A knowledge graph-based nuclear facility safety intelligent question and answer method and system

Abstract

The application relates to a nuclear facility safety intelligent question and answer method and system based on a knowledge graph, an initial data set of radiation protection knowledge is acquired, preprocessing is carried out, and a target data set is formed; feature extraction is carried out on the target data set, and multi-dimensional labeling is carried out, knowledge units are aligned through the features, and a knowledge graph database is formed; a natural language question of nuclear facility safety is acquired, semantic information is read, a feature vector corresponding to the semantic information is formed, label recognition is carried out on the feature vector and features in the knowledge graph database, and computer recognition language is output; the computer recognition language is input into a model with an attention mechanism, relevant description information is dynamically acquired in the knowledge graph database, and an intelligent answer to the natural language question is formed. The method achieves the technical effects that a knowledge graph is formed to associate various systems of information, subsequent query is facilitated, deep meanings of questions and answers are integrated and refined, and a better answer is given.

Description

Technical Field

[0001] This invention relates to the field of nuclear facility safety management technology, and in particular to a knowledge graph-based intelligent question-and-answer method and system for nuclear facility safety. Background Technology

[0002] Regarding personnel dosing and operational safety at nuclear facilities, most rely on staff experience to handle emergencies or to judge related issues that arise during nuclear facility operations based on personal experience. Some facilities utilize nuclear facility information management systems to upload and organize all information into record tables. When questions arise during nuclear facility operations, these tables are used to search for relevant events and obtain answers.

[0003] However, traditional nuclear facility information management systems use multiple independent relational databases (such as MySQL and Oracle) to store structured data related to dosage and operational safety, such as personnel information tables, equipment ledgers, and dosage record tables. Unstructured documents, such as regulations, standards, operating procedures, and accident reports, are stored on file servers or indexed using simple full-text search engines like Elasticsearch. This lack of deep, semantic-level connections between the databases and document repositories creates "data silos."

[0004] Meanwhile, traditional question-and-answer queries typically use keyword matching, which lacks true natural language understanding capabilities. It cannot analyze the deeper semantics of the question and the user's intent, and can only perform literal matching of the input keywords, resulting in the inability to find relevant content in a timely manner.

[0005] The above problems urgently need to be addressed. Summary of the Invention

[0006] This invention discloses a knowledge graph-based intelligent question-answering method and system for nuclear facility safety, aiming to solve the technical problems existing in the prior art.

[0007] The present invention adopts the following technical solution: On one hand, this invention provides a knowledge graph-based intelligent question-answering method for nuclear facility safety, comprising: acquiring an initial dataset of radiation protection knowledge; preprocessing the initial dataset to form a target dataset; extracting features from the target dataset and annotating the features in multiple dimensions; aligning the multi-dimensional annotated features with knowledge units to form a knowledge graph database; acquiring natural language questions about nuclear facility safety; reading semantic information from the natural language questions based on a BiLSTM-CRF joint model to form feature vectors corresponding to the semantic information; using the feature vectors and features in the knowledge graph database to perform label recognition and output computer-recognized language; inputting the computer-recognized language into a Seq2Seq model with an attention mechanism; dynamically acquiring descriptive information related to the computer-recognized language from the knowledge graph database; and sorting the descriptive information to form an intelligent answer to the natural language question.

[0008] Optionally, an initial dataset of radiation protection knowledge is obtained, and the initial dataset is preprocessed to form a target dataset, including: obtaining the initial dataset of radiation protection knowledge; using fuzzy matching to remove duplicates from the initial dataset; using a combination of ensemble learning and density clustering to handle outliers in the initial dataset; and using a context-aware imputation strategy to impute missing segments in the initial dataset.

[0009] Optionally, feature extraction is performed on the target dataset, and the features are labeled in multiple dimensions. After aligning the multi-dimensional labeled features into knowledge units, a knowledge graph database is formed. This includes: normalizing the target dataset and extracting features; introducing a triple labeling system to the features so that the features are labeled with tags, wherein the triple labeling system includes behavioral semantic labeling, temporal attribute labeling, and spatial topological labeling; automatically extracting entities from the tags using a combination of a pre-trained language model and a domain dictionary, establishing a domain thesaurus, and unifying all entities into the same terminology; and aligning the features, tags, and terms into knowledge units to form a knowledge graph database.

[0010] Optionally, based on the BiLSTM-CRF joint model, semantic information in the natural language question is read to form a feature vector corresponding to the semantic information. The feature vector and the features in the knowledge graph database are used for label recognition to output computer-recognized language. This includes: reading the natural language question in the BiLSTM layer, capturing the current word, the preceding speech information, and the following semantic information, and concatenating them to form complete contextual semantic information; extracting the feature vector from the complete contextual semantic information; receiving the feature vector in the CRF layer, and considering the transition probability between features, labels, and terms, and converting the feature vector into computer-recognized language according to the rule template.

