A method for grid device-oriented word collaborative entity recognition
By employing a word-based collaborative entity recognition method, and utilizing bidirectional long short-term memory networks and graph convolutional neural networks, the contextual dependencies of power grid equipment text are established, solving the accuracy problem of power grid equipment entity boundary recognition and achieving higher recognition accuracy and adaptability.
Patent Information
- Application Number
- CN202610859769.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-15
- Publication Date
- 2026-11-06
AI Technical Summary
In the power grid field, existing technologies struggle to accurately identify the physical boundaries of power grid equipment, and are affected by word segmentation errors and technical terms, resulting in low accuracy in entity recognition.
A word-cooperative entity recognition method is adopted. By constructing an embedding layer, a multi-layer network update layer and a decoding layer, and utilizing a bidirectional long short-term memory network and a graph convolutional neural network, the contextual dependencies of characters, self-matching words and word combination sequences are established. Finally, a graph attention neural network is used for feature fusion and label decoding.
It improves the accuracy of power grid equipment entity recognition, can adapt to different types of unstructured text, and has good scenario applicability and generalization ability.
Smart Images
Figure CN123050428A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of entity recognition technology for power grid equipment, and in particular to a word-based collaborative entity recognition method for power grid equipment. Background Technology
[0002] With the continuous advancement of new power systems, the scale of power grid equipment is constantly expanding, and the types of equipment are becoming increasingly complex. This results in a large amount of textual data generated during operation and maintenance, such as dispatch logs, defect records, maintenance reports, and repair work orders. These texts contain rich equipment and operational status information, which is of significant value for power grid safety analysis and intelligent operation and maintenance. Therefore, how to automatically extract relevant entities of power grid equipment from massive amounts of unstructured text has become one of the key issues in the development of power informatization and intelligentization.
[0003] Power grid equipment entity recognition is a specific application of named entity recognition in the power sector. Its goal is to identify entities with specific semantic categories from text, such as equipment names, equipment types, operating status, and geographical locations. However, compared to general domains, power grid text is significantly more specialized and complex. On the one hand, equipment names often exhibit complex structures, varying lengths, and flexible combinations, making it difficult to accurately define entity boundaries. On the other hand, the mixture of numerous technical terms, abbreviations, and symbols increases the difficulty of semantic understanding.
[0004] In recent years, with the development of deep learning technology, entity recognition methods based on neural networks have gradually become mainstream. For example, methods based on the BiLSTM-CRF structure can automatically learn contextual features, improving recognition performance to some extent. However, in Chinese power grid texts, due to the lack of explicit word boundaries, relying solely on character-level modeling easily overlooks semantic information at the lexical level; while methods relying on word segmentation results are easily affected by segmentation errors, leading to error propagation. Furthermore, the power grid field contains a large number of specialized terms composed of multiple characters, whose semantic information is difficult to fully express through single-granularity modeling. Summary of the Invention
[0005] To address the shortcomings of traditional word-assisted mechanisms in establishing sufficient contextual relationships between self-matching words and sentence sentences, and the lack of effective constraints in word boundary recognition, which can easily lead to blurred boundaries or splitting errors, this invention aims to provide a word-based collaborative entity recognition method for power grid equipment that enables words to establish continuous contextual dependencies with surrounding characters or other words, thereby improving the accuracy of entity recognition.
[0006] To achieve the above objectives, the present invention adopts the following technical solution: a word-based collaborative entity recognition method for power grid equipment, the method comprising the following sequential steps:
[0007] (1) Collect unstructured text data related to power grid equipment and preprocess it to obtain a power dataset; divide the power dataset into a training set, a validation set and a test set;
[0008] (2) Construct a power equipment named entity recognition model. The power equipment named entity recognition model includes an embedding layer, a multi-layer network update layer and a decoding layer. The embedding layer outputs a character vector sequence H, a self-matching vocabulary vector sequence U, and a global semantic vector set. The output of the multi-layer network update layer is then fed into the decoding layer, where the final character vector sequence R is output.
[0009] (3) Input the training set into the power equipment named entity recognition model for training to obtain the trained model;
[0010] (4) Input the unstructured text data to be identified into the trained model to obtain the recognition result.
