An intelligent alarm analysis method based on a topic model and a heterogeneous graph

Abstract

This invention discloses an intelligent alarm analysis method based on topic models and heterogeneous graphs. The method includes using a semantically related topic model to expand the topic features of alarm data, constructing a heterogeneous graph to model and extract potential semantic association features between different alarm information streams, using a neural network model to extract and process features from the alarm data, extracting text context sequence information features contained in the alarm text information, and concatenating and fusing the extracted potential semantic association features between different alarm information and the contained text context sequence information features. These features are then introduced into a self-attention model to highlight relevant features, and finally input into a fully connected layer and a Softmax layer for prediction output. This intelligent alarm analysis method based on topic models and heterogeneous graphs improves the accuracy and efficiency of intelligent alarm analysis.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically to an intelligent alarm analysis method based on topic models and heterogeneous graphs. Background Technology

[0002] The task of Natural Language Processing (NLP) is to develop the technological capabilities for computers and robots to communicate effectively with humans using natural language. Text analysis is an important research direction in NLP, involving the automatic extraction and classification of information from large amounts of text data. With the rapid development of artificial intelligence technologies, represented by deep neural networks, significant breakthroughs have been achieved in NLP using deep learning techniques, and text analysis technology is now widely applied in various fields such as news, healthcare, industry, and commerce. Depending on the application scenario, text analysis can be used for sentiment analysis, topic detection, and classification. Users can define rules based on different application scenarios and needs, and formulate different response strategies based on the information extracted from text analysis. For example, it can enable rapid and efficient analysis of sudden plant alarm information, generate intelligent alarm information, and provide solution suggestions. Governments can understand hot topics, grasp public opinion trends, and respond quickly to important public events.

[0003] Intelligent alarm analysis is a typical task of text analysis technology. With the rapid development of computer technology, digital system platforms are being introduced into more and more real-world applications, and the alarm analysis and processing module is a crucial component of these platforms, primarily used for the collection, display, analysis, and processing of alarm information. Currently, with the continuous expansion of data scale and the upgrading and transformation of the information technology industry, new requirements are being placed on the alarm analysis and processing module: how to achieve rapid and efficient information analysis and processing of a large amount of alarm information generated in a short period due to faults, extract important information, generate intelligent alarm information, and provide decision-making suggestions. Intelligent alarm analysis is based on artificial intelligence technology, using complex neural network models to model and process alarm information flows. It utilizes scientific neural network models and powerful computing power to achieve efficient information analysis and processing of alarm information, providing intelligent alarm information and solution suggestions for sudden problems, assisting users in making scientific and rapid decisions.

[0004] When analyzing and processing alarm information, intelligent alarm analysis cannot simply focus on individual alarm messages while ignoring the potential semantic relationships between multiple alarm messages generated within the same time period. Most intelligent alarm analysis methods primarily utilize machine learning algorithms or neural network models to process alarm messages one by one. However, this approach is less efficient than batch processing and yields incomplete information, lacking key details from other alarm messages within the same time period, ultimately leading to low efficiency. Summary of the Invention

[0005] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0006] Therefore, the purpose of this invention is to provide an intelligent alarm analysis method based on topic models and heterogeneous graphs, thereby improving the accuracy and efficiency of intelligent alarm analysis.

[0007] To address the aforementioned technical problems, according to one aspect of the present invention, the present invention provides the following technical solution:

[0008] An intelligent alarm analysis method based on topic modeling and heterogeneous graphs, comprising:

[0009] S1. A topic model based on semantic relevance is used to expand the topic features of the alarm data, and a heterogeneous graph is constructed to model and extract the potential semantic association features between different alarm information in the alarm information flow.

[0010] S2. Use a neural network model to extract and process features from alarm data, and extract the text context sequence information features contained in the alarm text information;

[0011] S3. The potential semantic association features and the implied text context sequence information features extracted from different alarm information are concatenated and fused, and introduced into the self-attention model to highlight relevant features. Then, they are input into the fully connected layer and the Softmax layer for prediction output.

[0012] As a preferred embodiment of the intelligent alarm analysis method based on topic model and heterogeneous graph described in this invention, in step S1, the semantically related topic model is generated by improving the BTM topic model from the construction word pair set and model training strategy.

