A nuclear power construction experience feedback method based on multi-dimensional cascading knowledge extraction

Abstract

The present application belongs to the technical field of nuclear power construction experience feedback, and particularly relates to a nuclear power construction experience feedback method based on multi-dimensional cascading knowledge extraction. The method breaks the traditional text recording mode, and presents the experience feedback knowledge result by using a knowledge graph carrier. A unified data experience feedback graph is constructed, and under the guidance of the key knowledge domain of experience feedback, the intelligent extraction of experience knowledge is carried out by using a natural language processing technology, and the structured knowledge is stored in a graph database to realize the knowledge of experience feedback. The method realizes the change of nuclear power construction experience feedback information from a document center to a knowledge center, and stores the experience feedback knowledge in the knowledge base in a unified and clear manner. In this way, when a quality event occurs in the process of nuclear power construction, the required knowledge can be quickly and accurately obtained by searching in the knowledge base, which is used for assisting the nuclear power professional technicians and / or managers in analysis and processing.

Description

Technical Field

[0001] This invention belongs to the field of nuclear power plant construction experience feedback technology, and particularly relates to a nuclear power plant construction experience feedback method based on multi-dimensional cascaded knowledge extraction. Background Technology

[0002] Nuclear power plant construction is a knowledge-intensive industry involving multiple disciplines, characterized by high safety requirements, long service life, and large-scale systems. Due to the extremely high costs associated with quality accidents, all sectors are making every effort to avoid them. Therefore, nuclear facility operators should establish a nuclear safety experience feedback system. Experience feedback refers to the collection, screening, evaluation, analysis, processing, and distribution of information on nuclear facility incidents, quality problems, and best practices. The aim is to summarize and promote best practices to prevent the recurrence of similar quality incidents and problems.

[0003] With the development of nuclear power engineering and the advancement of intelligent manufacturing technology, complex and ever-changing application scenarios have placed higher demands on the collaborative quality control of nuclear power construction. Nuclear power companies have begun to study how to further utilize and transmit experience feedback information, shifting from a traditional document-centric approach to a knowledge-centric one. To this end, Zhao Dong et al. developed an experience feedback management system, providing a unified platform for the archiving, summarizing, publishing, and reuse of nuclear power engineering design experience; Zou Xiang et al. proposed an AHP-based method for evaluating the importance of safety-related experience feedback information in nuclear power plants, assessing the relative importance of information to obtain feedback value and identify important experience feedback information. These platforms and methods have accelerated information sharing among departments and professional units to some extent, but the granularity of the transmitted experience feedback information is relatively coarse, and the key elements implicit in it still require further analysis and processing by professionals, making it difficult to directly and effectively utilize the experience feedback information. The nuclear power field urgently needs to leverage intelligent technology to analyze and extract experience feedback information and obtain important content, to organize and integrate the large number of quality event records generated during the construction process into structured experience feedback knowledge, assisting relevant personnel in quickly analyzing key elements, and realizing the efficient cross-enterprise flow and reuse of nuclear power construction quality information.

[0004] The significant achievements of machine learning in knowledge mining in recent years can be applied to nuclear power plant construction experience feedback research. Natural Language Processing (NLP) can use computers to process, understand, and standardize the expression of human language text, intelligently parsing important knowledge in experience feedback information, thereby improving its usability. Knowledge graphs proposed by SINGHAL et al. of Google can provide a unified visual representation of important knowledge in key domains of experience feedback, facilitating the efficient utilization and transmission of large-scale experience feedback information. Knowledge graphs based on NLP technology can effectively reduce the reliance on prior human knowledge in nuclear power plant experience feedback information processing, reducing human error to some extent, further enhancing the empowering role of data elements while meeting the needs of intelligent transmission across enterprises. Therefore, some scholars have applied natural language processing technology to the nuclear power field: Choi et al. extracted syntactic and semantic information from the operating procedures of nuclear power plants using rule-based information extraction methods; Deng Zhiguang et al. extracted fault features from on-site data of reactor outlet temperature in nuclear power plants by improving long short-term memory neural networks.

[0005] In summary, some scholars have made some attempts to mine nuclear power knowledge using machine learning and other methods. However, most of the research relies on manual methods, rule templates, or single algorithms. The generalization ability of the models is weak and the application support is insufficient. In particular, there is still a certain gap in the research on multi-dimensional correlation mining of nuclear power construction experience feedback information.

[0006] Therefore, how to effectively solve the problems of weak timeliness of experience feedback, inconsistent understanding of experience knowledge among various participants, and low reuse rate of experience feedback knowledge in the process of nuclear power plant construction has become an urgent issue to be addressed. Summary of the Invention

[0007] To address the shortcomings of the existing technologies, this invention provides a nuclear power plant construction experience feedback method based on multi-dimensional cascaded knowledge extraction, which can effectively solve problems such as weak timeliness of experience feedback, inconsistent understanding of experience knowledge among various participants, and low reuse rate of experience feedback knowledge during the nuclear power plant construction process.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0009] A nuclear power plant construction experience feedback method based on multidimensional cascaded knowledge extraction includes the following steps:

[0010] S1. Obtain experience feedback information from records of internally generated quality events throughout the entire lifecycle of nuclear power plant construction;

[0011] S2. Analyze the characteristics of experience feedback information to obtain key knowledge categories, hierarchically associate and couple various key knowledge categories, define the organizational structure of key knowledge domains of experience feedback, and obtain the ontology model of experience feedback graph.

[0012] S3. Based on the ontology model of S2, construct a multi-dimensional cascaded knowledge extraction model to obtain the filling data of the ontology model from the experience feedback information.

[0013] The multidimensional cascaded knowledge extraction model includes a knowledge mining model and a relation extraction model; the knowledge mining model is used to perceive and extract key knowledge of various categories in the experience feedback information; the relation extraction model is used to extract the relation features of key knowledge in the experience feedback information.

[0014] S4. Using the multidimensional knowledge extraction model built by S3, structured fill data is extracted from the experience feedback information obtained by S1, and the extracted fill data is mapped to the ontology model of S2 to generate the corresponding experience feedback graph and store it in the knowledge base.

[0015] S5. When a quality incident occurs during the construction of a nuclear power plant, the necessary knowledge is retrieved from the knowledge base to assist nuclear power professionals and / or managers in analysis and processing.

