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

Through the Eq2vec method, the CBOW model and graph comparison learning technology is used to solve the problem of inconsistent data quality and coding rules in the matching of similar equipment in nuclear power plants, and high-precision and low-cost equipment matching are achieved, which meets the needs of pre-dimensional strategy optimization.

CN119089211AActive Publication Date: 2024-12-06NUCLEAR POWER OPERATIONS RES INST (NPRI)
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202410858650.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-06-28
Publication Date
2024-12-06
Estimated Expiration
2044-06-28

AI Technical Summary

Technical Problem

The matching strategies of similar equipment in existing nuclear power plants are large data scale, poor data quality and inconsistent material coding rules of different power plants, resulting in poor matching effect, which cannot meet the requirements of pre-defense strategy optimization and research, and the NLP model directly applied lacks professionalism. Dictionary, low matching accuracy.

Method used

Using the Eq2vec method based on natural language processing, combined with the CBOW model, word shift distance WMD composition, graph attention GAT and graph comparison learning GCL technology, intelligent similar device matching is carried out, through vocabulary characterization learning, graph topology construction, graph data enhancement and Comparative learning to realize device embedded learning and similarity calculation.

Benefits of technology

It realizes high-precision, low-cost, fast and robust matching of similar equipment, solves the problems of professional vocabulary vectorization, multi-field information structure and lack of labeled data, and improves matching accuracy and efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119089211A_ABST
    Figure CN119089211A_ABST
Patent Text Reader

Abstract

The invention belongs to the technical field of artificial intelligence, and particularly relates to an intelligent matching method for similar equipment of a nuclear power plant based on natural language processing. Comprising the following steps: collecting related information for each nuclear power device and segmenting the information; a CBOW model is constructed and trained; constructing a GCL model; inputting the original image and the enhanced image into the constructed GCL model in multiple batches; calculating comparison loss according to the output graph representation, enabling the representation of the graph structure enhanced by the same graph to be close to each other, enabling the representation of the graph structure enhanced by different graphs to be far away from each other, and updating model parameters through back propagation; and according to the equipment embedding learned by the model, calculating the similarity degree between every two pieces of equipment by utilizing cosine similarity, and taking the similarity degree as a final similar equipment matching result. The method has the beneficial effects that a brand new Eq2vec method is provided, the technical scheme that graph attention GAT and a key feature splicing technology are introduced to innovate a graph contrast learning method GCL for intelligent similar equipment matching is adopted, the cost is lower, the speed is higher, and the precision is higher.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence, and in particular relates to an intelligent matching method for similar equipment in a nuclear power plant based on natural language processing. Background Art

[0002] In the nuclear power industry, a large amount of equipment operation and maintenance data has been accumulated during decades of operation. Due to the development of computer science and the advancement of computing power, how to apply this data more effectively and ensure the safety of power plants and maintain a high level of capacity in a more cost-effective way has become a key issue. Among them, the optimization of pre-maintenance strategies is an important direction for using artificial intelligence to improve equipment reliability, and the matching of similar equipment is one of the key technical issues.

[0003] The current similar equipment matching strategies for nuclear power plants are simply based on direct screening and searching based on rules. However, due to many reasons such as large data scale, poor data quality, and different material coding rules in different power plants at different times, the similar equipment matching results often cannot meet the needs 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 excellent results in many application fields. Applying NLP technology to the calculation of nuclear power plant equipment characterization can realize intelligent automatic equipment matching. However, due to the lack of professional dictionaries in the nuclear power industry, the direct application of existing NLP models has the problem of low matching accuracy, and there is no targeted optimization for actual application problems, which still cannot meet the needs of actual applications. Summary of the invention

[0005] The purpose of the present invention is to provide a method for intelligent matching of 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 a nuclear power plant based on natural language processing, comprising the following steps:

[0007] Step 1: Collect relevant information for each nuclear power equipment;

[0008] Step 2: Segment the text of all collected relevant information;

[0009] Step 3: Build the CBOW model;

[0010] Step 4: Train the CBOW model based on the vocabulary and Huffman tree constructed in step 2;

[0011] Step 5: After the CBOW model training is completed, the vocabulary representation is obtained, the representation of each field is summed and averaged to calculate the WMD, and the graph topology is constructed according to the σ value;

[0012] Step 6: Perform DropEdge operation on the constructed graph topology structure to obtain the corresponding enhanced graph structure;

[0013] Step 7: Construct the GCL model;

[0014] Step 8: Input the original image and enhanced image into the constructed GCL model in batches. Each layer is represented by a GAT computing node, and the key features, including device code, device name or manufacturer name, are copied and concatenated with the learned representation.

