Multi-feature recommendation method for power communication equipment fault information based on knowledge graph

Abstract

This invention discloses a multi-feature recommendation method for power communication equipment fault information based on a knowledge graph, comprising the following steps: obtaining user historical search information and power communication equipment fault information; constructing a user history search subgraph based on the user history search information; extracting user features based on the user history search subgraph; constructing a power communication equipment fault knowledge graph using Neo4j based on the power communication equipment fault information; extracting graph features based on target node domain information in the power communication equipment fault knowledge graph; inputting the user features and graph features into a cross-compression unit for cross-learning, and outputting power communication equipment fault prediction information. This method can achieve accurate retrieval and personalized recommendation of power communication equipment fault information, assisting in the intelligent operation and maintenance of power communication equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of power communication equipment fault inspection and repair, and more specifically to a multi-feature recommendation method for power communication equipment fault information based on a knowledge graph. Background Art

[0002] The expansion of maintenance and distribution scenarios for a large number of power communication devices presents new challenges for the intelligent operation and maintenance of power grids. Failures in power communication equipment can easily cause grid disturbances, impacting the perception and real-time assessment of grid operating status, leading to uncontrolled energy dispatch, distribution, and transmission, and seriously endangering the operational safety of the power system. Furthermore, the real-time collection of power communication equipment fault information has led to an exponential growth in fault data. This data is mostly stored in an unstructured format, making it difficult for power communication equipment maintenance personnel to directly utilize the implicit knowledge contained within. This makes fault location difficult and reduces maintenance efficiency. Therefore, an efficient solution is urgently needed to accurately locate equipment faults and achieve efficient maintenance and fault diagnosis of power equipment.

[0003] A knowledge graph is a graphical representation that uses visualization techniques to express knowledge and relational structures. Knowledge graphs refine vector representations through entity connections, effectively mining the underlying relationships between entities. This not only alleviates the data scarcity problem in the recommendation process but also makes the results more interpretable. However, current algorithms that use knowledge graphs to assist in recommendations are primarily based on embedding propagation, which underutilizes the structural information in the knowledge graph. Furthermore, the high-level interactions between graph features and user features that share latent features can easily introduce significant noise, resulting in low recommendation accuracy.

[0004] To address the complex structure of power communication equipment fault information and the inability to directly leverage the user behavior knowledge contained therein, which leads to difficulty locating power communication equipment faults and low maintenance efficiency, researchers have proposed a knowledge graph recommendation method that incorporates the influence effects of similar users. This method effectively utilizes the knowledge graph while expanding the interaction between users and the graph. Using a graph neural network neighborhood aggregation strategy and an attention mechanism, two high-level representations of users and items on the knowledge graph are captured. Based on the influence effects of similar users, an influence enhancement layer is designed to capture the latent representation of the similar user influence effects. These three representations are fed into a multilayer perceptron (MLP) to output a prediction score, enabling intelligent recommendation of communication equipment fault knowledge. This method can mine potential connections between users and the graph, capture potential user interests, and make recommendations based on user search intent. However, its performance at the fine-grained level of user-graph interaction needs improvement. Furthermore, the training of the MLP-based policy network is complex and difficult to converge or find an optimal solution.

[0005] Therefore, how to accurately locate equipment faults based on knowledge graphs, effectively solve the low accuracy problem caused by multi-feature high-order interactive noise in fault recommendations, realize accurate retrieval and personalized recommendation of power communication equipment fault information, and assist in the intelligent operation and maintenance of power communication equipment, is a problem that technical personnel in this field urgently need to solve. Summary of the Invention

[0006] In view of the above problems, the present invention provides a multi-feature recommendation method for power communication equipment fault information based on knowledge graph to at least solve some of the technical problems mentioned in the above background technology.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] The multi-feature recommendation method for power communication equipment fault information based on knowledge graph includes the following steps:

[0009] Obtain user history retrieval information and power communication equipment fault information;

[0010] Constructing a user history retrieval subgraph based on the user history retrieval information;

[0011] Retrieving a subgraph based on the user history to extract user features;

[0012] Based on the power communication equipment fault information, Neo4j is used to construct a power communication equipment fault knowledge graph;

[0013] Extracting graph features based on target node domain information in the power communication equipment fault knowledge graph;

[0014] The user features and the graph features are input into a cross-compression unit for cross-learning, and power communication equipment fault prediction information is output.