[0016] Compared with the prior art, the present invention has the following beneficial effects:

[0017] 1. This method can transform the experience feedback information from the records of quality events related to construction and governance into knowledge, and construct a unified experience feedback graph through multi-dimensional cascaded knowledge mining. Specifically, after acquiring the experience feedback information, its characteristics are analyzed, multiple types of high-value knowledge are hierarchically associated and coupled, and the organizational structure of key knowledge domains of experience feedback is defined, thereby constructing an ontology model (i.e., a framework model) for the experience feedback graph. Then, based on the model architecture of the ontology model, a multi-dimensional cascaded knowledge extraction model is constructed to extract key knowledge of each category from the experience feedback information, extract the relational features of each key knowledge, and map them into the ontology model to obtain the corresponding experience feedback graph, which is then stored in a knowledge base.

[0018] This method breaks away from traditional text-based recording methods, utilizing a knowledge graph to present the knowledge-based results of experience feedback. It constructs a unified experience feedback graph, and under the guidance of key knowledge domains in experience feedback, intelligently extracts experience knowledge using natural language processing technology. This structured knowledge is stored in a graph database to achieve the knowledge-based representation of experience feedback. This method transforms nuclear power plant construction experience feedback information from a document-centric to a knowledge-centric approach, storing experience feedback knowledge in a unified and clear manner in a knowledge base. Thus, when quality incidents occur during nuclear power plant construction, the necessary knowledge can be quickly and accurately retrieved from the knowledge base to assist nuclear power professionals and / or managers in analysis and processing.

[0019] 2. When constructing the experience feedback graph, this method employs a hybrid top-down and bottom-up approach. It organizes and reorganizes a large amount of complex and lengthy quality information, using key knowledge domains of experience feedback as the graph's schema layer. Simultaneously, based on the characteristics of nuclear power experience feedback texts, it utilizes natural language processing techniques to design named entity recognition (i.e., a knowledge mining model) and relation extraction methods. This extracts structured key knowledge groups as a data layer, mapping them into the ontology model and updating the knowledge structure. This ensures the high efficiency and effectiveness of the resulting experience feedback graph.

[0020] In summary, this method can effectively solve problems such as the weak timeliness of experience feedback, inconsistent understanding of experience knowledge among various stakeholders, and low reuse rate of experience feedback knowledge in the nuclear power plant construction process.

[0021] Preferably, S2 includes:

[0022] By analyzing the characteristics of experience feedback information, we can obtain key knowledge categories related to fault performance.

[0023] Each knowledge category is formed into a knowledge layer of the key knowledge domain for experience feedback, resulting in a domain knowledge spectrum. At the same time, the interaction between each knowledge layer is used to achieve correlation and coupling, resulting in a spatial knowledge structure with semantic association and logical reasoning capabilities, which serves as the organizational structure for the key knowledge domain for experience feedback.

[0024] Subsequently, for different knowledge layers in the key knowledge domain of experience feedback, as well as the relationships between each layer, class and object attributes are constructed to establish semantic relationships between node types and concepts or instances, obtain knowledge in the form of <entity-relationship-entity> structure, and explicitly represent the key knowledge domain of experience feedback to form an ontology model of experience feedback graph.

[0025] In this way, by defining unified data standards for key knowledge domains of experience feedback, we can clarify the objectives of knowledge mining and provide pattern support for the subsequent construction of experience feedback graphs.

[0026] Preferably, the knowledge categories include equipment information about the fault, the cause of the fault, the countermeasures, as well as the time of the fault, the stage to which it occurred, the related equipment, the unit to which it belongs, the main responsible personnel, and the enterprise department; the relationships include causal relationships, dependency relationships, hierarchical relationships, and temporal relationships.

[0027] This setup ensures the comprehensive effectiveness of key knowledge within the ontology model, thereby guaranteeing the effectiveness of the subsequent experience feedback graph.

[0028] Preferably, in S3, the knowledge mining model is a multi-layer neural network model that integrates attention mechanism and is built using a cascaded strategy; the knowledge mining model includes a word embedding layer, a knowledge perception layer and a knowledge prediction layer set in sequence.

[0029] To address the challenges of complex textual content and diverse relationship categories in nuclear power plant construction experience feedback information, single network models struggle to effectively identify and learn key features, making it difficult to mine, organize, and represent experience feedback knowledge. This new approach enables multi-dimensional feature mining of experience feedback information in the nuclear power field, efficiently extracting key domain knowledge and thus improving mining effectiveness in the niche area of ​​nuclear power.

[0030] Preferably, the word embedding layer is designed based on the BERT bidirectional encoder representation technology of Transformer. The working process of the word embedding layer includes: the input has three different embedding layers, which respectively represent the word information, sentence paragraph information and sentence position information of the input quality event text. The three layers of information vectors are concatenated as the input of BERT, and after passing through multiple Transformers, an output vector containing semantic features with multiple dimensions is obtained.

[0031] This allows computer models to recognize text in the nuclear power field, obtain output vectors containing semantic features of multiple dimensions, and provide numerical data input for subsequent models.

[0032] Preferably, the knowledge perception layer is constructed based on the Bidirectional Long Short-Term Memory Network (BiLSTM) fused with the Attention mechanism. The BiLSTM network is used to bidirectionally mine historical and future information in quality event segments, comprehensively linking long-term features. The Attention mechanism is used to calculate the degree of correlation between features between words in the experience feedback information, enabling the knowledge perception layer to focus its attention on the nuclear power-related area when processing information, so as to obtain more effective semantic information.

[0033] The texts of quality events generated during nuclear power plant construction are typically quite long, and models tend to forget past information during the learning and recognition process. This setting allows for the acquisition of more effective semantic information.

[0034] Preferably, the knowledge prediction layer is constructed based on conditional random fields and is used to optimize the output results of the knowledge perception layer. During prediction, the emission matrix and transition matrix are combined, and the highest-scoring labeled sequence corresponding to the sentence label is calculated based on the state score and transition score output by the upper layer, which is used as the corresponding key knowledge.

