Power transmission line text named entity recognition method and system based on GRTCN network

Abstract

The application discloses a power transmission line text named entity recognition method and system based on a GRTCN network. In view of the problems of semantic ambiguity and entity boundary ambiguity caused by flexible expression, non-standardized terms and strong context dependence in power transmission line construction text, a double-channel parallel architecture containing a bidirectional long short-term memory network and a gated residual time convolution network is constructed, wherein the gated residual time convolution network dynamically screens key context features through a gating mechanism and captures long-range dependence by combining a residual time convolution; a double-affine attention mechanism is further introduced to realize deep interaction and adaptive calibration of double-channel heterogeneous features, and finally, global structured prediction is performed through a conditional random field. The application can significantly improve the robustness to term variants and dynamic context, and provides reliable technical support for power transmission construction knowledge base construction.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing. Specifically, it relates to a method and system for identifying named entities in transmission line texts based on the GRTCN network. Background Art

[0002] The digital transformation process of the new power system continues to deepen, promoting a significant improvement in the intelligent level of power equipment. In this process, a large amount of text data containing domain knowledge is generated in the system operation and maintenance link, and the mining of its value has become the key breakthrough point for improving the intelligence of the power system. As the core technology of information extraction, named entity recognition (NER) provides basic support for applications such as intelligent question answering and information retrieval by locating and classifying key entities in unstructured texts. This technology has been successfully applied to multiple scenarios such as ancient books, industry and agriculture, network security, power safety monitoring, and dispatching log analysis, verifying the effectiveness of integrating domain knowledge and deep learning architectures (such as BERT and BiLSTM-CRF models) to solve vertical domain NER tasks.

[0003] Currently, NER research in the power field mainly focuses on highly structured normative texts, and these works rely on large-scale labeled data to drive complex models. However, it is worth noting that there are still significant research gaps in the texts generated in the transmission line construction scenario (such as construction plans and technical reports). Such texts exhibit three major characteristics: the expression has a high degree of flexibility, there are non-standard variants of professional terms, and the context semantics has a strong dependence. These characteristics pose severe challenges to traditional NER methods: the recognition mechanism based on fixed patterns and normative terms is difficult to adapt to dynamic contexts, resulting in the entity recognition accuracy being unable to meet the actual engineering requirements.

[0004] In view of the unique language characteristics and recognition bottlenecks of transmission line construction-generated texts, there is an urgent need to develop a dedicated NER solution. The purpose of this article is to design an optimized recognition method suitable for such texts, and by overcoming core problems such as expression diversity, term variation, and strong context correlation, to build a highly reliable information extraction technology support system for the intelligent upgrade of the power industry. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention provides a method and system for identifying named entities in transmission line texts based on the GRTCN network.

[0006] The technical solution adopted by the present invention is as follows: A method for identifying named entities in transmission line texts based on a gated residual temporal convolutional network, comprising the following steps: S1. Obtain the original construction text of the power transmission line. Input the original construction text of the power transmission line into the pre-trained BERT model. After processing by the deep bidirectional self-attention mechanism of Transformer, a distributed semantic word vector sequence is obtained.

[0007] This step generates context-aware word vector representations based on a pre-trained BERT model and utilizes the deep bidirectional attention mechanism of Transformer to parse lexical polysemy. This process transforms the original text into a distributed semantic word vector sequence, significantly enhancing its robust representation capability for variations and non-standard expressions in the field of power construction, thus providing high-quality input for subsequent steps.

[0008] S2. Input the semantic word vector sequence into a bidirectional long short-term memory network, and process it in parallel through a forward propagation layer and a backward propagation layer to obtain forward dependency features and backward context features; perform vector-level fusion of the forward dependency features and backward context features to obtain global semantic features.

[0009] This step utilizes a bidirectional long short-term memory (BiLSTM) network to perform forward and backward state propagation in parallel: forward propagation captures historical dependency features, and backward propagation extracts future contextual information. The dual-channel outputs are fused to construct global semantic associations, effectively modeling complex nonlinear semantic patterns in construction texts and addressing the semantic discrepancy problem caused by strong contextual associations.

[0010] S3. Input the semantic word vector sequence into a gated residual temporal convolutional network. After extracting multi-scale local semantic features through the temporal convolutional layer, generate a gated weight matrix through a gated unit. Multiply the multi-scale local semantic features with the gated weight matrix element by element to obtain the gated features. After zero-padding alignment, add the semantic word vector sequence to the gated features. After activation function processing, obtain the local semantic features.