[0013] As a preferred embodiment of the intelligent alarm analysis method based on topic model and heterogeneous graph described in this invention, the step of improving the BTM topic model from the construction of a set of word pairs is as follows:

[0014] Data preprocessing and word segmentation are performed on short texts in the corpus to obtain short texts d∈D;

[0015] Using the Jieba tool, all short texts d in the document set D are part-of-speech tagging, and nouns, verbs and adjectives are selected to form text d′. At the same time, a vocabulary table V of the corpus is constructed based on d′.

[0016] Based on d′, construct word pairs in the forms of noun-noun, verb-noun, and adjective-noun to obtain the basic word pair set. Where N B′ This represents word pair b in the basic word pair set B′. e =(w ie ,w je The quantity of );

[0017] The Word2vec model is used to obtain the vector representation of all words in the vocabulary V, the cosine similarity between words is calculated, and words with similarity greater than a predefined threshold η are selected and included in the set C.

[0018] Construct a set of related word pairs based on semantic similarity, iterate through all d′, and if the word w in d′ is... i If w exists in set C, then search for w in set C. i The set of similar words For sets In addition to w i Other words w j When w i and w j When not in the same text, w j Combine the remaining words in d′ with the words in the set B″ to form word pairs such as noun-noun, verb-noun, and adjective-noun, and put them into the set B″, thus obtaining the set of related word pairs. Where N B″ This represents word pair b in the set B″ of related word pairs. o =(w io ,w jo The quantity of );

[0019] The set of basic word pairs B′ and the set of related word pairs B″ are merged to obtain the set of word pairs. N B This represents the number of word pairs in the set B.

[0020] As a preferred embodiment of the intelligent alarm analysis method based on topic modeling and heterogeneous graphs described in this invention, the steps for improving the BTM topic model from the model training strategy are as follows:

[0021] Construct a semantically related set SR of word pairs b SR b It is a word pair b = (w i ,w j SR, a semantically related vocabulary set b SR collection b Including w i and w j The remaining words are related to w i and w jThe cosine similarity of all values ​​is greater than the predefined threshold γ, i.e., SR b ={w x |sim(w x ,w i )>γ,sim(w x ,w j )>γ},sim(w x ,w i ) indicates the calculation of w x and w i Corresponding word vectors and The cosine similarity is given by the formula (1);

[0022]

[0023] When word pair b is highly relevant to topic z, words semantically related to word pair b are selected for replacement. Formula (2) is used to calculate word pair b = (w i ,w j ) and all topics z k Correlation P(z) k |b), where k∈{1,...,K}, K represents the total number of topics, and formula (3) is used to calculate the maximum relevance P(z|b) between word pair b and all topics. max Formula (4) represents the degree of relevance between topic z and word pair b, and in formula (5) X b Follows a Bernoulli distribution, when X b When = 1, the word pair b and the corresponding semantically related word set SR will be used. b Replace the word in X when X b When = 0, only word pair b is replaced;

[0024]

[0025] P(z|b) max =max(P(z) k |b)) (3)

[0026]

[0027] X b ~Bern(χ b,z (5).

[0028] As a preferred embodiment of the intelligent alarm analysis method based on topic model and heterogeneous graph described in this invention, the steps for generating a semantically related topic model are as follows;

[0029] Generate a corpus-topic distribution θ ~ Dir(α) from a Dirichlet prior with parameter α;

[0030] For each topic z k (k = 1, ..., K), generate topic-word distributions from a Dirichlet prior with parameter β.

[0031] For any word pair b = (w) in the set of word pairs B i ,w j ), generate topic assignment z for word pairs b according to the multinomial distribution θ, satisfying z~Multi(θ), according to χ b,z Generate distribution X b ~Bern(χ b,z );

[0032] Where X b =0, according to the multinomial distribution Extract words w from topic z i and w j ,satisfy and

[0033] If X b =1, for set SR b All the words w x According to the multinomial distribution Extract words w from topic z x ,satisfy

[0034] As a preferred embodiment of the intelligent alarm analysis method based on topic model and heterogeneous graph described in this invention, the specific steps for constructing the heterogeneous graph include:

[0035] For any text d i After passing it through a semantically related BTM topic model, the topic feature vocabulary t is obtained. i Let T = {t0, t1, ..., t} be the topic feature vocabulary extracted from all texts using the topic model. m};

[0036] Construct a heterogeneous graph G = (V, E), where V is the set of all nodes, including all nodes of the text set, vocabulary set, and topic feature vocabulary set, and E is the set of edges between nodes;

[0037] Constructing an adjacency matrix A represents the relationship between the nodes, as specifically defined in formula (6):

[0038]

[0039] As a preferred embodiment of the intelligent alarm analysis method based on topic model and heterogeneous graph described in this invention, the step S2, in which a neural network model is used to extract and process features from the alarm data, and the text context sequence information features contained in the alarm text information are extracted as follows:

[0040] The alarm text information is vectorized using the ERNIE model to obtain C = [c1, c2, ..., c n ] T Then, the Bi-GRU model is used for feature extraction and processing to obtain global context features.

[0041] Compared with the prior art, the beneficial effects of this invention are: it uses a semantically related topic model to expand the topic features of alarm data, constructs a heterogeneous graph to model and extract the potential semantic associations between different alarm information in the alarm information stream, and uses a neural network model to perform comprehensive feature extraction and processing to generate intelligent alarm information.

[0042] This invention mines high-quality, representative thematic feature words from alarm data corpora from a semantic relevance perspective to enrich the semantic features of alarm data. In order to more comprehensively capture the potential semantic associations and information between different alarm information in the alarm information stream, this invention constructs a heterogeneous graph to model and learn the complex relationships between alarm information. It uses the node information propagation characteristics of the graph to extract information. In order to make full use of the serialized information of the text, this invention uses a neural network model to extract features from the word vector representation of alarm information, reducing the impact of the characteristics of the alarm information itself on the final analysis effect and further improving the accuracy of intelligent alarm analysis. Attached Figure Description

[0043] To more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and detailed embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0044] Figure 1 This is a flowchart of an intelligent alarm analysis method based on topic modeling and heterogeneous graphs according to the present invention;

[0045] Figure 2 This is a schematic diagram of the semantically related BTM topic model of the present invention;

[0046] Figure 3 This is a schematic diagram of the heterogeneous graph of the present invention;

[0047] Figure 4 This is a schematic diagram of the model training for the heterogeneous graph of the present invention. Detailed Implementation

[0048] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0049] Secondly, the present invention is described in detail with reference to the schematic diagrams. When describing the embodiments of the present invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not according to the usual scale. Furthermore, the schematic diagrams are merely examples and should not limit the scope of protection of the present invention. In addition, actual fabrication should include the three-dimensional spatial dimensions of length, width, and depth.

[0050] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0051] This invention provides an intelligent alarm analysis method based on topic models and heterogeneous graphs, which improves the accuracy and efficiency of intelligent alarm analysis.

[0052] Figure 1 The diagram shown is a flowchart of one embodiment of an intelligent alarm analysis method based on topic modeling and heterogeneous graphs according to the present invention. Please refer to [link / reference]. Figure 1 The intelligent alarm analysis method based on topic model and heterogeneous graph in this embodiment has the following specific steps:

[0053] S1. A topic model based on semantic relevance is used to expand the topic features of the alarm data, and a heterogeneous graph is constructed to model and extract the potential semantic association features between different alarm information in the alarm information flow.

[0054] S2. Use a neural network model to extract and process features from alarm data, and extract the text context sequence information features contained in the alarm text information;

[0055] S3. The potential semantic association features and the implied text context sequence information features extracted from different alarm information are concatenated and fused, and introduced into the self-attention model to highlight relevant features. Then, they are input into the fully connected layer and the Softmax layer for prediction output.

[0056] In step S1, based on the semantically relevant BTM topic model, such as Figure 1 As shown, the corpus contains N B A set of word pairs Where b i =(w i ,w j ) represents a word pair, K represents the number of latent topics in the model, and z represents word pair b. i =(w i ,w jThe theme of ) Let θ represent the lexical distribution under a topic, α represent the topic distribution in the corpus, α and β represent the hyperparameters of the "corpus-topic" and "topic-lexical" probability distributions, respectively, and χ represent the lexical distribution under a topic. b,z X is used to indicate the degree of relevance between topic z and word pair b. b The semantically related set SR used to determine whether to include word pairs b The word w in x Put it back. Represents the set SR b The number of words in the Chinese text.