[0035] The results from the knowledge-aware layer only learn each word in the text of nuclear power quality events, without considering the constraints between tags. For example, "pressure vessel" is a fixed technical term in the nuclear power field, and the order of the words cannot be changed. Therefore, the connection order of the tags needs to be constrained. Introducing Conditional Random Fields (CRF) to optimize the output of the upper layer can ensure that the empirical feedback knowledge in the final prediction is reasonable and effective.

[0036] Preferably, the relation extraction model is optimized from a segmented convolutional neural network (PCNN) based on multi-instance learning, including:

[0037] The vector representation layer is used to convert the input key knowledge into an embedding vector representation matrix; the embedding vector representation matrix includes word embedding vectors and position embedding vectors, the word embedding vectors are used to indicate the text words in the key knowledge, and the position embedding vectors are used to indicate the position information of each text word in the key knowledge.

[0038] Convolutional layers are used to extract local features from the embedding vector representation matrix in different regions, resulting in local feature maps for each region.

[0039] The piecewise max pooling layer is used to perform piecewise max pooling operations on the local feature maps of different regions to obtain the feature maps with the maximum pooling value for each region.

[0040] The fully connected layer is used to concatenate the pooled maximum feature maps corresponding to different regions, and then activate them using the Gaussian error linear unit (GELU) as the activation function to obtain the fully connected feature map of key knowledge.

[0041] The classification layer is used to make classification decisions and extract relationships from the fully connected feature maps of key knowledge, thereby obtaining the relational features of the textual information in the key knowledge and outputting them.

[0042] Because the nuclear power plant experience feedback information representing quality events is quite long, traditional methods for extracting relationships based on lexical and syntactic features have low accuracy and are prone to introducing noisy data. This setting can improve the generalization ability of the experience feedback information relationship extraction model and avoid the gradient vanishing problem.

[0043] Preferably, the activation function expression of the Gaussian error linear unit (GELU) is:

[0044]

[0045] Where x represents the input to the activation function, and π is the mathematical constant pi.

[0046] Preferably, in S5, the required knowledge is retrieved from the knowledge base using a reasoning engine.

[0047] This ensures the convenience and effectiveness of knowledge acquisition. Attached Figure Description

[0048] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0049] Figure 1 This is a flowchart illustrating the method.

[0050] Figure 2 This is a schematic diagram of the organizational structure of the key knowledge domains for experience feedback in Example 1;

[0051] Figure 3 This is a schematic diagram of the ontology model of the experience feedback graph in Example 1;

[0052] Figure 4 This is a schematic diagram of the knowledge mining model in Example 1;

[0053] Figure 5 This is a schematic diagram of the PCNN network model structure in Example 1;

[0054] Figure 6 This is a schematic diagram of the nuclear power plant construction experience feedback graph in Example 1;

[0055] Figure 7 This is a schematic diagram of the experience feedback knowledge application framework in Example 1;

[0056] Figure 8 This is a schematic diagram showing the performance comparison results of the knowledge mining model in Example 2;

[0057] Figure 9 This is a schematic diagram illustrating the association of quality issues in Example 3;

[0058] Figure 10 This is a schematic diagram of entity association knowledge in Example 3;

[0059] Figure 11 This is a schematic diagram of the fault knowledge spectrum in Example 3;

[0060] Figure 12 This is a schematic diagram of RPV-related knowledge retrieval in Example 3;

[0061] Figure 13 This is a schematic diagram of the experience feedback graph application process in Example 3. Detailed Implementation

[0062] The following detailed explanation illustrates the specific implementation methods:

[0063] Example 1

[0064] like Figure 1 As shown, this embodiment discloses a nuclear power plant construction experience feedback method based on multi-dimensional cascaded knowledge extraction, including the following steps:

[0065] S1. Obtain experience feedback information from records of internal quality events generated throughout the entire lifecycle of nuclear power plant construction.

[0066] Feedback on nuclear power plant construction comes from a wide range of sources, primarily internal and external. External experience mainly comes from the World Federation of Nuclear Owners' (WANO) Major Event Reports (SER) and Major Operating Event Reports (SOER). On the one hand, it involves root cause analysis of problems and events that have occurred at the plant itself and at other plants worldwide to prevent the recurrence of similar issues. For example, after the Fukushima nuclear accident, nuclear power companies learned from the experience, incorporated safety improvements into their design standards, put forward clearer requirements for the prevention and mitigation of serious accidents, and increased research on the Probabilistic Safety Assessment (PSA) technology system for nuclear power plant construction. On the other hand, it involves absorbing best practices from other plants worldwide, conducting internal comparative reviews, proposing improvement action plans, and tracking their implementation.

[0067] Compared to external feedback, internal experience feedback is more timely and effective during nuclear power plant construction. The construction process generates and accumulates a large amount of records, including DEN, UES, NCR, and FCR, covering all aspects and stages of the nuclear power project's lifecycle, from design, procurement and manufacturing, construction and installation, to commissioning and operation. These records, such as design drawings not meeting requirements or on-site quality verification failing, are escalated into quality events. These events provide detailed and accurate accounts of the process, causes, specific corrective measures, and improvement plans. This type of information is more relevant and usable, and requires timely processing and dissemination.

[0068] Therefore, in order to improve the efficiency of experience feedback, the scope of experience feedback knowledge in this method mainly includes the records of quality events generated internally throughout the entire life cycle of nuclear power plant construction.

[0069] S2. Analyze the characteristics of experience feedback information to obtain key knowledge categories, hierarchically associate and couple various key knowledge categories, define the organizational structure of key knowledge domains of experience feedback, and obtain the ontology model of experience feedback graph.

[0070] In specific implementation, S2 includes:

[0071] By analyzing the characteristics of experience feedback information, we can obtain key knowledge categories related to fault performance.

[0072] Each knowledge category is formed into a knowledge layer of the key knowledge domain for experience feedback, resulting in a domain knowledge spectrum. At the same time, the interaction between each knowledge layer is used to achieve correlation and coupling, resulting in a spatial knowledge structure with semantic association and logical reasoning capabilities, which serves as the organizational structure for the key knowledge domain for experience feedback.

[0073] Subsequently, for different knowledge layers in the key knowledge domain of experience feedback, as well as the relationships between each layer, class and object attributes are constructed to establish semantic relationships between node types and concepts or instances, obtain knowledge in the form of <entity-relationship-entity> structure, and explicitly represent the key knowledge domain of experience feedback to form an ontology model of experience feedback graph.

