subgraph, the service carrying status on different optical paths, and the impact of operations at each stage of the ticketing process on these statuses and carrying statuses. Finally, the global and local calculation results are merged and weighted to obtain a final knowledge graph representation with attention weights, ensuring that conflict resolution considers both the overall network structure and key local details.

[0063] In the dynamic attention weight update stage, it is necessary to monitor in real time the changes in the attributes of key nodes in the knowledge graph related to conflict judgment and the new relational data. Online learning algorithms and backpropagation algorithms are used to update the attention weights in real time and adjust model parameters so that the weights reflect the probability of conflict in the latest power communication system status. Meanwhile, due to the modeling of specific time periods, optical cables, optical paths, and service entities for each maintenance operation, the amount of graph information increases. Therefore, in the above stages, algorithms need to be optimized to reduce the computational load and complexity of traversal operations. For example, a pruning strategy can be used in global attention calculation to reduce unnecessary subgraph search ranges and focus only on regions that may generate conflicts; in local attention calculation, the calculation method of residual connections can be optimized to improve computational efficiency for more efficient conflict judgment.

[0064] According to the present invention, a knowledge graph-based method for identifying power communication maintenance conflicts, compared with the current method of determining maintenance conflicts by manually reviewing maintenance tickets, the present invention obtains power communication maintenance text information; then extracts entities from the power communication maintenance text information and mines the entity relationships between the entities; based on the entities and the entity relationships, a power communication maintenance knowledge graph is constructed, wherein each entity is a node in the power communication maintenance knowledge graph, and the entity relationships are edges between the nodes in the power communication maintenance knowledge graph; finally, power communication maintenance conflicts are identified based on the power communication maintenance knowledge graph. Therefore, by using images from power communication maintenance orders, a power communication maintenance knowledge graph is constructed. Ultimately, power communication maintenance conflicts are identified based on this knowledge graph, eliminating the need for manual intervention—that is, manual inspection of each maintenance order is unnecessary. This improves the automation and efficiency of power communication maintenance conflict identification. Furthermore, the knowledge graph accurately identifies complex entity relationships, such as dependencies between equipment and overlapping maintenance times. Humans may make mistakes and overlook details when identifying these complex relationships, thus this invention improves the accuracy of power communication maintenance conflict identification.

[0065] Furthermore, to better illustrate the above process of identifying power communication maintenance conflicts, as a refinement and extension of the above embodiments, this invention provides another knowledge graph-based method for identifying power communication maintenance conflicts, such as... Figure 3 As shown, the method includes:

[0066] 201. Obtain the text information of the power communication maintenance order.

[0067] Specifically, the power communication maintenance order text information is obtained by extracting various forms of maintenance-related documents, such as maintenance tickets, maintenance plans, and three measures and one plan, in PDF, Word text, and Excel spreadsheet formats.

[0068] 202. Extract entities and mine the entity relationships between entities from the power communication maintenance text information. Based on the entities and entity relationships, construct a power communication maintenance knowledge graph, where each entity is a node in the power communication maintenance knowledge graph, and entity relationships are edges between nodes in the power communication maintenance knowledge graph.

[0069] Specifically, after structuring the power communication maintenance-related data into power communication maintenance text information, each entity is extracted from the power communication maintenance text information, and the entity relationships between each entity are mined based on the context of each entity in the power communication maintenance text information. Then, each entity is used as a node in the power communication maintenance knowledge graph to be constructed, and the entity relationships between each entity are used as the edges between the corresponding nodes. Finally, the power communication maintenance knowledge graph is constructed based on each node and the edges between each node.

[0070] 203. Based on the knowledge graph of power communication maintenance, identify power communication maintenance conflicts.

[0071] In this embodiment of the invention, after constructing the power communication maintenance knowledge graph, it is necessary to identify power communication maintenance conflicts based on the power communication maintenance knowledge graph. Therefore, step 203 specifically includes: based on the power communication maintenance knowledge graph, identifying at least one of the following conflicts: time conflict identification, resource conflict identification, and logical conflict identification.