[0011] Optionally, a BiLSTM layer is used to read natural language questions, capture the current word, the preceding speech information, and the following semantic information, and concatenate them to form complete contextual semantic information. This includes: using a forward LSTM in the BiLSTM layer to read the sentence from left to right of the natural language question, capturing the semantic information of the current word and the preceding text; using a backward LSTM in the BiLSTM layer to read the natural language question from right to left of the question, capturing the speech information of the current word and the following text; and concatenating the semantic information of the current word and the preceding text, as well as the speech information of the current word and the following text, to form complete contextual semantic information.

[0012] Optionally, a CRF layer is used to receive the feature vector, and considering the transition probabilities between features, labels, and terms, the feature vector is converted into computer-recognized language according to a rule template, including: The transition probability is calculated as follows: Where X is the observation sequence, Y is the labeled sequence predicted by the CRF layer after processing the observation sequence, i is the i-th observation, and Z(x) is the normalization factor. For features, For feature vectors, The weights corresponding to the features, These are the weights corresponding to the feature vectors.

[0013] Optionally, the computer-recognized language is input into a Seq2Seq model with an attention mechanism. Descriptive information related to the computer-recognized language is dynamically acquired from the knowledge graph database. Based on this descriptive information, an intelligent answer to the natural language question is formed. This includes: using an encoder in the Seq2Seq model with an attention mechanism to encode the computer-recognized language into a sequence of context vectors; using a decoder in the Seq2Seq model with an attention mechanism to search for and generate an answer in the knowledge graph database, wherein the generated answer includes multiple feature words; the decoder calculates the weight distribution of each feature word when generating the answer using an attention mechanism; the Seq2Seq model dynamically monitors descriptive information related to the currently generated feature word in the encoder output sequence; and all descriptive information is integrated, sorted, and converted into natural language to form an intelligent answer.

[0014] According to another aspect of the present invention, a knowledge graph-based intelligent question-answering system for nuclear facility safety is also provided, comprising: a multi-source data governance module, which acquires an initial dataset of radiation protection knowledge, preprocesses the initial dataset to form a target dataset; a knowledge graph management module, which extracts features from the target dataset, annotates the features in multiple dimensions, aligns the multi-dimensional annotated features with knowledge units, and forms a knowledge graph database; a natural language question engine module, which acquires natural language questions about nuclear facility safety, reads semantic information from the natural language questions based on a BiLSTM-CRF joint model, forms feature vectors corresponding to the semantic information, performs label recognition using the feature vectors and features in the knowledge graph database, and outputs computer-recognized language; and a natural language answer engine module, which inputs the computer-recognized language into a Seq2Seq model with an attention mechanism, dynamically acquires descriptive information related to the computer-recognized language from the knowledge graph database, and sorts the descriptive information to form an intelligent answer to the natural language question.

[0015] According to another aspect of the present invention, a non-volatile storage medium is also provided, the non-volatile storage medium storing a plurality of instructions, the instructions being adapted to be loaded by a processor and executed any one of the knowledge graph-based intelligent question-answering methods for nuclear facility security.

[0016] According to another aspect of the present invention, a computer program product is also provided, including a computer program that, when executed by a processor, implements the steps of any one of the knowledge graph-based intelligent question-answering methods for nuclear facility security.

[0017] The technical solution adopted in this invention can achieve at least one of the following beneficial effects: In this embodiment of the invention, a knowledge graph database is constructed, and all data is aligned to facilitate subsequent queries. Simultaneously, a BiLSTM-CRF joint model is employed to extract key points from the questions and answers, enabling a deeper analysis of user intent. Furthermore, a Seq2Seq model with an attention mechanism is used to search the knowledge graph database for descriptive information matching the user intent, forming an intelligent answer and effectively outputting the desired result to the user. This achieves the goal of forming a knowledge graph that connects information from multiple systems, facilitating subsequent queries, while also realizing the technical effect of integrating and refining the deeper meaning of questions and answers to provide a superior response. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below, forming part of the present invention. The illustrative embodiments of the present invention and their descriptions explain the present invention and do not constitute an improper limitation of the present invention. In the accompanying drawings: Figure 1 This is a flowchart of a knowledge graph-based intelligent question-answering method for nuclear facility safety in Embodiment 1 of the present invention; Figure 2 This is a flowchart of the knowledge graph construction process in a knowledge graph-based intelligent question-answering method for nuclear facility safety according to Embodiment 1 of the present invention; Figure 3 This is a flowchart of the question-answering process in a knowledge graph-based intelligent question-answering method for nuclear facility safety according to Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of the BiLSTM model structure in a knowledge graph-based intelligent question-answering method for nuclear facility safety in Embodiment 1 of the present invention; Figure 5 This is a schematic diagram of the CRF layer structure in a knowledge graph-based intelligent question-answering method for nuclear facility safety according to Embodiment 1 of the present invention; Figure 6 This is a schematic diagram of the structure of a knowledge graph-based intelligent question-and-answer system for nuclear facility safety in Embodiment 2 of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. In the description of this invention, it should be noted that the term "or" is generally used to include the meaning of "and / or," unless otherwise expressly indicated.

[0020] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or a magnetic connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. Furthermore, in the description of this application, the terms "first," "second," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance. In the description of this invention, "a plurality of" means at least two, such as two, three, or more, unless otherwise explicitly specified.