[0011] This step designs a gated residual temporal convolutional network (GRTCN) as a complementary feature extraction channel: temporal convolutional layers extract multi-scale local semantic features, capturing the combinatorial patterns of terms; the gating mechanism dynamically activates key context fragments, suppressing noise interference caused by non-standard term variants; residual connections ensure lossless transmission of long-distance semantics, avoiding the gradient vanishing problem. This step specializes in enhancing the model's robustness to non-standardized representations.

[0012] S4. Input the global semantic features and the local semantic features into a dual affine attention network. After processing with the first weight matrix and transpose operation, the first attention weighted result of the global semantic features is obtained. After processing with the second weight matrix and transpose operation, the second attention weighted result of the local semantic features is obtained. The first attention weighted result and the second attention weighted result are fused to obtain the fused semantic features.

[0013] This step introduces a dual affine attention BiAffine mechanism to achieve dual-channel deep interaction: the symmetric dependency between the global semantic features of BiLSTM and the local semantic features of GRTCN is modeled through a learnable parameter matrix; an attention weight map is also generated, adaptively calibrated and weighted to increase the contribution of the two types of heterogeneous features; finally, the fusion vector is output to enhance the key information for entity boundary discrimination and semantic ambiguity resolution, thus solving the deficiency of insufficient representation by a single feature source.

[0014] S5. Input the fused semantic features into a conditional random field, process them with a label transition constraint matrix and a Viterbi algorithm to obtain the optimal label path, and output the named entity recognition result.

[0015] This step utilizes Conditional Random Fields (CRF) for structured prediction at the global semantic feature level. Power domain knowledge is embedded through predefined label transfer rules; optimal label paths are derived through joint inference, eliminating boundary splits and label conflicts caused by word-by-word annotation; finally, entities conforming to the semantic specifications of power construction are output, providing a reliable structured data foundation for knowledge base construction.

[0016] Furthermore, step S2 specifically includes: The semantic word vector sequence is input into a bidirectional long short-term memory network (LSTM), and processed in parallel by forward and backward propagation layers to obtain forward dependency features and backward context features. Specifically, the forward propagation layer processes each word vector sequentially to capture historical dependency information, while the backward propagation layer processes each word vector in reverse order to capture future context information. The forward dependency features and backward context features are then fused at the vector level to obtain the global semantic features. , means as follows:

[0017] In the formula, This represents the input word vector. This indicates that the global semantic features of the text are captured using a BiLSTM network.

[0018] Furthermore, in step 3, dynamic dilated convolution is employed. The actual dilation coefficient is generated through learnable parameters. The semantic word vector sequence is input into the learnable parameter layer and normalized using the Sigmoid activation function to obtain normalized coefficients. The maximum dilation coefficient is multiplied by the normalized coefficient, rounded, and then incremented by one to obtain the actual dilation coefficient. This actual dilation coefficient controls the padding amount of the convolution kernel. The actual dilation coefficient is obtained using the following formula:

[0019] In the formula, These are learnable parameters; Indicates the maximum expansion coefficient; The gating unit processes the output feature map after two dynamic dilation convolutions using the Sigmoid activation function to obtain a gating weight matrix. The elements in the gating weight matrix are used to control the transmission ratio of features at corresponding positions. Specifically as follows: 1) Output local semantic features after the first-level dynamic dilated convolution. :

[0020] In the formula, Represents the input word vector; Indicates the weights of the convolution kernel; This is the actual coefficient of thermal expansion; The kernel size; 2) Similarly, the local semantic features are output again after a second-level dynamic dilation convolution. :

[0021] In the formula, Indicates the weights of the convolution kernel; Local semantic features; This is the actual coefficient of thermal expansion; The kernel size; 3) The output F after two dynamic dilation convolutions is used by a gating unit to generate a gating weight matrix. :

[0022] In the formula, For convolution kernel; Indicates the activation function; 4) Multiply the output F of the two-stage dynamically dilated convolution with the gating weight matrix G element-wise to obtain the gating features. :

[0023] This process achieves feature selection through element-wise multiplication. When the elements in the gate weight matrix approach 0, i.e. Suppressing the corresponding positional features in real time, approaching 1, i.e. Enhance the corresponding location features in time; 5) Fusion with residual paths: The semantic word vector sequence X is zero-padded and aligned, then added to the gated features, and processed by the ReLU activation function to obtain local semantic features. :

[0024] In the formula, This is a zero-filling operation to ensure that the residual path is time-aligned with the main path; Represents the input word vector; This represents the activation function.

