Nuclear power experience feedback search method and system based on non-limited entities and cross-stage

Abstract

The present invention proposes a nuclear power experience feedback search method and system based on non-limited entities and cross-stages, relates to the field of nuclear power technology, searches for experience feedback cases related to a text to be searched from an experience feedback library, metadata of the experience feedback cases include case description text, classification labels, entity names and semantic feature vectors, and is divided into three levels of screening, specifically: predicting the classification labels of the text to be searched, and performing a first level of screening on the cases in the experience feedback library through the classification labels; identifying the entity names in the text to be searched, and performing a second level of non-limited entity screening on the cases after the first level of screening through the entity names; similarity calculation based on non-limited entities and a third level of cross-stage screening, and obtaining a final search result from the cases after the second level of screening; the present invention uses non-limited entities and cross-stages to no longer limit the search to specific entities or a single stage, thereby realizing complete, accurate and limited nuclear power experience feedback search.

Description

Technical Field

[0001] The present invention belongs to the field of nuclear power technology, and in particular relates to a method and system for searching nuclear power experience feedback based on non-limited entities and across stages. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Experience feedback is a professional name for knowledge management in the field of nuclear power. It is an important means to continuously improve the quality level of nuclear power plant equipment, avoid recurrence of problems, and ensure the implementation of nuclear safety responsibilities. It is an important guarantee for the safety of nuclear power plants. In essence, experience feedback is a reuse process of various problem solutions (including good practices). It is a management activity with a clear purpose. The ultimate goal of experience feedback is to avoid the occurrence of similar or similar adverse events and accidents under the current circumstances, and to prevent other invalid and inefficient activities. Being able to effectively reuse various experience feedback data is an effective guarantee for achieving its ultimate goal.

[0004] The reuse of experience feedback data is mainly to search for similar cases based on the text fragments provided by the user's current business process, so as to provide users with reference and reference for historical experience feedback content; because the experience feedback data of the entire database has the characteristic of massive dimensions, it is inefficient and difficult for users to actively search for historical experience feedback cases. From the technical point of view of traditional search or content push, people generally require that the search results should have a result response to the query keywords, and the search results should be listed one by one. The results returned by some search engine services with strong technical capabilities (such as Google) will be further sorted from high to low according to similarity. Traditional search engines may return tens of thousands of content at a time, and it is necessary to further filter out the desired results manually. The goal of nuclear power experience feedback content retrieval is to find all the data related to each substantive content and ensure that the experience feedback items of important levels are not omitted; at the same time, the content found and returned should be highly relevant, and too much irrelevant data should not be returned. In summary, the requirements for content search are: the search results should be complete, accurate and limited in number. This kind of experience feedback search has higher requirements than the search technology of traditional search engines. Therefore, the current search methods such as combined keyword filtering and screening are not suitable for searching nuclear power experience feedback cases, and conventional search methods will lead to poor search results. Summary of the invention

[0005] In order to overcome the deficiencies of the above-mentioned prior art, the present invention provides a nuclear power experience feedback search method and system based on non-limited entities and cross-stages. Through non-limited entities and cross-stages, the search is no longer limited to specific entities or a single stage, thereby achieving a complete, accurate and limited nuclear power experience feedback search.

[0006] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0007] The first aspect of the present invention provides a nuclear power experience feedback search method based on non-limited entities and across stages.

[0008] Based on the non-limited entity and cross-stage nuclear power experience feedback search method, experience feedback cases related to the text to be searched are searched from the experience feedback library. The metadata of the experience feedback cases includes case description text, classification label, entity name and semantic feature vector, which are divided into three levels of screening, specifically:

[0009] Predict the classification labels of the text to be searched, and use the classification labels to perform the first level of screening on the cases in the experience feedback library;

[0010] Identify entity names in the text to be searched, and perform a second level of screening of non-qualified entities on the cases after the first level of screening by using the entity names;

[0011] The identified entity names are deleted from the text to be searched, and the deleted text is encoded into a semantic feature vector. Based on the similarity between the semantic feature vectors, a third cross-stage screening is performed on the cases after the second screening to obtain the final search results.

[0012] Furthermore, the process of entering the experience feedback cases into the experience feedback database is as follows:

[0013] Obtain the experience feedback cases that have passed the review and are waiting to be stored;

[0014] Perform classification inference on the case description text of the case to obtain the classification label of the case;

[0015] Recognize entity names in the case description text of the case to obtain entity names contained in the case;

[0016] Encode the case description text of the case to obtain the feature vector of the case;

[0017] The case description text, classification label, entity name and feature vector are saved in the experience feedback library as metadata of the experience feedback case.