[0021] Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0022] First, to facilitate understanding of the embodiments of the present invention, some terms or nouns involved in the present invention will be explained below: A knowledge graph is a semantic knowledge representation and storage system with a graph structure at its core. It describes the relationships between things in the real world through triples of entities, relations, and attributes, and organizes fragmented information into a structured, reasonable knowledge network.

[0023] To address the problems existing in related technologies, this application provides a knowledge graph-based intelligent question-answering method and system for nuclear facility safety.

[0024] Example 1 This embodiment provides a knowledge graph-based intelligent question-answering method for nuclear facility safety, such as... Figure 1 As shown, Figure 1 This is a flowchart of a knowledge graph-based intelligent question-answering method for nuclear facility safety according to Embodiment 1 of the present invention. The method includes: Step S102: Obtain the initial dataset of radiation protection knowledge, preprocess the initial dataset to form the target dataset; like Figure 2 As shown, Figure 2This is a flowchart of the knowledge graph construction process in a knowledge graph-based intelligent question-answering method for nuclear facility safety according to Embodiment 1 of the present invention. The knowledge graph construction process integrates a data connector (multiple datasets are placed together to form an initial dataset), a data cleaning engine (preprocessing the initial dataset), and an annotation pipeline (annotating the target dataset). It also interacts with a data warehouse or data lake (external database) to output high-quality training and knowledge source data. By constructing a dose and operational safety knowledge graph and performing strict entity alignment such as standardization, synonymization, disambiguation, and completion, the originally scattered structured and unstructured data are integrated into a semantically interconnected knowledge network. This enables the system to understand and utilize the deep-level relationships between data, realizing the deep integration and intelligent association of nuclear facility dose and operational safety knowledge.

[0025] In some preferred embodiments, the initial dataset of radiation protection knowledge is obtained, and the initial dataset is preprocessed to form a target dataset, including: obtaining the initial dataset of radiation protection knowledge; using fuzzy matching to remove duplicates from the initial dataset; using a combination of ensemble learning and density clustering to handle outliers in the initial dataset; and using a context-aware imputation strategy to impute missing segments in the initial dataset.

[0026] Optionally, the initial dataset needs to be preprocessed, including data cleaning and normalization, specifically data deduplication, outlier detection and handling, and missing value imputation.

[0027] Optionally, data deduplication adopts a hybrid deduplication strategy that combines exact matching and fuzzy matching. For data with globally unique identifiers (such as overhaul numbers and equipment IDs), exact matching is used for deduplication. For text description data such as work order descriptions and event reports, fuzzy matching is used for deduplication. A similarity threshold is set by calculating the intersection-union ratio (Jaccard similarity coefficient) of the text. Deduplication logic is triggered when the similarity exceeds the threshold.

[0028] Optionally, outlier detection and processing employ a combination of ensemble learning and density clustering. For numerical monitoring data such as dose rates, the Isolation Forest algorithm is used to quickly identify outliers that deviate significantly from the data distribution by randomly partitioning the feature space. For data with spatial or temporal correlations, such as personnel trajectories, the Local Outlier Factor algorithm is used to detect anomalies by comparing the local density of a data point with its k nearest neighbors. Detected outliers will be removed, corrected, or marked according to their business importance. It should be noted that corrected or marked outliers will not be used for model training.

[0029] Optionally, missing value imputation adopts a context-aware intelligent imputation strategy. For numerical fields, if the missing rate is less than 5% and the distribution is uniform, the median is used to imput the interference of extreme values. If it is time series data, spline interpolation based on time series analysis is used for smooth imputation. For categorical fields or fields with high missing rates, a prediction model based on random forest is constructed, using other complete fields as features to predict missing values, thereby achieving the maximum information preservation.

[0030] Step S104: Extract features from the target dataset and annotate the features in multiple dimensions. Align the multi-dimensional annotated features with knowledge units to form a knowledge graph database. Optionally, the process of building a knowledge graph database uses a knowledge extraction service, ontology manager, Neo4j graph database instance, and API gateway to quickly and accurately identify features in the target dataset and achieve accurate construction of the knowledge graph database.

[0031] In some preferred embodiments, feature extraction is performed on the target dataset, and the features are labeled in multiple dimensions. After aligning the multi-dimensional labeled features into knowledge units, a knowledge graph database is formed. This includes: normalizing the target dataset and extracting features; introducing a triple labeling system to label the features, wherein the triple labeling system includes behavioral semantic labeling, temporal attribute labeling, and spatial topological labeling; automatically extracting entities from the labels using a combination of a pre-trained language model and a domain dictionary, establishing a domain thesaurus, and unifying all entities into the same terminology; and aligning the features, labels, and terms into knowledge units to form a knowledge graph database.

[0032] Optionally, the target dataset can be transformed into a standardized knowledge graph database, which requires data normalization and multi-dimensional annotation. Specifically, in data normalization, the cleaned structured data is normalized using the Z-Score normalization method to transform numerical features of different dimensions, such as dose and activity, to the same scale; for unstructured text data, word segmentation and stop word removal are performed to transform the text into a dense vector representation and construct a unified feature representation space.