[0072] Specifically, in one embodiment of the present invention, a power communication maintenance knowledge graph can be input into a preset conflict identification model for conflict identification to obtain power maintenance conflicts. The construction process of the preset conflict identification model is as follows: an initial conflict identification model is constructed, and a sample dataset is obtained, wherein the sample dataset includes a sample power communication maintenance knowledge graph with conflict annotation information; the sample dataset is divided into training data and test data, and the initial conflict identification model is trained using the training data, and tested using the test data; finally, the initial conflict identification model that meets the test conditions is determined as the preset conflict identification model. In another embodiment of the present invention, conflict identification rules can also be formulated based on professional knowledge and experience in the field of power communication maintenance; the entities and relationships in the power maintenance knowledge graph are used to match the rules to identify potential conflicts. For example, if there are multiple contradictory fault handling measures in the graph, a conflict can be considered to exist. The embodiments of the present invention use knowledge graph technology to model and analyze text data, thereby enabling accurate early identification and warning of potential maintenance conflicts.

[0073] 204. Efficiency optimization of power communication maintenance conflict identification based on attention mechanism.

[0074] In this embodiment of the invention, to improve the efficiency and accuracy of maintenance conflict identification, an attention mechanism needs to be introduced for optimization. Based on this, step 204 specifically includes: an input layer that converts data into vector form for input; an attention layer that uses the attention mechanism to calculate the weights between different data vectors, highlighting key information; and a conflict identification layer that uses specific algorithms or rules to determine conflicts. In the conflict detection phase, relevant optical path services and conflict-prone areas are prioritized based on service priority and historical data; in the conflict localization phase, conflict nodes and edges are quickly located in the knowledge graph using the attention mechanism, and the location and connection relationships of equipment are clarified by combining location coding; in the conflict resolution phase, different resolution strategies are weighted according to service real-time performance, importance, and maintenance resource status, and then comprehensively evaluated.

[0075] 205. Issue early warnings for identified power communication maintenance conflicts.

[0076] Specifically, when a conflict is detected during power communication maintenance, an early warning of the power communication maintenance conflict is issued based on information such as the conflict type.

[0077] According to another knowledge graph-based method for identifying power communication maintenance conflicts provided by the present invention, compared with the current method of determining maintenance conflicts by manually reviewing maintenance tickets, the present invention obtains power communication maintenance text information, extracts entities from the power communication maintenance text information and mines the entity relationships between the entities, and constructs a power communication maintenance knowledge graph based on the entities and the entity relationships, wherein each entity is a node in the power communication maintenance knowledge graph, and the entity relationships are edges between the nodes in the power communication maintenance knowledge graph; finally, power communication maintenance conflicts are identified based on the power communication maintenance knowledge graph. This invention constructs a knowledge graph for power communication maintenance, and ultimately identifies power communication maintenance conflicts based on this knowledge graph. No manual intervention is required, meaning there's no need for manual inspection of each maintenance order, thus improving the automation and efficiency of power communication maintenance conflict identification. Furthermore, the knowledge graph accurately identifies complex entity relationships, such as carrier relationships, time associations, and conflict relationships. Humans may make mistakes and overlook details when identifying these complex relationships, thus this invention improves the accuracy of power communication maintenance conflict identification.

[0078] Furthermore, as Figure 1 In a specific implementation, this invention provides a knowledge graph-based power communication maintenance conflict identification device, such as... Figure 4 As shown, the device includes: an acquisition unit 31, a construction unit 32, and an identification unit 33.

[0079] The acquisition unit 31 can be used to acquire power communication maintenance text information.

[0080] The construction unit 32 can be used to extract entities and mine entity relationships between entities in the power communication maintenance text information, and construct a power communication maintenance knowledge graph based on the entities and entity relationships, wherein each entity is a node in the power communication maintenance knowledge graph, and the entity relationship is an edge between the nodes in the power communication maintenance knowledge graph.

[0081] In specific application scenarios, in order to extract entities from power communication maintenance text information, such as Figure 5 As shown, the construction unit 32 includes an acquisition module 321 and an entity extraction module 322.