[0011] Step (1) specifically includes the following steps in sequence:
[0012] (1a) Collect unstructured text data related to power grid equipment, including inspection records, defect reports, operation and maintenance logs and technical specifications;
[0013] (1b) Preprocess the unstructured text data, including cleaning, deduplication and annotation. After cleaning and deduplication, a high-quality text corpus is obtained. Each character of the text in the text corpus is annotated with BMEO tags to obtain the corresponding power dataset. In BMEO tags, B represents the starting character of the power grid equipment entity, M represents the middle character of the power grid equipment entity, E represents the ending character of the power grid equipment entity, and O represents a non-entity character.
[0014] (1c) Divide the power dataset into training set, validation set and test set.
[0015] In step (2), the embedding layer includes a bidirectional long short-term memory network and a word combination sequence construction module. The processing method of the embedding layer includes the following steps:
[0016] (3a) Take the character sequence from the training set as input, denoted as ={ , ,… }, Let c be a sequence of characters, and n be the number of characters in c. It is the last character in c; at the same time, substring matching is performed on the character sequence to extract all characters that can be matched with the character sequence. Find words that completely match consecutive substrings and construct a self-matching word sequence. ={ , ,… }, where m is a self-matching word sequence The number of self-matching words in the text. It is the last self-matching word in l; each is accessed through a character embedding lookup table. Lexical Embedding Lookup Table , character sequence Each character in the sequence of self-matching words Each word in the text is mapped to a low-dimensional vector, resulting in a character vector. and vocabulary vectors :
[0017] ;
[0018] ;
[0019] in, It is the i-th character in c. It is the j-th self-matching word in l;
[0020] (3b) Character vector The contextual features are extracted from the input bidirectional long short-term memory network, which includes a forward long short-term memory network and a backward long short-term memory network. The hidden states are calculated using the forward long short-term memory network and the backward long short-term memory network, respectively.
[0021] = LSTM1( , );
[0022] = LSTM2( , );
[0023] In the formula, LSTM1 is a forward long short-term memory network, and LSTM2 is a backward long short-term memory network. It is the first forward long short-term memory network. The hidden state of the character vector output at each position. It is the backward long short-term memory network in the first... The hidden state of the character vector output at each position. It is the first in the feedforward long short-term memory network The hidden state of the character vector at each position. This represents the first [unit / item] in the backward long short-term memory network. The hidden state of the character vector at each position;
[0024] Will and By splicing, we obtain the first... The first context feature vector at each position :
[0025] ;
[0026] In the formula, This indicates vector concatenation;
[0027] The final character vector sequence H is obtained as follows: , ,…, ,…, );
[0028] vocabulary vectors The contextual features are extracted using a bidirectional long short-term memory network, and the hidden states are calculated using a forward long short-term memory network and a backward long short-term memory network, respectively.
[0029] = LSTM1( , );
[0030] = LSTM2( , );
[0031] In the formula, It is the first forward long short-term memory network. The hidden state of the word vector output at each position. It is the backward long short-term memory network in the first... The hidden state of the word vector output at each position. It is the first in the feedforward long short-term memory network The hidden state of the word vector at each position. This represents the first [unit / item] in the backward long short-term memory network. The hidden state of the word vector at each position;
[0032] Will and By splicing, we obtain the first... The second context feature vector of each character :
[0033] ;
[0034] The final result is a self-matching vocabulary vector sequence U = ( ) containing contextual features. , ,…, ,…, );
[0035] (3c) Use H and U in the word combination sequence construction module to construct the word combination vector sequence V={ , ,…, ,…, }, s≤n, where s is the number of elements in V. It is the first in V There are elements, among which... Satisfy: When a continuous segment of characters is matched by a self-matching word When full coverage is achieved, the second context feature vector is used. Replace this continuous character segment, that is = Otherwise, the original character vector representation is retained, i.e. = ; ∈U, while maintaining the original order, achieves the fusion expression of information at different granularities.
[0036] The word combination vector sequence V is input into a bidirectional long short-term memory network to extract contextual features. The hidden states are first calculated through a forward long short-term memory network and a backward long short-term memory network, respectively.
