Thermal power generating unit feed pump fault auxiliary decision-making method and system based on knowledge graph

Abstract

The invention discloses a knowledge graph-based auxiliary fault decision-making method and system for a water feeding pump of a thermal power generating unit, and the method comprises the steps: carrying out the automatic marking of a text data set of the water feeding pump of the thermal power generating unit based on expert experience and a deep learning model, and dividing the marked text into a training set and a test set; preprocessing the training set and the test set; performing entity recognition and relation extraction on the processed text through a deep learning neural network, and extracting the relation between entities; a knowledge fusion method is adopted to eliminate redundant, repeated or contradictory information of different text sources; the method comprises the following steps of: automatically extracting entity and relational knowledge by adopting a trained model by arranging a power plant water supply pump document, constructing triple data, and importing the triple data into a database; according to the method, by constructing the equipment operation inspection intelligent application based on the knowledge graph, knowledge inheritance can be promoted, knowledge learning can be promoted, knowledge value can be released, burden reduction and efficiency improvement of a thermal power plant basic level can be served, and the text marking accuracy and efficiency can be improved.

Description

Technical Field

[0001] The present invention relates to a method and system for auxiliary decision-making on a feedwater pump failure of a thermal power unit, and in particular to a method and system for auxiliary decision-making on a feedwater pump failure of a thermal power unit based on a knowledge graph. Background Art

[0002] With the advancement of new power systems, the operating complexity of equipment has increased significantly. In actual production, thermal power units face many new challenges, such as participating in deep peak regulation of the power grid and ultra-low emission transformation. On the one hand, these problems affect the safety and reliability of the equipment, and on the other hand, they bring considerable pressure to the economic operation and environmental protection of the units. At the same time, the labor intensity of operators continues to increase, but their experience tends to be lost.

[0003] As a traditional industry, the power generation industry has formed a large amount of theoretical knowledge in the process of continuous development. At the same time, power generation companies have accumulated a large amount of data and information in their daily operation and maintenance. These data and information contain a lot of important information that can be mined and sorted. The use of manual labeling is limited by low efficiency, slow speed, and high requirements for professional skills. The use of deep learning models to automatically label text data sets can greatly improve the labeling efficiency, but its accuracy is biased. Summary of the invention

[0004] Purpose of the invention: The purpose of the present invention is to provide a method and system for auxiliary decision-making of feedwater pump failure of thermal power units based on knowledge graph, so as to improve the accuracy and efficiency of text annotation.

[0005] Technical solution: The present invention comprises the following steps:

[0006] Based on expert experience and deep learning models, the text dataset of the feedwater pump of the thermal power plant is automatically annotated, and the annotated text is divided into a training set and a test set;

[0007] Preprocess the training set and test set;

[0008] Through deep learning neural networks, the processed text is subjected to entity recognition and relationship extraction to extract entities and relationships between entities;

[0009] Use knowledge fusion methods to eliminate redundant, repeated or contradictory information from different text sources;

[0010] By organizing the power plant feed water pump documents, the trained model is used to automatically extract entity and relationship knowledge, construct triple data, and import it into the database.

[0011] The training set and the test set are obtained by dividing the corpus of entity and relational knowledge in the annotated text.

[0012] The text data set includes the operation log of the feedwater pump of the thermal power unit, equipment product documents, equipment defect files, and failure cases.

[0013] The preprocessing includes word segmentation and stop word filtering, and vector conversion is performed on the processed text, the preprocessed text is converted into a vector form, and the basic text is obtained.

[0014] The entity recognition and relationship extraction of the processed text is specifically: using a deep learning neural network to extract features from the obtained word vectors, and then obtaining the contextual relationship corresponding to each word vector.

[0015] The entity word vector in the word vector is determined and extracted according to the context relationship, and the entity word vector is used as the first entity.

[0016] The specific method for obtaining the first entity is: by parsing the context relationship corresponding to each word vector, determining whether the word vector is an entity word vector or a relationship word vector, and extracting and collecting the word vectors determined as entity word vectors as the first entity, providing a parameter basis for the subsequent retrieval process.

