An intelligent question-answering method for outdoor substation and a robot placed outdoors
By constructing an entity recognition BERT+FLAT model and a knowledge graph model, the problems of low accuracy and high latency in handling complex issues in the power industry by existing intelligent question answering systems have been solved. This has enabled rapid and accurate question answering of substation equipment faults, improving the processing efficiency of grassroots personnel.
Patent Information
- Application Number
- CN202210748904.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-06-29
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2042-06-29
AI Technical Summary
Existing intelligent question-answering systems are unable to effectively handle complex questions involving multiple entities or relationships in the power sector, resulting in low accuracy and delayed response, failing to meet the needs of frontline personnel for rapid judgment of substation equipment faults.
A BERT+FLAT entity recognition model is constructed, which combines character-integrated vocabulary information to extract device entity elements and fill slots. A knowledge graph model is used for subgraph matching, and the knowledge graph triple with the highest similarity is selected as the question answer to achieve intelligent question answering.
It improved the accuracy of machine-based question-and-answering based on text information of power equipment fault cases, reduced the response latency of the question-and-answer robot, and enabled it to quickly and accurately identify and provide feedback on the fault location, cause, and suggested measures, thereby improving the progress of equipment fault handling.
Smart Images

Figure CN115129842B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to an intelligent question-answering method for an outdoor substation and a robot placed outdoors, belonging to the technical field of outdoor robots. Background Art
[0002] In the long-term operation of the power system, a large number of historical equipment failure cases are stored in the substation. The case fully records the equipment failure information of different periods, such as the ledger information (including equipment model specifications, voltage level, factory date, commissioning date, etc.), fault phenomenon description, fault cause, solution measures, etc., which contains rich equipment hidden fault problem information and expert handling experience. In fact, when grassroots personnel carry out operation and maintenance, overhaul and other operations, they must strictly follow the process specifications of technical standards to carry out operations. However, the equipment failure situation in the substation is complicated, and the handling process is often limited by the practical experience, knowledge reserves and professional capabilities of grassroots personnel. Therefore, based on the structuring, knowledge and intelligence of substation equipment failure cases, building an intelligent question-and-answer system can effectively improve the work efficiency of grassroots personnel.
[0003] The existing intelligent question-answering systems designed for the power sector mainly include power grid customer service question-answering systems, transformer operation and inspection intelligent question-answering systems, and power business question-answering systems. These systems have insufficient semantic understanding capabilities and have good effects on typical simple questions such as single entity, entity-entity, entity-relationship-entity, but cannot solve the complexity of a question containing multiple entities or relationships. Summary of the invention
[0004] In view of the defects of the prior art, the first object of the present invention is to provide a method for extracting device entity elements in natural language questions by constructing an entity recognition BERT+FLAT model, and filling the slots of natural language questions; and based on the method of integrating characters into vocabulary information, performing preliminary extraction of device entity elements to obtain preliminary entity extraction results; after deduplication and screening of the preliminary entity extraction results, constructing an entity pinyin homonym library; then converting the entity pinyin homonym library into Chinese character pinyin data with tones; obtaining the final entity extraction result according to the matching of the Chinese character pinyin data; and at the same time, filling the blank of the absolute position coding information of the natural language questions according to the relative position coding information of the natural language questions, so as to realize natural language The slots of the question are filled; then a knowledge graph model is constructed to perform sub-graph matching on the question graph to obtain knowledge graph triples; the similarity between the knowledge graph triples and the question graph is calculated, and the knowledge graph triples with the highest similarity are selected as the answers to the questions and answers, thereby realizing intelligent question and answering of outdoor substations, thereby effectively improving the accuracy of the machine's question and answering based on the text information of substation equipment failure cases, reducing the response delay of the question and answer robot, and being able to quickly and accurately judge and provide feedback to the grassroots personnel based on the questioning needs of grassroots personnel (questioners) in different descriptive forms. Effective content such as the fault location, cause of the fault, and recommended measures of the substation equipment can be fed back to the grassroots personnel, thereby speeding up the progress of equipment fault handling in the station, and realizing an intelligent question and answer method for outdoor substations.