[0025] Furthermore, the dual affine attention network described in step 4 models the symmetric dependency between global semantic features and local semantic features through a learnable first weight matrix and a second weight matrix, respectively. The first attention weighting result is obtained using the following formula:

[0026] The second attention-weighted result is obtained using the following formula:

[0027] In the formula, and These represent sentence context information constructed using GRTCN and BILSTM networks, respectively. and These are all weight matrices, obtained through network learning; This indicates the transpose operation.

[0028] Further, the fused semantic features obtained in step S4 are input into a conditional random field (CRF). A predefined label transition constraint matrix is ​​used in the CRF, embedding label transition rules from the power construction field to constrain the legal transition relationships between adjacent labels. The maximum likelihood labeling path is solved using the Viterbi algorithm, and the label state score and transition probability are jointly optimized to obtain the optimal label path, outputting the named entity recognition result. Its expression is:

[0029] In the formula, Indicates a connection operation. This indicates that a conditional random field operation is used. This indicates the final output; and These represent sentence context information constructed using GRTCN and BILSTM networks, respectively.

[0030] Another technical solution adopted by the present invention is: A named entity recognition system for power transmission line text based on gated residual temporal convolutional networks includes: The first module is used to obtain the original construction text of the power transmission line. The original construction text of the power transmission line is input into the pre-trained BERT model and processed by the deep bidirectional self-attention mechanism of Transformer to obtain a distributed semantic word vector sequence. The second module includes a bidirectional long short-term memory network subunit and a gated residual temporal convolutional network subunit. The bidirectional long short-term memory network subunit is used to process the semantic word vector sequence in parallel through forward and backward propagation layers to obtain forward dependency features and backward context features. The forward dependency features and backward context features are then fused at the vector level to obtain global semantic features. The gated residual temporal convolutional network subunit is used to extract multi-scale local semantic features from the semantic word vector sequence through temporal convolutional layers. After that, a gated weight matrix is ​​generated through a gated unit. The multi-scale local semantic features are multiplied element-wise with the gated weight matrix to obtain gated features. The semantic word vector sequence is zero-padded and aligned before being added to the gated features. After processing with an activation function, local semantic features are obtained. The third module is used to input the global semantic features and the local semantic features into a dual affine attention network, process them with a first weight matrix and transpose operation to obtain a first attention weighted result of the global semantic features; process them with a second weight matrix and transpose operation to obtain a second attention weighted result of the local semantic features; and fuse the first attention weighted result and the second attention weighted result to obtain a fused semantic feature. The fourth module is used to input the fused semantic features into a conditional random field, process them with a label transition constraint matrix and a Viterbi algorithm to obtain the optimal label path, and output the named entity recognition result.

[0031] Furthermore, the gated unit in the gated residual temporal convolutional network subunit processes the output feature map after two dynamic dilation convolutions using the Sigmoid activation function to obtain a gated weight matrix. The elements in the gated weight matrix are used to control the transmission ratio of features at corresponding positions.

[0032] Furthermore, the gated residual temporal convolutional network subunit is equipped with a dynamic dilation convolution module. This module generates the actual dilation coefficient through learnable parameters, inputs the semantic word vector sequence into the learnable parameter layer and normalizes it through the Sigmoid activation function to obtain the normalized coefficient, multiplies the maximum dilation coefficient with the normalized coefficient, rounds it down and adds one to obtain the actual dilation coefficient, and uses the actual dilation coefficient to control the filling amount of the convolution kernel.

[0033] Furthermore, the dual affine attention network models the symmetric dependency between global semantic features and local semantic features using a learnable first weight matrix and a second weight matrix, respectively, and generates an attention weight map for weighting the two types of heterogeneous features.

[0034] Furthermore, the semantic word vector sequence output by the first module is used as the shared input of the bidirectional long short-term memory network subunit and the gated residual temporal convolutional network subunit in the second module, and the output of the bidirectional long short-term memory network subunit and the output of the gated residual temporal convolutional network subunit are connected to the third module in parallel.