[0017] The knowledge fusion is specifically as follows: using an entity alignment method to match or align items representing the same entity to eliminate duplication or redundancy of entity information; using relationship fusion to integrate relationship information from different data sources and map the relationship to a unified representation space; using attribute fusion to integrate attribute information from different data sources and map the attribute to a unified representation space; using semantic fusion to integrate semantic information from different data sources and map the semantic information to a unified semantic space; evaluating the fused knowledge based on expert experience to verify whether the fusion result meets the expected needs and goals.

[0018] The construction of triple data is specifically as follows:

[0019] All available documents on the power plant's feedwater pumps are sorted out, and erroneous or incomplete document records in the original document data are filtered out. The trained entity extraction model and relationship extraction model are used to automatically extract the corresponding types of entity knowledge and relationship knowledge in the text set to construct the triple data of the feedwater pump fault knowledge graph.

[0020] A knowledge graph-based auxiliary decision-making system for feedwater pump failure in thermal power plants, comprising:

[0021] Text annotation module: Automatically annotate text data sets based on manual annotation based on expert experience and deep learning, and divide the annotated text into training sets and test sets;

[0022] Text processing module: performs text preprocessing on the annotated text to obtain preprocessed basic data;

[0023] Entity recognition and relationship extraction module: Through deep learning neural network, the processed text is subjected to entity recognition and relationship extraction to extract entities and the relationships between entities;

[0024] Knowledge fusion module: uses knowledge fusion methods to eliminate redundant, repeated or contradictory information from different text sources;

[0025] Triple data construction module: By organizing the documents of the power plant feed water pump, the trained model is used to automatically extract entity and relationship knowledge, construct triple data, and import it into the database.

[0026] Beneficial effects: The present invention extracts structured storage knowledge from unstructured texts such as operation and maintenance data, water pump design drawings, equipment defect data, fault cases, etc. through statistical methods and deep learning methods based on artificial neural networks, and constructs a knowledge graph of water pump equipment failures, which can realize intelligent data knowledge management. At the same time, the recommendation algorithm based on the knowledge graph can also better fit the recommendation scenario of intelligent operation and maintenance of thermal power plants. By building intelligent applications for equipment operation and maintenance business based on knowledge graphs, it can promote knowledge inheritance, promote knowledge learning, and release knowledge value in equipment acceptance, operation, inspection and maintenance, so as to serve the burden reduction and efficiency increase of the grassroots level of thermal power plants and improve the accuracy and efficiency of text annotation. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 is a flow chart of the present invention;

[0028] Figure 2 A flowchart for text annotation based on expert experience manual annotation combined with deep learning;

[0029] Figure 3 Flowchart of feature extraction of word vectors for the BiLSTM model. DETAILED DESCRIPTION

[0030] The present invention will be further described below in conjunction with the accompanying drawings.

[0031] Example 1

[0032] The knowledge graph-based auxiliary decision-making method for feedwater pump failure of thermal power units in this embodiment adopts technologies such as statistics, artificial intelligence and knowledge graphs to meet the needs of intelligent data knowledge management. At the same time, the recommendation algorithm based on the knowledge graph can also better fit the recommendation scenario of intelligent operation and maintenance of thermal power plants. By building intelligent applications for equipment operation and inspection business based on knowledge graphs, it is possible to promote knowledge inheritance, promote knowledge learning, release knowledge value, and serve the burden reduction and efficiency increase of the grassroots level of thermal power plants in the links of equipment acceptance, operation, inspection and maintenance. Specifically, the following steps are included:

[0033] Step 1: Perform text annotation based on expert experience combined with a deep learning model

[0034] Automatically annotate the entity and relationship types of the text dataset of the feed water pump of a thermal power unit based on expert experience and a deep learning model, and divide the corpus of entity and relationship knowledge in the annotated text into a training text set and a test text set, providing key support for subsequent deep learning recognition and processing.