[0005] The second purpose of the present invention is to provide an intelligent question-and-answer robot for outdoor substations, which effectively improves the accuracy of question-and-answer based on text information of substation equipment failure cases by constructing a knowledge base module and an intelligent question-and-answer module, reduces the response delay of the question-and-answer robot, and can quickly and accurately judge and feedback to the grassroots personnel the fault location, cause, suggested measures and other effective content of the substation equipment according to the different description forms of question questions asked by the grassroots personnel (questioners), thereby speeding up the progress of equipment failure handling in the station.
[0006] To achieve one of the above purposes, the first technical solution of the present invention is:
[0007] An intelligent question-answering method for an outdoor substation comprises the following steps:
[0008] The first step is to receive natural language questions from external input;
[0009] The second step is to identify the intent of the natural language questions in the first step, analyze the questioner's inquiry intention, and classify them into inquiries about fault location, fault cause, and recommended measures;
[0010] The third step is to build an entity recognition BERT+FLAT model based on the intent recognition results in the second step, extract device entity elements in natural language questions, and fill in the slots of natural language questions.
[0011] The entity recognition BERT+FLAT model performs preliminary extraction of device entity elements based on a method of integrating characters into vocabulary information to obtain preliminary entity extraction results;
[0012] After removing duplicates and filtering the preliminary entity extraction results, an entity pinyin homonym database is constructed;
[0013] Then convert the entity pinyin homonym database into Chinese character pinyin data with tones;
[0014] The final entity extraction result is obtained by matching the Chinese character pinyin data;
[0015] At the same time, according to the relative position coding information of the natural language question, the blank of the absolute position coding information of the natural language question is filled to realize the slot filling of the natural language question;
[0016] The device entity element at least includes a name and / or a position and / or a quantifier and / or a time and / or a device subcomponent and / or device status information;
[0017] The fourth step is to determine whether there is a relationship between the device entity elements in the third step and what kind of relationship exists, and obtain the relationship information between the device entity elements;
[0018] Step 5: Based on the relationship information in step 4, the hidden relationships between the device entity elements are mined to obtain a question sentence graph;
[0019] Step 6: Build a knowledge graph model to perform subgraph matching on the question-sentence graph in step 5 to obtain knowledge graph triples.
[0020] In the seventh step, the similarity between the knowledge graph triple in the sixth step and the question-sentence graph is calculated, and the knowledge graph triple with the highest similarity is selected as the answer to the question and answer, thus realizing intelligent question and answer for outdoor substations.
[0021] The present invention constructs an entity recognition BERT+FLAT model through continuous exploration and experimentation, extracts device entity elements in natural language questions, and performs slot filling on natural language questions; the entity recognition BERT+FLAT model performs preliminary extraction of device entity elements based on a method of integrating characters into vocabulary information to obtain preliminary entity extraction results; removes duplicates and filters the preliminary entity extraction results to construct an entity pinyin homonym library; then converts the entity pinyin homonym library into Chinese character pinyin data with tones; obtains the final entity extraction result according to the matching of the Chinese character pinyin data; and at the same time, according to the natural language questions
[0022] Relative position coding information fills the gap of absolute position coding information of natural language questions and realizes slot filling of natural language questions; then a knowledge graph model is constructed to perform sub-graph matching on the question graph to obtain knowledge graph triples; the similarity between the knowledge graph triples and the question graph is calculated, and the knowledge graph triples with the highest similarity are selected as the answer to the question and answer, so as to realize intelligent question and answer of outdoor substations, thereby effectively improving the accuracy of the machine's question and answer based on the text information of substation equipment failure cases, reducing the response delay of the question and answer robot, and being able to quickly and accurately judge and feedback to the grassroots personnel (questioners) according to the different description forms of questioning needs of the grassroots personnel (questioners) and other effective contents such as the fault location, fault cause, and recommended measures of the substation equipment, thereby accelerating the progress of equipment fault handling in the station.
[0023] At the same time, the present invention adopts the entity recognition BERT+FLAT model, and performs entity extraction and slot filling based on the method of integrating characters into vocabulary information. The relative position encoding information of the text is introduced in the encoding layer, filling the gap of only absolute position information in the original converter transformer. This method will not affect the extraction results due to word segmentation errors, and can comprehensively utilize the information of characters and words to further improve the efficiency of intent recognition, and then quickly and accurately judge and feedback to grassroots personnel the effective content such as the fault location, fault cause, and recommended measures of the substation equipment, thereby accelerating the progress of equipment fault handling within the station.
