A nuclear power plant similar equipment intelligent matching method based on natural language processing

Abstract

The application belongs to the technical field of artificial intelligence, and particularly relates to a nuclear power plant similar equipment intelligent matching method based on natural language processing. The method comprises the following steps: collecting relevant information for each nuclear power equipment and segmenting; constructing and training a CBOW model; constructing a GCL model; inputting original pictures and enhanced pictures in batches into the constructed GCL model; calculating a contrast loss according to the output picture representation, wherein the representations of the graph structures enhanced from the same picture are close to each other, the representations of the graph structures enhanced from different pictures are far away from each other, and the model parameters are updated through back propagation; and calculating the similarity degree between two pieces of equipment according to the equipment embedding learned by the model, and taking the similarity degree as the final similar equipment matching result. The beneficial effect is that the application proposes a new Eq2vec method, introduces a graph attention GAT and a key feature splicing technology to innovate a graph contrast learning method GCL for the technical scheme of intelligent similar equipment matching, and the cost is lower, the speed is faster, and the precision is higher.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to an intelligent matching method for similar equipment in nuclear power plants based on natural language processing. Background Technology

[0002] In the nuclear power industry, decades of operation have accumulated a vast amount of equipment operation and maintenance data. With advancements in computer science and computing power, a key challenge has become how to more effectively utilize this data and ensure power plant safety and maintain high levels of capability in a more cost-effective manner. Among these challenges, optimizing pre-maintenance strategies is a crucial direction for leveraging artificial intelligence to improve equipment reliability, and matching similar equipment is one of the key technical issues within this process.

[0003] Current matching strategies for similar equipment in nuclear power plants are simply based on rules for direct filtering and searching. However, due to the large scale of data, poor data quality, and different material coding rules for different power plants at different times, the matching effect of similar equipment often fails to meet the requirements and cannot support subsequent important related tasks and research such as pre-maintenance strategy optimization.

[0004] In recent years, Natural Language Processing (NLP), as an important research direction in the development of artificial intelligence, has achieved remarkable results in many application fields. Applying NLP technology to the computation of nuclear power plant equipment representation can achieve intelligent automatic equipment matching. However, due to the lack of a professional dictionary for the nuclear power industry, directly applying existing NLP models results in low matching accuracy, and without targeted optimization for practical application problems, it still cannot meet the needs of practical applications. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent matching method for similar equipment in nuclear power plants based on natural language processing, which can solve the problem of achieving high-precision matching of similar equipment in nuclear power plants using NLP methods.

[0006] The technical solution of the present invention is as follows: A method for intelligent matching of similar equipment in nuclear power plants based on natural language processing, comprising the following steps: Step 1: Collect relevant information for each nuclear power plant; Step 2: Segment the text of all collected relevant information; Step 3: Construct the CBOW model; Step 4: Train the CBOW model based on the vocabulary and Huffman tree constructed in Step 2; Step 5: After the CBOW model is trained, the vocabulary representation is obtained. The WMD is calculated by summing and averaging the representation of each field, and the graph topology is constructed based on the σ value. Step 6: Perform the DropEdge operation on the constructed graph topology to obtain the corresponding enhanced graph structure; Step 7: Construct the GCL model; Step 8: Input the original image and the enhanced image into the constructed GCL model in multiple batches. Each layer is represented by a GAT computing node, and key features, including device code, device name or manufacturer name, are copied and spliced ​​with the learned representation. Step 9: Calculate the contrastive loss based on the output graph representations. Graph representations augmented from the same graph are close to each other, while graph representations augmented from different graphs are far apart. Backpropagate to update the model parameters. Step 10: Based on the device embeddings learned by the model, calculate the similarity between pairs of devices using cosine similarity, and use this as the final result of similar device matching.

[0007] The relevant information mentioned in step 1 includes power plant, unit, system, equipment name, equipment description, manufacturer, equipment code, equipment type, and equipment classification.

[0008] In step 2, all the relevant information collected, including information about a certain device, includes multiple fields. The text under a certain field of a certain device constitutes a segment. Each segment of text is segmented into words, and word frequencies are counted and a vocabulary is constructed. A Huffman tree is then constructed based on the vocabulary.

[0009] In step 3, the CBOW model includes an input layer, a mapping layer, and an output layer. The input layer takes the representations of the preceding and following words as input. After mapping by the mapping layer, the output layer takes the probability distribution of the center word as input. The difference between the output probability distribution and the actual center word is used as the loss function to obtain the initial CBOW model structure.