[0037] =LSTM1( , );
[0038] =LSTM2( , );
[0039] In the formula, It is the first forward long short-term memory network. The hidden state of the word combination vector sequence V output at each position. It is the backward long short-term memory network in the first... The hidden state of the word combination vector sequence V output at each position. It is the first in the feedforward long short-term memory network The hidden state of the word combination vector sequence V at each position. This represents the first [unit / item] in the backward long short-term memory network. The hidden state of the word combination vector sequence V at each position;
[0040] Will and By splicing, we can obtain the first The third context feature vector at each position :
[0041] ;
[0042] Based on the third context feature vector Finally, a new word combination vector sequence Z = ( , ,…, ,…, );
[0043] Perform vector concatenation on the final hidden states at both ends of the new word combination vector sequence Z to obtain the global semantic representation of the new word combination vector sequence Z. :
[0044] ;
[0045] In the formula, It is the hidden state output by the feedforward long short-term memory network at the last moment of the new word combination vector sequence Z. It is the hidden state output by the backward long short-term memory network at the last moment of the new word combination vector sequence Z;
[0046] Use multiple different Construct a global semantic vector set Then H, U, It is output to the multi-layer network update layer.
[0047] The multi-layer network update layer includes a graph convolutional neural network and a graph attention neural network, which construct an interactive graph adjacency matrix of self-matching words and word combination sequences. Self-matching vocabulary and character containment graph adjacency matrix Transformation graph adjacency matrix Adjacency matrix of Lattice graph ;
[0048] In step (2), the processing method of the multi-layer network update layer includes the following steps:
[0049] (4a) Put U and Perform vector concatenation to obtain Initial node features :
[0050] = ;
[0051] U and As input to the graph convolutional neural network, the representation after feature extraction and information fusion is obtained through the graph convolutional neural network. :
[0052] ;
[0053] In the formula, For graph convolutional neural networks;
[0054] from Extract the corresponding self-matching vocabulary parts to obtain a new self-matching vocabulary vector sequence. :
[0055] ;
[0056] Where m is the self-matching word sequence The number of self-matching words in the text;
[0057] (4b) The new self-matching word vector sequence Concatenating the character vector sequence H with the input feature Node of the self-matching vocabulary and word interaction graph is obtained:
[0058] Node = (H, );
[0059] Use 3 independent graph attention neural networks , and The node information of three different self-matching words and character interaction graphs is fused. The initial node input of the three graphs is Node:
[0060] = (Node, );
[0061] = (Node, );
[0062] = (Node, );
[0063] in, This is the full node representation of the output of a graph attention neural network. ; yes The output node representation, yes The output node representation, yes The output node representation; F is The size of the vector dimension, C is The corresponding self-matching vocabulary and character containment graph, T is The corresponding transformation graph of self-matching words and characters, L is The corresponding Lattice diagram of self-matching words and characters; e is C, T, The enumeration identifier; m is the self-matching word sequence. The number of self-matching words in the sequence, where n is the number of characters in character sequence c;
[0064] from Extracting character-level representations :
[0065] ;
[0066] (4c) Introduce a fusion mechanism to uniformly model information from different sources, and... The final character vector sequence is obtained by weighting and fusing multiple features using a weight matrix with H. , :
[0067] = H+ + + ;
[0068] in, , , and All of these are weight parameter matrices to be learned; yes The character representation, yes The character representation, yes The character representation;
[0069] Then Output to the decoding layer.
[0070] In step (2), the decoding layer includes a conditional random field, and the specific processing method of the decoding layer refers to:
[0071] Label decoding of R is performed using a conditional random field, assuming the input character sequence is... ={ , ,… The corresponding target label sequence is d = ( , ,…, Given the input conditions, what is the conditional probability of d occurring? Represented as:
[0072] = ;
[0073] in, Represents the candidate label sequence; yes The Middle Tag The corresponding launch parameter weights; yes From the previous tag Move to current tag The transfer score; yes The Middle Tag The corresponding launch parameter weights; yes From the previous tag Move to current tag The transfer score; yes Global overall feature representation; It is an exponential function; yes The Middle Location tags; yes The Middle Location tags; yes The Middle A character vector;
[0074] Select the label sequence with the highest probability, label each character with a BMEO label, and output the final power grid equipment entity recognition result.