[0035] The text dataset includes the operation logs of the feed water pump of a thermal power unit, equipment product documents, equipment defect files, fault cases, etc.

[0036] Step 2: Preprocess the training text set and the test text set, including Chinese word segmentation and stop word filtering. Specifically:

[0037] Adopt a method that combines dictionary matching and rule matching, and call the jieba word segmentation library to implement Chinese word segmentation. First, use jieba to load a large Chinese dictionary, which contains a large number of Chinese words and their corresponding parts of speech, and supplement professional vocabulary in the field of thermal power generation according to professional experience. Then, according to this dictionary, the input text is segmented. Two methods of prefix dictionary and suffix dictionary are adopted to match each word in the text with the words in the dictionary, and the word with the largest match is used as the word segmentation result. If no matching word is found in the dictionary, segmentation is performed based on rules.

[0038] After Chinese word segmentation, stop word removal processing is performed. Stop word removal mainly filters out words with little meaning in the text, such as the word "de" for forming words, the conjunction "and", the adverb "even", the modal particle "ba", etc., so as to reduce the dimensionality of the feature vector and improve the accuracy of extracting knowledge.

[0039] Step 3: Through a deep learning neural network, perform entity recognition and relationship extraction on the processed text, and extract the entities and the relationships between entities. Specifically:

[0040] Adopt the BiLSTM model to extract features from the obtained word vectors, and then obtain the context relationship corresponding to each word vector. The BiLSTM model uses a gated recurrent unit to replace the LSTM, effectively avoiding the problem of gradient disappearance. The gated recurrent unit, that is, the GRU, contains two LSTM layers, a forward LSTM and a backward LSTM. The forward LSTM is used to learn the forward features of the text sequence, and the backward LSTM is used to learn the backward features of the text sequence. At the same time, through forward propagation and backward propagation, feature extraction is performed on each word vector to obtain the context relationship corresponding to each word vector, that is, the context feature representation.

[0041] Step 4: Use knowledge fusion methods to eliminate redundant, repeated or contradictory information from different text sources and improve the accuracy and consistency of knowledge.

[0042] Step 5: By organizing the documents of the power plant feed water pump, the trained model is used to automatically extract entity and relationship knowledge, construct triple data, and import it into the Neo4j graph database.

[0043] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0044] like Figure 1 As shown, the knowledge graph-based auxiliary decision-making method for feedwater pump failure of thermal power units in this embodiment first collects the operation and maintenance data of feedwater pumps of thermal power units, feedwater pump design drawings, equipment defect data, fault cases and other editable texts, and annotates them, combining expert experience manual annotation with deep learning models to improve the accuracy and efficiency of text annotation. Specifically, it includes the following steps:

[0045] Step 1: Use manual annotation based on expert experience combined with deep learning to annotate text, specifically:

[0046] The collected text data sets such as the operation log of the feed water pump of the thermal power unit, equipment product documents, equipment defect files, and failure cases are manually annotated, and the deep learning model is used to automatically annotate the text data sets. The annotated objects are the entity types and relationship types to be identified. The corpus of entity and relationship knowledge in the annotated text is divided into a training text set and a test text set.

[0047] like Figure 2 As shown in the figure, the deep learning model of this example uses the CNN-BiLSTM-CRF model to train the text annotation model for the relevant text of the water pump. The CNN-BiLSTM-CRF model is a deep learning model that combines convolutional neural network (CNN), bidirectional long short-term memory network (BiLSTM) and conditional random field (CRF). It has strong feature extraction, context understanding and label dependency capture capabilities in the text annotation model training of complex texts, and can effectively capture important information in the text and improve the accuracy and consistency of text annotation tasks. The CNN-BiLSTM-CRF model is significantly better than LSTM, LSTM-CRF, BiLSTM-CRF and other models in complex text processing performance.