[0015] Furthermore, constructing a user history retrieval subgraph based on the user history retrieval information specifically includes:

[0016] Extracting user knowledge triples based on the user history retrieval information;

[0017] Based on the user knowledge triples, a user history retrieval subgraph is constructed.

[0018] Furthermore, the extracting user features by searching the subgraph based on the user history specifically includes:

[0019] Taking the user history search information as a seed set for preference propagation, expanding outward along the entity-oriented relationship in the user history search subgraph with the target entity node as the center, to form a z-th hop entity set corresponding to the z-th hop domain information;

[0020] A graph convolutional neural network is used to extract user features from the z-th hop entity set.

[0021] Furthermore, in the process of extracting user features based on the user history retrieval subgraph, a graph convolutional neural network is used to aggregate any two-hop domain information of entities in the user history retrieval subgraph.

[0022] Furthermore, the power communication equipment fault information includes unstructured text data and structured text data.

[0023] Furthermore, the method of constructing a knowledge graph of power communication equipment faults using Neo4j based on the power communication equipment fault information specifically includes:

[0024] Extracting fault knowledge triples based on the unstructured text data;

[0025] According to the fault knowledge triples and structured text data, a power communication equipment fault knowledge graph is constructed on Neo4j using the Cypher language.

[0026] Furthermore, the extraction of fault knowledge triples based on the unstructured text data specifically includes:

[0027] Preprocessing the unstructured text data to obtain a power communication equipment fault corpus;

[0028] The MASK-LM method is used to serialize the fault short text data in the power communication equipment fault corpus;

[0029] For the serialized fault short text data, the ALBERT model is used to obtain the global semantic features of the fault short text data; and the CNN model is used to obtain the local semantic features of the fault short text data;

[0030] According to the global semantic features and the local semantic features, fault knowledge triples are extracted.

[0031] Furthermore, the ALBERT model includes a multi-layer bidirectional Transformer encoder;

[0032] The multi-layer bidirectional Transformer encoder is used to extract the global semantic features of the fault short text data; wherein the Self-attention module in the multi-layer bidirectional Transformer encoder is used to extract the semantic vector features of the fault short text data and generate a word vector.

[0033] Furthermore, the CNN model includes an input layer, a convolutional layer, a pooling layer and a softmax layer;

[0034] The input layer is used to input the word vector;

[0035] The convolution layer and pooling layer are used to extract local semantic features of the fault short text data;

[0036] The softmax layer is used to perform classification prediction on the entity relationships contained in the fault short text data.

[0037] It can be seen from the above technical solution that, compared with the prior art, the present invention discloses a method for recommending multi-features of fault information of power communication equipment based on knowledge graph, which has the following beneficial effects:

[0038] 1. The present invention automatically shares potential features through cross-compression units, completes high-order interaction between user features and graph features, and achieves accurate prediction of conventional fault retrieval.

[0039] 2. The present invention adopts the ALBERT model + CNN model, which can improve the level of semantic feature extraction, thereby effectively constructing fault knowledge triples.

[0040] 3. The present invention uses position, entity tag and part of speech as feature input to extract entities and relationships in the fault text information of power communication equipment, reducing the loss of semantic information features and effectively utilizing equipment fault data.

[0041] 4. The present invention constructs a knowledge graph of fault information of power communication equipment to realize structured storage and visual display of fault information. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0043] Figure 1 The accompanying figure is a schematic diagram of the framework of the multi-feature recommendation method for power communication equipment fault information based on knowledge graph provided by the present invention.

[0044] Figure 2 The accompanying figure is a schematic diagram of part of the content of the power communication equipment fault knowledge graph provided by the present invention.

[0045] Figure 3 The accompanying figure is a schematic diagram of evaluation indicators for the multi-feature recommendation method for power communication equipment fault information based on knowledge graph provided by the present invention.