[0018] Furthermore, the classification label is generated by using a natural language processing classification algorithm mechanism to perform classification prediction on the text to be classified;

[0019] The encoding uses the word embedding technology of the BERT model to obtain the semantic feature vector of the text to be encoded.

[0020] Furthermore, the entity recognition is performed by using a naming entity name recognition technology to recognize the device names and component names contained in the text.

[0021] Furthermore, the cases after the first screening are screened by the identified entity names, including two methods: not limiting the entity at all and limiting only part of the entity;

[0022] The completely unqualified entity is not filtered by entity name;

[0023] The limiting only some entities is to select some entity names for screening.

[0024] Furthermore, the similarity between the semantic feature vectors is calculated by using a hierarchical navigable small-world algorithm and a similarity measurement formula.

[0025] Furthermore, the third cross-stage screening is to sort the cases after the second screening by similarity, and delete irrelevant cases whose similarity does not exceed a threshold value based on the sorting results. By setting the threshold value, irrelevant cases are distinguished from cases in the same stage and across stages.

[0026] The second aspect of the present invention provides a nuclear power experience feedback search system based on non-limited entities and across stages.

[0027] Based on the non-limited entity and cross-stage nuclear power experience feedback search system, the experience feedback cases related to the text to be searched are searched from the experience feedback library. The metadata of the experience feedback cases include case description text, classification label, entity name and semantic feature vector, including the first screening module, the second screening module and the third screening module:

[0028] The first screening module is configured to: predict the classification label of the text to be searched, and perform the first screening of the cases in the experience feedback library through the classification label;

[0029] The second screening module is configured to: identify entity names in the text to be searched, and perform a second screening of non-qualified entities on the cases after the first screening by using the entity names;

[0030] The third screening module is configured to delete the identified entity names from the text to be searched, encode the deleted text into a semantic feature vector, and perform a third screening across stages on the cases after the second screening based on the similarity between the semantic feature vectors to obtain the final search results.

[0031] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of the non-limited entity-based, cross-stage nuclear power experience feedback search method as described in the first aspect of the present invention.

[0032] The fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of the non-limited entity-based, cross-stage nuclear power experience feedback search method as described in the first aspect of the present invention are implemented.

[0033] One or more of the above technical solutions have the following beneficial effects:

[0034] The present invention uses non-limited entities and cross-stages to no longer limit the search to specific entities or a single stage, thereby improving the search capability of semantic understanding and achieving complete, accurate and limited nuclear power experience feedback search.

[0035] The present invention utilizes named entity recognition technology to identify entity names in text, and implements screening of unrestricted entities by not restricting entities completely or restricting only parts of entities, and not restricting or partially restricting entities for filtering.

[0036] The present invention adopts a screening method that combines classification labels and similarity sorting to distinguish irrelevant cases from cases in the same stage and across stages, thereby realizing cross-stage screening.

[0037] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0039] Figure 1 This is a flow chart of the method of the first embodiment.

[0040] Figure 2 Schematic diagram of the text classification decision matrix of the first embodiment.

[0041] Figure 3 This is a flowchart for storing experience feedback cases in the first embodiment. DETAILED DESCRIPTION

[0042] It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0043] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0044] Embodiment 1

[0045] In one embodiment of the present disclosure, a method for searching nuclear power experience feedback based on non-limited entities and across stages is provided, wherein experience feedback cases related to a text to be searched are searched from an experience feedback library, wherein metadata of the experience feedback cases include case description text, classification labels, entity names, and semantic feature vectors, and are divided into three levels of screening, including the following steps:

[0046] Step S1: predict the classification label of the text to be searched, and perform the first level of screening on the cases in the experience feedback library through the classification label;

[0047] Step S2: Identify entity names in the text to be searched, and perform a second screening of non-qualified entities on the cases after the first screening by using the entity names;

[0048] Step S3: Delete the identified entity names from the text to be searched, encode the deleted text into a semantic feature vector, and perform a third cross-stage screening on the cases after the second screening based on the similarity between the semantic feature vectors to obtain the final search results.

[0049] The following is a detailed description of the implementation process of the non-limited entity and cross-stage nuclear power experience feedback search method of this embodiment.

[0050] The following three business scenarios for nuclear power experience feedback search are provided:

[0051] 1) When viewing a specific report in an experience feedback library, you need to search for relevant historical cases in the experience feedback library based on the case description text.