[0035] Compared with the prior art, the present invention has the following advantages: 1) This invention proposes a gated residual temporal convolutional network to capture potential information from the context. By dynamically filtering key features through a gating mechanism and combining residual connections, it ensures long-range semantic transmission and effectively suppresses noise interference caused by non-standard terminology variants. 2) This invention introduces a dual affine attention mechanism to achieve deep feature interaction, establishes a learnable parameter matrix, models the symmetric mapping relationship between the global semantic features output by the bidirectional long short-term memory network and the local semantic features extracted by the gated residual temporal convolutional network, adaptively calibrates the contribution of the two types of heterogeneous features, and significantly improves the ability to resolve semantic ambiguity and distinguish entity boundaries. Attached Figure Description

[0036] The present invention will be further described below with reference to the accompanying drawings and specific embodiments: Figure 1 This is a flowchart illustrating the method in an embodiment of the present invention; Figure 2 This is a diagram of the overall network structure used in an embodiment of the present invention; Figure 3 This is a schematic diagram of the dynamic dilated convolution module according to an embodiment of the present invention; Figure 4 This is a model diagram of the gated residual temporal convolutional network GRTCN according to an embodiment of the present invention. Detailed Implementation

[0037] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0038] Please see Figure 1 The present invention provides a method for named entity recognition of power transmission line text based on GRTCN network, comprising the following steps: Step 1: Dynamic semantic encoding.

[0039] Let the original power transmission line construction text be represented as Q, which consists of a series of... Composed of 1 word, represented as { , ..., The original text about power transmission line construction is input into a pre-trained BERT model. After processing through the Transformer's deep bidirectional self-attention mechanism, a distributed semantic word vector sequence is obtained. This process transforms each word into a fixed-dimensional word vector containing rich contextual semantics, expressed as:

[0040] In the formula, This indicates that BERT word embeddings are used to obtain the first word in the text sentence. The word vectors corresponding to each word.

[0041] In this embodiment, BERT-based word embedding technology is used to capture the semantics of individual words in a sentence, which helps in the subtle understanding of word meaning, especially in the field of natural language processing. The model uses an embedding process to convert individual words into fixed-dimensional vector representations; each word in the input text is encoded by a pre-trained BERT model, transforming it into a fixed-dimensional word vector containing rich contextual semantics, thereby obtaining a deep semantic representation sequence of the text.

[0042] Step 2: Two-way long-range dependency modeling.

[0043] The semantic word vector sequence obtained in step 1 is input into a bidirectional long short-term memory network, and processed in parallel through forward and backward propagation layers to obtain forward dependency features and backward context features; wherein, the forward propagation layer follows the order from... arrive Each word vector is processed sequentially to capture historical dependency information; the backpropagation layer processes each word vector in turn according to the order from... arrive Each word vector is processed sequentially to capture future contextual information; forward dependency features and backward contextual features are then fused at the vector level to obtain global semantic features. , means as follows:

[0044] In the formula, This represents the input word vector. This indicates that the global semantic features of the text are captured using a BiLSTM network.

[0045] Step 3: Local semantic feature enhancement and noise suppression.

[0046] The semantic word vector sequence obtained in step 1 is input into a gated residual temporal convolutional network, which first uses a dynamically dilated convolutional module to extract multi-scale local semantic features. For example... Figure 2 As shown, the dynamically dilated convolution module parameterizes the dilation coefficient, enabling the model to adaptively adjust the receptive field of the convolution kernel based on the context. Specifically, it transforms the semantic word vector sequence... The input is a learnable parameter layer, which is then processed by a Sigmoid activation function to generate normalized coefficients. :

[0047] In the formula, These are learnable parameters; For activation functions; Actual expansion coefficient Determined by the mapping function:

[0048] This value controls the padding amount of causal convolutions in real time. ( (where is the kernel size), enabling the convolution operation to adapt to the receptive field; where, This represents the maximum expansion coefficient.