[0046] Figure 4The accompanying figure is a schematic diagram of the loss rate results of the model of the multi-feature recommendation method for power communication equipment fault information based on the knowledge graph provided by the present invention. DETAILED DESCRIPTION

[0047] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0048] An embodiment of the present invention discloses a multi-feature recommendation method for power communication equipment fault information based on a knowledge graph, comprising the following steps: obtaining user history retrieval information and power communication equipment fault information; constructing a user history retrieval subgraph based on the user history retrieval information; extracting user features based on the user history retrieval subgraph; constructing a power communication equipment fault knowledge graph using Neo4j based on the power communication equipment fault information; extracting graph features based on the power communication equipment fault knowledge graph; inputting the user features and the graph features into a cross-compression unit for cross-learning, and outputting power communication equipment fault prediction information.

[0049] The following three parts will explain the above steps in detail.

[0050] 1. Extract user features:

[0051] Using the entity relationship extraction model mentioned below, we complete the knowledge extraction of user historical search information, obtain user knowledge triples, and use the preference propagation algorithm to construct the user historical search subgraph;

[0052] Then, the preference propagation algorithm is used to obtain the z-th hop entity set of the target entity node in the user history retrieval subgraph: the user history retrieval information is used as the seed set for preference propagation, and the target entity node is expanded outward along the entity-directed relationship in the user history retrieval subgraph to form the z-th hop entity set corresponding to the z-th (z=1,2,…n) hop domain information. The relevant entity set is defined as:

[0053]

[0054] in, Represents a set of related entities; Represents the entity set contained in the user's historical search information; z represents the z-th hop domain information in the entity set; v i Indicates the vth iusers; e′ represents the head entity in the triple information in the user history retrieval subgraph; r represents the relationship in the triple information in the user history retrieval subgraph; e represents the tail entity in the triple information in the user history retrieval subgraph; It represents the entity set of the z-1th hop. When z=1, the entity set of item 0 represents the target entity node, which indicates the user's historical search information.

[0055] A graph convolutional neural network is used to extract user features from the obtained z-th hop;

[0056] To avoid introducing a lot of noise, a graph convolutional neural network is used to aggregate the adjacent two-hop neighborhood information of entities in the user history retrieval subgraph. The initial representation of user features and their neighborhood representation are aggregated using an aggregator. The user feature representation containing neighborhood information is defined as:

[0057]

[0058] in, Indicates the vth i The retrieval problem of users; b represents the vector bias; represents the z-th hop entity set obtained;

[0059] 2. Extracting graph features:

[0060] Obtaining power communication equipment fault information; the power communication equipment fault information includes unstructured text data and structured text data;

[0061] Preprocess the unstructured text data to obtain a corpus of power communication equipment faults: Use an expert database to segment and supplement unstructured text data that exceeds a preset byte length. Perform preprocessing on the unstructured text data based on its byte length, including word segmentation, pause word removal, and sentence segmentation. The preprocessed unstructured text data is stored as short fault text data, serving as the corpus of power communication equipment faults.

[0062] The MASK-LM method is used to analyze the fault short text data X={X1,X2...X n} is serialized; wherein any word in the fault short text data may be randomly masked or replaced, and then the meaning of the masked or replaced word is inferred based on the meaning of the adjacent words. Then the fault short text data is encoded as P = {P1, P2...P n}, where P n Indicates the nth serialized encoded character of the fault short text data.

[0063] For the serialized fault short text data, the ALBERT model is used to obtain the global semantic features of the fault short text data: the ALBERT model has the advantage of capturing global features in semantic extraction. The serialized fault short text data is input into the semantic extraction layer of the ALBERT model, and then the multi-layer bidirectional Transformer encoder in the ALBERT model is used to generate the global semantic features of the fault text W = {W0, W2...W n The self-attention module in the multi-layer bidirectional Transformer encoder is used to extract the semantic vector features of the fault short text data and generate word vectors. The global semantic feature extraction formula is defined as:

[0064]

[0065] Among them, U W represents the feature weight calculated based on context semantics; d represents the vector dimension;

[0066] For the serialized fault short text data, a CNN model is used to obtain the local semantic features of the fault short text data. The CNN model includes an input layer, a convolutional layer, a pooling layer, and a softmax layer. The input layer is used to input word vectors; the convolutional layer and the pooling layer are used to extract the local semantic features of the fault short text data; and the softmax layer is used to classify and predict the entity relationships contained in the fault short text data.