[0010] In step 4, the word vector representation of the CBOW model adopts the one-hot method, the learning rate is initially set to 0.025, the number of words before and after (i.e. the window size) is set to 5, the minimum word frequency is ignored and set to 1, and the weight parameters are updated by backpropagation through the loss function to obtain the trained CBOW model.

[0011] In step 5, σ is a hyperparameter in the algorithm, used to define the density of the graph structure. The initial value of σ is set to 0.5.

[0012] In step 6, the α value of DropEdge is set to 0.05.

[0013] In step 7, a batch of graph structures and corresponding augmented graphs are randomly sampled, hidden layers are constructed, and the final representation of each graph is calculated through graph convolution, and a contrastive loss function is constructed.

[0014] The beneficial effects of this invention are as follows: This invention proposes a novel Eq2vec method, which is a technical solution for intelligent similar equipment matching based on the CBOW algorithm to learn nuclear power equipment vocabulary data, utilizing Word Shift Distance (WMD) graph construction, and introducing Graph Attention Neural Network (GAT) and key feature concatenation techniques to employ the innovative Graph Comparative Learning (GCL) method. This approach offers lower cost, faster speed, higher accuracy, and stronger robustness. The CBOW algorithm solves the problem of vectorizing specialized nuclear power equipment vocabulary; WMD graph construction solves the problem of structuring multi-field information in nuclear power equipment; Graph Attention Neural Network (GAT) and feature concatenation solve the problem of efficiently learning the embedding of important fields; and Graph Data Augmentation and Comparative Learning (GCL) solve the problem of lacking effective data labels. Ultimately, the Eq2vec method achieves superior similar equipment matching results. Attached Figure Description

[0015] Figure 1 The overall architecture for the device embedding learning (Eq2vec) method; Figure 2 The CBOW model structure; Figure 3 This refers to the process of learning vocabulary representations. Detailed Implementation

[0016] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] The technical problem this invention aims to solve is how to achieve high-precision matching of similar equipment in nuclear power plants using NLP methods. Traditional rule-based search methods are complex, often requiring expert input to set rules for each search, and suffer from poor robustness. Data quality issues can lead to either failing to find the desired equipment or finding too many dissimilar devices, impacting the final matching result and failing to meet the actual needs of engineers. Furthermore, conventional artificial intelligence methods are not specifically optimized for similar equipment matching tasks, generally suffering from poor vectorization of specialized terminology, inability to effectively utilize multi-field information, and low model performance due to a lack of manually labeled data.

[0018] The present invention provides a novel intelligent matching method for similar nuclear power plant equipment based on natural language processing. This method employs the Eq2vec approach, which is a technical solution that uses the CBOW algorithm to learn lexical data of nuclear power equipment, constructs graphs using word shift distance (WMD), introduces graph attention (GAT) and key feature splicing technology, and uses the innovative graph contrast learning method GCL to perform intelligent similar equipment matching. This method is lower in cost, faster in speed, higher in accuracy, and more robust.

[0019] The intelligent similar device matching method (Eq2vec) used in this invention combines a Continuous Bag-of-Words (CBOW) structure based on a neural network language model, a graph topology construction method based on Word Mover's Distance (WMD), and a device embedding learning method based on Graph Contrastive Learning (GCL) and Graph Attention Network (GAT). The overall architecture is described in [link to architecture]. Figure 1 .

[0020] This invention first uses the CBOW model structure to learn device-related lexical representations. This model includes an input layer, a mapping layer, and an output layer (see Figure 1). The core idea of ​​the CBOW model is to predict the center word using the preceding and following words. Therefore, the input layer consists of four words before and after the preceding words. After summing and mapping by the intermediate mapping layer, the output is the probability of the center word. The complete lexical representation learning process can be represented as follows: Figure 3 First, the equipment inventory data is collected and organized, mainly including information such as the power plant, unit, system, equipment name, description, manufacturer, equipment code, equipment type, and equipment classification. Then, based on this, the text data is segmented into sentences and words to construct a dataset. Simultaneously, a vocabulary list is built, word frequencies are calculated, and Huffman trees are constructed. After the above data preparation and processing are completed, the CBOW model can be trained. Upon successful training, the final distributed representation of the vocabulary is obtained.