[0048] Convolutional neural network (CNN) is used to extract text character-level features. Considering that the water pump text information contains a mixture of Chinese and English, CNN is used to process English words composed of more fine-grained letters. For example, CNN is used to extract character-level features of entities such as water pump operation logs and fault cases. Different random character vectors are set for different types of characters to distinguish the case of characters and character types (letters, numbers, punctuation, special characters). For example, uppercase A and lowercase a correspond to two different sets of character vectors. Each word is converted into a corresponding character vector to form a character vector matrix for the word. The longest word is used as the baseline to avoid the problem of different sizes of character vector matrices caused by different word lengths, and placeholders are filled on the left and right ends of the word. In the convolution operation of extracting character-level feature vectors, an output feature map is obtained through an activation function. Each output map can be a combination of the convolution values ​​of multiple input maps. Next, a maximum pooling operation is applied to the feature map, and the average or maximum value is used as the feature corresponding to the filter. Finally, the representation of the word is obtained by connecting all feature maps.

[0049] z (cnn) =f(W (c) *x+b (c) )

[0050] Among them, z (cnn) represents the feature map after convolution, f is the activation function, W (c) is the convolution kernel parameter, b (c) is the bias term, * represents the convolution operation, and x is the input text sequence.

[0051] Bidirectional long short-term memory network (BiLSTM) is used to understand the contextual information of text and can effectively capture the temporal relationship in text sequences. BiLSTM controls the flow of information through gating units and has memory units that can save and transmit long-term dependency information. Considering that LSTM only obtains the previous information, the following information may also be crucial. In order to effectively utilize the contextual information, bidirectional LSTM (BiLSTM) is used to extract features. BiLSTM uses the sequential and reverse order of the input sequence to calculate two different hidden layer representations. The final hidden layer representation is then obtained by vector concatenation. BiLSTM receives the character-level feature vector sequence extracted by CNN as input at each time state, and then concatenates the feature vector sequence forward LSTM and the feature vector sequence backward LSTM for analysis to obtain a complete feature vector sequence.

[0052] Forward sequence: in, is the forward feature vector sequence of the forward LSTM, It is the feature map after convolution.

[0053] Reverse sequence: in, is the reverse feature vector sequence of the reverse LSTM, It is the feature map after convolution.

[0054] The complete eigenvector: Concatenate the feature vector states of the forward and backward LSTMs.

[0055] Conditional Random Field (CRF) is used for sequence labeling tasks to improve the accuracy of labeling by modeling the dependencies between sequence labels. CRF models usually include a state transition matrix and an emission matrix to represent the transition probability and emission probability between labels. CRF is accessed after BiLSTM, and CRF is used to consider the transition probability between labels and calculate the probability of the entire label sequence.

[0056]

[0057] Among them, S(x,y) is the score function, the score of the input sequence x and the label sequence y.

[0058] When calculating the label score corresponding to a character, first extract the information features and contextual relationship features of the manually annotated text set based on expert experience, then score and evaluate the character features in the manually annotated text set, and add its score value to the label score to improve the score of the function. The higher the score, the higher the accuracy of the prediction. Then, the CRF layer is used to calculate the transition probability between labels and the probability of the entire label sequence. In the training phase, the goal of the model is to maximize the logarithmic probability of the correct label sequence. The Viterbi algorithm is used to calculate the optimal label sequence for inference to obtain the optimal label sequence.

[0059] Step 2: Processing of water pump equipment related texts, specifically:

[0060] In this embodiment, the annotated text is preprocessed to obtain preprocessed basic data. These preprocessed data will serve as basic text and play a key role in subsequent deep learning recognition processing.

[0061] The text processing method adopted in this embodiment is specifically as follows:

[0062] The text to be processed is processed with word segmentation and stop word removal in turn, and the processed text is vectorized, the preprocessed text is converted into a vector form to obtain the basic text.