[0074] The knowledge categories include equipment information about the fault, the cause of the fault, the countermeasures, as well as the time of the fault, the stage to which it occurred, the related equipment, the unit to which it belongs, the main responsible personnel, and the enterprise department; the relationships include causal relationships, dependency relationships, hierarchical relationships, and temporal relationships.

[0075] The records of quality events generated internally throughout the entire life cycle of nuclear power plant construction contain a wide variety of content and categories, but not all information is meaningful. In order to realize the knowledge-based feedback of experience, it is necessary to select and define the elements with feedback value, form a unified and standardized knowledge domain, and ensure the consistency and interpretability of information.

[0076] The key knowledge domain of experience feedback is a collection of key knowledge from experience feedback information, and it is a hierarchical description of quality events in the construction process. Its content mainly includes key domain entities and domain relationships. Key domain entities refer to entities that have important significance and role in the nuclear power construction field, including events, objects, locations, causes, actions, personnel, etc. Domain relationships describe the connections, dependencies, or influences between them, such as causal relationships, dependency relationships, hierarchical relationships, and temporal relationships.

[0077] Statistical analysis of experience feedback information from nuclear power companies reveals that descriptive information about malfunctions appears frequently in documented quality events, possessing extremely high data value. This data is a key element that all stakeholders need to grasp during construction and is crucial to the efficiency and quality of nuclear power plant construction. Based on this analysis and in conjunction with expert opinions in related fields, we preliminarily define the key knowledge domain of experience feedback as revolving around malfunctions, forming a malfunction manifestation layer. Further analysis and evaluation of various key information closely related to malfunction manifestations suggests that collecting equipment information about malfunctions helps in identifying key areas of focus and continuous tracking in subsequent stages, while also allowing for horizontal comparison with other units in the project to prevent similar equipment from recurring quality problems. Extracting information on the causes of malfunctions helps companies or departments handle problems more effectively, preventing potential problems at their source and reducing the number of quality events. Centralizing information on malfunction-related measures facilitates quick access and resolution for relevant personnel when similar problems occur, improving problem-solving efficiency. Furthermore, the time of malfunction occurrence, the stage to which it occurred, the related equipment, the unit to which it occurred, the key responsible personnel, and the relevant company department are all essential information with communicable value, as shown in Table 1.

[0078] Table 1 Knowledge Categories and Their Functions

[0079]

[0080] The key knowledge elements from different dimensions identified above are grouped into knowledge layers within the key knowledge domain of experience feedback, including the fault equipment layer, fault cause layer, and corresponding measures layer, forming a domain knowledge hierarchy. Simultaneously, inter-layer interactions are used to achieve interconnected coupling, transforming the structure into a spatial knowledge structure with semantic association and logical reasoning capabilities. This integrates nuclear power plant construction experience feedback resources, with a specific organizational structure as follows: Figure 2 As shown.

[0081] Based on the above classification and association of experience feedback knowledge, the key knowledge domains of experience feedback are taken as the pattern layer of the experience feedback graph and explicitly represented by ontology modeling methods, providing a framework for knowledge extraction.

[0082] Using Protégé software, different levels of domain knowledge in the key knowledge domain of experience feedback, including unit level, equipment level, fault level, cause level, measure level, stage level, time level, enterprise level, and personnel level, as well as the relationships between each level, are constructed using classes (owl) and object properties (owl). This establishes semantic relationships between node types and concepts or instances, obtaining knowledge in an <entity-relationship-entity> structure. The key knowledge domain of experience feedback is then explicitly represented, forming an ontology model of the experience feedback graph, such as... Figure 3 As shown.

[0083] By defining unified data standards for key knowledge domains of experience feedback, we can clarify the objectives of knowledge mining and provide pattern support for the subsequent construction of experience feedback graphs.

[0084] S3. Based on the ontology model of S2, a multi-dimensional cascaded knowledge extraction model is constructed to obtain the ontology model's fill data from the experience feedback information. The multi-dimensional cascaded knowledge extraction model includes a knowledge mining model and a relation extraction model; the knowledge mining model is used to perceive and extract key knowledge of various categories from the experience feedback information; the relation extraction model is used to extract the relational features of each key knowledge in the experience feedback information.

[0085] To further construct a complete experience feedback graph, under the guidance of the key knowledge domain framework of experience feedback, we used knowledge extraction methods to analyze and mine important semantic features of quality information in the nuclear power field, and then filled the constructed experience feedback graph ontology model with data.

[0086] To address the challenges of complex textual content and diverse relationship categories in nuclear power plant construction experience feedback information, single network models are insufficient for effectively identifying and learning key features, making it difficult to mine, organize, and represent experience feedback knowledge. To improve the mining results in the niche field of nuclear power, this method proposes a multi-dimensional cascaded knowledge extraction approach: constructing a multi-layer neural network model to perform multi-dimensional feature mining on experience feedback information in the nuclear power field, efficiently extracting key domain knowledge.

[0087] Key knowledge elements contained in nuclear power quality incidents need to be automatically identified and extracted from corpus text using Named Entity Recognition (NER). Methods based on manually written rule templates or feature-based statistical models rely heavily on knowledge bases and dictionaries, lacking portability and having limited research specifically for the nuclear power field. Furthermore, building new dictionaries and rule templates is time-consuming. Therefore, an automated entity extraction method is adopted, embedding text as vectors into a neural network. End-to-end named entity recognition is achieved through deep learning, enabling named entity recognition of large amounts of text in a short time, thus saving mining costs and time.

[0088] Considering the complexity of experience feedback information in the nuclear power field, this method employs deep learning to construct a three-layer functional network for word embedding, knowledge perception, and knowledge prediction, resulting in a knowledge mining model. This model perceives and captures key elements in nuclear power quality events from multiple dimensions, integrates an attention mechanism to enhance focus on key words in the nuclear power field, and effectively addresses the problem of poor knowledge recognition performance in this area. The structure of the knowledge mining model is as follows: Figure 4 As shown.

[0089] The word embedding layer is based on the BERT bidirectional encoder representation technology of Transformer.