[0024] As the preferred technical measures:
[0025] In the second step, the method of intent recognition is as follows:
[0026] Step 21, vectorizing the natural language question to obtain a word vector;
[0027] Step 22, construct a pre-trained model BERT to optimize the word vector in step 21 to obtain an optimized word vector;
[0028] Step 23, constructing a neural network layer Bi-LSTM to semantically encode the optimized word vector in step 22 to obtain a semantic encoding vector;
[0029] Step 24, calculating the weight of the semantic encoding vector in step 23 through the attention layer Attention;
[0030] Step 25, multiply all semantic coding vectors by corresponding weights and sum them up to obtain a feature vector;
[0031] Step 26, input the feature vector in step 25 into the logistic regression function softmax to obtain the questioning intention recognition result of the questioner.
[0032] As the preferred technical measures:
[0033] In the third step, the method for constructing the entity recognition BERT+FLAT model is as follows: Step 31, vectorizing the inputted substation equipment historical fault case text to obtain a case word vector;
[0034] Step 32, using the pre-trained model BERT training case word vector to generate a word tree, starting from the head of each sentence, matching all words under the corresponding characters according to the maximum matching principle, and placing the word information directly at the end of the sentence to obtain the training text;
[0035] In the pre-trained model BERT, the maximum text length that can be processed is ZD, and each token is represented as a BS-dimensional vector;
[0036] The identification token is a single character or word, which is converted into an identification token-id and mapped to an address id to obtain a vector matrix representation of (ZD, BS);
[0037] Step 33, inputting the training text in step 32 into the position encoding layer for encoding, the encoding is divided into absolute position encoding and relative position encoding;
[0038] The absolute position encoding is obtained according to the order of the training text sequence;
[0039] Relative position coding has four types of position coding. Each character has two identical head and tail position codings, and each word has two different head and tail position codings. Four relative distances are calculated between every two nodes. The specific calculation formula is as follows:
[0040]
[0041]
[0042]
[0043]
[0044] In the formula head [ i ]、 tail [ i ] represent the head and tail codes of a character or word respectively, , , , Respectively represent i The first character or word and the j The distance between the heads of individual characters or words;
[0045] The calculation results are fused through the fully connected layer to obtain the final relative position encoding, and the calculation formula is as follows:
[0046]
[0047]
[0048]
[0049] In the formula represents the learnable parameters, represents the concatenation operator, d yes , , , and k express ij Dimension index of position encoding;
[0050] Step 34, using the multi-head self-attention mechanism, combined with the relative position encoding in step 33, extracts text information features. The calculation formula of text information features is as follows:
[0051]
[0052]
[0053]
[0054]
[0055] In the formula , , , and u, These are all learnable parameters in the pre-trained model BERT;
[0056] Step 35: Input the text information features in step 34 into the discriminative probability model CRF layer to predict and output the text sequence with entity labels. Y ={ Y 1 ,Y 2 ,Y 3 ,...,Y n}, and obtain the preliminary entity extraction results of substation equipment failure cases.
[0057] As the preferred technical measures:
[0058] In the fourth step, the method for obtaining the relationship information is as follows:
[0059] Step 41, vectorizing the device entity elements to convert them into index vectors of the same length;
[0060] Step 42: The word vector corresponding to the index vector in step 41 c ijt Transmitted to the neural network layer Bi-GRU for semantic encoding to obtain the semantic encoding vector h ijt ;
[0061] Step 43: convert the semantic encoding vector in step 42 h ijt Input into the logistic regression softmax classifier to obtain the relationship classification result.
[0062] As the preferred technical measures:
[0063] Semantic Encoding Vector h ijt The calculation formula is as follows:
[0064]
[0065] In the formula c ijt Represents the word vector of the tth word in the jth sentence input at time i. Bi-GRU is a neural network function, which includes forward and backward neural network GRU, which can fully learn contextual relationships.