[0052] 2) When processing a technical report with a problem description, it is necessary to search for relevant historical cases in the experience feedback library based on the problem description text.

[0053] 3) When the user deems it necessary, a text snippet is needed to search for relevant cases in the experience feedback library.

[0054] The three scenarios use case description text or problem description text or text fragments as the text to be searched, and search for related experience feedback cases from the experience feedback library. Compared with the existing solutions, the innovation of this solution lies in non-restricted entities and cross-stages. The following first explains non-restricted entities and cross-stages, and then explains the new case entry and case search based on the text to be searched based on non-restricted entities and cross-stages.

[0055] 1. Non-qualified entities:

[0056] The entity here, that is, the entity name, includes equipment names and component names. For example, steam generator, main pipeline, manual shut-off valve, and waste heat discharge heat exchanger are equipment names, and V-groove and nut are component names.

[0057] The non-limited entity does not limit the specific entity in the specific report text when searching for related cases of a specific text. The "limited specific entity" means that only the case of a certain device or a certain component can be found; that is, all entity names or part of the entity names are deleted from the text, thereby removing the influence of the entity name on the search or comparison results. Several examples are used to further explain:

[0058] Example 1

[0059] The text to be searched is: "When welding the first steam generator of Unit 2, welding started from the outer wall. After welding about 18mm, the inner wall was cleaned about 5mm. Welding stress release occurred after cleaning, but due to insufficient strength (thickness) of the remaining weld, the parallelism between the main flange surface of the pump casing and the steam generator support reference surface increased sharply from 1.0mm to 2.62mm.", which includes the equipment name "steam generator".

[0060] The search for cases will not be limited to the equipment name "steam generator", but will provide welding cases such as other equipment or parts, that is, the entity is not limited at all.

[0061] The fully unqualified entity is not filtered by entity name:

[0062] Specifically, even if entity names are identified from the text to be searched, but the entity names are not used for filtering and screening, and the entities used for filtering are not limited, then all cases of entity names are included in the filtering results.

[0063] Example 2

[0064] The text to be searched is: "Surface liquid penetration inspection and radiographic inspection of the main pipeline confirmed that the grinder had mistakenly ground off part of the pre-stacked edge of the V-groove of the main pipeline. Afterwards, the pre-stacked edge area that was ground thinner has been repaired during the welding process.", which includes the equipment name "main pipeline" and the component name "V-groove".

[0065] The search for cases will not be limited to the grinding cases of the "V-shaped groove" of the "main pipeline", but will provide grinding cases of the "V-shaped groove" of other equipment such as "manual shut-off valve" and "waste heat discharge heat exchanger", that is, only some entities are limited.

[0066] To limit only some entities is to select some entity names for filtering:

[0067] Specifically, in the case where multiple entity names are identified from the text to be searched, one of the entity names is selected for screening, thereby eliminating the possibility that related cases may be excluded due to joint screening using multiple entity names.

[0068] The recognition of entity names in text is achieved through a perceptron-based sequence labeling model in named entity recognition (NER). The perceptron-based sequence labeling model is trained based on a corpus of device names and component names to identify the device names and component names contained in the text to be searched. The specific steps are as follows:

[0069] Step 1: Model training: Use a manually annotated corpus of named entity recognition in a general field that can be used for training purposes (such as the MSRA Named Entity Corpus) to train a perceptron-based sequence labeling model. This model can label continuous text sequences with part-of-speech tags, and adjacent text segments labeled as noun parts of speech are separated and recognized as named entity names.

[0070] Step 2: Incremental learning: Taking advantage of the perceptron sequence tagging model's ability to incrementally learn new parts of speech at any time, the pre-sorted training set of named entity names in the nuclear power field and their new part-of-speech tags are input into the model as incremental learning materials for incremental learning. The "new parts of speech" here are specially designed parts of speech for the nuclear power experience feedback field, including two types: n-equip (representing noun parts of speech and belonging to equipment names) and n-parts (representing noun parts of speech and belonging to component names).

[0071] Step 3: Online recognition: Use the sequence labeling model after incremental learning to perform named entity recognition on the search text. From the recognized named entities, select text fragments with the parts of speech of n-equ ip and n-parts, that is, select the equipment name and part name as the entity name.

[0072] Therefore, the algorithm model has the ability to identify the names of nuclear power professional equipment and parts, and the entity names after identification are used for the second level of screening.