[0015] Step 9: Calculate the contrast loss based on the output graph representation. The representations of graph structures enhanced by the same graph are close to each other, while the representations of graph structures enhanced by different graphs are far away from each other. Back propagation is used to update the model parameters.

[0016] Step 10: Based on the device embedding learned by the model, the cosine similarity is used to calculate the similarity between two devices as the final similar device matching result.

[0017] The relevant information described in step 1 includes power plant, unit, system, equipment name, equipment description, manufacturer, equipment code, equipment type, and equipment classification.

[0018] All the relevant information collected in step 2, including the relevant information of a certain device, includes multiple fields. The text under a certain field of a certain device constitutes a paragraph. Each paragraph of text is segmented, and the word frequency is counted, a word list is constructed, and a Huffman tree is constructed based on the word list.

[0019] In step 3, the CBOW model includes an input layer, a mapping layer, and an output layer. The content of the input layer is the representation of the previous and next words. After being mapped by the mapping layer, the content of the output layer is the probability distribution of the central word. The difference between the output probability distribution and the actual central word is used as the loss function to obtain the initial CBOW model structure.

[0020] In step 4, the word vector representation of the CBOW model adopts the one-hot method, the initial learning rate is set to 0.025, the number of previous and next words, that is, the window size is set to 5, the minimum word frequency is ignored and set to 1, and the weight parameters are updated through back propagation through the loss function calculation to obtain the trained CBOW model.

[0021] 6. A method for intelligent matching of similar equipment in a nuclear power plant based on natural language processing as described in claim 1, characterized in that: in step 5, σ is a hyperparameter in the algorithm, which is used to define the density of the graph structure, and the initial σ value is set to 0.5.

[0022] The alpha value of DropEdge in step 6 is set to 0.05.

[0023] In step 7, a batch of graph structures and corresponding enhanced graphs are randomly sampled, a hidden layer is constructed, the representation of each graph is finally output through graph convolution calculation, and a contrast loss function is constructed.

[0024] The beneficial effects of the present invention are as follows: the present invention proposes a brand-new Eq2vec method, which is a technical solution for intelligent similar equipment matching based on the CBOW algorithm to learn the vocabulary data of nuclear power equipment, use the word shift distance WMD composition, introduce the graph attention GAT and key feature splicing technology to innovate the graph comparative learning method GCL, with lower cost, faster speed, higher accuracy and stronger robustness. The CBOW algorithm solves the problem of vectorization of professional vocabulary of nuclear power equipment, the word shift distance (WMD) composition solves the problem of structuring multi-field information of nuclear power equipment, the graph attention neural network (GAT) and feature splicing solve the problem of efficient learning of important field embedding, and the graph data enhancement and comparative learning (GCL) solve the problem of lack of effective data labels. Finally, the Eq2vec method achieves similar equipment matching results with better performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 The overall architecture of the device embedding learning (Eq2vec) method;

[0026] Figure 2 It is the CBOW model structure;

[0027] Figure 3 For the vocabulary representation learning process. DETAILED DESCRIPTION

[0028] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0029] The technical problem to be solved by the present invention is how to use NLP methods to achieve high-precision matching of similar equipment in nuclear power plants. The traditional direct rule-based search method is relatively complicated. Experts may need to provide experience to set rules for each search, and the method has poor robustness. It is very likely that the desired equipment cannot be found due to data quality reasons or too many dissimilar equipment will be found, affecting the final matching effect and failing to meet the actual needs of engineers. In addition, conventional artificial intelligence methods cannot optimize similar equipment matching tasks in a targeted manner. There are many common problems such as poor vectorization of professional vocabulary, inability to comprehensively and effectively utilize multi-field information, and lack of manually labeled data leading to poor model performance.

[0030] The method for intelligent matching of similar equipment in nuclear power plants based on natural language processing provided by the present invention adopts a new Eq2vec method. It is a technical solution for intelligent matching of similar equipment based on the CBOW algorithm to learn the vocabulary data of nuclear power equipment, use the word shift distance WMD to compose the graph, introduce the graph attention GAT and key feature splicing technology to innovate the graph comparison learning method GCL, which has lower cost, faster speed, higher accuracy and stronger robustness.

[0031] The intelligent similar device matching method (Eq2vec) adopted in the present invention combines the Continuous Bag-of-Words (CBOW) structure based on the neural network language model, the graph topology structure construction method based on the Word Mover's Distance (WMD), and the device embedding learning method based on Graph Contrastive Learning (GCL) and Graph Attention Network (GAT). The overall architecture is shown in Figure 1 .