[0063] This embodiment does not further limit the method of word segmentation processing. In this example, the jieba word segmentation library is preferably used to perform text preprocessing on the text to be processed. First, jieba is used to load a large Chinese dictionary, which contains a large number of Chinese words and their corresponding parts of speech, and relevant professional words in the field of feed pumps for thermal power units are supplemented according to professional experience. Then, the input text is segmented according to this dictionary. Two methods, the prefix dictionary and the suffix dictionary, are adopted to match each word in the text with the words in the dictionary, and the word with the largest match is used as the word segmentation result. If no matching word is found in the dictionary, segmentation is performed based on rules. Finally, the preprocessed text is divided into meaningful words and phrases. If the text to be processed is a fault report, the meaningful words may include: words corresponding to types such as feed pump fault description, feed pump fault location, feed pump fault component, feed pump fault repair record, feed pump technical parameters, feed pump model and serial number, feed pump maintenance suggestions, feed pump fault cause analysis, safety warnings, feed pump fault handling opinions, etc.

[0064] This embodiment performs stop word removal processing on the processed text after word segmentation processing and keyword extraction, specifically to remove meaningless words in the text, such as: the word "de" (的), "shi" (是), "zai" (在) for forming words, the conjunction "yiji" (以及), the adverb "shenzhi" (甚至), the modal particle "ba" (吧), etc. Removing these words from the text will not cause any semantic changes. This can not only reduce the dimension of the feature vector and improve the accuracy of extracting knowledge, but also extract the key parts of the text to be processed input by the user through stop word removal, facilitating subsequent vector transformation of the extracted text to obtain the basic text.

[0065] This embodiment obtains the corresponding basic text by converting the processed text into a vector form, providing a data basis for subsequent deep learning recognition processing of the basic text and facilitating subsequent deep learning recognition processing of the basic text.

[0066] This embodiment does not further limit the specific implementation method of vector transformation. In this example, it is preferably to convert the text into a vector form through the Word2Vec model to obtain the basic text.

[0067] The operation logs, equipment product documents, equipment defect files, fault cases and other texts related to the feedwater pumps in the thermal power units are divided into keywords, and after the stop words are removed, they are converted into a format that can be recognized by the Word2Vec algorithm. Considering that fault texts usually involve a wealth of technical terms and domain-specific terms, the Skip-gram model can better capture the semantic relationship and contextual information between words when processing large-scale corpora, so as to better learn the representation of word vectors. Therefore, the Skip-gram model in the Word2Vec algorithm is used to train the processed text. During the training process, the context or target word of each word is predicted according to the given context window size, and then the word vector is updated by optimizing the objective function. After the training is completed, each word is mapped to a vector space of fixed dimension. These word vectors usually have semantic information, so that similar words are close in the vector space, while words with different semantics are far away in the vector space.

[0068] Step 3: Entity recognition and relationship extraction: Through deep learning neural networks, entity recognition and relationship extraction are performed on the processed text to extract entities and relationships between entities, specifically:

[0069] Reference Figure 3 , this embodiment adopts the BiLSTM model, i.e., a deep learning neural network, to perform feature extraction on the word vectors obtained above, and then obtain the contextual relationship corresponding to each word vector. Among them, the BiLSTM model is specifically a bidirectional long short-term memory network (Bi-directional Long Short-Term Memory) model. As a bidirectional recurrent neural network model, feature information can be extracted for the contextual relationship between two word vectors. The BiLSTM model uses gated recurrent units (Gated Recurrent Units) instead of LSTM, which effectively avoids the problem of gradient disappearance. The gated recurrent unit, i.e., GRU, includes two LSTM layers, a forward LSTM and a reverse LSTM, wherein the forward LSTM is used to learn the forward features of the text sequence, and the reverse LSTM is used to learn the reverse features of the text sequence. At the same time, through forward propagation and back propagation, feature extraction is performed on each word vector to obtain the contextual relationship corresponding to each word vector, i.e., the contextual characteristic representation.

[0070] Determine and extract the entity word vector in the word vector according to the contextual relationship, and use the entity word vector as the first entity. By analyzing the contextual relationship corresponding to each word vector, determine whether the word vector is an entity word vector or a relationship word vector, and extract the word vector determined as the entity word vector as the first entity, providing a parameter basis for the subsequent retrieval process.