[0073] 2. Cross-stage

[0074] When searching for relevant cases of a specific text, this method does not limit the searched cases to a specific stage, and thus can provide cross-stage case search results.

[0075] For example, the problem description of a certain case is "The noise of the DC-1 rectifier input transformer is 101.6dB, and the noise of the rectifier input transformer of other chargers is 81.3dB. The noise of the DC-1 rectifier input transformer is much louder than that of the same model charger, and the original design is unreasonable." This case is a description of the designer's unreasonable design problem and should be classified as a problem description in the design stage.

[0076] For cases in the experience feedback database that do not belong to the design stage, there are texts such as "During the factory test and acceptance of the excitation transformer of unit 1#, the design noise value was found to be <78dB(A), and the actual noise value under AN conditions was 64dB(A), which did not meet the noise requirements for the excitation transformer in the TG package contract and the excitation system technical agreement between our company and ABB (<55dB(A))". This text belongs to the description of the "test" stage.

[0077] The "cross-stage" search is intended to achieve the following effect: when a user wants to search for the design of "noise (acoustics)" as a topic, the search results go beyond the content of the "design" stage and include noise problems of such equipment in the "test" and other stages.

[0078] In order to achieve the above-mentioned cross-stage effect, in this embodiment, a classification list of technical points involving nuclear power experience feedback is maintained to define the types of technical points of the case text. The identified case technical points are stored as a classification label of the case data in the core data element of the experience feedback. Here, the various technical points in the classification list are defined according to the meaning that does not involve the stage. For example, in the above example, the technical point should be defined as "noise (acoustics)" instead of "noise reduction design"; in this way, the cases in the above "experimental" stage and the cases in the "design" stage can all be labeled as "noise (acoustics)" in the training set for model training, and identified as "noise (acoustics)" by the classifier.

[0079] Here, the classification label is identified by using the text classification algorithm (model) of Natural Language Processing (NLP) to infer which category of the classification list the text to be classified belongs to, and obtain the classification label.

[0080] The overall process of text classification is generally as follows:

[0081] If there is a set of label categories (C) and a set of documents (D), and there is the following expression:

[0082] C={c1,c2,...,c m},D={d1,d2,...,d n}

[0083] Text classification can be defined as a method for determining each element a in the matrix ij Assigning a Boolean value to a task, such as Figure 2 Specifically, if a ij The value of 1 represents document d j Belongs to category c i On the contrary, if the matrix a ij A value of 0 means document d j Not in category c i From a computer perspective, the process of text classification is to solve the problem of estimating the mapping between the category set C and the document set D, thereby forming a classifier, which can be expressed as:

[0084] f:{c i ,d j}→C×D

[0085] In the specific process, the development of the classifier uses machine learning methods to use the approximate mapping (f′) learned from the training set as a model for inference to perform the prediction task of text classification.

[0086] This embodiment converts the text into a vector representation through the BERT model, inputs it into a specific CNN (convolutional neural network) model for training, and obtains a deep learning classifier based on a neural network. The final output of the neural network is a specific classification category. When the text to be classified is input into this deep learning classifier, the model infers that the text is "noise (acoustics)" and "firefighting" through the approximation (f') of such a mapping. With such a definition, it paves the way for the feature vector search algorithm below.

[0087] Given that the extracted classification labels are defined as not containing or involving the semantics or information of a specific stage, when searching for related cases, the recall of the results of the "cross-stage" step is achieved through library-wide label filtering and sorting of feature vector similarity scores.

[0088] Therefore, based on the above classification labels, the specific implementation method across stages can be found in the specific steps of “4. Case search based on the text to be searched”.

[0089] 3. New cases added to the database

[0090] The following is an explanation of the process of entering experience feedback cases into the experience feedback database. Figure 3 As shown, specifically including:

[0091] (1) Obtain the experience feedback cases that have passed the review and are waiting to be stored;

[0092] First, the quality of the newly entered case content is manually reviewed and judged, mainly to check whether there are any duplications in the case history and whether the semantics of the problem description is understandable enough; cases that fail the review will not proceed to the following process.

[0093] (2) Classify and infer the case description text of the case to obtain the classification label of the case;

[0094] From the newly entered experience feedback content (i.e., case description text), the Natural Language Processing (NLP) classification algorithm mechanism is used to infer the category to which it belongs, and then assisted by manual correction. The corrected classification is used as the classification label for the case.

