A power transmission line naming recognition method based on bidirectional enhanced nonlinear pulse neural network

Abstract

The application provides a power transmission line naming recognition method based on a bidirectional enhanced nonlinear pulse nerve, and the core innovation is that a BiENSNP module is constructed to deeply simulate a dynamic information processing mechanism of a biological nerve system, the model is endowed with powerful basic representation capability, and is especially good at modeling complex nonlinear semantic patterns contained in text. The fusion architecture significantly improves the robustness and accuracy of the model in identifying entities in the power transmission line construction field (especially generative text), thereby laying a solid and good scalable technical foundation for constructing high-quality power transmission line construction knowledge base and other key application scenarios.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to a method for naming and recognizing power transmission lines based on bidirectional enhanced nonlinear spiking neural networks. Background Technology

[0002] With the accelerated digital transformation of new power systems, the digitalization and intelligence levels of power equipment are continuously improving. In this process, the operation and maintenance of power systems generate massive amounts of text data rich in domain knowledge. Effectively mining the value of this data is crucial for improving the system's intelligence level. Named Entity Recognition (NER), as a core technology for information extraction, can locate and classify key entities from unstructured or semi-structured text, serving as the foundation for downstream applications such as intelligent question answering and information retrieval. In recent years, NER technology has been widely applied in various fields, including ancient books, mining, agriculture, industrial manufacturing, cybersecurity, and even power safety, power grid dispatching, and equipment defect analysis, fully demonstrating the effectiveness of combining domain knowledge with advanced model architectures (such as BERT, BiLSTM, and CRF) to solve specific NER problems.

[0003] However, existing NER research in the power sector primarily focuses on documents with relatively fixed structures and standardized expressions. These studies typically rely on large-scale, high-quality labeled data to drive complex models to achieve the desired performance. A significant research gap lies in texts generated in the field of transmission line construction, such as construction reports and technical solutions. These texts are characterized by flexible and varied expressions, non-standard variations of technical terminology, and strong contextual dependencies, posing a serious challenge to traditional NER methods that rely on fixed patterns and standardized terminology, resulting in recognition accuracy that fails to meet practical needs. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of the prior art by providing a transmission line naming and recognition method based on bidirectional enhanced nonlinear pulse neural networks, in order to provide more reliable information extraction technology support for the intelligent upgrading of the power industry.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a method for naming and recognizing transmission lines based on bidirectional enhanced nonlinear pulse neural networks, comprising the following steps: S1. Based on the pre-trained BERT model, the input text is dynamically encoded to generate word vector representations with context awareness. The deep bidirectional attention mechanism of the Transformer architecture is used to effectively capture the multiple features of words in a specific context and transform the original text into a distributed vector sequence containing semantic information. S2. Deep contextual modeling of word vectors is performed using the bidirectional enhanced nonlinear spiking neural system BiENSNP. The bio-inspired computing module simulates the characteristics of neuronal spiking transmission and performs forward and backward information transmission; forward propagation captures historical dependencies, and backward propagation extracts future features. S3. A multi-head self-attention mechanism is used to adaptively weight the encoded semantic features. By calculating the interaction importance weights between words, the bio-inspired computing module dynamically focuses on key contextual information related to entity recognition, while suppressing irrelevant semantic noise. S4. The refined feature input conditional random field decoding layer is used for structured prediction. The feature input conditional random field solves the optimal transition path at the global sequence level by modeling the transition constraint relationship between labels.

[0006] Furthermore, in step S2, a bidirectional enhanced nonlinear spiking neural system is used to extract text context. Specifically: ; in, Text sentences are transformed into word vectors using the BERT model word embedding technique.