[0066] As the preferred technical measures:
[0067] In the sixth step, the method for constructing the knowledge graph model is as follows:
[0068] Step 61, vectorizing the device entity elements to obtain index vectors with the same length;
[0069] Step 62: The word vector corresponding to the index vector in step 61 c ijt Transmitted to the neural network layer Bi-GRU for semantic encoding to obtain the semantic encoding vector h ijt ;
[0070] Step 63: convert the semantic encoding vector in step 62 h ijt Input into the logistic regression softmax classifier to obtain the relationship classification result;
[0071] Step 64: Based on the relationship classification result in step 63, the extracted equipment entity elements and relationships are stored as knowledge using the graph database Neo4j to form a knowledge graph about substation equipment failures consisting of nodes: equipment entity elements and edges: relationships;
[0072] Each device entity element pair and the relationship between them in the knowledge graph form a triple< h , r , t >,<h,r,t> Respectively represent <head entity, relationship, tail entity>;
[0073] Step 65, using the relationship prediction model to complete the paths of different device entity elements, and find potential directed edges for different nodes in the knowledge graph in step 64.
[0074] As the preferred technical measures:
[0075] The completion method of the relationship prediction model is as follows:
[0076] Step 651, using the vector translation algorithm TransH, for each relationship r , giving a hyperplane W r ;
[0077] Step 652, calculate the relationship vector on the hyperplane Wr in step 651 d r ;
[0078] Step 653, according to the relationship vector in step 652 d r , the original head entity h and tail entity t Mapped onto the hyperplane is h r , t r ;
[0079] Step 654, requiring the triple to satisfy h r + d r = t r , for triples with the same head entity and relation ( h , r , t 1 )and( h , r , t 2 ) Through the relationship r The hyperplane mappings are:
[0080] h+r=t 1
[0081] h+r=t 2 ;
[0082] Implement path completion for entity elements of different devices.
[0083] As the preferred technical measures:
[0084] In the sixth step, the subgraph matching method is as follows:
[0085] The question-sentence graph uses the search tree model in the graph-speaking subgraph isomorphism algorithm VF2 to match the subgraph with the knowledge graph. The matching relationship is as follows:
[0086] G1=(E1,R1),G2=(E2,R2),where G1 is the question sentence graph, G2 is the knowledge graph, E is the set of nodes, i.e., equipment entity elements, and R is the set of edges, i.e., relationships;
[0087] Each node in the subgraph isomorphism algorithm VF2 is s.
[0088] s contains a series of mappings of nodes in G1 and G2<key,value> ,
[0089] Among them, key represents the node in G1, and value represents the node in G2;
[0090] As the height of the search tree increases, the number of key-value pairs in s will gradually increase;
[0091] When all node pairs in s satisfy the constraint formula, the state s is called a consistent state, and subgraph matching is completed;
[0092] The constraint formula is as follows:
[0093]
[0094]
[0095]
[0096]
[0097]
[0098]
[0099] M E1XE2 represents the corresponding relationship set of nodes in the question sentence graph and knowledge graph. ,but Denotes G corresponding to node u 2 Nodes in ,but Denotes G corresponding to node v 1 Nodes in .
[0100] As the preferred technical measures:
[0101] In the seventh step, the similarity calculation formula is as follows:
[0102] In the formula A and B They represent the knowledge graph triple and question sentence graph respectively. The output interval of the similarity value is [0,1]. The larger the value, the higher the similarity between the knowledge graph triple and the question sentence graph.
[0103] To achieve one of the above purposes, the second technical solution of the present invention is:
[0104] An intelligent question-answering robot placed in an outdoor substation applies the above-mentioned intelligent question-answering method for an outdoor substation; the robot comprises a knowledge base module, an intelligent question-answering module, a battery, several circuit modules, and a body.