[0075] As can be seen from the above technical solution, the present invention has the following beneficial effects: First, by constructing a word combination sequence, the present invention directly embeds words into the original temporal structure, enabling words to establish continuous contextual dependencies with surrounding characters or other words, thereby obtaining a representation that is closer to the actual semantic environment; Second, the present invention improves the boundary determination capability of self-matching words. Multiple word combination forms, through different granularity construction methods, provide multi-perspective boundary information support for words, thereby helping the model to more accurately determine the entity range; Third, the present invention introduces graph convolutional neural networks and graph attention neural networks, and builds various interactive graph structures, fully... The invention explores the relationships between characters, self-matching words, and word combination sequences to achieve deep interaction and fusion of multi-dimensional information, effectively enhancing the text feature expression capability. This allows the model to fully capture the complex semantic relationships and structural features in the text of power grid equipment, further improving the accuracy of entity recognition. Fourth, the invention adopts a multi-graph collaborative feature extraction method, analyzing text information from different perspectives such as attribution relationships, neighborhood semantics, and multi-granular structure. Combined with a learnable weight matrix, it completes multi-source feature adaptive fusion, which can adapt to different types of unstructured text such as power grid operation and maintenance records, defect reports, and inspection logs, and has good scenario applicability and generalization ability. Attached Figure Description
[0076] Figure 1 This is a diagram of the overall network architecture of the present invention;
[0077] Figure 2 A sequence diagram of word combinations;
[0078] Figure 3 This is a word / phrase collaboration diagram. Detailed Implementation
[0079] like Figure 1 As shown, a word-based collaborative entity recognition method for power grid equipment includes the following sequential steps:
[0080] (1) Collect unstructured text data related to power grid equipment and preprocess it to obtain a power dataset; divide the power dataset into a training set, a validation set and a test set;
[0081] (2) Construct a power equipment named entity recognition model. The power equipment named entity recognition model includes an embedding layer, a multi-layer network update layer and a decoding layer. The embedding layer outputs a character vector sequence H, a self-matching vocabulary vector sequence U, and a global semantic vector set. The output of the multi-layer network update layer is then fed into the decoding layer, where the final character vector sequence R is output.
[0082] (3) Input the training set into the power equipment named entity recognition model for training to obtain the trained model;
[0083] (4) Input the unstructured text data to be identified into the trained model to obtain the recognition result.
[0084] Step (1) specifically includes the following steps in sequence:
[0085] (1a) Collect unstructured text data related to power grid equipment, including inspection records, defect reports, operation and maintenance logs and technical specifications;
[0086] (1b) Preprocess the unstructured text data, including cleaning, deduplication and annotation. After cleaning and deduplication, a high-quality text corpus is obtained. Each character of the text in the text corpus is annotated with BMEO tags to obtain the corresponding power dataset. In BMEO tags, B represents the starting character of the power grid equipment entity, M represents the middle character of the power grid equipment entity, E represents the ending character of the power grid equipment entity, and O represents a non-entity character.
[0087] (1c) Divide the power dataset into training set, validation set and test set.
[0088] In step (2), the embedding layer includes a bidirectional long short-term memory network (BiLSTM) and a word combination sequence construction module. The processing method of the embedding layer includes the following steps:
[0089] (3a) Take the character sequence from the training set as input, denoted as ={ , ,… }, Let c be a sequence of characters, and n be the number of characters in c. It is the last character in c; at the same time, substring matching is performed on the character sequence to extract all characters that can be matched with the character sequence. Find words that completely match consecutive substrings and construct a self-matching word sequence. ={ , ,… }, where m is a self-matching word sequence The number of self-matching words in the text. It is the last self-matching word in l; each is accessed through a character embedding lookup table. Lexical Embedding Lookup Table , character sequence Each character in the sequence of self-matching words Each word in the text is mapped to a low-dimensional vector, resulting in a character vector. and vocabulary vectors :
[0090] ;
[0091] ;
[0092] in, It is the i-th character in c. It is the j-th self-matching word in l;
[0093] (3b) Character vector The contextual features are extracted from the input bidirectional long short-term memory network, which includes a forward long short-term memory network and a backward long short-term memory network. The hidden states are calculated using the forward long short-term memory network and the backward long short-term memory network, respectively.
[0094] = LSTM1( , );
[0095] = LSTM2( , );
[0096] In the formula, LSTM1 is a forward long short-term memory network, and LSTM2 is a backward long short-term memory network. It is the first forward long short-term memory network. The hidden state of the character vector output at each position. It is the backward long short-term memory network in the first... The hidden state of the character vector output at each position. It is the first in the feedforward long short-term memory network The hidden state of the character vector at each position. This represents the first [unit / item] in the backward long short-term memory network. The hidden state of the character vector at each position;
[0097] Will and By splicing, we obtain the first... The first context feature vector at each position :
[0098] ;
[0099] In the formula, This indicates vector concatenation;
[0100] The final character vector sequence H is obtained as follows: , ,…, ,…, );
[0101] vocabulary vectors The contextual features are extracted using a bidirectional long short-term memory network, and the hidden states are calculated using a forward long short-term memory network and a backward long short-term memory network, respectively.