[0090] To enable computer models to recognize text in the nuclear power field, a word embedding layer based on Bidirectional Encoder Representations from Transformers (BERT) was designed to map words to real-valued vectors. BERT is a pre-trained network model with strong generalization capabilities. Its input has three different embedding layers, representing word information (token embedding), sentence / segment information (segment embedding), and sentence / position information (position embedding) of the input quality event text. These three layers of information vectors are concatenated as the input to BERT, and after passing through multiple Transformer layers, an output vector containing semantic features with multiple dimensions is obtained, providing numerical data input for subsequent models.

[0091] The knowledge perception layer is constructed based on the Bidirectional Long Short-Term Memory (BiLSTM) network and the Attention mechanism.

[0092] The texts of quality events generated during nuclear power plant construction are typically quite long, and models tend to forget past information during the learning and recognition process. Therefore, a bidirectional Long Short-Term Memory (BiLSTM) network is used to bidirectionally mine historical and future information in the quality event texts, comprehensively linking long-term features. To further enhance the text's sensitivity to nuclear power-related elements, this method incorporates an attention mechanism to calculate the degree of correlation between features between words in nuclear power experience feedback information. This allows the model to focus its attention on nuclear power-related areas when processing information, thereby obtaining more effective semantic information.

[0093] The knowledge prediction layer is built on Conditional Random Fields (CRF) and is used to optimize the output results of the knowledge perception layer.

[0094] The results from the knowledge-aware layer only learn each word in the text of nuclear power quality events, without considering the constraints between tags. For example, "pressure vessel" is a fixed technical term in the nuclear power field, and the order of the words cannot be changed; therefore, the connection order of tags needs to be constrained. Conditional Random Fields (CRF) are introduced to optimize the output of the upper layer. During prediction, the emission matrix and transition matrix are combined, and the highest-scoring label sequence corresponding to the sentence tag is calculated using the state score and transition score output from the upper layer. This ensures that the empirical feedback knowledge in the final prediction is reasonable and effective. The specific calculation process is as follows:

[0095]

[0096] in For label y i Transfer to label y i+1 The score, S i ,y i The i-th character of the input text sequence is predicted as the label y. i The score.

[0097] The probability of the output label sequence Y is calculated using the normalized exponential function Softmax:

[0098]

[0099] Where Y X For the set of all possible label sequences, This is the actual label sequence.

[0100] Finally, the Viterbi algorithm is used to obtain the globally optimal label sequence:

[0101]

[0102] The relation extraction model is optimized from a segmented convolutional neural network (PCNN) based on multi-instance learning.

[0103] To address the complex and difficult-to-organize relationships in nuclear power plant construction experience feedback information, relation extraction (RE) technology is needed to extract relation features from nuclear power quality text. Since the nuclear power experience feedback information expressing quality events is quite long, traditional methods based on lexical and syntactic features for relation extraction have low accuracy. To avoid introducing noisy data, a Piece-Wise-CNN (PCNN) method based on multi-instance learning is adopted for relation extraction. The feature extraction scheme is adjusted based on the Convolutional Neural Network (CNN) model, treating relation extraction as a multi-instance problem, selecting only the instance closest to the true category from each bag as the output. To reduce error propagation or accumulation, position vectors are introduced into the sentence vector representation. Sentences are divided into three segments based on the position of key information in the nuclear power experience feedback information, and max pooling is performed separately for each segment to extract more fine-grained relation features. Finally, the output of the segmented pooling layer is fed into the softmax layer to obtain the classification of nuclear power experience feedback information relationships. The specific structure is as follows: Figure 5 As shown.

[0104] Based on the highly specialized nature of nuclear power plant experience feedback information, the PCNN relationship extraction model is improved by introducing the concept of regularization and using the Gaussian error linear unit GELU with smoother nonlinear characteristics as the activation function. This improves the generalization ability of the experience feedback information relationship extraction model and avoids the gradient vanishing problem.

[0105] Specifically, such as Figure 5 As shown, the optimized relation extraction model includes:

[0106] The Vector Representation layer is used to convert the input key knowledge into an embedding vector representation matrix. The embedding vector representation matrix includes word embedding vectors and position embedding vectors. The word embedding vectors are used to indicate the text words in the key knowledge, and the position embedding vectors are used to indicate the position information of each text word in the key knowledge.

[0107] Convolutional layers are used to extract local features from the embedding vector representation matrix in different regions, resulting in local feature maps for each region.

[0108] The piecewise max pooling layer is used to perform piecewise max pooling operations on the local feature maps of different regions to obtain the feature maps with the maximum pooling value for each region.

[0109] The fully connected layer is used to concatenate the pooled maximum feature maps corresponding to different regions, and then activate them using the Gaussian error linear unit GELU as the activation function to obtain the fully connected feature map of key knowledge.

[0110] The classification layer (Softmax classifier) ​​is used to make classification decisions and extract relationships from the fully connected feature maps of key knowledge, thereby obtaining the relational features of the textual information in the key knowledge and outputting them.

[0111] Let the input be x, and the specific expression be:

[0112]

[0113] Where x represents the input to the activation function, and π is the mathematical constant pi.

[0114] S4 uses the multidimensional knowledge extraction model built in S3 to extract structured fill data from the experience feedback information obtained in S1, and maps the extracted fill data to the ontology model in S2 to generate the corresponding experience feedback graph, which is then stored in the knowledge base.

[0115] The purpose of knowledge-based experience feedback is to enable the key information from experience feedback to be transmitted, retrieved, and shared in actual nuclear power plant construction.

[0116] After the experience feedback information is extracted through a multi-dimensional cascaded knowledge extraction model, the original large amount of unstructured quality event text is transformed into experience feedback key domains rich in structured semantic knowledge. Based on this, the Neo4j graph database is further used to construct the nodes and edges in the structure, completing ontology model mapping and knowledge specification storage, and realizing the visual integration of experience feedback knowledge. The resulting nuclear power plant construction experience feedback graph is as follows: Figure 6 As shown.

[0117] By centrally managing the key knowledge domains of organized nuclear power plant construction experience feedback using a knowledge base, their independent and dispersed nature can be changed, ensuring that experience feedback knowledge is presented in a unified and clear manner. Based on this, further establishing experience feedback knowledge reasoning rules transforms experience feedback information into a knowledge structure with semantic association and logical reasoning capabilities, achieving coupled interaction between knowledge points. That is, triggering corresponding reasoning processes based on specific conditions allows related knowledge to be acquired. The application framework is as follows: Figure 7 As shown. For example, if a certain failure mode occurs, what specific solutions or improvement actions should be taken, etc.