[0082] The acquisition module 321 can be used to acquire a preset entity extraction model, wherein the preset entity extraction model includes a text encoder for extracting feature vectors, a deep self-attention network for learning context-dependent vector representations, and a conditional random field for labeling; the preset entity extraction model is pre-trained based on a sample text dataset.

[0083] The entity extraction module 322 can be used to input the power communication maintenance text information into the preset entity extraction model, extract the feature vector of the power communication maintenance text information through the text encoder to obtain the feature vector, extract the text context features of the feature vector through the deep self-attention network to obtain the context features, and perform label recognition on the context features through the conditional random field to obtain the sequence label, wherein the sequence label contains each entity and its corresponding entity category.

[0084] In specific application scenarios, in order to discover entity relationships, the construction unit 32 also includes a relationship mining module 323.

[0085] The acquisition module 321 can also be used to acquire a preset relationship recognition model, wherein the preset relationship recognition model includes an input representation layer for information embedding, a low-level feature extraction layer for low-level feature extraction, a dynamic routing capsule network feature clustering layer with word attention mechanism for global relationship feature extraction, and a relationship prediction layer for entity relationship prediction; the preset relationship recognition model is pre-trained based on an entity text dataset.

[0086] The relationship mining module 323 can be used to input the power communication maintenance text information and each entity into the preset relationship recognition model. The input representation layer performs word embedding, character embedding, part-of-speech embedding, and word position embedding on the power communication maintenance text information, and fuses the word embedding, character embedding, part-of-speech embedding, and word position embedding to obtain fused embedding. The low-level feature extraction layer extracts low-level features from the fused embedding to obtain low-level features. The dynamic routing capsule network feature clustering layer of the word attention mechanism extracts features from the low-level features to obtain global relationship features. The relationship prediction layer predicts relationships between the global relationship features to obtain entity relationships between the entities.

[0087] In a specific application scenario, the low-level feature extraction layer includes a bidirectional long short-term memory neural network, which includes a forward network layer and a backward network layer. To perform low-level feature extraction, the relationship mining module 323 can be used to input the fused embedding into the bidirectional long short-term memory neural network for global sequence feature extraction, to obtain the forward global sequence features corresponding to the forward network layer and the backward global sequence features corresponding to the backward network layer, and to fuse the forward global sequence features and the backward global sequence features to obtain the low-level features.

[0088] In specific application scenarios, the dynamic routing capsule network feature clustering layer of the word attention mechanism includes: a low-level capsule network, a word attention mechanism, and a high-level capsule network. In order to extract global relation features, the relation mining module 323 can be used to input the low-level features into the low-level capsule network for feature extraction to obtain local features, input the local features into the word attention mechanism for weight allocation to obtain weighted local features, and input the weighted local features into the high-level capsule network for feature extraction to obtain global relation features.

[0089] In specific application scenarios, in order to perform conflict identification, the identification unit 33 can be used to identify at least one of the following conflicts in power communication maintenance: time conflict identification, resource conflict identification, and logical conflict identification, based on the power communication maintenance knowledge graph.

[0090] In specific application scenarios, in order to provide early warning of power communication maintenance conflicts, the device also includes an early warning unit 34.

[0091] The early warning unit 34 can be used to provide early warning for identified power communication maintenance conflicts.

[0092] In specific application scenarios, in order to optimize the knowledge graph of power communication maintenance, the device also includes an optimization unit 35.

[0093] The nodes and edges in the power communication maintenance knowledge graph are converted into vector representations. A multi-head attention mechanism is used to calculate the vector representation of the power communication maintenance knowledge graph. In the multi-head attention mechanism, multiple attention heads are set, and each head focuses on and calculates the vectors of the power communication maintenance knowledge graph from different perspectives. The outputs of the multiple attention heads are concatenated and linearly transformed to obtain a power communication maintenance knowledge graph with attention weights.

[0094] It should be noted that other corresponding descriptions of the functional modules involved in the knowledge graph-based power communication maintenance conflict identification device provided in this embodiment of the invention can be found in [reference]. Figure 1 The corresponding description of the method shown will not be repeated here.