[0102] = LSTM1( , );
[0103] = LSTM2( , );
[0104] In the formula, It is the first forward long short-term memory network. The hidden state of the word vector output at each position. It is the backward long short-term memory network in the first... The hidden state of the word vector output at each position. It is the first in the feedforward long short-term memory network The hidden state of the word vector at each position. This represents the first [unit / item] in the backward long short-term memory network. The hidden state of the word vector at each position;
[0105] Will and By splicing, we obtain the first... The second context feature vector of each character :
[0106] ;
[0107] The final result is a self-matching vocabulary vector sequence U = ( ) containing contextual features. , ,…, ,…, );
[0108] (3c) such as Figure 2 As shown, in the word combination sequence construction module, H and U are used to construct the word combination vector sequence V={ , ,…, ,…, }, s≤n, where s is the number of elements in V. It is the first in V There are elements, among which... Satisfy: When a continuous segment of characters is matched by a self-matching word When full coverage is achieved, the second context feature vector is used. Replace this continuous character segment, that is = Otherwise, the original character vector representation is retained, i.e. = ; ∈U;
[0109] The word combination vector sequence V is input into a bidirectional long short-term memory network to extract contextual features. The hidden states are first calculated through a forward long short-term memory network and a backward long short-term memory network, respectively.
[0110] =LSTM1( , );
[0111] =LSTM2( , );
[0112] In the formula, It is the first forward long short-term memory network. The hidden state of the word combination vector sequence V output at each position. It is the backward long short-term memory network in the first... The hidden state of the word combination vector sequence V output at each position. It is the first in the feedforward long short-term memory network The hidden state of the word combination vector sequence V at each position. This represents the first [unit / item] in the backward long short-term memory network. The hidden state of the word combination vector sequence V at each position;
[0113] Will and By splicing, we can obtain the first The third context feature vector at each position :
[0114] ;
[0115] Based on the third context feature vector Finally, a new word combination vector sequence Z = ( , ,…, ,…, );
[0116] Perform vector concatenation on the final hidden states at both ends of the new word combination vector sequence Z to obtain the global semantic representation of the new word combination vector sequence Z. :
[0117] = ;
[0118] In the formula, It is the hidden state output by the feedforward long short-term memory network at the last moment of the new word combination vector sequence Z. It is the hidden state output by the backward long short-term memory network at the last moment of the new word combination vector sequence Z;
[0119] Use multiple different Construct a global semantic vector set Then H, U, It is output to the multi-layer network update layer.
[0120] The multi-layer network update layer includes a graph convolutional neural network and a graph attention neural network, which construct an interactive graph adjacency matrix of self-matching words and word combination sequences. Self-matching vocabulary and character containment graph adjacency matrix Transformation graph adjacency matrix Adjacency matrix of Lattice graph Composition method such as Figure 3 As shown.
[0121] The processing method for the multi-layer network update layer includes the following steps:
[0122] (4a) Put U and Perform vector concatenation to obtain Initial node features :
[0123] = ;
[0124] U and As input to the graph convolutional neural network, the representation after feature extraction and information fusion is obtained through the graph convolutional neural network. :
[0125] ;
[0126] In the formula, For graph convolutional neural networks;
[0127] from Extract the corresponding self-matching vocabulary parts to obtain a new self-matching vocabulary vector sequence. :
[0128] ;
[0129] Where m is the self-matching word sequence The number of self-matching words in the text;
[0130] (4b) The new self-matching word vector sequence Concatenating the character vector sequence H with the input feature Node of the self-matching vocabulary and word interaction graph is obtained:
[0131] Node = (H, );
[0132] Use 3 independent graph attention neural networks , and The node information of three different self-matching words and character interaction graphs is fused. The initial node input of the three graphs is Node:
[0133] = (Node, );
[0134] = (Node, );
[0135] = (Node, );
[0136] in, This is the full node representation of the output of a graph attention neural network. ; yes The output node representation, yes The output node representation, yes The output node representation; F is The size of the vector dimension, C is The corresponding self-matching vocabulary and character containment graph, T is The corresponding transformation graph of self-matching words and characters, L is The corresponding Lattice diagram of self-matching words and characters; e is C, T, The enumeration identifier; m is the self-matching word sequence. The number of self-matching words in the sequence, where n is the number of characters in character sequence c;
[0137] from Extracting character-level representations :
[0138] ;
[0139] (4c) Introduce a fusion mechanism to uniformly model information from different sources, and... The final character vector sequence is obtained by weighting and fusing multiple features using a weight matrix with H. , :
[0140] = H+ + + ;
[0141] in, , , and All of these are weight parameter matrices to be learned; yes The character representation, yes The character representation, yes The character representation;
[0142] Then Output to the decoding layer.