[0118] By combining machine learning and data mining techniques, and analyzing large amounts of experience feedback data, potential patterns and rules can be automatically discovered, which can help establish more intelligent and efficient reasoning rules. By using professional knowledge graph or expert system technologies, a reasoning engine can be built to transform the reasoning rules into a form that computers can understand and execute, thereby improving the interpretability of knowledge and realizing the automatic reasoning and application of nuclear power plant construction experience feedback knowledge.

[0119] S5. When a quality incident occurs during the construction of a nuclear power plant, the reasoning engine searches the knowledge base to obtain the necessary knowledge to assist nuclear power professionals and / or managers in analysis and processing.

[0120] This method breaks away from traditional text recording methods, utilizing knowledge graphs to present the knowledge-based results of experience feedback. It constructs a unified experience feedback graph, and under the guidance of key knowledge domains in experience feedback, uses natural language processing technology to intelligently extract experience knowledge. Structured knowledge is stored in a graph database, and an inference engine is built to provide knowledge retrieval tools, thereby realizing the knowledge-based representation of experience feedback.

[0121] This method can transform the experience feedback information recorded in the construction quality event records into knowledge, and construct a unified experience feedback graph through multi-dimensional cascaded knowledge mining. Specifically, after acquiring the experience feedback information, its characteristics are analyzed, multiple types of high-value knowledge are hierarchically associated and coupled, and the organizational structure of key knowledge domains of experience feedback is defined, thereby constructing an ontology model (i.e., a framework model) for the experience feedback graph. Then, based on the model architecture of the ontology model, a multi-dimensional cascaded knowledge extraction model is constructed to extract key knowledge of each category from the experience feedback information, extract the relational features of each key knowledge, and map them onto the ontology model to obtain the corresponding experience feedback graph, which is then stored in a knowledge base. This method breaks away from the traditional text recording method, using a knowledge graph carrier to present the knowledge-based results of experience feedback. It constructs a unified data experience feedback graph, and under the guidance of key knowledge domains of experience feedback, uses natural language processing technology for intelligent extraction of experience knowledge, storing structured knowledge through a graph database to achieve the knowledge-based transformation of nuclear power plant construction experience feedback information from a document-centric to a knowledge-centric approach, storing experience feedback knowledge in a unified and clear manner in a knowledge base. In this way, when a quality incident occurs during the construction of a nuclear power plant, the necessary knowledge can be quickly and accurately retrieved from the knowledge base to assist nuclear power professionals and / or managers in analysis and processing.

[0122] In addition, this method employs a hybrid top-down and bottom-up approach when constructing the experience feedback graph. It organizes and reorganizes a large amount of complex and lengthy quality information, using key knowledge domains of experience feedback as the graph's schema layer. Simultaneously, based on the characteristics of nuclear power experience feedback texts, it utilizes natural language processing techniques to specifically design named entity recognition and relation extraction methods. This extracts structured key knowledge groups as data layers and maps them into the ontology model, updating the knowledge structure. This ensures the high efficiency and effectiveness of the resulting experience feedback graph. This method effectively addresses issues such as the weak timeliness of experience feedback during nuclear power plant construction, inconsistent understanding of experience knowledge among various stakeholders, and low reuse rate of experience feedback knowledge.

[0123] Example 2

[0124] To illustrate the effectiveness of the knowledge mining model and relation extraction model in this method, the following explanation is provided:

[0125] Quality event records from nuclear power company information systems were collected as raw training data. After cleaning, the data length was standardized according to the characteristics of the information text. Following preliminary rule-based collection, screening, and pre-experimentation verification, labels and relationship annotations were performed according to commonly used annotation strategies in natural language processing tasks, further completing the data preprocessing work and providing a data foundation for text mining. This dataset consists of 350 quality event records, which were divided into training, test, and validation sets in a 7:1.5:1.5 ratio to calculate specific evaluation metrics for the model.

[0126] Model parameter settings

[0127] Based on literature review and preliminary experiments, the main hyperparameter settings in the knowledge mining model are shown in Table 2 below. Other parameters are set to default values. The preprocessed experience feedback dataset mentioned above is then fed into the model for training.

[0128] Table 2 Model Parameter Settings

[0129]

[0130]

[0131] Knowledge extraction effect

[0132] To verify the effectiveness of knowledge extraction and relation extraction models in extracting key semantic features from experiential feedback information, several traditional knowledge mining models were compared with the BERT-BiLSTM-CRF model proposed in this study on the same dataset. The results are as follows. Figure 8 As shown.

[0133] From the performance comparison results, it can be seen that in the entity recognition task of empirical feedback information, the BERT-BiLSTM-CRF model with the fusion attention mechanism used in this method has achieved the best performance in various indicators. The precision, recall, and F1 value (a composite indicator of precision and recall) have reached 91.49%, 93.14%, and 92.31% respectively, and the model recognition effect is good.

[0134] Similarly, when comparing the improved PCNN model in this method with the ordinary PCNN, the results are shown in Table 3.

[0135] Table 3 Performance comparison results of relation extraction models

[0136]

[0137] According to the performance comparison experiment results of the relation extraction task in Table 3, the precision, recall, and F1 value of the improved PCNN relation extraction model proposed in this method are all higher than those of the ordinary PCNN, reaching 93.18%, 93.36%, and 92.87% respectively. It can better identify the relation features of nuclear power quality texts and provide data support for the next step of constructing a complete empirical feedback map.

[0138] Example 3

[0139] Based on the empirical feedback map generated by the neo4j graph database, cypher language instructions can be used for further in-depth knowledge mining and application.

[0140] Query and reuse of quality problem handling measures

[0141] By retrieving a specific knowledge base through the empirical feedback map, it is possible to specifically query the handling measures for specific quality problems. For example, in the quality event of gear rubbing during nuclear power construction, the instruction in the empirical feedback map can be:

[0142] “MATCH(n)-[r]-()

[0143] WHERE n.gz CONTAINS 'rubbing'

[0144] RETURN n,r”

[0145] Retrieve past empirical knowledge, obtain the fault nodes related to "rubbing", expand the nodes to further obtain the corresponding empirical feedback map and node information statistics, specifically as Figure 9 shown.