[0033] Optionally, for machine parsing of the deep meaning of data, a triple annotation system is introduced for multi-dimensional semantic annotation. Specifically: Behavioral semantic annotation is based on a predefined ontology library of dose and operation safety to identify and label key behavioral events in the text, such as tagging "dose exceeding the standard", "abnormal start / stop of equipment", "illegal entry of personnel", etc.; Temporal attribute annotation uses a rule engine and time expression recognition technology to extract and standardize time information from the text, marking the start time, duration, and end time for each event or state to form a complete event time sequence chain; Spatial topology annotation associates equipment codes, area numbers with a three-dimensional plant model (BIM) to label the data with accurate three-dimensional geographical coordinates and equipment topology relationships (such as "Equipment A is located in Room B and is a subsystem of System C").

[0034] Finally, all the data that has been cleaned, normalized, and annotated is integrated and stored in a data warehouse with a distributed architecture, forming a highly available and multi-modal dataset for dose and operation safety to support upper-layer applications.

[0035] Optionally, the dataset is promoted to a structured knowledge network with semantic association and reasoning capabilities. Specifically, knowledge extraction and entity alignment methods are used. From the constructed dataset, a pre-trained language model and a domain dictionary are combined to automatically extract entities (such as radioactive sources, staff, protection doors), attributes (such as activity, job type, shielding thickness), and relationships (such as located in, operated by, required to comply with).

[0036] Optionally, to ensure the consistency of knowledge, strict entity alignment is carried out. Among them, entity standardization is to establish a domain synonym thesaurus to unify all expressions to standard terms. For example, "Radiation" and "irradiation" are unified to "radiation". Entity synonymization is to calculate the similarity of entity vectors in a low-dimensional space and merge entities with similar semantics (such as "receiving a dose" and "being irradiated"). Entity disambiguation is to use context analysis to distinguish polysemy. Entity completion is to predict and complete potential missing entities through a graph neural network using the topological structure of existing entities and relationships.

[0037] Optionally, after the dataset alignment, ontology modeling and knowledge storage are required. Specifically, an upper ontology in the field of dose and operation safety management is constructed to clearly define core concepts (such as personnel, equipment, area, regulations, events) and their hierarchical and associative relationships (such as is-a, part-of, located-in). The aligned knowledge units are stored in a Neo4j graph database according to the ontology model. Using its native graph storage engine, millisecond-level responses for complex association queries can be achieved, and it provides a basis for path-based reasoning calculations.

[0038] Finally, the generated native graph storage engine is encapsulated as a knowledge service, meaning the knowledge graph capabilities are provided externally via a RESTful API. The interfaces are divided into two categories: a data layer API, which provides CRUD operations on the knowledge graph and accepts and executes Cypher query statements; and a business layer API, which encapsulates common question-and-answer scenarios, such as initializing sessions, submitting questions, and retrieving history, enabling convenient calls from front-end applications.

[0039] Step S106: Obtain natural language questions about nuclear facility safety; read semantic information from natural language questions based on the BiLSTM-CRF joint model; form feature vectors corresponding to semantic information; use feature vectors and features in the knowledge graph database to perform label recognition; and output computer-recognized language. like Figure 3 As shown, Figure 3 This is a flowchart of the question-answering process in a knowledge graph-based intelligent question-answering method for nuclear facility safety according to Embodiment 1 of the present invention. The natural language question and answer are independent microservices, internally encapsulating a BiLSTM-CRF model, a rule template engine, and a Seq2Seq generation model. It receives questions and returns answers via RPC or HTTP interfaces. By employing a BiLSTM-CRF joint model for question parsing, it can fully utilize contextual information to accurately identify entity boundaries and types, and combine rule templates to accurately transform semantic intent into knowledge graph queries. This ensures that the system understands that "EPD," "electronic personal dosimeter," and "personal dosimeter" refer to the same entity, and can distinguish the meaning of "dosage" in different contexts, thus directly returning accurate answers, greatly improving the accuracy and efficiency of retrieval, and enhancing the precision of information retrieval and the depth of semantic understanding.

[0040] In some preferred embodiments, the semantic information in natural language questions is read based on the BiLSTM-CRF joint model to form feature vectors corresponding to the semantic information. The feature vectors are then used to perform label recognition with features in the knowledge graph database to output computer-recognized language. This includes: reading natural language questions in the BiLSTM layer, capturing the current word, the preceding speech information, and the following semantic information, and concatenating them to form complete contextual semantic information; extracting feature vectors from the complete contextual semantic information; and using the CRF layer to receive the feature vectors, considering the transition probabilities between features, labels, and terms, and converting the feature vectors into computer-recognized language according to the rule template.

[0041] Optionally, based on the parsed entities and intents, a corresponding rule template is matched. For example, the template "Inquiry into the historical activity areas of the person querying" is matched for "Zhang San entered which high radiation areas in 2023?". This template fills the entities into the predefined Cypher statement skeleton and automatically generates the following Cypher query statement. The statement is sent to the knowledge graph service interface, executed in the graph database, and the results are returned.