[0032] The present invention first uses the CBOW model structure to learn the device-related vocabulary representation. The model includes an input layer, a mapping layer, and an output layer. The core idea of ​​the CBOW model is to use the previous and next words to predict the central word. Therefore, the input layer is four words before and after, and after the middle mapping layer sums and maps, the output is the probability of the central word. The complete vocabulary representation learning process can be expressed as Figure 3 . First, the equipment list data is collected and sorted, mainly collecting the power plant, unit, system, equipment name, equipment description, manufacturer, equipment code, equipment type, equipment classification and other related content of the equipment. Then, on this basis, the text data is divided into sentences and words to build a data set. At the same time, a word list is also built, word frequency is counted, and a Huffman tree is built. After the above data preparation and processing work is completed, the CBOW model can be trained. After the training is completed, the distributed representation of the final vocabulary can be obtained.

[0033] After obtaining the vocabulary representation, the word shift distance WMD is used to calculate the similarity between different fields of the same device. Word shift distance is a method to measure the semantic similarity between two texts. Its core idea is to "move" the vocabulary in one text to the vocabulary in another text to minimize the sum of the distances generated in this "movement" process. The smaller the distance, the higher the similarity between the two texts; conversely, the larger the distance, the lower the similarity. The advantage of WMD is that it does not require any hyperparameters, can be calculated directly, and has high interpretability and accuracy. Even if the same words do not appear in two sentences, WMD can effectively identify whether their semantics are the same.

[0034] The graph topology structure G = {V, E} can be constructed based on the similarity calculated by WMD, where V represents a set of nodes, the number of nodes is equal to the number of fields, and E represents a set of edges. For any two nodes, if WMD>σ, there is a connection, and σ is a threshold set in advance to control the density of the graph.

[0035] After the graph topology is constructed, the graph data is enhanced, that is, all graph structures are randomly dropped (DropEdge) operation is performed, and the hyperparameter α is used to control the ratio of dropped edges. Then the original graph and the enhanced graph are used as the input of GCL. Among them, the node representation is the mean of the sum of the representations after each field is segmented. Each layer in GCL uses the attention mechanism to participate in the operation, that is, for the node pair ij, the attention coefficient α ij It can be expressed as:

[0036]

[0037] In the above formula, is the attention layer, and Then they represent the embedding of node i and node j respectively, N i Then is the set of neighbor nodes of node i, and w is the parameter matrix. In the process of attention calculation, key feature splicing operation will be performed, with R1 and R2 representing the node embedding matrix before the attention layer and the node embedding matrix after the attention layer respectively, and n representing the total number of nodes, then:

[0038] R2=concatenate(GAT(R1),R1 copy )

[0039] R1 copy =copy(R1[:1,:],n)

[0040] like Figure 1As shown, the key features of R1 are copied n times and concatenated with the matrix after graph attention to obtain R2. This step makes the key features more effective in the process of device embedding learning. The key features here are node features. In similar device matching tasks, the key features are usually device codes, device names or manufacturer names. There is no doubt that these fields are more important than other fields and can even determine the final matching results. Therefore, the key feature concatenation technology is introduced in the model architecture to help the model learn the final device embedding more accurately.

[0041] The output result of the output layer is the representation of each graph structure, that is, the embedding of the device, where the calculation part of the contrast loss function of GCL can be expressed as:

[0042]

[0043]

[0044] s ij =cos_sim(f i , f j )

[0045] In the above formula, the 1st, 3rd, 5th, ... images are the original images, the 2nd, 4th, 6th, ... images are the corresponding enhanced images, and N is the total number of original image structures. This loss function can be understood as the logarithm of the normalized similarity between the original image and the enhanced image and then the inverse. Optimizing this loss function will reduce the gap between the original image and the enhanced image, and amplify the gap between the original image and other images, thereby achieving the effect of self-supervision, so that the model can accurately learn the device embedding without manual labels.

[0046] After learning the device embedding, for any pair of devices, the similarity between the devices can be measured by calculating the cosine similarity, thus completing the task of intelligent matching of similar devices.

[0047] A method for intelligent matching of similar equipment in a nuclear power plant based on natural language processing comprises the following steps:

[0048] Step 1: Collect relevant information for each nuclear power equipment;

[0049] The relevant information includes power plant, unit, system, equipment name, equipment description, manufacturer, equipment code, equipment type, and equipment classification.

[0050] Step 2: Segment the text of all collected relevant information, including the relevant information of 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, build a word list, and build a Huffman tree based on the word list.

[0051] Step 3: Construct a CBOW model. The CBOW model includes an input layer, a mapping layer, and an output layer. The content of the input layer is the representation of the previous and next words. After being mapped by the mapping layer, the content of the output layer is the probability distribution of the central word. The difference between the output probability distribution and the actual central word is used as the loss function to obtain the initial CBOW model structure.