[0143] The decoding layer includes a conditional random field, and the specific processing method of the decoding layer refers to:
[0144] Label decoding of R is performed using a conditional random field, assuming the input character sequence is... ={ , ,… The corresponding target label sequence is d = ( , ,…, Given the input conditions, what is the conditional probability of d occurring? Represented as:
[0145] = ;
[0146] in, Represents the candidate label sequence; yes The Middle Tag The corresponding launch parameter weights; yes From the previous tag Move to current tag The transfer score; yes The Middle Tag The corresponding launch parameter weights; yes From the previous tag Move to current tag The transfer score; yes Global overall feature representation; It is an exponential function; yes The Middle Location tags; yes The Middle Location tags; yes The Middle A character vector;
[0147] Select the label sequence with the highest probability, label each character with a BMEO label, and output the final power grid equipment entity recognition result.
[0148] like Figure 2 As shown, firstly, based on the input sentence, the original character sequence and the self-matched words obtained through dictionary matching are processed by BiLSTM to extract contextual features. Then, a word combination sequence is constructed. Character information preserves syntactic structure in a fine-grained manner, while self-matched word information supplements contextual expression with higher semantic granularity. Based on this, BiLSTM is used to extract contextual features from the word combination sequence, obtaining a contextual representation containing both forward and backward semantic information. Simultaneously, an interaction graph structure is further constructed to connect character nodes and self-matched word nodes, explicitly modeling the semantic association between them. Through information propagation and feature fusion in the graph structure, the model can comprehensively utilize multi-granular semantic information to achieve accurate identification of entity boundaries and categories in the sentence. Finally, the sequence labeling results are output through the fused feature representation.
[0149] like Figure 3 As shown, the collaboration graph, from top to bottom, includes an interaction graph of self-matching words and word combination sequences, an inclusion graph of self-matching words and characters, a transformation graph, and a Lattice graph.
[0150] In the self-matching vocabulary and word combination sequence interaction graph, taking the input sentence "water, electricity, station, generate, electricity, machine, group, sustain, continue, of, transmit, electricity, power" as an example, the dictionary matching obtains words such as "hydropower station", "power generation", "generator set", "unit", "sustained", "transmission", and "electricity", and constructs multiple word combination sequences. Different sequences establish cross-sequence connections through shared characters and vocabulary, thus forming a multi-path interaction structure.
[0151] In the inclusion graph, self-matching word nodes are connected to the character nodes they contain. For example, “hydropower station” is included with “water”, “electricity”, and “station”, and “generator set” is connected with “generator”, “electricity”, “machine”, and “set”. This is used to explicitly model the semantic coverage relationship of self-matching words to characters.
[0152] In the transformation graph, by introducing global connections between self-matching words and characters, the positional relationships in the character sequence are aligned with the semantics of self-matching words, thereby realizing the propagation and mapping of self-matching word information to the character layer and enhancing the semantic expressive power of character-level representation.
[0153] In the Lattice graph, a lattice-like graph representation is constructed by combining the character sequence structure and the self-matching word matching results. This integrates character sequence connections with self-matching word span connections, enabling the model to introduce long-distance word-level dependencies while preserving the original sequence order.
[0154] In summary, this invention constructs word combination sequences, directly embedding words into the original temporal structure, enabling words to establish continuous contextual dependencies with surrounding characters or other words, thereby obtaining a representation that more closely resembles the actual semantic environment. This invention enhances the boundary determination capability of self-matching words; various word combination forms, constructed with different granularities, provide multi-perspective boundary information support for words, thus helping the model to more accurately determine the entity range. This invention introduces graph convolutional neural networks and graph attention neural networks, and constructs various interactive graph structures to fully explore the relationships between characters, self-matching words, and word combination sequences, achieving deep interaction and fusion of multi-dimensional information, effectively strengthening the text feature expression capability, allowing the model to fully capture the complex semantic relationships and structural features in power grid equipment text, further improving the accuracy of entity recognition. This invention adopts a multi-graph collaborative feature extraction method, analyzing text information from different perspectives such as attribution relationships, neighborhood semantics, and multi-granularity structures, and combining a learnable weight matrix to complete multi-source feature adaptive fusion, adapting to different types of unstructured text such as power grid operation and maintenance records, defect reports, and inspection logs, possessing good scenario applicability and generalization ability.