[0007] Furthermore, S3 specifically refers to: Context weight information is obtained through a self-attention mechanism. : ; ; ; in, Text context extracted for bidirectional enhanced nonlinear spiking neural systems; Mean value is calculated from the text context extracted from a bidirectional enhanced nonlinear spiking neural system. The length of the text; The weights used for training a neural network can be obtained through model training; The bias can also be obtained through network training: To enhance the nonlinear spiking neural system for extracting text context; The correlation coefficient; Obtain the weights between each word in the context; multiply these weights by the context words to get the importance of each word in the context. ; It is a bidirectional enhanced nonlinear pulse nerve.

[0008] Furthermore, the context-weighted coefficients are calculated, and the final output is processed through a CRF: ; The feature input is a conditional random field.

[0009] Furthermore, it also includes: The first module implements deep semantic initialization of text. It converts the original input into dynamic word vector representation through a pre-trained BERT model, and uses the multi-layer self-attention mechanism of Transformer to parse the contextual polysemy of words and generate distributed feature vectors containing global semantic associations. The second module performs bidirectional contextual semantic modeling. Based on a bidirectional enhanced nonlinear spiking neural system, it simulates the spiking characteristics of neurons to process forward or backward semantic flows in parallel. By fusing historical and future contextual features, it explicitly captures long-distance dependencies and nonlinear patterns in construction texts, enhancing the robust parsing capability for flexible expressions. The third module: to achieve key feature enhancement and noise suppression, adopt a multi-head self-attention mechanism to recalibrate the encoded features, and dynamically amplify the entity-related context signal by calculating word-level interaction weights to weaken redundant information; The fourth module completes the global optimization of entity labeling decisions, uses conditional random fields to model label transfer constraints, and searches for the optimal labeling path in the global sequence space.

[0010] The beneficial effects of this invention are: bidirectional enhancement of nonlinear spiking neural network to capture potential information from context; a multi-head self-attention mechanism is proposed to establish connections between contexts; thereby enriching the understanding of the interaction between entities and context. Attached Figure Description

[0011] Figure 1 This is a network structure diagram of a power transmission line naming and recognition method based on bidirectional enhanced nonlinear spiking neural networks. Figure 2 The model diagram of ENSNP; Figure 3 The model diagram of BiENSNP; Figure 4 This is a diagram showing the effect of different numbers of neurons. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0013] Please see Figure 1A method for naming and recognizing power transmission lines based on bidirectional enhanced nonlinear pulse neural networks includes the following steps: S1. Based on the pre-trained BERT model, the input text is dynamically encoded to generate word vector representations with context awareness. The deep bidirectional attention mechanism of the Transformer architecture is used to effectively capture the multiple features of words in a specific context and transform the original text into a distributed vector sequence containing semantic information. S2. Deep contextual modeling of word vectors is performed using the bidirectional enhanced nonlinear spiking neural system BiENSNP. The bio-inspired computing module simulates the characteristics of neuronal spiking transmission and performs forward and backward information transmission; forward propagation captures historical dependencies, and backward propagation extracts future features. S3. A multi-head self-attention mechanism is used to adaptively weight the encoded semantic features. By calculating the interaction importance weights between words, the bio-inspired computing module dynamically focuses on key contextual information related to entity recognition, while suppressing irrelevant semantic noise. S4. The refined feature input conditional random field decoding layer is used for structured prediction. The feature input conditional random field solves the optimal transition path at the global sequence level by modeling the transition constraint relationship between labels.

[0014] In step S2, a bidirectional enhanced nonlinear spiking neural system is used to extract text context. Specifically: ; in, Text sentences are transformed into word vectors using the BERT model word embedding technique.

[0015] Specifically, S3 is: Context weight information is obtained through a self-attention mechanism. : ; ; ; in, Text context extracted for bidirectional enhanced nonlinear spiking neural systems; Mean value is calculated from the text context extracted from a bidirectional enhanced nonlinear spiking neural system. The length of the text; The weights used for training a neural network can be obtained through model training; The bias can also be obtained through network training: To enhance the nonlinear spiking neural system for extracting text context; The correlation coefficient; Obtain the weights between each word in the context; multiply these weights by the context words to get the importance of each word in the context. ; It is a bidirectional enhanced nonlinear pulse nerve.