[0095] Based on the above, Figure 1 Accordingly, this embodiment of the invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the following steps: acquiring power communication maintenance text information; extracting entities from the power communication maintenance text information and mining the relationships between the entities; constructing a power communication maintenance knowledge graph based on the entities and their relationships, wherein each entity serves as a node in the power communication maintenance knowledge graph, and the entity relationships serve as edges between the nodes in the power communication maintenance knowledge graph; and identifying power communication maintenance conflicts based on the power communication maintenance knowledge graph.

[0096] Based on the above, Figure 1 The method shown and as Figure 4 The embodiment of the device shown in the invention also provides a physical structure diagram of a computer device, such as... Figure 6 As shown, the computer device includes: a processor 41, a memory 42, and a computer program stored in the memory 42 and executable on the processor. Both the memory 42 and the processor 41 are mounted on a bus 43. When the processor 41 executes the program, it performs the following steps: acquiring power communication maintenance text information; extracting entities from the power communication maintenance text information and mining the relationships between the entities; constructing a power communication maintenance knowledge graph based on the entities and their relationships, wherein each entity serves as a node in the power communication maintenance knowledge graph, and the entity relationships serve as edges between the nodes in the power communication maintenance knowledge graph; and identifying power communication maintenance conflicts based on the power communication maintenance knowledge graph.

[0097] The present invention acquires power communication maintenance text information; then extracts entities and mines the entity relationships between these entities from the text information; and constructs a power communication maintenance knowledge graph based on these entities and their relationships. Each entity serves as a node in the knowledge graph, and the entity relationships serve as edges between these nodes. Finally, power communication maintenance conflicts are identified based on the knowledge graph. Thus, by constructing a power communication maintenance knowledge graph from power communication maintenance order images, and ultimately identifying power communication maintenance conflicts based on this knowledge graph, manual intervention is eliminated, meaning there is no need for manual item-by-item inspection of the maintenance order. This improves the automation and efficiency of power communication maintenance conflict identification. Furthermore, the knowledge graph can accurately identify complex entity relationships, such as dependencies between equipment and overlapping maintenance times. Manual identification of these complex relationships may result in errors and oversights, thus the present invention improves the accuracy of power communication maintenance conflict identification.

[0098] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0099] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A knowledge graph-based method for identifying power communication maintenance conflicts, characterized in that, include: Obtain power communication maintenance text information; Entities are extracted from the power communication maintenance text information, and the relationships between the entities are mined. Based on the entities and their relationships, a power communication maintenance knowledge graph is constructed, wherein each entity is a node in the power communication maintenance knowledge graph, and the entity relationships are edges between the nodes in the power communication maintenance knowledge graph. Based on the aforementioned power communication maintenance knowledge graph, power communication maintenance conflicts are identified; Among them, entities are extracted from power communication maintenance text information, including: A preset entity extraction model is obtained, wherein the preset entity extraction model includes a text encoder for extracting feature vectors, a deep self-attention network for learning context-dependent vector representations, and a conditional random field for labeling; the preset entity extraction model is pre-trained based on a sample text dataset; the power communication maintenance text information is input into the preset entity extraction model, the feature vectors of the power communication maintenance text information are extracted by the text encoder to obtain feature vectors, the text context features are extracted by the deep self-attention network to obtain context features, and the label recognition is performed on the context features by the conditional random field to obtain sequence label labels, wherein the sequence label labels contain each entity and its corresponding entity category; The identification of power communication maintenance conflicts based on the power communication maintenance knowledge graph includes: Based on the power communication maintenance knowledge graph, the method identifies at least one of the following conflicts: time conflict identification, resource conflict identification, and logical conflict identification. After identifying the power communication maintenance conflicts based on the power communication maintenance knowledge graph, the method further includes: issuing an early warning for the identified power communication maintenance conflicts. Before identifying power communication maintenance conflicts based on the power communication maintenance knowledge graph, the method further includes: The nodes and edges in the power communication maintenance knowledge graph are converted into vector representations. A multi-head attention mechanism is used to calculate the vector representation of the power communication maintenance knowledge graph. In the multi-head attention mechanism, multiple attention heads are set, and each head focuses on and calculates the vectors of the power communication maintenance knowledge graph from different perspectives. The outputs of the multiple attention heads are concatenated and linearly transformed to obtain a power communication maintenance knowledge graph with attention weights.