[0049] Based on this, the gated residual temporal convolutional network generates a gated weight matrix through two levels of dynamic dilation convolution followed by gated units. For example... Figure 3 As shown, the input X is fed into the residual path and the main path in parallel. The main path contains two levels of dynamically dilated convolutions; specifically as follows: 1) Local semantic features are output after the first-level dynamic dilation convolution, i.e., dynamic convolution 1. :

[0050] In the formula, Represents the input word vector; Indicates the weights of the convolution kernel; This is the actual coefficient of thermal expansion; The kernel size; 2) Similarly, the local semantic features are output again after a second-level dynamic dilation convolution, i.e., dynamic convolution 2. :

[0051] In the formula, Indicates the weights of the convolution kernel; Local semantic features; This is the actual coefficient of thermal expansion; The kernel size; 3) The output F after two dynamic dilation convolutions is used by a gating unit to generate a gating weight matrix. :

[0052] In the formula, For convolution kernel; Indicates the activation function; 4) Multiply the output F of the two-stage dynamically dilated convolution with the gating weight matrix G element-wise to obtain the gating features. :

[0053] This process achieves feature selection through element-wise multiplication. When the elements in the gate weight matrix approach 0, i.e. Suppressing the corresponding positional features in real time, approaching 1, i.e. Enhance the corresponding location features in time; 5) Fusion with residual paths: The semantic word vector sequence X is zero-padded and aligned, then added to the gated features, and processed by the ReLU activation function to obtain local semantic features. :

[0054] In the formula, This is a zero-filling operation to ensure that the residual path is time-aligned with the main path; Represents the input word vector; This represents the activation function. This design retains the advantages of residual learning while also implementing a feature-level attention mechanism.

[0055] In this embodiment, word vector sequences are input into a Temporal Convolutional Network (TCN). Its causal dilated convolutional structure is used to initially extract temporal features of the text. This process expands the receptive field using a preset dilation coefficient and maintains training stability through residual connections. However, this architecture has significant limitations: a fixed dilation pattern is difficult to adapt to complex temporal features, and the traditional direct addition of features by residual blocks leads to noise being transmitted along with key information. Therefore, to further optimize feature extraction, the system introduces a dynamic dilated convolution module. This module parameterizes the dilation coefficient, enabling the model to adaptively adjust the receptive field of the convolution kernel according to the specific context, thus more flexibly capturing dependencies at different distances. Furthermore, a gated residual block is constructed based on this, using a gated unit to weight the feature map generated after convolution element-wise, thereby enhancing key features and suppressing noise features. The selected features are then fused with the aligned original residual path features to output refined local semantic features.

[0056] Step 4: Collaborative optimization of heterogeneous features.

[0057] The global semantic features obtained in step 2 Compared with the local semantic features obtained in step 3 Input a dual affine attention network; this network models the symmetric dependencies between two types of features using a learnable first and second weight matrix, respectively; the specific processing procedure is as follows:

[0058]

[0059] In the formula, and These represent sentence context information constructed using GRTCN and BILSTM networks, respectively. and These are all weight matrices, obtained through network learning; Indicates the transpose operation; The first attention weighted result and the second attention weighted result are fused to obtain the fused semantic features.

[0060] In this embodiment, to fuse global and local semantic features, the system uses a BiLSTM network in parallel to capture the global semantic features of the text. Furthermore, it interacts with the local semantic features obtained in the previous steps through a dual affine attention mechanism to generate an enhanced feature representation that integrates global and local semantics. It employs a dual affine attention network to model the symmetric dependency between the global semantic features of BiLSTM and the local semantic features of GRTCN through a learnable parameter matrix. It also generates an attention weight map, adaptively calibrates and weights the contributions of the two types of heterogeneous features, thereby obtaining refined contextual information of the sentence.

[0061] Step 5: Domain knowledge-driven structured decoding.

[0062] The fused semantic features obtained in step 4 are input into a conditional random field (CRF). A predefined label transition constraint matrix is ​​used in the CRF, embedding label transition rules from the power construction field to constrain the legal transition relationships between adjacent labels. The maximum likelihood labeling path is solved using the Viterbi algorithm, and the label state score and transition probability are jointly optimized to obtain the optimal label path, outputting the named entity recognition result. Its expression is:

[0063] In the formula, Indicates a connection operation. This indicates that a conditional random field operation is used. This indicates the final output; and These represent sentence context information constructed using GRTCN and BILSTM networks, respectively.

[0064] In this embodiment, the fused feature sequence is input into the Conditional Random Field (CRF) decoding layer. The CRF uses the learned inter-label transition constraint rules to infer the optimal named entity label sequence, thus completing the entity recognition task for the entire text.