[0016] The context-weighted coefficients are calculated, and the final output is processed through a CRF: ; The feature input is a conditional random field.

[0017] Also includes: The first module implements deep semantic initialization of text. It converts the original input into dynamic word vector representation through a pre-trained BERT model, and uses the multi-layer self-attention mechanism of Transformer to parse the contextual polysemy of words and generate distributed feature vectors containing global semantic associations. The second module performs bidirectional contextual semantic modeling. Based on a bidirectional enhanced nonlinear spiking neural system, it simulates the spiking characteristics of neurons to process forward or backward semantic flows in parallel. By fusing historical and future contextual features, it explicitly captures long-distance dependencies and nonlinear patterns in construction texts, enhancing the robust parsing capability for flexible expressions. The third module: to achieve key feature enhancement and noise suppression, adopt a multi-head self-attention mechanism to recalibrate the encoded features, and dynamically amplify the entity-related context signal by calculating word-level interaction weights to weaken redundant information; The fourth module completes the global optimization of entity labeling decisions, uses conditional random fields to model label transfer constraints, and searches for the optimal labeling path in the global sequence space.

[0018] Example: Input a text segment as This text consists of a series of words, represented as ; BERT-based word embeddings were used, and the embedding process was employed to convert individual words into fixed-dimensional vector representations. ; This indicates that BERT word embedding technology is used on the text sentence.

[0019] See details Figure 2The proposed EVSNP network embodies a cyclic paradigm and has proven effective in processing sequential data, particularly in addressing the challenges of time series forecasting. The model's basic operating mechanism comprises three nonlinear gates: a reset gate, a consumption gate, and a generation gate. These gates play a crucial role in regulating the information flow within the network. Specifically, the reset gate controls the degree to which past information is retained or discarded, the consumption gate manages the amount of information absorbed from the current input, and the generation gate controls how much information generates a spike in the output. These nonlinear gating mechanisms significantly enhance the system's memory retention, state regulation, and time series data processing capabilities. ; ; ; in, , and These represent the reset gate, the consumption gate, and the generation gate, respectively. , , , , and These represent the weights of each gate, obtained through training; , and The corresponding bias is also obtained through training; , and Both represent activation functions; The intermediate output generated by the neuron is: ; in, , , and These represent the weights and biases, respectively, which are obtained through network training. Indicates an intermediate state; This indicates element-wise multiplication; Finally, the state equation can be updated to: ; in, express The state that is constantly transmitted to the next moment; This indicates the output at the current moment; This indicates element-wise multiplication; Drawing on the core idea of ​​Bidirectional Long Short-Term Memory (BiLSTM) networks, a bidirectional structure is proposed to enhance contextual feature extraction capabilities, such as... Figure 3As shown, this structure maintains two independent hidden state sequences, modeling the forward and backward dependencies of the input sequence respectively, enabling the network to integrate global contextual information of the sequence at each time step. To effectively fuse the semantic features captured bidirectionally, the forward and backward state sequences are finally concatenated along the feature dimension to form a comprehensive input representation for subsequent processing modules. ; in, and These represent positive and negative context information, respectively. This indicates a concatenation operation, which yields the context state information. .

[0020] By employing an attention network and leveraging the autocorrelation of text sentences, fine-grained contextual information is obtained through filtering. Contextual weight information is then acquired through a self-attention mechanism. : ; ; ; By obtaining the weights of each word in the context, and multiplying these weights by the weights of the context words, we can obtain the importance of each word in the context. ; Context and aspect information are calculated using weighted coefficients and the results are output through a CRF network: ; Training and evaluation were performed on a dataset in the field of power transmission lines.