[0105] The knowledge base module is used to process and analyze unstructured substation equipment failure case text information, and after a series of unit processing in the module, a structured, easy-to-operate, and easy-to-use knowledge set is formed, which includes a substation equipment failure case text information input unit, an entity extraction unit, a relationship extraction unit, a knowledge storage unit, a relationship prediction unit, and a knowledge update unit;
[0106] A substation equipment failure case text information input unit, used to receive a substation equipment failure text data set input from outside the system;
[0107] The entity extraction unit is built based on the entity recognition BERT+FLAT model, which is used to extract information elements such as equipment name, location, quantifier, time, equipment sub-components, equipment status, etc. in the substation equipment fault text dataset;
[0108] The relation extraction unit is used to determine whether there is a relationship between entities in the text and what kind of relationship exists;
[0109] Knowledge storage unit, used to store entity and relationship information extracted from text to form a knowledge graph;
[0110] The relationship prediction unit is used to mine hidden relationships between entities and improve the existing knowledge graph;
[0111] The knowledge update unit is used to subsequently update new nodes: entities and edges: relationships on the existing knowledge graph;
[0112] The intelligent question-answering module is used to process and analyze the questions input by grassroots personnel and feedback relevant answers, and includes a question input unit, a question intention recognition unit, a slot filling unit, a subgraph matching unit, an answer sorting unit, and an answer encapsulation output unit;
[0113] A question input unit, used for receiving natural language questions input from outside the system;
[0114] The intention recognition unit is used to analyze the inquiry intentions of grassroots personnel and classify them according to inquiries about fault locations, fault causes, and recommended measures;
[0115] The slot filling unit is built based on the entity recognition BERT+FLAT model, which is used to extract the device name, location, quantifier, time, device sub-components, and device status information elements in the question;
[0116] The subgraph matching unit is used to perform subgraph matching between the constructed question-sentence graph and the knowledge graph in the knowledge base, and return the matching knowledge graph triples;
[0117] The answer ranking unit is used to select the subgraph matching results with the highest similarity;
[0118] The answer encapsulation output unit is used to encapsulate the subgraph matching results with similarity;
[0119] The battery is sealed with a metal member;
[0120] The circuit module is encapsulated with metal parts with electromagnetic shielding function, and the cables connecting each circuit module are also inserted into a metal shielding sleeve for protection; and the connections and gaps of each component are covered with silver-copper conductive paint paper with an inert metal plated on the surface; the entire outer surface of the fuselage is covered with a Micros waterproof pressure relief film.
[0121] The present invention aims at solving the problem of substation equipment failure and provides an intelligent question-answering robot for assisting decision-making on the substation site, which has the following advantages:
[0122] 1) The entity extraction unit and slot filling unit of the present invention are constructed using the entity recognition BERT+FLAT model. Entity extraction and slot filling are performed based on the method of integrating characters into vocabulary information. The relative position encoding information of the text is introduced in the encoding layer, filling the gap of only absolute position information in the original transformer. This method will not affect the extraction results due to word segmentation errors, and can comprehensively utilize the information of characters and words.
[0123] 2) The present invention adopts the entity recognition BERT+FLAT model, which can support parallel computing, has fast inference speed, and reduces the response delay of the question-answering robot.
[0124] 3) The intelligent question-answering module of the present invention can significantly improve the accuracy of answering complex questions containing multiple entities or relationships in a question sentence.
[0125] 4) The present invention adds a knowledge update unit, so there is no need to retrain all the text data. When there are new fault cases or new knowledge needs to be supplemented, the inspection personnel only need to input the new knowledge text into the knowledge update unit to continuously expand the knowledge of the system, realize the automatic update of the knowledge base during system service, and improve the robot's question-answering ability.
[0126] Compared with the prior art, the present invention has the following beneficial effects:
[0127] The present invention constructs an entity recognition BERT+FLAT model through continuous exploration and experimentation, extracts device entity elements in natural language questions, and performs slot filling on natural language questions; the entity recognition BERT+FLAT model performs preliminary extraction of device entity elements based on a method of integrating characters into vocabulary information to obtain preliminary entity extraction results; removes duplicates and filters the preliminary entity extraction results to construct an entity pinyin homonym library; then converts the entity pinyin homonym library into Chinese character pinyin data with tones; obtains the final entity extraction result according to the matching of the Chinese character pinyin data; and at the same time, according to the natural language questions
[0128] Relative position coding information fills the gap of absolute position coding information of natural language questions and realizes slot filling of natural language questions; then a knowledge graph model is constructed to perform sub-graph matching on the question graph to obtain knowledge graph triples; the similarity between the knowledge graph triples and the question graph is calculated, and the knowledge graph triples with the highest similarity are selected as the answer to the question and answer, so as to realize intelligent question and answer of outdoor substations, thereby effectively improving the accuracy of the machine's question and answer based on the text information of substation equipment failure cases, reducing the response delay of the question and answer robot, and being able to quickly and accurately judge and feedback to the grassroots personnel (questioners) according to the different description forms of questioning needs of the grassroots personnel (questioners) and other effective contents such as the fault location, fault cause, and recommended measures of the substation equipment, thereby accelerating the progress of equipment fault handling in the station.