[0065] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the principles and essence of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for named entity recognition of power transmission line text based on GRTCN network, characterized in that, Includes the following steps: S1. Obtain the original construction text of the power transmission line. Input the original construction text of the power transmission line into the pre-trained BERT model. After processing by the deep bidirectional self-attention mechanism of Transformer, a distributed semantic word vector sequence is obtained. S2. Input the semantic word vector sequence into a bidirectional long short-term memory network, and process it in parallel through a forward propagation layer and a backward propagation layer to obtain forward dependency features and backward context features. The forward dependency features and the backward context features are fused at the vector level to obtain global semantic features; S3. Input the semantic word vector sequence into a gated residual temporal convolutional network. After extracting multi-scale local semantic features through the temporal convolutional layer, generate a gated weight matrix through a gated unit. Multiply the multi-scale local semantic features with the gated weight matrix element by element to obtain the gated features. After zero-padding alignment, add the semantic word vector sequence to the gated features. After activation function processing, obtain the local semantic features. S4. Input the global semantic features and the local semantic features into a dual affine attention network, and process them through the first weight matrix and transpose operation to obtain the first attention weighted result of the global semantic features; After processing with the second weight matrix and transpose, the second attention weighted result of the local semantic features is obtained; the first attention weighted result and the second attention weighted result are fused to obtain the fused semantic features; S5. Input the fused semantic features into a conditional random field, process them with a label transition constraint matrix and a Viterbi algorithm to obtain the optimal label path, and output the named entity recognition result.

2. The method for named entity recognition of power transmission lines based on GRTCN networks according to claim 1, characterized in that, Step S2 is as follows: The semantic word vector sequence is input into a bidirectional long short-term memory network (LSTM), and processed in parallel by forward and backward propagation layers to obtain forward dependency features and backward context features. Specifically, the forward propagation layer processes each word vector sequentially to capture historical dependency information, while the backward propagation layer processes each word vector in reverse order to capture future context information. The forward dependency features and backward context features are then fused at the vector level to obtain the global semantic features. , means as follows: In the formula, This represents the input word vector. This indicates that the global semantic features of the text are captured using a BiLSTM network.

3. The method for named entity recognition of power transmission lines based on GRTCN network according to claim 1, characterized in that, Step 3 employs dynamic dilated convolution. The actual dilation coefficient is generated using learnable parameters. The semantic word vector sequence is input into the learnable parameter layer and normalized using the Sigmoid activation function to obtain normalized coefficients. The maximum dilation coefficient is multiplied by the normalized coefficient, rounded, and then incremented by one to obtain the actual dilation coefficient. This actual dilation coefficient controls the padding amount of the convolution kernel. The actual dilation coefficient is obtained using the following formula: In the formula, These are learnable parameters; Indicates the maximum expansion coefficient; The gating unit processes the output feature map after two dynamic dilation convolutions using the Sigmoid activation function to obtain a gating weight matrix. The elements in the gating weight matrix are used to control the transmission ratio of features at corresponding positions. Specifically as follows: 1) Output local semantic features after the first-level dynamic dilated convolution. : In the formula, Represents the input word vector; Indicates the weights of the convolution kernel; This is the actual coefficient of thermal expansion; The kernel size; 2) Similarly, the local semantic features are output again after a second-level dynamic dilation convolution. : In the formula, Indicates the weights of the convolution kernel; Local semantic features; This is the actual coefficient of thermal expansion; The kernel size; 3) The output F after two dynamic dilation convolutions is used by a gating unit to generate a gating weight matrix. : In the formula, For convolution kernel; Indicates the activation function; 4) Multiply the output F of the two-stage dynamically dilated convolution with the gating weight matrix G element-wise to obtain the gating features. : This process achieves feature selection through element-wise multiplication. When the elements in the gate weight matrix approach 0, i.e. Suppressing the corresponding positional features in real time, approaching 1, i.e. Enhance the corresponding location features in time; 5) Fusion with residual paths: The semantic word vector sequence X is zero-padded and aligned, then added to the gated features, and processed by the ReLU activation function to obtain local semantic features. : In the formula, This is a zero-filling operation to ensure that the residual path is time-aligned with the main path; Represents the input word vector; This represents the activation function.