[0095] (3) Identify entity names in the case description text of the case to obtain the entity names contained in the case;

[0096] Named Entity Recognition technology is used to identify entity names, including equipment names and component names, from the newly entered experience feedback content (i.e., case description text).

[0097] (4) Encode the case description text of the case to obtain the feature vector of the case;

[0098] The newly entered experience feedback content (i.e., case description text) is input into the BERT pre-trained model, and the word embedding technology of the BERT model is used to obtain a semantic feature vector; the semantic feature vector is a high-dimensional vector, and each component of the vector is a high-precision decimal.

[0099] (5) The case description text, classification label, entity name and feature vector are saved as metadata of the experience feedback case in the experience feedback library.

[0100] The classification labels, entity names, and BERT semantic feature vectors identified from the newly entered experience feedback are stored in the system database as metadata of the experience feedback case, waiting to be used in subsequent searches.

[0101] 4. Case search based on the text to be searched

[0102] Search the experience feedback cases related to the text to be searched from the experience feedback library. The metadata of the experience feedback cases include case description text, classification label, entity name and semantic feature vector. The specific steps of case search based on non-restricted entities and cross-stage are as follows:

[0103] (1) Using the Natural Language Processing (NLP) classification algorithm mechanism, the classification label of the text to be searched is inferred;

[0104] (2) Use named entity recognition technology on the search text to identify and delete entity names from the text, including device names and component names;

[0105] (3) The text to be searched after deleting the entity name is input into the BERT pre-trained model for encoding to obtain a specific semantic feature vector.

[0106] (4) Only cases with the same classification label are selected from the entire database to obtain the results of the first screening.

[0107] (5) In the results of the first screening, the entity name is selectively added as a new screening filter condition to obtain the results of the second screening, which contains the basic range data of each stage.

[0108] (6) Using the semantic feature vectors generated by the text to be searched, the Hierarchical Navigable Small World graphs (HNSW) algorithm is used to perform a third-stage cross-stage screening from the basic range data (i.e., the results of the second screening). (Available steps)

[0109] (7) After the third screening, the similarity measurement is calculated between each case vector screened out and the semantic feature vector generated by the text to be searched, and the similarity result is used for the subsequent score sorting. The similarity calculation formula is:

[0110]

[0111] Among them, the semantic feature vector generated by the text to be searched is V A The semantic feature vector of each case screened by the HNSW algorithm is V X To express.

[0112] HNSW is used together with the similarity measurement formula. The former is used to accelerate the search speed, reduce the time and space cost of the search, and improve the search efficiency. The latter is responsible for calculating the similarity between high-dimensional vector data points. According to the similarity calculation results, the cases screened by each HNSW algorithm are sorted according to the calculation score. Since the basic range data contains cases of various stages, the top ranked cases are non-restricted entities and experience feedback cases of the same stage, the cases with the following scores are non-restricted entities and cross-stage experience feedback cases, and the cases with the lowest scores are those irrelevant cases that are very different from the semantics of the newly entered cases to be retrieved. The irrelevant cases with scores below a certain limited score (i.e., the threshold) are removed from the sorting list, and the cases of non-restricted entities, cross-stages (also including the current stage), similar professional fields and semantically similar are obtained, which are used to provide searchers with valuable reference use cases for nuclear power experience feedback on a specific matter.

[0113] Embodiment 2

[0114] In one embodiment of the present disclosure, a nuclear power experience feedback search system based on non-limited entities and across stages is provided, which searches for experience feedback cases related to a text to be searched from an experience feedback library, wherein the metadata of the experience feedback cases includes case description text, classification labels, entity names, and semantic feature vectors, and includes a first screening module, a second screening module, and a third screening module:

[0115] The first screening module is configured to: predict the classification label of the text to be searched, and perform the first screening of the cases in the experience feedback library through the classification label;

[0116] The second screening module is configured to: identify entity names in the text to be searched, and perform a second screening of non-qualified entities on the cases after the first screening by using the entity names;

[0117] The third screening module is configured to delete the identified entity names from the text to be searched, encode the deleted text into a semantic feature vector, and perform a third screening across stages on the cases after the second screening based on the similarity between the semantic feature vectors to obtain the final search results.

[0118] Embodiment 3

[0119] The purpose of this embodiment is to provide a computer-readable storage medium.

[0120] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the non-limited entity-based, cross-stage nuclear power experience feedback search method as described in the first embodiment of the present disclosure.

[0121] Embodiment 4

[0122] The purpose of this embodiment is to provide an electronic device.