[0155] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.
Claims
1. A method for grid device oriented word co-entity recognition, the method comprising: The method includes the following steps in sequence: (1) Collect unstructured text data related to power grid equipment and preprocess it to obtain a power dataset; divide the power dataset into a training set, a validation set and a test set; (2) Construct a power equipment named entity recognition model. The power equipment named entity recognition model includes an embedding layer, a multi-layer network update layer and a decoding layer. The embedding layer outputs a character vector sequence H, a self-matching vocabulary vector sequence U, and a global semantic vector set. The output of the multi-layer network update layer is then fed into the decoding layer, where the final character vector sequence R is output. (3) Input the training set into the power equipment named entity recognition model for training to obtain the trained model; (4) Input the unstructured text data to be identified into the trained model to obtain the recognition result.
2. The word-based collaborative entity recognition method for power grid equipment according to claim 1, characterized in that: Step (1) specifically includes the following steps in sequence: (1a) Collect unstructured text data related to power grid equipment, including inspection records, defect reports, operation and maintenance logs and technical specifications; (1b) Preprocess the unstructured text data, including cleaning, deduplication and annotation. After cleaning and deduplication, a high-quality text corpus is obtained. Each character of the text in the text corpus is annotated with BMEO tags to obtain the corresponding power dataset. In BMEO tags, B represents the starting character of the power grid equipment entity, M represents the middle character of the power grid equipment entity, E represents the ending character of the power grid equipment entity, and O represents a non-entity character. (1c) Divide the power dataset into training set, validation set and test set.
3. The word-based collaborative entity recognition method for power grid equipment according to claim 1, characterized in that: In step (2), the embedding layer includes a bidirectional long short-term memory network and a word combination sequence construction module. The processing method of the embedding layer includes the following steps: (3a) Take the character sequence from the training set as input, denoted as ={ , ,… }, Let c be a sequence of characters, and n be the number of characters in c. It is the last character in c; at the same time, substring matching is performed on the character sequence to extract all characters that can be matched with the character sequence. Find words that completely match consecutive substrings and construct a self-matching word sequence. ={ , ,… }, where m is a self-matching vocabulary sequence The number of self-matching words in the text. It is the last self-matching word in l; each is accessed through a character embedding lookup table. Lexical Embedding Lookup Table , character sequence Each character in the sequence of self-matching words Each word in the text is mapped to a low-dimensional vector, resulting in a character vector. and vocabulary vectors : ； ； in, It is the i-th character in c. It is the j-th self-matching word in l; (3b) Character vector The contextual features are extracted from the input bidirectional long short-term memory network, which includes a forward long short-term memory network and a backward long short-term memory network. The hidden states are calculated using the forward long short-term memory network and the backward long short-term memory network, respectively. = LSTM1( , ); = LSTM2( , ); In the formula, LSTM1 is a forward long short-term memory network, and LSTM2 is a backward long short-term memory network. It is the first forward long short-term memory network. The hidden state of the character vector output at each position. It is the backward long short-term memory network in the first... The hidden state of the character vector output at each position. It is the first in the feedforward long short-term memory network The hidden state of the character vector at each position. This represents the first [unit / item] in the backward long short-term memory network. The hidden state of the character vector at each position; Will and By splicing, we obtain the first... The first context feature vector at each position : ； In the formula, This indicates vector concatenation; The final character vector sequence H is obtained as follows: , ,…, ,…, ); vocabulary vectors The contextual features are extracted using a bidirectional long short-term memory network, and the hidden states are calculated using a forward long short-term memory network and a backward long short-term memory network, respectively. = LSTM1( , ); = LSTM2( , ); In the formula, It is the first forward long short-term memory network. The hidden state of the word vector output at each position. It is the backward long short-term memory network in the first... The hidden state of the word vector output at each position. It is the first in the feedforward long short-term memory network The hidden state of the word vector at each position. This represents the first [unit / item] in the backward long short-term memory network. The hidden state of the word vector at each position; Will and By splicing, we obtain the first... The second context feature vector of each character : ； The final result is a self-matching vocabulary vector sequence U = ( ) containing contextual features. , ,…, ,…, ); (3c) Use H and U in the word combination sequence construction module to construct the word combination vector sequence V={ , ,…, ,…, }, s≤n, where s is the number of elements in V. It is the first in V There are elements, among which... Satisfy: When a continuous