[0067] The association between the word vectors output by the previous layer is performed using a one-dimensional convolutional neural network containing K convolution kernels to extract information and perform convolution operations. The output dimension of the convolution layer depends on the length of the sentence. To facilitate downstream tasks, the output of the convolution layer must be independent of the sequence length. Pooling operations are used, and maximum pooling is used to capture the structural semantic information characteristics of the entity. The local semantic feature extraction formula in this embodiment of the present invention is defined as:

[0068]

[0069] in, Represents the convolution kernel Q c The weight in row a and column k; C i，j Represents the jth eigenvalue of the i-th vector in the convolutional layer and the eigenvector sequence; d Q Represents the size of the convolution kernel; d represents the dimension of the feature vector after convolution of a single convolution kernel;

[0070] Extract fault knowledge triples based on global and local semantic features. Use Cypher to build a power communication equipment fault knowledge graph on Neo4j based on the fault knowledge triples and structured text data. Extract graph features based on the target node domain information in the power communication equipment fault knowledge graph. Structured storage and visualization of fault information can be achieved based on Neo4j, and direct query of fault knowledge can be achieved using Cypher.

[0071] Based on the constructed knowledge graph of power communication equipment failure, the neighborhood information of the target node of the user retrieval entity is aggregated to obtain the graph characteristics; in order to avoid the neighboring data of the retrieval target node being too sparse, the knowledge graph of power communication equipment failure is first converted into an undirected graph. Since the number of neighborhood entities of different target nodes varies depending on the target node, in order to maintain uniformity in the research process, each target node is set to retain the same number of neighborhood entities. In order to obtain the graph characteristics, the degree to which the target nodes corresponding to different user retrieval questions in the graph are affected by their neighbor nodes is first studied. In the embodiment of the present invention, the neighbor score of the user retrieval entity as the target node is calculated. The score is defined as:

[0072]

[0073] in, represents the problem neighbor score; Indicates the vth i The search question of a user; e represents the entity contained in the knowledge graph of power communication equipment failure; G(e) represents the entity set in the knowledge graph of power communication equipment failure;

[0074] In order to characterize the neighborhood topology of entity e, the first-order neighborhood representation is calculated in the embodiment of the present invention, which is specifically defined as:

[0075]

[0076] S(e) represents the first-order domain entity set;

[0077] Aggregate the knowledge graph neighborhood information of the question entity retrieved by the user, and obtain the graphical feature representation of the aggregated neighborhood information as follows:

[0078]

[0079] Among them, Y represents the feature weight and b represents the vector bias.

[0080] 3. Input user features and graph features into the cross-compression unit for cross-learning:

[0081] To improve generalization and reduce fitting noise, a cross-compression unit is used to explicitly exchange high-level feature information between user and graph representations. Latent features in both user and graph representations are automatically shared to enable cross-transfer and mutual complementation of knowledge, and to update historical retrieval records and atlas entity neighborhood information.

[0082] In the embodiment of the present invention, the cross-compression unit includes a cross-compression mapping layer, a potential feature sharing learning module, and a prediction function layer. In the cross-compression mapping, user features and graph features are cross-compressed in the feature cross matrix C. In the W module, the potential features are shared to obtain the shared user features and graph features, and information prediction is performed according to the prediction function. Specifically:

[0083] The feature intersection matrix C is defined as:

[0084]

[0085] Where T represents the matrix transpose operation; Represents the characteristics of the graph; Represents user characteristics;

[0086] Project C into the potential representation space through cross compression, and update the user features and graph features through the potential feature sharing unit. The specific feature vectors are expressed as:

[0087]

[0088]