[0057] Compared to traditional BTM topic models, semantically related BTM topic models improve upon traditional models in two aspects: constructing the word pair set and the model training strategy. The specific steps for constructing the word pair set B based on part-of-speech tagging and semantic similarity are as follows:

[0058] Data preprocessing and word segmentation are performed on short texts in the corpus to obtain short texts d∈D;

[0059] Using the Jieba tool, all short texts d in the document set D are part-of-speech tagging, and nouns, verbs and adjectives are selected to form text d′. At the same time, a vocabulary table V of the corpus is constructed based on d′.

[0060] Based on d′, construct word pairs in the forms of noun-noun, verb-noun, and adjective-noun to obtain the basic word pair set. Where N B′ This represents word pair b in the basic word pair set B′. e =(w ie ,w je The quantity of );

[0061] The Word2vec model is used to obtain the vector representation of all words in the vocabulary V, the cosine similarity between words is calculated, and words with similarity greater than a predefined threshold η are selected and included in the set C.

[0062] Construct a set of related word pairs based on semantic similarity, iterate through all d′, and if the word w in d′ is... i If w exists in set C, then search for w in set C. i The set of similar words For sets In addition to w i Other words w j When w i and w j When not in the same text, w j Combine the remaining words in d′ with the words in the set B″ to form word pairs such as noun-noun, verb-noun, and adjective-noun, and put them into the set B″, thus obtaining the set of related word pairs. Where N B″ This represents word pair b in the set B″ of related word pairs. o =(w io ,w jo The quantity of );

[0063] The set of basic word pairs B′ and the set of related word pairs B″ are merged to obtain the set of word pairs. N B This represents the number of word pairs in the set B.

[0064] The specific steps for improving model training strategies based on semantic similarity are as follows:

[0065] Construct a semantically related set SR of word pairs b SR b It is a word pair b = (w i ,w j SR, a semantically related vocabulary set b SR collection b Including w i and w j The remaining words are related to w i and w j The cosine similarity of all values ​​is greater than the predefined threshold γ, i.e., SR b ={w x |sim(w x ,w i )>γ,sim(w x ,w j )>γ},sim(w x ,w i ) indicates the calculation of w x and w i Corresponding word vectors and The cosine similarity is given by the formula (1);

[0066]

[0067] When word pair b is highly relevant to topic z, words semantically related to word pair b are selected for replacement. Formula (2) is used to calculate word pair b = (w i ,w j ) and all topics z k Correlation P(z) k |b), where k∈{1,...,K}, K represents the total number of topics, and formula (3) is used to calculate the maximum relevance P(zb) of word pair b with all topics. max Formula (4) represents the degree of relevance between topic z and word pair b, and in formula (5) X b Follows a Bernoulli distribution, when X bWhen = 1, the word pair b and the corresponding semantically related word set SR will be used. b Replace the word in X when X b When = 0, only word pair b is replaced;

[0068]

[0069] P(z|b) max =max(P(z) k |b)) (3)

[0070]

[0071] X b ~Bern(χ b,z (5)

[0072] The semantically related BTM topic model constructs a set B of word pairs based on part-of-speech tagging and semantic similarity on the training corpus, and then performs a topic generation process, as shown below:

[0073] Generate a corpus-topic distribution θ ~ Dir(α) from a Dirichlet prior with parameter α;

[0074] For each topic z k (k = 1, ..., K), generate topic-word distributions from a Dirichlet prior with parameter β.

[0075] For any word pair b = (w) in the set of word pairs B i ,w j ), generate topic assignment z for word pairs b according to the multinomial distribution θ, satisfying z~Multi(θ), according to χ b,z Generate distribution X b ~Bern(χ b,z );

[0076] Where X b =0, according to the multinomial distribution Extract words w from topic z i and w j ,satisfy and

[0077] If X b =1, for set SR b All the words w x According to the multinomial distribution Extract words w from topic z x ,satisfy

[0078] To more comprehensively capture the potential semantic relationships and information between different alarm messages in the alarm information stream, this invention constructs a heterogeneous graph to model and learn the complex relationships between alarm messages, and uses a GCN neural network model for training on the graph. Before constructing the heterogeneous graph, a semantically related BTM topic model is used to obtain the topic feature vocabulary of the input text. The specific steps include:

[0079] Data preprocessing and word segmentation are performed on short texts in the corpus to obtain short texts d∈D;

[0080] Using the Jieba tool, all short texts d in the document set D are part-of-speech tagging, and nouns, verbs and adjectives are selected to form text d′. At the same time, a vocabulary table V of the corpus is constructed based on d′.