[0146] Compared to traditional text-based recording of quality incidents, a standardized graphical structure is more conducive to managers organizing and managing experience feedback knowledge. This means all quality incident records are organized and stored according to the same structure and standards, ensuring that complex information is expressed and stored consistently. Users do not need to read large amounts of text documents; even non-professionals can quickly understand key content, obtain solutions to quality problems, and other relevant knowledge by browsing knowledge nodes and connections. Furthermore, it avoids cognitive biases caused by different individuals or teams using varying writing styles in experience feedback, improving the comprehensibility and reusability of experience feedback, thereby enhancing the efficiency and quality of experience feedback.

[0147] Quality issue related entity query and push

[0148] In the actual construction of nuclear power plants, experience feedback engineers are not archival professionals and have limited information retrieval capabilities. They often rely on keywords to search for resources, resulting in low recall and precision, and coarse granularity. This makes it difficult to uncover the tacit knowledge in experience feedback information and leads to delays in its dissemination. Relational knowledge retrieval based on experience feedback maps breaks away from traditional keyword retrieval methods. Managers can quickly query and push various entities related to specific quality issues based on conditions such as quality problem nodes and relationships.

[0149] For example, there might not be a direct relationship between the equipment's unit and the stage at which the failure occurred, but by using a knowledge graph multi-hop retrieval method, the implicit path relationships between entities can be queried. Input command:

[0150] “MATCH q=(a)-[*1..3]-n

[0151] WHRER n.jd CONTAINS 'Construction Phase'

[0152] AND a.jz CONTAINS'Unit 3'

[0153] RETURN q”

[0154] A multi-level jump query revealed the indirect relationship between "Unit 3" and "Construction Phase," as shown in the following results. Figure 10 As shown.

[0155] As shown in the graph, three pieces of equipment in Unit 3 experienced different types of quality issues during the construction phase. Further analysis by selecting specific nodes reveals the corresponding causes of the failures, personnel involved, and other information. This helps management clarify relevant quality-related business interfaces and personnel responsibilities across enterprises, facilitating feedback to the appropriate departments. During the subsequent commissioning phase, relevant personnel need to focus on the quality status of these pieces of equipment to reduce the possibility of the same quality issues recurring or other quality problems arising from the same equipment. Pushing the query results to report generation tools and data analysis platforms for further data analysis helps understand the scope of the quality issues and the status of related entities. Visualizing the data allows for broader consideration of factors in the analysis of related issues, supporting decision-making and the implementation of improvement measures, making collaborative quality control more precise, comprehensive, and efficient.

[0156] Experience feedback spectrum knowledge statistics and prevention

[0157] By clustering the same label class recorded in the experience feedback graph, an experience feedback knowledge spectrum can be obtained to represent a knowledge base for a specific domain. Based on the experience feedback knowledge spectrum, various types of experience knowledge can be quickly accessed and retrieved, eliminating the need for experience feedback engineers to spend a lot of time and effort searching and collecting them.

[0158] For example, enter the command: "MATCH(n:gz)RETURN n"

[0159] Obtain the fault knowledge spectrum as follows Figure 11 As shown, by statistically analyzing the number of quality events that occur in a project, and identifying the most frequent failures of the same type, it is helpful to identify potential risks and early warning signals, and to take corresponding preventive measures to reduce the occurrence of quality problems.

[0160] Based on experience feedback spectrum, key knowledge management and statistics provide users with powerful search and filtering functions, which helps to reveal problems and trends in the nuclear power construction process, help managers identify common quality events, failure modes and risk factors, thereby providing key insights to assist decision-makers and teams in better understanding and responding to problems.

[0161] Case Studies

[0162] By integrating experience feedback mapping and deep knowledge mining capabilities into the experience feedback system of nuclear power companies, the system provides functions such as raw text recording, mapping display, correlation retrieval, and data analysis to assist managers in making nuclear power construction decisions. This transforms traditional decision-making based on human experience into a knowledge-driven intelligent decision-making model, effectively improving the interpretability of experience feedback information during the nuclear power construction process.

[0163] Taking the reactor pressure vessel (RPV) as an example, one of the main pieces of equipment in nuclear power plants, the RPV has numerous components, and its quality is crucial to the safety of the nuclear power plant. During installation or commissioning, when a quality problem occurs in a component of the reactor pressure vessel, it can be recorded by on-site personnel in the project experience feedback system. Through intelligent analysis of the text database using a multi-dimensional cascaded knowledge extraction method, a visualized experience feedback graph can be generated in real time in the "Graph Display" module. Based on a standardized graph, relevant management personnel can use the inference engine to query commands and conduct in-depth analysis of the reactor pressure vessel in the "Association Search" module, thereby obtaining the associated knowledge of RPV failures. Specifically, for example... Figure 12 As shown. Managers can quickly understand structured correlation knowledge from the RPV experience feedback map. Specifically, in the project's historical quality events, five different types of failures occurred in seven reactor pressure vessel-related equipment (including sub-equipment) belonging to Units 3 and 4, including the insulation support, workstation monitoring software, top cover assembly, and nozzle assembly. This allows for the identification of the root causes and corresponding measures. Further, the "Data Analysis" module provides statistical information on the failure stage, categories of components with frequent quality problems, response measures for past quality events, and personnel involved. Statistical reports are generated and fed back to relevant departments, achieving accurate transmission of experience feedback information. This enables efficient, scientific, and comprehensive planning and development of nuclear power equipment construction schemes, ensuring efficient and safe collaborative quality control of nuclear power equipment. For specific application procedures, see [link to application process]. Figure 13 .