[0071] Step 4: Knowledge fusion: Use knowledge fusion methods to eliminate redundant, repeated or contradictory information from different text sources and improve the accuracy and consistency of knowledge. Specifically:

[0072] Use entity alignment methods to match or align items representing the same entity to eliminate duplication or redundancy of entity information; use relationship fusion to integrate relationship information from different data sources, resolve differences in relationship semantics, and map relationships to a unified representation space; use attribute fusion to integrate attribute information from different data sources, resolve differences in relationship semantics, and map attributes to a unified representation space; use semantic fusion to integrate semantic information from different data sources, resolve differences in relationship semantics, and map semantic information to a unified semantic space.

[0073] Finally, the fused knowledge is evaluated based on expert experience to verify whether the fusion results meet the expected needs and goals.

[0074] Step 5: Import and store the Neo4j graph database. By organizing the documents of the power plant feed water pump, the trained model is used to automatically extract entity and relationship knowledge, construct triple data, and import it into the Neo4j graph database. Specifically:

[0075] First, all available documents for the power plant feedwater pump are sorted out, including equipment product documents, technical system standards, equipment defect files, failure cases, management regulations, operating procedures, maintenance procedures, maintenance task work orders, etc., and the document records with errors or incompleteness in the original document data are filtered out. The specific operations include clearing null value data and invalid data, checking data consistency, etc. Then, the trained entity extraction model and relationship extraction model are used to automatically extract the corresponding types of entity knowledge and relationship knowledge in the text set, and construct the triple data of the feedwater pump fault knowledge graph. Finally, the Cypher language is used to batch import and store the knowledge triple data into the Neo4j graph database.

[0076] Example 2

[0077] The knowledge graph-based auxiliary decision-making system for feedwater pump failure of a thermal power plant in this embodiment includes:

[0078] Text annotation module: Automatically annotate text datasets based on manual annotation based on expert experience combined with deep learning. The annotation objects are the entity types and relationship types to be identified. The corpus of entity and relationship knowledge in the annotated text is divided into a training text set and a test text set.

[0079] Text processing module: preprocess the annotated text to obtain preprocessed basic data. These preprocessed data will serve as basic text and play a key role in subsequent deep learning recognition processing.

[0080] Entity recognition and relationship extraction module: Through deep learning neural network, entity recognition and relationship extraction are performed on the processed text to extract entities and the relationships between entities.

[0081] Knowledge fusion module: Use knowledge fusion methods to eliminate redundant, repeated or contradictory information from different text sources and improve the accuracy and consistency of knowledge.

[0082] Triple data construction module: By organizing the documents of the power plant feed water pump, the trained model is used to automatically extract entity and relationship knowledge, construct triple data, and import it into the Neo4j graph database.

[0083] The present invention provides a method and system for auxiliary decision-making of feedwater pump faults of thermal power plants based on knowledge graphs, which obtains basic texts by preprocessing texts related to feedwater pumps of thermal power plants, and obtains entities and entity relationships by using deep learning methods to identify and process texts; and obtains corresponding information by retrieving entities and entity relationships through retrieval models. The present invention performs secondary simplification and clarification of user input questions through text preprocessing and deep learning models, thereby improving the accuracy and efficiency of obtaining entities and entity relationships. The knowledge graph fault diagnosis of feedwater pumps of thermal power plants is of great significance in improving fault diagnosis efficiency, preventive maintenance, optimizing maintenance strategies, improving equipment safety and stability, and promoting technological innovation and knowledge sharing, and plays a key role in ensuring the safe and stable operation of thermal power plant equipment and improving production efficiency.