[0081] Based on d′, construct word pairs in the forms of noun-noun, verb-noun, and adjective-noun to obtain the basic word pair set. Where N B′ This represents word pair b in the basic word pair set B′. e =(w ie ,w je The quantity of );

[0082] The Word2vec model is used to obtain the vector representation of all words in the vocabulary V, the cosine similarity between words is calculated, and words with similarity greater than a predefined threshold η are selected and included in the set C.

[0083] Construct a set of related word pairs based on semantic similarity, iterate through all d′, and if the word w in d′ is... i If w exists in set C, then search for w in set C. i The set of similar words For sets In addition to w i Other words w j When w i and w j When not in the same text, w j Combine the remaining words in d′ with the words in the set B″ to form word pairs such as noun-noun, verb-noun, and adjective-noun, and put them into the set B″, thus obtaining the set of related word pairs. Where N B″ This represents word pair b in the set B″ of related word pairs. o =(w io ,w jo The quantity of ).

[0084] The set of basic word pairs B′ and the set of related word pairs B″ are merged to obtain the set of word pairs. N B This represents the number of word pairs in the set B.

[0085] The heterogeneous graph is constructed based on topic models and alarm information flows. The specific steps include:

[0086] For any text d i After passing it through a semantically related BTM topic model, the topic feature vocabulary t is obtained. i Let T = {t0, t1, ..., t} be the topic feature vocabulary extracted from all texts using the topic model. m};

[0087] Construct a heterogeneous graph G = (V, E), such as Figure 3 As shown, V is the set of all nodes, including all nodes in the text set, vocabulary set, and topic feature vocabulary set, and E is the set of edges between nodes;

[0088] Constructing an adjacency matrix A represents the relationship between the nodes, as specifically defined in formula (6):

[0089]

[0090] In step S2, a neural network model is used to extract and process features from the alarm data. The steps for extracting the text context sequence information features contained in the alarm text information are as follows: The word sequence of the alarm text information is vectorized using the ERNIE model to obtain C = [c1, c2, ..., c n ] T Then, the Bi-GRU model is used for feature extraction and processing to obtain global context features.

[0091] This invention constructs a heterogeneous graph based on topic models and alarm information flows, and introduces a two-layer GCN model for training and learning on the graph, such as... Figure 4 As shown, messages are propagated between nodes such as text, topics, and words, and high-level semantic relationships between nodes are learned, as shown in formula (7), where H l Represents the state of the l-th layer, with the initial state H. 0 =L, degree matrix D ii =∑ j A ij σ is the ReLU activation function, W l It is a trainable linear transformation weight, which, after being trained by the GCN model, ultimately outputs a high-level feature representation H. G .

[0092]

[0093] The features processed by GCN and Bi-GRU are concatenated and fused, and a self-attention model is introduced to highlight relevant features. These features are then input into a fully connected layer and a softmax layer for prediction output. The specific formula is shown in the figure, where W... fc and b fc These are the parameter matrix and bias of the fully connected layer.

[0094]

[0095] H′=H·α T (9)

[0096] P = softmax(W) fc H′+b fc (10)

[0097] In order to achieve the best model performance, the cross-entropy loss function is introduced to constrain the parameters and L2 regularization term is added to reduce the risk of overfitting. The loss function is shown in formula (11), where λ is the regularization term coefficient and θ represents all parameters.