4. The method for named entity recognition of power transmission lines based on GRTCN network according to claim 1, characterized in that, The dual affine attention network described in step 4 models the symmetric dependency between global semantic features and local semantic features using a learnable first weight matrix and a second weight matrix, respectively. The first attention weighting result is obtained using the following formula: The second attention-weighted result is obtained using the following formula: In the formula, and These represent sentence context information constructed using GRTCN and BILSTM networks, respectively. and These are all weight matrices, obtained through network learning; This indicates the transpose operation.

5. The method for named entity recognition of power transmission lines based on GRTCN network according to claim 4, characterized in that, The fused semantic features obtained in step S4 are input into a conditional random field. A predefined label transition constraint matrix is ​​used in the conditional random field. This matrix embeds the label transition rules in the field of power construction to constrain the legal transition relationship between adjacent labels. The maximum likelihood labeling path is solved by the Viterbi algorithm. The label state score and transition probability are jointly optimized to obtain the optimal label path and output the named entity recognition result. Its expression is: In the formula, Indicates a connection operation. This indicates that a conditional random field operation is used. This indicates the final output; and These represent sentence context information constructed using GRTCN and BILSTM networks, respectively.

6. A text-based named entity recognition system for power transmission lines based on GRTCN networks, characterized in that, include: The first module is used to obtain the original construction text of the power transmission line. The original construction text of the power transmission line is input into the pre-trained BERT model and processed by the deep bidirectional self-attention mechanism of Transformer to obtain a distributed semantic word vector sequence. The second module includes a bidirectional long short-term memory network subunit and a gated residual temporal convolutional network subunit. The bidirectional long short-term memory network subunit is used to process the semantic word vector sequence in parallel through forward and backward propagation layers to obtain forward dependency features and backward context features. The forward dependency features and backward context features are then fused at the vector level to obtain global semantic features. The gated residual temporal convolutional network subunit is used to extract multi-scale local semantic features from the semantic word vector sequence through temporal convolutional layers. After that, a gated weight matrix is ​​generated through a gated unit. The multi-scale local semantic features are multiplied element-wise with the gated weight matrix to obtain gated features. The semantic word vector sequence is zero-padded and aligned before being added to the gated features. After processing with an activation function, local semantic features are obtained. The third module is used to input the global semantic features and the local semantic features into a dual affine attention network, and after processing by the first weight matrix and transpose operation, obtain the first attention weighted result of the global semantic features; After processing with the second weight matrix and transpose, the second attention weighted result of the local semantic features is obtained; the first attention weighted result and the second attention weighted result are fused to obtain the fused semantic features; The fourth module is used to input the fused semantic features into a conditional random field, process them with a label transition constraint matrix and a Viterbi algorithm to obtain the optimal label path, and output the named entity recognition result.

7. A text-based named entity recognition system for power transmission lines based on a GRTCN network according to claim 5, characterized in that, The gated unit in the gated residual temporal convolutional network subunit processes the output feature map after two dynamic dilation convolutions using the Sigmoid activation function to obtain a gated weight matrix. The elements in the gated weight matrix are used to control the transmission ratio of features at corresponding positions.

8. A text-based named entity recognition system for power transmission lines based on a GRTCN network according to claim 5, characterized in that, The gated residual temporal convolutional network subunit is equipped with a dynamic dilation convolution module. This module generates the actual dilation coefficient through learnable parameters. The semantic word vector sequence is input into the learnable parameter layer and normalized by the Sigmoid activation function to obtain the normalized coefficient. The maximum dilation coefficient is multiplied by the normalized coefficient, rounded, and then incremented by one to obtain the actual dilation coefficient. The actual dilation coefficient is used to control the filling amount of the convolution kernel.

9. A text-based named entity recognition system for power transmission lines based on a GRTCN network according to claim 5, characterized in that, The dual affine attention network models the symmetric dependency between global semantic features and local semantic features using a learnable first weight matrix and a second weight matrix, respectively, and generates an attention weight map to weight the two types of heterogeneous features.

10. A text-based named entity recognition system for power transmission lines based on a GRTCN network according to claim 5, characterized in that, The semantic word vector sequence output by the first module is used as the shared input of the bidirectional long short-term memory network subunit and the gated residual temporal convolutional network subunit in the second module. The output of the bidirectional long short-term memory network subunit and the output of the gated residual temporal convolutional network subunit are connected to the third module in parallel.

Images