[0021] After obtaining the lexical representations, Word Shift Distance (WMD) is used to calculate the similarity between different fields on the same device. WMD is a method for measuring the semantic similarity between two texts. Its core idea is to "move" words from one text to words in another text to minimize the total distance generated during this "movement." The smaller the distance, the higher the similarity between the two texts; conversely, the larger the distance, the lower the similarity. The advantages of WMD are that it does not require any hyperparameters, can be directly calculated, and has high interpretability and accuracy. Even if two sentences do not contain the same words, WMD can effectively identify whether their semantics are the same.

[0022] The graph topology G = {V, E} can be constructed based on the similarity calculated by WMD, where V represents the set of nodes, the number of nodes is equal to the number of fields, and E represents the set of edges. For any two nodes, WMD > σ, there is a connection. σ is a pre-set threshold used to control the density of the graph.

[0023] After the graph topology is constructed, graph data augmentation is performed, which involves randomly dropping edges (DropEdge) on all graph structures. The hyperparameter α controls the proportion of edges dropped. Then, the original graph and the augmented graph are used as inputs to the GCL. Here, the node representation is the mean of the summation of each tokenized field. Each layer in the GCL uses an attention mechanism in its computation; that is, for node pairs ij, the attention coefficient is... It can be represented as: , In the above formula, This is the attention layer. and These represent the embedding of node i and node j, respectively. Then it is the set of neighboring nodes of node i. Let R1 and R2 represent the node embedding matrix before and after the attention layer, respectively, and n represent the total number of nodes. Then: , like Figure 1 As shown, the key features of R1 are copied n times and concatenated with the matrix after graph attention to obtain R2. This step strengthens the role of key features in the device embedding learning process. Here, key features refer to the features of a node. In similar device matching tasks, key features are usually device codes, device names, or manufacturer names. Undoubtedly, these fields are of significant importance compared to other fields and can even determine the final matching result. Therefore, the key feature concatenation technique is introduced into the model architecture to help the model learn the final device embedding more accurately.

[0024] The output layer outputs a representation of each graph structure, i.e., the device embedding. The contrastive loss function calculation part of GCL can be expressed as: , , , In the above formula, images 1, 3, 5… are the original images, images 2, 4, 6… are the corresponding augmented images, and N is the total number of original image structures. This loss function can be understood as taking the logarithm of the normalized similarity between the original image and the augmented image and then inverting it. Optimizing this loss function will reduce the gap between the original image and the augmented image and amplify the gap between the original image and other images, thereby achieving a self-supervised effect, enabling the model to accurately learn the device embedding even without human labels.

[0025] After learning the device embedding, for any pair of devices, the cosine similarity can be calculated to measure the degree of similarity between devices, thus completing the intelligent matching task of similar devices.

[0026] A method for intelligent matching of similar equipment in nuclear power plants based on natural language processing includes the following steps: Step 1: Collect relevant information for each nuclear power plant; The relevant information includes power plant, generating unit, system, equipment name, equipment description, manufacturer, equipment code, equipment type, and equipment classification.

[0027] Step 2: Segment the text of all collected relevant information, including information about a certain device. The relevant information contains multiple fields. The text under a certain field of a certain device constitutes a segment. Segment each segment of text, count the word frequency, construct a word list, and construct a Huffman tree based on the word list.

[0028] Step 3: Construct the CBOW model. The CBOW model includes an input layer, a mapping layer, and an output layer. The input layer takes the representations of the preceding and following words as input. After mapping by the mapping layer, the output layer takes the probability distribution of the center word as input. The difference between the output probability distribution and the actual center word is used as the loss function to obtain the initial CBOW model structure.

[0029] Step 4: Train the CBOW model using the constructed vocabulary and Huffman tree. The word vector representation of the model adopts the one-hot encoding method. The initial learning rate is set to 0.025, the number of preceding and following words (i.e., the window size) is set to 5, the minimum word frequency to be ignored is set to 1, and the vocabulary size is unlimited. Backpropagation is used to update the weight parameters through the loss function to obtain the trained CBOW model.

[0030] Step 5: After the CBOW model is trained, the vocabulary representations are obtained. The WMD is calculated by summing and averaging the representations of each field, and the graph topology is constructed based on the σ value. σ is a hyperparameter in the algorithm, used to define the density of the graph structure. Based on practical experience, the initial σ value is set to 0.5.