[0123] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps of the non-limited entity-based, cross-stage nuclear power experience feedback search method as described in the first embodiment of the present disclosure are implemented.

[0124] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A nuclear power experience feedback search method based on non-limited entities and across stages, characterized in that: Search the experience feedback cases related to the text to be searched from the experience feedback library. The metadata of the experience feedback cases includes case description text, classification label, entity name and semantic feature vector. The screening is divided into three steps, specifically: Predict the classification labels of the text to be searched, and use the classification labels to perform the first level of screening on the cases in the experience feedback library; The extracted classification labels are defined as those that do not contain or involve the semantics or information of a specific stage; Identify entity names in the text to be searched, and perform a second level of screening of non-qualified entities on the cases after the first level of screening by using the entity names; The second screening of the non-limited entities includes two methods: not limiting the entity at all and limiting only part of the entity; The completely unqualified entity is not filtered by entity name; The limiting of only some entities is to select some entity names for screening; The non-limited entity does not limit the specific entity in the specific report text when searching for related cases of a specific text; The identified entity names are deleted from the text to be searched, and the deleted text is encoded into a semantic feature vector. Based on the similarity between the semantic feature vectors, a third cross-stage screening is performed on the cases after the second screening to obtain the final search results.

2. The non-limited entity-based, cross-stage nuclear power experience feedback search method according to claim 1, characterized in that: The process of entering experience feedback cases into the experience feedback database is as follows: Obtain the experience feedback cases that have passed the review and are waiting to be stored; Perform classification inference on the case description text of the case to obtain the classification label of the case; Recognize entity names in the case description text of the case to obtain entity names contained in the case; Encode the case description text of the case to obtain the feature vector of the case; The case description text, classification label, entity name and feature vector are saved in the experience feedback library as metadata of the experience feedback case.

3. The non-limited entity-based, cross-stage nuclear power experience feedback search method according to claim 2, characterized in that: The classification label is to use the natural language processing classification algorithm mechanism to perform classification prediction on the text to be classified; The encoding uses the word embedding technology of the BERT model to obtain the semantic feature vector of the text to be encoded.

4. The non-limited entity-based, cross-stage nuclear power experience feedback search method according to claim 1, characterized in that: The entity name recognition is to recognize the device name and component name contained in the text through the recognition technology of the named entity name.

5. The non-limited entity-based, cross-stage nuclear power experience feedback search method according to claim 1, characterized in that: The similarity between the semantic feature vectors is calculated by using a hierarchical navigable small-world algorithm and a similarity measurement formula.

6. The non-limited entity-based, cross-stage nuclear power experience feedback search method according to claim 1, characterized in that: The third cross-stage screening is to sort the cases after the second screening by similarity, and delete irrelevant cases whose similarity does not exceed a threshold value based on the sorting results. By setting the threshold, irrelevant cases are distinguished from cases in the same stage and across stages.

7. A nuclear power experience feedback search system based on non-limited entities and across stages, characterized by: Searching for experience feedback cases related to the text to be searched from the experience feedback library, wherein the metadata of the experience feedback cases includes case description text, classification label, entity name and semantic feature vector, including a first screening module, a second screening module and a third screening module: The first screening module is configured to: predict the classification label of the text to be searched, and perform the first screening of the cases in the experience feedback library through the classification label; The extracted classification labels are defined as those that do not contain or involve the semantics or information of a specific stage; The second screening module is configured to: identify entity names in the text to be searched, and perform a second screening of non-qualified entities on the cases after the first screening by using the entity names; The second screening of the non-limited entities includes two methods: not limiting the entity at all and limiting only part of the entity; The completely unqualified entity is not filtered by entity name; The limiting of only some entities is to select some entity names for screening; The non-limited entity does not limit the specific entity in the specific report text when searching for related cases of a specific text; The third screening module is configured to delete the identified entity names from the text to be searched, encode the deleted text into a semantic feature vector, and perform a third screening across stages on the cases after the second screening based on the similarity between the semantic feature vectors to obtain the final search results.

8. An electronic device, comprising: a memory for non-transitory storage of computer readable instructions; a processor for executing the computer-readable instructions; Wherein, when the computer-readable instructions are executed by the processor, the method according to any one of claims 1 to 6 is executed.

9. A storage medium, characterized in that: The computer-readable instructions are non-transitorily stored, wherein when the computer-readable instructions are executed by a computer, the method according to any one of claims 1 to 6 is performed.

Images