[0098]

[0099] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. An intelligent alarm analysis method based on topic modeling and heterogeneous graphs, characterized in that, include: S1. A topic model based on semantic relevance is used to expand the topic features of the alarm data, and a heterogeneous graph is constructed to model and extract the potential semantic association features between different alarm information in the alarm information flow. S2. Use a neural network model to extract and process features from alarm data, extracting text context sequence information features contained in alarm text information; S3. Concatenate and fuse the potential semantic association features between different alarm information and the text context sequence information features contained therein, introduce them into the self-attention model to highlight relevant features, and then input them into the fully connected layer and the Softmax layer for prediction output; In step S1, the semantically related topic model is generated by improving the BTM topic model from the construction word pair set and model training strategy; The specific steps for constructing a heterogeneous graph include: For any text After passing it through a semantically related BTM topic model, the topic feature vocabulary is obtained. The topic feature words extracted from all texts using the topic model are denoted as... ; Constructing heterogeneous graphs ,in It is the collection of all nodes, including all nodes in the text set, vocabulary set, and topic feature vocabulary set. It is the set of edges between nodes; Constructing an adjacency matrix , The relationships between nodes are represented, as specifically defined in formula (6): （6）。 2. The intelligent alarm analysis method based on topic modeling and heterogeneous graphs according to claim 1, characterized in that, The steps to improve the BTM topic model from the construction of the word pair set are as follows: Data preprocessing and word segmentation are performed on short texts in the corpus to obtain short texts. ; Using Jieba tool to process document collections All short texts Perform part-of-speech tagging, selecting nouns, verbs, and adjectives to form the text. At the same time, based on Vocabulary for corpus construction ; according to Construct word pairs in the form of nouns with nouns, verbs with nouns, and adjectives with nouns to obtain the basic set of word pairs. ,in Represents the set of basic terms Chinese words Quantity; The vocabulary was obtained using the Word2vec model. The vector representations of all words are used to calculate the cosine similarity between words, and words with similarity greater than a predefined threshold are selected. The words are included in the set middle; Construct a set of related word pairs based on semantic similarity, and iterate through all of them. ,if words in There exist sets In the set, then Search in The set of similar words For sets Except Other words ,when and When not in the same text, and The remaining words are used to construct word pairs in the form of nouns and nouns, verbs and nouns, and adjectives and nouns, and then placed into a set. In the end, a set of related word pairs is obtained. ,in Indicates the set of related words Chinese words Quantity; Basic word pair set Related word pair set Merging yields a set of word pairs. , Representation of sets of words The number of word pairs.

3. The intelligent alarm analysis method based on topic modeling and heterogeneous graphs according to claim 1, characterized in that, The steps to improve the BTM topic model based on the model training strategy are as follows: Construct a semantically related set of word pairs , It is a word pair semantic related word set ,gather include and The rest of the words and and The cosine similarity of all of them is greater than the predefined threshold. ,Right now , Indicates calculation and Corresponding word vectors and The cosine similarity is given by the formula (1); （1） When word pair With the topic When highly relevant, choose word pairs. Semantic word replacement, formula (2) is used to calculate word pairs. With all topics correlation ,in , Representing the total number of topics, formula (3) is used to calculate word pairs. Maximum relevance to all topics Formula (4) represents the theme Pair of words The degree of correlation, in formula (5) Follows Bernoulli distribution, when Word pairing and corresponding semantic related word sets Words in the text are replaced when Only word pairs Put it back; （2） （3） （4） （5）。 4. The intelligent alarm analysis method based on topic modeling and heterogeneous graphs according to claim 2, characterized in that, The steps to generate a semantically relevant topic model are as follows; From the parameter is Generating Corpus-Topic Distribution in Dirichlet Priors ; For each topic From the parameter is Generating topic-word distributions in Dirichlet priors ; For the set of word pairs any word pair According to the multinomial distribution Generate word pairs Topic Allocation ,satisfy ,according to Generate distribution ; Among them, if According to the multinomial distribution From the theme Extracting words and ,satisfy and ; like For sets All words According to the multinomial distribution From the theme Extracting words ,satisfy .

5. The intelligent alarm analysis method based on topic modeling and heterogeneous graphs according to claim 1, characterized in that, In step S2, the neural network model is used to extract and process features from the alarm data. The steps for extracting the text context sequence information features contained in the alarm text information are as follows: The word sequence of the alarm text information is vectorized using the ERNIE model to obtain... Then, the Bi-GRU model is used for feature extraction and processing to obtain global context features.

Images