[0042] In some preferred embodiments, a BiLSTM layer is used to read natural language questions, capture the current word, the preceding speech information, and the following semantic information, and then concatenate them to form complete contextual semantic information. This includes: using a forward LSTM in the BiLSTM layer to read the sentence from left to right of the natural language question, capturing the semantic information of the current word and the preceding text; using a backward LSTM in the BiLSTM layer to read the natural language question from right to left of the natural language question, capturing the speech information of the current word and the following text; and concatenating the semantic information of the current word and the preceding text, as well as the speech information of the current word and the following text, to form complete contextual semantic information.

[0043] Optionally, the user inputs a natural language question, such as "Which high-radiation areas did Zhang San enter in 2023?", which is then parsed using the BiLSTM layer in the BiLSTM-CRF joint model. The BiLSTM layer consists of a forward LSTM and a backward LSTM. The forward LSTM reads the sentence from left to right, capturing the current word and information from the preceding context; the backward LSTM reads from right to left, capturing the current word and information from the following context. Finally, the forward and backward hidden states at each time step are concatenated to form a feature vector representation containing complete contextual semantic information.

[0044] like Figure 4 As shown, Figure 4 This is a schematic diagram of the BiLSTM model structure in a knowledge graph-based intelligent question-answering method for nuclear facility safety according to Embodiment 1 of the present invention. and BiLSTM is a two-layer feature extraction layer constructed using neurons from LSTM, processing the sequence input from both positive and negative directions. It memorizes and outputs features from two directions, and integrates the output vectors from the two layers to obtain an output vector that contains the overall features. In this way, BiLSTM's ability to extract features from text sequences is greatly improved. It learns from two directions of the sentence, acquiring RWP (Reference Work Request) forms and query record contextual information, thereby extracting semantically richer sentence features. This enables the model to accurately understand that "Zhang San" is a person's name, "2023" is a time, and "high radiation zone" is a region category.

[0045] In some preferred embodiments, a CRF layer is used to receive feature vectors and consider the transition probabilities between features, labels, and terms. The feature vectors are then converted into computer-recognized language according to a rule template, including the following calculation of transition probabilities: Where X is the observation sequence, Y is the labeled sequence predicted by the CRF layer after processing the observation sequence, i is the i-th observation, and Z(x) is the normalization factor. For features, For feature vectors, The weights corresponding to the features, These are the weights corresponding to the feature vectors.

[0046] Optionally, the CRF layer acts as a decoder, receiving the feature sequence output by the BiLSTM. The CRF is a discriminative undirected graph model, the specific structure of which is as follows: Figure 5 As shown, Figure 5 This is a schematic diagram of the CRF layer structure in a knowledge graph-based intelligent question-answering method for nuclear facility safety according to Embodiment 1 of the present invention. X represents the observation sequence, and Y represents the labeled sequence predicted by the model after processing the observation sequence; both have the same structure. A conditional random field is constructed under the observation sequence X.

[0047] Then P(Y|X) can be called a linear-chain conditional random field. Linear-chain conditional random fields are common type of conditional random fields. P(Y|X) depends only on X and the preceding and following states of its output label. For dose and job data annotation tasks, linear-chain conditional random fields can calculate the probability of the entire text annotation sequence, considering the relationship between Yi and its preceding and following text based on the global semantic features of the text. The conditional probability is shown in the formula: Where Z(X) is the normalization factor, used to convert the calculated product into a probability. For features, For feature vectors, The weights corresponding to the features, These are the weights corresponding to the feature vectors.

[0048] Optionally, the entity information and question type information obtained from the text can be transformed into query statements using rule templates and submitted to the knowledge graph database to obtain answers. Based on the semantic information of dosage and job-related consultations, the questions are transformed according to the rule templates, ultimately converting them into query statements that the knowledge graph database can understand. For example, if a user enters "How to collect EPD", the semantic parsing model will extract the entity "EPD" and the question intent "collection method".

[0049] Optionally, the CRF layer considers the transition probability between tags, such as "person's name" being more likely to be followed by "verb" than "place name". The Viterbi algorithm is used to solve for the globally optimal tag sequence, thereby accurately identifying the entity boundaries and types in the question and accurately determining that the user's intent is "to query the historical activity area of ​​the person".

[0050] Step S108: Input the computer-recognized language into the Seq2Seq model with an attention mechanism, dynamically obtain descriptive information related to the computer-recognized language from the knowledge graph database, and form intelligent answers to natural language questions based on the sorting of descriptive information.

[0051] Optionally, leveraging the inherent associative query and reasoning capabilities of knowledge graph databases, comprehensive information related to the question can be proactively presented. Simultaneously, using Seq2Seq with an attention mechanism to generate answers, fragmented information can be integrated into a coherent, natural, and focused textual narrative. Furthermore, triple-annotated data can support more complex analyses. This provides nuclear facility workers, especially in emergency response scenarios, with direct and efficient comprehensive decision-making support, transforming fragmented presentation methods.