[0164] This method focuses on the collaborative quality control needs of nuclear power equipment, addressing issues such as weak timeliness of experience feedback among enterprises, inconsistent understanding of experience feedback knowledge among various parties, and low reuse rate of experience feedback information. It constructs a unified data model of experience feedback graphs based on quality event text data. The characteristics of experience feedback information during nuclear power construction and the challenges faced under a "human-centric" approach are analyzed. A method for knowledge-based experience feedback is proposed, detailing the construction scheme for key knowledge domains. For data mining tasks within the experience feedback graph, a multi-dimensional information perception model is designed, fully utilizing context and key semantic features to improve the mining effect of named entity recognition and relation extraction in the nuclear power field. Data validation was achieved using actual experience feedback information from a nuclear power construction company, resulting in a visualized graph and deep knowledge mining. Finally, a case study of reactor pressure vessels is conducted, obtaining associated knowledge and statistical information. This provides relevant quality assurance personnel with intelligent and refined quality knowledge management tools to assist management decisions, helping nuclear power companies to deeply reuse quality data resources, breaking down barriers between quality information from geographically distributed enterprises, and thus improving the quality and efficiency of nuclear power construction.

[0165] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit the technical solutions. Those skilled in the art should understand that any modifications or equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.

Claims

1. A nuclear power plant construction experience feedback method based on multi-dimensional cascaded knowledge extraction, characterized in that, Includes the following steps: S1. Obtain experience feedback information from records of internally generated quality events throughout the entire lifecycle of nuclear power plant construction; S2. Analyze the characteristics of experience feedback information to obtain key knowledge categories, hierarchically associate and couple various key knowledge categories, define the organizational structure of key knowledge domains of experience feedback, and obtain the ontology model of experience feedback graph. S3. Based on the ontology model of S2, a multi-dimensional cascaded knowledge extraction model is constructed to obtain the filling data of the ontology model from the experience feedback information. The multi-dimensional cascaded knowledge extraction model includes a knowledge mining model and a relation extraction model. The knowledge mining model is used to perceive and capture key knowledge of each category in the experience feedback information. The relation extraction model is used to extract the relation features of each key knowledge in the experience feedback information. S4. Using the multidimensional knowledge extraction model built by S3, structured fill data is extracted from the experience feedback information obtained by S1, and the extracted fill data is mapped to the ontology model of S2 to generate the corresponding experience feedback graph and store it in the knowledge base. S5. When a quality incident occurs during the construction of a nuclear power plant, the knowledge base is searched to obtain the necessary information to assist nuclear power professionals and / or managers in analysis and processing. In S3, the knowledge mining model is a multi-layer neural network model that integrates attention mechanism and is built using a cascaded strategy; the knowledge mining model includes a word embedding layer, a knowledge perception layer and a knowledge prediction layer set in sequence. The relation extraction model is optimized from a segmented convolutional neural network (PCNN) based on multi-instance learning, including: The vector representation layer is used to convert the input key knowledge into an embedding vector representation matrix; the embedding vector representation matrix includes word embedding vectors and position embedding vectors, the word embedding vectors are used to indicate the text words in the key knowledge, and the position embedding vectors are used to indicate the position information of each text word in the key knowledge. Convolutional layers are used to extract local features from the embedding vector representation matrix in different regions, resulting in local feature maps for each region. The piecewise max pooling layer is used to perform piecewise max pooling operations on the local feature maps of different regions to obtain the feature maps with the maximum pooling value for each region. The fully connected layer is used to concatenate the pooled maximum feature maps corresponding to different regions, and then activate them using the Gaussian error linear unit (GELU) as the activation function to obtain the fully connected feature map of key knowledge. The classification layer is used to make classification decisions and extract relationships from the fully connected feature maps of key knowledge, obtain the relationship features presented in the text information of key knowledge, and output them. The word embedding layer is designed based on the BERT bidirectional encoder representation technology of Transformer. The working process of the word embedding layer includes: the input has three different embedding layers, which respectively represent the word information, sentence paragraph information and sentence position information of the input quality event text. The three layers of information vectors are concatenated as the input of BERT, and after passing through multiple Transformers, an output vector containing semantic features with multiple dimensions is obtained. The knowledge perception layer is constructed based on the Bidirectional Long Short-Term Memory Network (BiLSTM) fused with the Attention mechanism. The BiLSTM network is used to bidirectionally mine historical and future information in quality event segments, comprehensively linking long-term features. The Attention mechanism is used to calculate the degree of correlation between features between words in the experience feedback information, enabling the knowledge perception layer to focus its attention on the nuclear power-related areas when processing information, so as to obtain effective semantic information. The activation function expression for the Gaussian error linear unit (GELU) is: ； Where x represents the input to the activation function, and π is the mathematical constant pi.

2. The nuclear power plant construction experience feedback method based on multi-dimensional cascaded knowledge extraction as described in claim 1, characterized in that: S2 includes: By analyzing the characteristics of experience feedback information, we can obtain key knowledge categories related to fault performance. Each knowledge category is formed into a knowledge layer of the key knowledge domain for experience feedback, resulting in a domain knowledge spectrum. At the same time, the interaction between each knowledge layer is used to achieve correlation and coupling, resulting in a spatial knowledge structure with semantic association and logical reasoning capabilities, which serves as the organizational structure for the key knowledge domain for experience feedback. Subsequently, for different knowledge layers in the key knowledge domain of experience feedback, as well as the relationships between each layer, class and object attributes are constructed to establish semantic relationships between node types and concepts or instances, obtain knowledge in the form of <entity-relationship-entity> structure, and explicitly represent the key knowledge domain of experience feedback to form an ontology model of experience feedback graph.

3. The nuclear power plant construction experience feedback method based on multi-dimensional cascaded knowledge extraction as described in claim 2, characterized in that: The knowledge categories include equipment information about the fault, the cause of the fault, the countermeasures, as well as the time of the fault, the stage to which it occurred, the related equipment, the unit to which it belongs, the main responsible personnel, and the enterprise department; the relationships include causal relationships, dependency relationships, hierarchical relationships, and temporal relationships.

4. The nuclear power plant construction experience feedback method based on multi-dimensional cascaded knowledge extraction as described in claim 1, characterized in that: The knowledge prediction layer is built on conditional random fields and is used to optimize the output results of the knowledge perception layer. During prediction, the emission matrix and transition matrix are combined, and the highest-scoring labeled sequence corresponding to the sentence label is calculated based on the state score and transition score output by the upper layer, which is used as the corresponding key knowledge.

5. The nuclear power plant construction experience feedback method based on multi-dimensional cascaded knowledge extraction as described in claim 1, characterized in that: In S5, the reasoning engine searches the knowledge base to obtain the required knowledge.

Images