Claims

1. A knowledge graph-based auxiliary decision-making method for feedwater pump failure in thermal power plants, characterized in that: The following steps are involved: Based on expert experience and deep learning models, the text dataset of the feedwater pump of the thermal power plant is automatically annotated, and the annotated text is divided into a training set and a test set; Preprocess the training set and test set; Through deep learning neural networks, the processed text is subjected to entity recognition and relationship extraction to extract entities and relationships between entities; Use knowledge fusion methods to eliminate redundant, repeated or contradictory information from different text sources; By organizing the power plant feed water pump documents, the trained model is used to automatically extract entity and relationship knowledge, construct triple data, and import it into the database.

2. According to the knowledge graph-based auxiliary decision-making method for feedwater pump failure of a thermal power plant unit according to claim 1, it is characterized in that: The training set and the test set are obtained by dividing the corpus of entity and relational knowledge in the annotated text.

3. According to claim 2, a knowledge graph-based auxiliary decision-making method for feedwater pump failure of a thermal power plant, characterized in that: The text data set includes the operation log of the feedwater pump of the thermal power unit, equipment product documents, equipment defect files, and failure cases.

4. The method for auxiliary decision-making of feedwater pump failure of a thermal power unit based on knowledge graph according to claim 1 or 2, characterized in that: The preprocessing includes word segmentation and stop word filtering, and vector conversion is performed on the processed text, the preprocessed text is converted into a vector form, and the basic text is obtained.

5. According to claim 4, a knowledge graph-based auxiliary decision-making method for feedwater pump failure of a thermal power plant is characterized in that: The entity recognition and relationship extraction of the processed text is specifically: using a deep learning neural network to extract features from the obtained word vectors, and then obtaining the contextual relationship corresponding to each word vector.

6. The method for auxiliary decision-making of feedwater pump failure of thermal power unit based on knowledge graph according to claim 5 is characterized in that: The entity word vector in the word vector is determined and extracted according to the context relationship, and the entity word vector is used as the first entity.

7. The method for auxiliary decision-making of feedwater pump failure of thermal power plant based on knowledge graph according to claim 6 is characterized in that: The specific method for obtaining the first entity is: by parsing the context relationship corresponding to each word vector, determining whether the word vector is an entity word vector or a relationship word vector, and extracting and collecting the word vectors determined as entity word vectors as the first entity, providing a parameter basis for the subsequent retrieval process.

8. The method for auxiliary decision-making of feedwater pump failure of thermal power plant based on knowledge graph according to claim 7 is characterized in that: The knowledge fusion is specifically as follows: using an entity alignment method to match or align items representing the same entity to eliminate duplication or redundancy of entity information; using relationship fusion to integrate relationship information from different data sources and map the relationship into a unified representation space; Select attribute fusion to integrate attribute information from different data sources and map the attributes into a unified representation space; Semantic fusion is used to integrate semantic information from different data sources and map the semantic information into a unified semantic space; the fused knowledge is evaluated based on expert experience to verify whether the fusion results meet the expected needs and goals.

9. The method for auxiliary decision-making of feedwater pump failure of thermal power plant based on knowledge graph according to claim 8, characterized in that: The construction of triple data is specifically as follows: All available documents on the power plant's feedwater pumps are sorted out, and erroneous or incomplete document records in the original document data are filtered out. The trained entity extraction model and relationship extraction model are used to automatically extract the corresponding types of entity knowledge and relationship knowledge in the text set to construct the triple data of the feedwater pump fault knowledge graph.

10. A knowledge graph-based auxiliary decision-making system for feedwater pump failure in thermal power plants, characterized in that: include: Text annotation module: Automatically annotate text data sets based on manual annotation based on expert experience and deep learning, and divide the annotated text into training sets and test sets; Text processing module: performs text preprocessing on the annotated text to obtain preprocessed basic data; Entity recognition and relationship extraction module: Through deep learning neural network, the processed text is subjected to entity recognition and relationship extraction to extract entities and the relationships between entities; Knowledge fusion module: uses knowledge fusion methods to eliminate redundant, repeated or contradictory information from different text sources; Triple data construction module: By organizing the documents of the power plant feed water pump, the trained model is used to automatically extract entity and relationship knowledge, construct triple data, and import it into the database.

Images