[0052] In some preferred embodiments, computer-recognized language is input into a Seq2Seq model with an attention mechanism. Descriptive information related to the computer-recognized language is dynamically acquired from a knowledge graph database. Based on the ordered descriptive information, an intelligent answer to a natural language question is formed. This includes: using an encoder in the Seq2Seq model with an attention mechanism to encode the computer-recognized language into a sequence of context vectors; using a decoder in the Seq2Seq model with an attention mechanism to search for and generate an answer in the knowledge graph database, wherein the generated answer includes multiple feature words; the decoder calculates the weight distribution of each feature word when generating the answer using an attention mechanism; the Seq2Seq model dynamically monitors descriptive information related to the currently generated feature word in the encoder output sequence; and all descriptive information is integrated, ordered, and converted into natural language to form an intelligent answer.

[0053] Optionally, for direct factual queries, the system formats the query results and returns them directly; for answers that require explanation or generation, the system uses a Seq2Seq model with an attention mechanism, which includes an encoder and a decoder. The encoder encodes the key information or original question into a sequence of context vectors. When generating each word of the answer, the decoder calculates a weight distribution through the attention mechanism, dynamically focusing on the part of the encoder's output sequence that is most relevant to the current generation. For example, when generating "high radiation area", it focuses on the descriptive information about the area in the knowledge base, combines all the descriptive information, converts it into natural language, and forms an intelligent answer.

[0054] Optionally, the above mechanism ensures that the generated answer is grammatically correct and highly relevant to the query content. For example, it can generate: "Zhang San entered two high-radiation areas in 2023, namely the 'RX plant core area' (May 10, 2023) and the 'spent fuel pool refueling area' (November 23, 2023). Specific dose records have been sent."

[0055] Through steps S102 to S108 above, a high-quality multimodal dose and work safety knowledge graph is constructed. Based on this, a question-answering intelligent algorithm with deep semantic understanding and efficient interaction capabilities is built to realize the systematic management and intelligent application of radiation protection knowledge, thereby improving the safe operation level and decision support capabilities of nuclear facilities.

[0056] Example 2 This embodiment also provides a knowledge graph-based intelligent question-and-answer system for nuclear facility safety, which is used to implement the above embodiments and preferred embodiments. Details already described will not be repeated. As used below, the terms "module" and "system" can refer to a combination of software and / or hardware that performs a predetermined function. Although the systems described in the following embodiments are preferably implemented in software, hardware implementations, or combinations of software and hardware, are also possible and contemplated.

[0057] According to embodiments of the present invention, a system embodiment for implementing the above-described knowledge graph-based intelligent question-answering method for nuclear facility safety is also provided. Figure 6 This is a schematic diagram of the structure of a knowledge graph-based intelligent question-and-answer system for nuclear facility safety in Embodiment 2 of the present invention, as shown below. Figure 6 As shown, the above system includes: a multi-source data governance module 201, a knowledge graph management module 202, a natural language question engine module 203, and a natural language answer engine module 204, wherein: The multi-source data governance module 201 acquires the initial dataset of radiation protection knowledge, preprocesses the initial dataset, and forms the target dataset. The knowledge graph management module 202 extracts features from the target dataset and annotates the features in multiple dimensions. After aligning the multi-dimensional annotated features into knowledge units, a knowledge graph database is formed. Natural Language Question Engine Module 203 acquires natural language questions about nuclear facility safety, reads semantic information from the natural language questions based on the BiLSTM-CRF joint model, forms feature vectors corresponding to the semantic information, uses the feature vectors and features in the knowledge graph database to perform label recognition, and outputs computer-recognized language. The natural language answer engine module 204 takes computer-recognized language as input into a Seq2Seq model with an attention mechanism, dynamically retrieves descriptive information related to the computer-recognized language from a knowledge graph database, and forms intelligent answers to natural language questions based on the sorting of descriptive information.

[0058] It should be noted that the above modules can be implemented by software or hardware. For example, for the latter, it can be implemented in the following ways: the above modules can be located in the same processor; or the above modules can be located in different processors in any combination.

[0059] It should be noted that the multi-source data governance module 201, knowledge graph management module 202, natural language question engine module 203, and natural language answer engine module 204 mentioned above correspond to steps S102 to S108 in the embodiments. The instances and application scenarios implemented by the above modules and their corresponding steps are the same, but are not limited to the content disclosed in the above embodiments. It should be noted that the above modules, as part of the system, can run on a computer terminal.

[0060] It should be noted that the optional or preferred implementation methods of this embodiment can be found in the relevant descriptions in the embodiments, and will not be repeated here.

[0061] The aforementioned knowledge graph-based intelligent question-and-answer system for nuclear facility safety may also include a processor and a memory. The aforementioned multi-source data governance module 201, knowledge graph management module 202, natural language question engine module 203, and natural language answer engine module 204 are all stored in the memory as program modules, and the processor executes the aforementioned program modules stored in the memory to realize the corresponding functions.

[0062] The processor contains a core that retrieves the corresponding program modules from memory. One or more cores may be configured. Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, like read-only memory (ROM) or flash RAM. Memory includes at least one memory chip.