[0089] Among them, M represents the perceptron learning method; represents the user features after compression mapping; e r represents the entity representation in the graph; v represents the user; Indicates averaging in the prediction function; V F represents the user features after cross-compression shared learning of latent features; E F Represents the graph features after cross-compression shared learning of latent features;

[0090] The scoring function is used to calculate the fault prediction information of power communication equipment, which is expressed as:

[0091]

[0092] Among them, w vg Represents the weight including user information and graph information;

[0093] This power communication equipment fault prediction information is a retrieval of fault information, based on a constructed knowledge graph, and recommends answers to retrieval questions. For example, when searching for the current symptoms of a certain fault on the graph, a solution and related information about the fault are obtained.

[0094] In summary, the embodiments of this invention provide a method for recommending multi-feature fault information of power communication equipment based on knowledge graphs. This method solves the current problems of difficult fault location and low maintenance efficiency of power communication equipment through knowledge graph technology, realizes accurate retrieval and personalized recommendation of fault information of power communication equipment, and assists in the intelligent operation and maintenance of power communication equipment. Knowledge joint extraction is performed on communication equipment fault information, and unstructured fault data is converted into triples of entities and relationships, fully mining the implicit knowledge between information; a knowledge graph of equipment faults is constructed and visualized, intuitively describing the rich correlations between faults, and based on the user's historical retrieval data, fully understanding the user's retrieval intention, so as to provide personalized recommendation services to the user; fault recommendation based on multiple features aggregates neighborhood information based on historical records and fault knowledge, improves the accuracy of retrieval results, and realizes intelligent recommendation of power communication equipment fault information.

[0095] In other embodiments, the fusion knowledge graph assisted recommendation method in the above embodiment can be replaced by an embedding-based recommendation method or a path-based recommendation method. Among them, the embedding-based recommendation method uses the knowledge graph node embedding technology to learn the entity embedding in the knowledge graph, and uses the learned entity representation as the prior information of the project. However, this type of model fails to capture the high-order semantic information of the project on the knowledge graph; the path-based recommendation method mines the high-order connection relationship between users and projects in the knowledge graph (heterogeneous graph), which is called meta-path.

[0096] Next, the above content will be described through specific embodiments.

[0097] First, the extracted fault text is segmented, with the sentence length set to a uniform 200 characters. The remaining sentences are then segmented and padded with missing or redundant characters to a uniform length. Missing text and blank lines are also deleted. The text sequence annotation format for the corpus uses the BIO annotation strategy to label each character in the user behavior sample set, marking each element as "BX," "IX," "EX," or "O." B represents the beginning of an entity or relationship, I represents the middle, E represents the end of an entity or relationship, and O indicates that the text does not belong to any entity or relationship type. An example of the annotation of the specific text data corpus is shown in Table 1.

[0098] Table 1. Corpus sample annotation examples

[0099]

[0100]

[0101] The design of the power communication equipment fault information ontology layer focuses on fault knowledge and historical record analysis, emphasizing the relationship between demand and behavior. The ontology of the equipment fault information knowledge graph is divided into a three-layer structure. For each fault scenario in the map, relationships and attributes are extracted based on this structure. This redesigned ontology layer enables more detailed solutions during grid equipment fault retrieval and helps personnel analyze the causes of topology failures, as shown in Table 2.

[0102] Table 2 Design of ontology layer of knowledge graph of power communication equipment fault

[0103]

[0104] Use the graph database to store the fault knowledge data of power communication equipment and form a fault knowledge graph. Figure 2 This is a partial view of a graphical visualization of power communication equipment fault operation and maintenance data. Existing fault knowledge is integrated through ontology-level reconstruction. The stored fault data is connected to dispatch specifications through subordinate or inclusion relationships. The nodes in the diagram represent equipment fault entities, and the phrases on the arrows between them represent the relationship type between the entities. Different colors correspond to different entity categories. Entities include equipment fault knowledge, such as device name, fault type, fault level, fault description, fault cause, and solution.

[0105] Finally, the proposed method was experimentally compared with the multi-class model, and the precision P, recall R and F1 value were used to evaluate the recommendation results. The specific results are as follows: Figure 3 and Figure 4 shown.