segment of characters is matched by a self-matching word When full coverage is achieved, the second context feature vector is used. Replace this continuous character segment, that is = Otherwise, the original character vector representation is retained, i.e. = ; ∈U; The word combination vector sequence V is input into a bidirectional long short-term memory network to extract contextual features. The hidden states are first calculated through a forward long short-term memory network and a backward long short-term memory network, respectively. =LSTM1( , ); =LSTM2( , ); In the formula, It is the first forward long short-term memory network. The hidden state of the word combination vector sequence V output at each position. It is the backward long short-term memory network in the first... The hidden state of the word combination vector sequence V output at each position. It is the first in the feedforward long short-term memory network The hidden state of the word combination vector sequence V at each position. This represents the first [unit / item] in the backward long short-term memory network. The hidden state of the word combination vector sequence V at each position; Will and By splicing, we can obtain the first The third context feature vector at each position : ； Based on the third context feature vector Finally, a new word combination vector sequence Z = ( , ,…, ,…, ); Perform vector concatenation on the final hidden states at both ends of the new word combination vector sequence Z to obtain the global semantic representation of the new word combination vector sequence Z. : ； In the formula, It is the hidden state output by the feedforward long short-term memory network at the last moment of the new word combination vector sequence Z. It is the hidden state output by the backward long short-term memory network at the last moment of the new word combination vector sequence Z; Use multiple different Construct a global semantic vector set Then H, U, It is output to the multi-layer network update layer.
4. The word-based collaborative entity recognition method for power grid equipment according to claim 1, characterized in that: In step (2), the multi-layer network update layer includes a graph convolutional neural network and a graph attention neural network, constructing a self-matching vocabulary and word combination sequence interaction graph adjacency matrix. Self-matching vocabulary and character containment graph adjacency matrix Transformation graph adjacency matrix Adjacency matrix of Lattice graph ; The processing method for the multi-layer network update layer includes the following steps: (4a) Put U and By concatenating the vectors, we obtain Initial node features : = ； U and As input to the graph convolutional neural network, the representation after feature extraction and information fusion is obtained through the graph convolutional neural network. : ； In the formula, For graph convolutional neural networks; from Extract the corresponding self-matching vocabulary parts to obtain a new self-matching vocabulary vector sequence. : ； Where m is the self-matching word sequence The number of self-matching words in the text; (4b) The new self-matching word vector sequence Concatenating the character vector sequence H with the input feature Node of the self-matching vocabulary and word interaction graph is obtained: Node = (H, ); Use 3 independent graph attention neural networks , and The node information of three different self-matching words and character interaction graphs is fused. The initial node input of the three graphs is Node: = (Node, ); = (Node, ); = (Node, ); in, This is the full node representation of the output of a graph attention neural network. ; yes The output node representation, yes The output node representation, yes The output node representation; F is The size of the vector dimension, C is The corresponding self-matching vocabulary and character containment graph, T is The corresponding transformation graph of self-matching words and characters, L is The corresponding Lattice diagram of self-matching words and characters; e is C, T, The enumeration identifier; m is the self-matching word sequence. The number of self-matching words in the sequence, where n is the number of characters in character sequence c; from Extracting character-level representations : ； (4c) Introduce a fusion mechanism to uniformly model information from different sources, and... The final character vector sequence is obtained by weighting and fusing multiple features using a weight matrix with H. , : = H+ + + ； in, , , and All of these are weight parameter matrices to be learned; yes The character representation, yes The character representation, yes The character representation; Then Output to the decoding layer.
5. The word-based collaborative entity recognition method for power grid equipment according to claim 1, characterized in that: In step (2), the decoding layer includes a conditional random field, and the specific processing method of the decoding layer refers to: Using conditional random fields Perform tag decoding, assuming the input character sequence is... ={ , ,… The corresponding target label sequence is = ( , ,…, Given the input conditions, Conditional probability of occurrence Represented as: = ； in, Represents the candidate label sequence; yes The Middle Tag The corresponding launch parameter weights; yes From the previous tag Move to current tag The transfer score; yes The Middle Tag The corresponding launch parameter weights; yes From the previous tag Move to current tag The transfer score; yes Global, overall, and holistic feature representation; It is an exponential function; yes The Middle Location tags; yes The Middle Location tags; yes The Middle A character vector; Select the label sequence with the highest probability, label each character with a BMEO label, and output the final power grid equipment entity recognition result.