[0063] According to an embodiment of this application, an embodiment of a non-volatile storage medium is also provided. Optionally, in this embodiment, the non-volatile storage medium includes a stored program, wherein, when the program runs, it controls the device where the non-volatile storage medium is located to execute any of the aforementioned knowledge graph-based intelligent question-answering methods for nuclear facility security.

[0064] Optionally, in this embodiment, the non-volatile storage medium may be located in any computer terminal in a group of computer terminals in a computer network, or in any mobile terminal in a group of mobile terminals, and the non-volatile storage medium includes stored programs.

[0065] Optionally, during program execution, the device containing the non-volatile storage medium can be controlled to perform the following functions: acquire an initial dataset of radiation protection knowledge, preprocess the initial dataset to form a target dataset; extract features from the target dataset and annotate the features in multiple dimensions, align the multi-dimensional annotated features with knowledge units to form a knowledge graph database; acquire natural language questions about nuclear facility safety, read the semantic information in the natural language questions based on the BiLSTM-CRF joint model, form feature vectors corresponding to the semantic information, use the feature vectors and features in the knowledge graph database to perform label recognition, and output computer-recognized language; input the computer-recognized language into a Seq2Seq model with an attention mechanism, dynamically acquire descriptive information related to the computer-recognized language in the knowledge graph database, and form intelligent answers to the natural language questions based on the sorting of descriptive information.

[0066] According to an embodiment of this application, an embodiment of a processor is also provided. Optionally, in this embodiment, the processor is used to run a program, wherein the program executes any of the above-described knowledge graph-based intelligent question-answering methods for nuclear facility security.

[0067] According to an embodiment of this application, an embodiment of a computer program product is also provided. Optionally, in this embodiment, the computer program product includes a computer program that, when executed by a processor, implements the steps of any of the above-described knowledge graph-based intelligent question-answering method methods for nuclear facility security.

[0068] Optionally, when the aforementioned computer program product is executed on a data processing device, it is suitable to execute an initialization program with the following method steps: acquiring an initial dataset of radiation protection knowledge; preprocessing the initial dataset to form a target dataset; extracting features from the target dataset and annotating the features in multiple dimensions; aligning the multi-dimensional annotated features with knowledge units to form a knowledge graph database; acquiring natural language questions about nuclear facility safety; reading semantic information from the natural language questions based on a BiLSTM-CRF joint model to form feature vectors corresponding to the semantic information; using the feature vectors and features in the knowledge graph database for label recognition; outputting computer-recognized language; inputting the computer-recognized language into a Seq2Seq model with an attention mechanism; dynamically acquiring descriptive information related to the computer-recognized language from the knowledge graph database; and forming intelligent answers to the natural language questions based on the sorting of the descriptive information.

[0069] This invention provides an electronic device, which includes a processor, a memory, and a program stored in the memory and executable on the processor. When the processor executes the program, it performs the following steps: acquiring an initial dataset of radiation protection knowledge; preprocessing the initial dataset to form a target dataset; extracting features from the target dataset and annotating the features in multiple dimensions; aligning the multi-dimensional annotated features with knowledge units to form a knowledge graph database; acquiring natural language questions about nuclear facility safety; reading semantic information from the natural language questions based on a BiLSTM-CRF joint model to form feature vectors corresponding to the semantic information; using the feature vectors and features in the knowledge graph database for label recognition; outputting computer-recognized language; inputting the computer-recognized language into a Seq2Seq model with an attention mechanism; dynamically acquiring descriptive information related to the computer-recognized language from the knowledge graph database; and forming intelligent answers to the natural language questions based on the sorted descriptive information.

[0070] The order of the above embodiments of the present invention is merely for description and does not represent the superiority or inferiority of the embodiments.

[0071] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0072] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The system embodiments described above are merely illustrative; for example, the division of modules described above can be a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between modules, and may be electrical or other forms.

[0073] The modules described above as separate components may or may not be physically separate. Similarly, the components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple modules. Some or all of the modules can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0074] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0075] If the aforementioned integrated modules are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable non-volatile storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a non-volatile storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned non-volatile storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0076] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A knowledge graph-based intelligent question-answering method for nuclear facility safety, characterized in that, include: An initial dataset of radiation protection knowledge is obtained, and the initial dataset is preprocessed to form a target dataset; Feature extraction is performed on the target dataset, and the features are labeled in multiple dimensions. After aligning the multi-dimensional labeled features into knowledge units, a knowledge graph database is formed. Natural language questions about nuclear facility safety are obtained, semantic information in the natural language questions is read based on the BiLSTM-CRF joint model, feature vectors corresponding to the semantic information are formed, and the feature vectors are used to perform label recognition with the features in the knowledge graph database to output computer-recognized language. The computer-recognized language is input into a Seq2Seq model with an attention mechanism. Descriptive information related to the computer-recognized language is dynamically obtained from the knowledge graph database. Based on the sorting of the descriptive information, an intelligent answer to the natural language question is formed.