[0052] Step 4: Train the CBOW model based on the constructed vocabulary and Huffman tree. The word vector representation of the model uses the one-hot method. The initial learning rate is set to 0.025, the number of previous and next words, that is, the window size is set to 5, the minimum word frequency is set to 1, and the vocabulary size is unlimited. The weight parameters are updated through back propagation through the loss function calculation to obtain the trained CBOW model.

[0053] Step 5: After the CBOW model training is completed, the vocabulary representation is obtained, the representation of each field is summed and averaged to calculate the WMD, and the graph topology is constructed according to the σ value. σ is a hyperparameter in the algorithm, which is used to define the density of the graph structure. Based on practical experience, the initial σ value is set to 0.5.

[0054] Step 6: Perform DropEdge operation on the constructed graph topology to obtain the corresponding enhanced graph structure. The α value of DropEdge is set to 0.05.

[0055] Step 7: Build the GCL model, randomly sample a batch of graph structures and corresponding enhanced graphs, build the hidden layer, calculate the final output of each graph representation through graph convolution, and build the contrast loss function.

[0056] Step 8: Input the original image and enhanced image into the constructed GCL model in 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 concatenated with the learned representation.

[0057] Step 9: Calculate the contrast loss based on the output graph representation. The representations of graph structures enhanced by the same graph are close to each other, while the representations of graph structures enhanced by different graphs are far away from each other. Back-propagation is used to update the model parameters.

[0058] Step 10: Based on the device embedding learned by the model, the cosine similarity is used to calculate the similarity between two devices as the final similar device matching result.

Claims

1. A method for intelligent matching of similar equipment in nuclear power plants based on natural language processing, characterized in that: The steps include: Step 1: Collect relevant information for each nuclear power equipment; Step 2: Segment the text of all collected relevant information; Step 3: Build 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 training is completed, the vocabulary representation is obtained, the representation of each field is summed and averaged to calculate the WMD, and the graph topology is constructed according to the σ value; Step 6: Perform DropEdge operation on the constructed graph topology structure to obtain the corresponding enhanced graph structure; Step 7: Construct the GCL model; Step 8: Input the original image and enhanced image into the constructed GCL model in batches. Each layer is represented by a GAT computing node, and the key features, including device code, device name or manufacturer name, are copied and concatenated with the learned representation. Step 9: Calculate the contrast loss based on the output graph representation. The representations of graph structures enhanced by the same graph are close to each other, while the representations of graph structures enhanced by different graphs are far away from each other. Back propagation is used to update the model parameters. Step 10: Based on the device embedding learned by the model, the cosine similarity is used to calculate the similarity between two devices as the final similar device matching result.

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

3. The method for intelligent matching of similar equipment in a nuclear power plant based on natural language processing according to claim 1, characterized in that: All the relevant information collected in step 2, including the relevant information of a certain device, includes multiple fields. The text under a certain field of a certain device constitutes a paragraph. Each paragraph of text is segmented, and the word frequency is counted, a word list is constructed, and a Huffman tree is constructed based on the word list.

4. The method for intelligent matching of similar equipment in a nuclear power plant based on natural language processing according to claim 1, characterized in that: In step 3, the CBOW model includes an input layer, a mapping layer, and an output layer. The content of the input layer is the representation of the previous and next words. After being mapped by the mapping layer, the content of the output layer is the probability distribution of the central word. The difference between the output probability distribution and the actual central word is used as the loss function to obtain the initial CBOW model structure.

5. The method for intelligent matching of similar equipment in a nuclear power plant based on natural language processing according to claim 1, characterized in that: In step 4, the word vector representation of the CBOW model adopts the one-hot method, the initial learning rate is set to 0.025, the number of previous and next words, that is, the window size is set to 5, the minimum word frequency is ignored and set to 1, and the weight parameters are updated through back propagation through the loss function calculation to obtain the trained CBOW model.

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

5.

7. The method for intelligent matching of similar equipment in a nuclear power plant based on natural language processing according to claim 1, characterized in that: The alpha value of DropEdge in step 6 is set to 0.

05.

8. The method for intelligent matching of similar equipment in a nuclear power plant based on natural language processing according to claim 1, characterized in that: In step 7, a batch of graph structures and corresponding enhanced graphs are randomly sampled, a hidden layer is constructed, the representation of each graph is finally output through graph convolution calculation, and a contrast loss function is constructed.

Citation Information

Patent Citations

  • Text structuring method aiming at power grid fault cases

    CN107992597A

  • Text similarity measure method combining word clustering and word combination semantic features

    CN108399163A

  • An intelligent operation and maintenance statement similarity matching method based on natural language processing

    CN109902159A

  • Knowledge graph fusion method based on entity alignment

    CN110941722A

  • Heterogeneous device interaction integration system and method based on OPC UA

    CN117411946A