[0021] See details Figure 4 This study investigated the impact mechanism of the number of neurons on model performance through systematic experiments. Experimental results showed that as the number of neurons increased, the model's learning ability significantly improved, and performance exhibited a clear upward trend; however, when the number of neurons exceeded a critical value, the performance gain showed a diminishing marginal effect. Through comprehensive evaluation using multiple sets of controlled experiments, a configuration of 200 neurons was ultimately determined as the optimal architecture, achieving the best balance between model expressive power and generalization performance.

[0022] The embodiments described above are merely illustrative of implementation methods of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be defined by the appended claims.

Claims

1. A method for naming and recognizing power transmission lines based on bidirectional enhanced nonlinear pulse neural networks, characterized in that, Includes the following steps: S1. Based on the pre-trained BERT model, the input text is dynamically encoded to generate word vector representations with context awareness. The deep bidirectional attention mechanism of the Transformer architecture is used to effectively capture the multi-meaning features of words in specific contexts and transform the original text into a distributed vector sequence containing semantic information. S2. Deep contextual modeling of word vectors is performed using the bidirectional enhanced nonlinear spiking neural system BiENSNP. The bio-inspired computing module simulates the characteristics of neuronal spiking transmission and performs forward and backward information transmission; forward propagation captures historical dependencies, and backward propagation extracts future features. S3. A multi-head self-attention mechanism is used to adaptively weight the encoded semantic features. By calculating the interaction importance weights between words, the bio-inspired computing module dynamically focuses on key contextual information related to entity recognition, while suppressing irrelevant semantic noise. S4. The refined feature input conditional random field decoding layer is used for structured prediction. The feature input conditional random field solves the optimal transition path at the global sequence level by modeling the transition constraint relationship between labels.

2. The method for naming and recognizing power transmission lines based on bidirectional enhanced nonlinear pulse neural networks according to claim 1, characterized in that: In step S2, a bidirectional enhanced nonlinear spiking neural system is used to extract text context. Specifically: ； in, Text sentences are transformed into word vectors using the BERT model word embedding technique.

3. The method for naming and recognizing transmission lines based on bidirectional enhanced nonlinear pulse neural networks according to claim 1, characterized in that, Specifically, S3 is: Context weight information is obtained through a self-attention mechanism. : ； ； ； in, Text context extracted for bidirectional enhanced nonlinear spiking neural systems; Mean value is calculated from the text context extracted from a bidirectional enhanced nonlinear spiking neural system. The length of the text; The weights used for training a neural network can be obtained through model training; The bias can also be obtained through network training: To enhance the nonlinear spiking neural system for extracting text context; The correlation coefficient; Obtain the weights between each word in the context; multiply these weights by the context words to get the importance of each word in the context. ； It is a bidirectional enhanced nonlinear pulse nerve.

4. The method for naming and recognizing transmission lines based on bidirectional enhanced nonlinear pulse neural networks according to claim 3, characterized in that, The context-weighted coefficients are calculated, and the final output is processed through a CRF: ； The feature input is a conditional random field.

5. The method for naming and recognizing transmission lines based on bidirectional enhanced nonlinear pulse neural networks according to claim 3, characterized in that, Also includes: The first module implements deep semantic initialization of text. It converts the original input into dynamic word vector representation through a pre-trained BERT model, and uses the multi-layer self-attention mechanism of Transformer to parse the contextual polysemy of words and generate distributed feature vectors containing global semantic associations. The second module performs bidirectional contextual semantic modeling. Based on a bidirectional enhanced nonlinear spiking neural system, it simulates the spiking characteristics of neurons to process forward or backward semantic flows in parallel. By fusing historical and future contextual features, it explicitly captures long-distance dependencies and nonlinear patterns in construction texts, enhancing the robust parsing capability for flexible expressions. The third module: to achieve key feature enhancement and noise suppression, adopt a multi-head self-attention mechanism to recalibrate the encoded features, and dynamically amplify the entity-related context signal by calculating word-level interaction weights to weaken redundant information; The fourth module completes the global optimization of entity labeling decisions, uses conditional random fields to model label transfer constraints, and searches for the optimal labeling path in the global sequence space.

Images