2. The intelligent question-answering method for nuclear facility safety based on knowledge graphs according to claim 1, characterized in that, An initial dataset of radiation protection knowledge is obtained, and the initial dataset is preprocessed to form a target dataset, including: An initial dataset of radiation protection knowledge was obtained, and fuzzy matching was used to remove duplicates from the initial dataset. An outlier processing method is adopted to handle the data that deviates in the initial dataset by combining ensemble learning and density clustering; A context-aware imputation strategy is used to impute missing segments in the initial dataset.

3. The knowledge graph-based intelligent question-answering method for nuclear facility safety according to claim 1, characterized in that, Feature extraction is performed on the target dataset, and the features are labeled in multiple dimensions. After aligning the multi-dimensional labeled features into knowledge units, a knowledge graph database is formed, including: The target dataset is normalized and then feature extraction is performed. A triple annotation system is introduced for the features so that the features are labeled, wherein the triple annotation system includes behavioral semantic annotation, temporal attribute annotation, and spatial topological annotation; By combining a pre-trained language model with a domain dictionary, entities are automatically extracted from the tags, a domain thesaurus is built, and all entities are unified into the same terminology. Features, tags, and terms are aligned into knowledge units to form a knowledge graph database.

4. The knowledge graph-based intelligent question-answering method for nuclear facility safety according to claim 1, characterized in that, The semantic information of the natural language question is read based on the BiLSTM-CRF joint model to form a feature vector corresponding to the semantic information. The feature vector is then used to perform label recognition with the features in the knowledge graph database to output computer-recognized language, including: The BiLSTM layer is used to read natural language questions, capture the current word, the preceding speech information and the following semantic information, and splice them together to form complete contextual semantic information. Extract the feature vector from the complete contextual semantic information; The feature vector is received by a CRF layer, and the transition probabilities between features, labels and terms are considered. The feature vector is then converted into computer-recognized language according to a rule template.

5. The intelligent question-answering method for nuclear facility safety based on knowledge graphs according to claim 4, characterized in that, The BiLSTM layer is used to read natural language questions, capturing the current word, preceding speech information, and following semantic information, and then concatenating them to form complete contextual semantic information, including: The forward LSTM in the BiLSTM layer is used to read the sentence from left to right in the natural language question, capturing the semantic information of the current word and the preceding text; The backward LSTM in the BiLSTM layer is used to read from right to left in a natural language question, capturing the speech information of the current word and the following words; Based on the semantic information of the current word and the preceding text, as well as the phonetic information of the current word and the following text, complete contextual semantic information is formed by splicing them together.

6. The knowledge graph-based intelligent question-answering method for nuclear facility safety according to claim 4, characterized in that, The feature vector is received using a CRF layer, and the transition probabilities between features, labels, and terms are considered. The feature vector is then converted into computer-recognized language according to a rule template, including: The transition probability is calculated as follows: Where X is the observation sequence, Y is the labeled sequence predicted by the CRF layer after processing the observation sequence, i is the i-th observation, and Z(x) is the normalization factor. For features, For feature vectors, The weights corresponding to the features, These are the weights corresponding to the feature vectors.

7. The intelligent question-answering method for nuclear facility safety based on knowledge graphs according to claim 1, characterized in that, The computer-recognized language is input into a Seq2Seq model with an attention mechanism. Descriptive information related to the computer-recognized language is dynamically acquired from the knowledge graph database. Based on this descriptive information, an intelligent answer to the natural language question is formed, including: The encoder in the Seq2Seq model with attention mechanism encodes computer-recognized language into a sequence of context vectors; The decoder in the Seq2Seq model with attention mechanism is used to search for the generated answer in the knowledge graph database, where the generated answer includes multiple feature words; When generating each feature word in the answer, the decoder calculates the weight distribution of each feature word through an attention mechanism; The Seq2Seq model dynamically focuses on descriptive information related to the currently generated feature words in the encoder output sequence; All descriptive information is integrated and sorted, then converted into natural language to form an intelligent response.

8. A knowledge graph-based intelligent question-and-answer system for nuclear facility safety, characterized in that, include: The multi-source data governance module acquires an initial dataset of radiation protection knowledge, preprocesses the initial dataset, and forms a target dataset. The knowledge graph management module extracts features from the target dataset and annotates the features in multiple dimensions. After aligning the multi-dimensional annotated features into knowledge units, a knowledge graph database is formed. The natural language question engine module acquires natural language questions about nuclear facility security, reads semantic information from the natural language questions based on the BiLSTM-CRF joint model, forms feature vectors corresponding to the semantic information, uses the feature vectors and features in the knowledge graph database to perform label recognition, and outputs computer-recognized language. The natural language answering engine module inputs the computer-recognized language into a Seq2Seq model with an attention mechanism, dynamically acquires descriptive information related to the computer-recognized language from the knowledge graph database, and forms an intelligent answer to the natural language question based on the sorting of the descriptive information.

9. A non-volatile storage medium, characterized in that, The non-volatile storage medium stores multiple instructions, which are adapted to be loaded and executed by a processor as described in any one of claims 1 to 7: a knowledge graph-based intelligent question-answering method for nuclear facility security.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the knowledge graph-based intelligent question-answering method for nuclear facility safety as described in any one of claims 1 to 7.

Images