2. The method according to claim 1, characterized in that, Mining the entity relationships between the entities, including: A preset relationship recognition model is obtained, wherein the preset relationship recognition model includes an input representation layer for information embedding, a low-level feature extraction layer for low-level feature extraction, a dynamic routing capsule network feature clustering layer with word attention mechanism for global relationship feature extraction, and a relationship prediction layer for entity relationship prediction; the preset relationship recognition model is pre-trained based on an entity text dataset. The power communication maintenance text information and each entity are input into the preset relationship recognition model. The input representation layer performs word embedding, character embedding, part-of-speech embedding, and word position embedding on the power communication maintenance text information, and then fuses the word embedding, character embedding, part-of-speech embedding, and word position embedding to obtain fused embedding. The low-level feature extraction layer extracts low-level features from the fused embedding to obtain low-level features. The dynamic routing capsule network feature clustering layer of the word attention mechanism extracts features from the low-level features to obtain global relationship features. The relationship prediction layer predicts relationships between the entities by using the global relationship features.

3. The method according to claim 2, characterized in that, The low-level feature extraction layer includes a bidirectional long short-term memory neural network, which includes a forward network layer and a backward network layer. The step of extracting low-level features from the fused embedding through the low-level feature extraction layer to obtain low-level features includes: The fused embedding is input into the bidirectional long short-term memory neural network for global sequence feature extraction, to obtain the forward global sequence features corresponding to the forward network layer and the backward global sequence features corresponding to the backward network layer. The forward global sequence features and the backward global sequence features are fused to obtain the low-level features.

4. The method according to claim 2, characterized in that, The dynamic routing capsule network feature clustering layer of the word attention mechanism includes: a low-level capsule network, a word attention mechanism, and a high-level capsule network; The dynamic routing capsule network feature clustering layer using the word attention mechanism extracts features from the low-level features to obtain global relation features, including: The low-level features are input into the low-level capsule network for feature extraction to obtain local features. The local features are then input into the word attention mechanism for weight allocation to obtain weighted local features. Finally, the weighted local features are input into the high-level capsule network for feature extraction to obtain global relation features.

5. A knowledge graph-based power communication maintenance conflict identification device, characterized in that, include: The acquisition unit is used to acquire power communication maintenance text information; A construction unit is used to extract entities and mine entity relationships between entities from the power communication maintenance text information. Based on the entities and entity relationships, a power communication maintenance knowledge graph is constructed. Each entity is a node in the power communication maintenance knowledge graph, and the entity relationships are edges between nodes. Extracting entities from the power communication maintenance text information includes: obtaining a preset entity extraction model, which includes a text encoder for extracting feature vectors, a deep self-attention network for learning context-dependent vector representations, and a conditional random field for labeling; the preset entity extraction model is pre-trained based on a sample text dataset; the power communication maintenance text information is input into the preset entity extraction model; the text encoder extracts feature vectors from the power communication maintenance text information to obtain feature vectors; the deep self-attention network extracts text context features from the feature vectors to obtain context features; and the conditional random field identifies labels on the context features to obtain sequence labels, wherein the sequence labels contain each entity and its corresponding entity category. The optimization unit is used to convert the nodes and edges in the power communication maintenance knowledge graph into vector representations, and to calculate the vector representations of the power communication maintenance knowledge graph using a multi-head attention mechanism. In the multi-head attention mechanism, multiple attention heads are set, and each head focuses on and calculates the vectors of the power communication maintenance knowledge graph from different perspectives. The outputs of the multiple attention heads are concatenated and linearly transformed to obtain the power communication maintenance knowledge graph with attention weights. The identification unit is used to identify power communication maintenance conflicts based on the power communication maintenance knowledge graph. The identification of power communication maintenance conflicts based on the power communication maintenance knowledge graph includes: identifying at least one of the following conflicts: time conflict identification, resource conflict identification, and logical conflict identification based on the power communication maintenance knowledge graph. The early warning unit is used to issue early warnings for identified power communication maintenance conflicts.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 4.

7. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 4.

Images