[0106] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0107] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-feature recommendation method for power communication equipment fault information based on knowledge graph, characterized by: The steps include: Obtain user history retrieval information and power communication equipment fault information; Constructing a user history retrieval subgraph based on the user history retrieval information; Retrieving a subgraph based on the user history to extract user features; Based on the power communication equipment fault information, Neo4j is used to construct a power communication equipment fault knowledge graph; Extracting graph features based on target node neighborhood information in the power communication equipment fault knowledge graph; Inputting the user features and the graph features into a cross-compression unit for cross-learning, and outputting power communication equipment fault prediction information; The step of retrieving the subgraph based on the user history and extracting user features specifically includes: Taking the user history retrieval information as a seed set for preference propagation, expanding outward along the entity-directed relationship in the user history retrieval subgraph with the target entity node as the center, to form a z-th hop entity set corresponding to the z-th hop neighborhood information; Extracting user features from the z-th hop entity set using a graph convolutional neural network; The extracting graph features according to the target node neighborhood information in the power communication equipment fault knowledge graph specifically includes: Converting the power communication equipment fault knowledge graph into an undirected graph; and setting each target node in the power communication equipment fault knowledge graph to retain the same number of neighboring entities; Obtain the degree of influence of the corresponding neighboring entity when the user search entity is the target node, which is recorded as the neighbor score; Based on the neighbor scores, the knowledge graph neighborhood information of the user search entity is aggregated to obtain graph features of the aggregated neighborhood information.

2. The method for recommending multi-features of power communication equipment fault information based on knowledge graph according to claim 1 is characterized in that: The step of constructing a user history search subgraph based on the user history search information specifically includes: Extracting user knowledge triples based on the user history retrieval information; Based on the user knowledge triples, a user history retrieval subgraph is constructed.

3. The method for recommending multi-features of power communication equipment fault information based on knowledge graph according to claim 1 is characterized in that: In the process of extracting user features according to the user history retrieval subgraph, a graph convolutional neural network is used to aggregate any two-hop neighborhood information of entities in the user history retrieval subgraph.

4. The method for recommending multi-features of power communication equipment fault information based on knowledge graph according to claim 1 is characterized in that: The power communication equipment fault information includes unstructured text data and structured text data.

5. The method for recommending multi-features of power communication equipment fault information based on knowledge graph according to claim 4 is characterized in that: The method of constructing a knowledge graph of power communication equipment faults using Neo4j based on the power communication equipment fault information specifically includes: Extracting fault knowledge triples based on the unstructured text data; According to the fault knowledge triples and structured text data, a power communication equipment fault knowledge graph is constructed on Neo4j using the Cypher language.

6. The method for recommending multi-features of power communication equipment fault information based on knowledge graph according to claim 5 is characterized in that: Extracting fault knowledge triples based on the unstructured text data specifically includes: Preprocessing the unstructured text data to obtain a power communication equipment fault corpus; The MASK-LM method is used to serialize the fault short text data in the power communication equipment fault corpus; For the serialized fault short text data, the ALBERT model is used to obtain the global semantic features of the fault short text data; and the CNN model is used to obtain the local semantic features of the fault short text data; According to the global semantic features and the local semantic features, fault knowledge triples are extracted.

7. The method for recommending multi-features of power communication equipment fault information based on knowledge graph according to claim 6 is characterized in that: The ALBERT model includes a multi-layer bidirectional Transformer encoder; The multi-layer bidirectional Transformer encoder is used to extract the global semantic features of the fault short text data; wherein the Self-attention module in the multi-layer bidirectional Transformer encoder is used to extract the semantic vector features of the fault short text data and generate a word vector.

8. The method for recommending multi-features of power communication equipment fault information based on knowledge graph according to claim 7 is characterized in that: The CNN model includes an input layer, a convolutional layer, a pooling layer, and a softmax layer; The input layer is used to input the word vector; The convolution layer and pooling layer are used to extract local semantic features of the fault short text data; The softmax layer is used to perform classification prediction on the entity relationships contained in the fault short text data.

Images