[0031] Step 6: Perform the DropEdge operation on the constructed graph topology to obtain the corresponding enhanced graph structure. The α value of DropEdge is set to 0.05.

[0032] Step 7: Construct the GCL model, randomly sample a batch of graph structures and corresponding augmented graphs, construct hidden layers, calculate the final output representation of each graph through graph convolution, and construct a contrastive loss function.

[0033] Step 8: Input the original image and the enhanced image into the constructed GCL model in multiple batches. Each layer is represented by a GAT computing node, and the key features (usually device code, device name or manufacturer name) are copied and spliced ​​with the learned representation.

[0034] Step 9: Calculate the contrast loss based on the output graph representations. Graph representations augmented from the same graph are close to each other, while graph representations augmented from different graphs are far apart. Backpropagate to update the model parameters.

[0035] Step 10: Based on the device embeddings learned by the model, calculate the similarity between pairs of devices using cosine similarity, and use this as the final result of similar device matching.

Claims

1. A method for intelligent matching of similar equipment in nuclear power plants based on natural language processing, characterized in that, Includes the following steps: Step 1: Collect relevant information for each nuclear power plant; Step 2: Segment the text of all collected relevant information; Step 3: Construct the CBOW model; Step 4: Train the CBOW model based on the vocabulary and Huffman tree constructed in Step 2; Step 5: After the CBOW model is trained, the vocabulary representation is obtained. The WMD is calculated by summing and averaging the representation of each field, and the graph topology is constructed based on the σ value. Step 6: Perform the DropEdge operation on the constructed graph topology to obtain the corresponding enhanced graph structure; Step 7: Construct the GCL model; Step 8: Input the original image and the enhanced image into the constructed GCL model in multiple batches. Each layer is represented by a GAT computing node, and key features, including device code, device name or manufacturer name, are copied and spliced ​​with the learned representation. Step 9: Calculate the contrastive loss based on the output graph representations. Graph representations augmented from the same graph are close to each other, while graph representations augmented from different graphs are far apart. Backpropagate to update the model parameters. Step 10: Based on the device embeddings learned by the model, calculate the similarity between pairs of devices using cosine similarity, and use this as the final result of similar device matching.

2. The intelligent matching method for similar equipment in nuclear power plants based on natural language processing as described in claim 1, characterized in that: The relevant information mentioned in step 1 includes power plant, unit, system, equipment name, equipment description, manufacturer, equipment code, equipment type, and equipment classification.

3. The intelligent matching method for similar equipment in nuclear power plants based on natural language processing as described in claim 1, characterized in that: In step 2, all the relevant information collected, including information about a certain device, includes multiple fields. The text under a certain field of a certain device constitutes a segment. Each segment of text is segmented into words, and word frequencies are counted and a vocabulary is constructed. A Huffman tree is then constructed based on the vocabulary.

4. The intelligent matching method for similar equipment in nuclear power plants based on natural language processing as described in claim 1, characterized in that: In step 3, the CBOW model includes an input layer, a mapping layer, and an output layer. The input layer takes the representations of the preceding and following words as input. After mapping by the mapping layer, the output layer takes the probability distribution of the center word as input. The difference between the output probability distribution and the actual center word is used as the loss function to obtain the initial CBOW model structure.

5. The intelligent matching method for similar equipment in nuclear power plants based on natural language processing as described in claim 1, characterized in that: In step 4, the word vector representation of the CBOW model adopts the one-hot method, the learning rate is initially set to 0.025, the window size for the number of words before and after is set to 5, the minimum word frequency is ignored is set to 1, and the weight parameters are updated by backpropagation through the loss function to obtain the trained CBOW model.

6. The intelligent matching method for similar equipment in nuclear power plants based on natural language processing as described in claim 1, characterized in that: In step 5, σ is a hyperparameter in the algorithm, used to define the density of the graph structure. The initial value of σ is set to 0.

5.

7. The intelligent matching method for similar equipment in nuclear power plants based on natural language processing as described in claim 1, characterized in that: In step 6, the α value of DropEdge is set to 0.

05.

8. The intelligent matching method for similar equipment in nuclear power plants based on natural language processing as described in claim 1, characterized in that: In step 7, a batch of graph structures and corresponding augmented graphs are randomly sampled, hidden layers are constructed, and the final representation of each graph is calculated through graph convolution, and a contrastive loss function is constructed.

Images