[0129] At the same time, the present invention adopts the entity recognition BERT+FLAT model, and performs entity extraction and slot filling based on the method of integrating characters into vocabulary information. The relative position encoding information of the text is introduced in the encoding layer, filling the gap of only absolute position information in the original converter transformer. This method will not affect the extraction results due to word segmentation errors, and can comprehensively utilize the information of characters and words to further improve the efficiency of intent recognition, and then quickly and accurately judge and feedback to grassroots personnel the effective content such as the fault location, fault cause, and recommended measures of the substation equipment, thereby accelerating the progress of equipment fault handling within the station.
[0130] Furthermore, after continuous exploration and experimentation, the present invention provides an intelligent question-and-answer robot placed in an outdoor substation by constructing a knowledge base module and an intelligent question-and-answer module, which effectively improves the accuracy of the machine's question-and-answer based on the text information of substation equipment failure cases, reduces the response delay of the question-and-answer robot, and can quickly and accurately judge and provide feedback to the grassroots personnel on the fault location, cause, suggested measures and other effective content of the substation equipment based on the questioning needs of the grassroots personnel (questioners) in different descriptive forms, thereby accelerating the progress of equipment failure handling within the station. BRIEF DESCRIPTION OF THE DRAWINGS
[0131] Figure 1 A flow chart of the method for constructing a knowledge base module of the present invention;
[0132] Figure 2 It is a flow chart of the operation of the entity extraction unit of the present invention;
[0133] Figure 3 This is a flow chart showing the operation of the relationship extraction unit of the present invention;
[0134] Figure 4 It is a partial schematic diagram of the knowledge graph of power transformation equipment failure of the present invention;
[0135] Figure 5 This is a schematic diagram of the function of the relationship prediction unit of the present invention;
[0136] Figure 6 This is the TransH principle diagram of the present invention;
[0137] Figure 7 This is a flow chart of the method for constructing the intelligent question-answering module of the present invention;
[0138] Figure 8 It is a flowchart of the operation of the recognition unit of the present invention;
[0139] Fig. 9 This is a hardware configuration diagram of the intelligent question-answering robot placed in an outdoor substation according to the present invention. DETAILED DESCRIPTION
[0140] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.
[0141] On the contrary, the present invention covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present invention as defined by the claims. Further, in order to make the public have a better understanding of the present invention, some specific details are described in detail in the detailed description of the present invention below. Those skilled in the art can fully understand the present invention without the description of these details.
[0142] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which the present invention pertains. The terms used herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "or / and" as used herein includes any and all combinations of one or more of the associated listed items.
[0143] A specific embodiment of the intelligent question-answering method for outdoor substations of the present invention:
[0144] An intelligent question-answering method for an outdoor substation comprises the following steps:
[0145] The first step is to receive natural language questions from external input;
[0146] The second step is to identify the intent of the natural language questions in the first step, analyze the questioner's inquiry intention, and classify them into inquiries about fault location, fault cause, and recommended measures;
[0147] The third step is to build an entity recognition BERT+FLAT model based on the intent recognition results in the second step, extract device entity elements in natural language questions, and fill in the slots of natural language questions.
[0148] The entity recognition BERT+FLAT model performs preliminary extraction of device entity elements based on a method of integrating characters into vocabulary information to obtain preliminary entity extraction results;
[0149] After removing duplicates and filtering the preliminary entity extraction results, an entity pinyin homonym database is constructed;
[0150] Then convert the entity pinyin homonym database into Chinese character pinyin data with tones;
[0151] The final entity extraction result is obtained by matching the Chinese character pinyin data;
[0152] At the same time, according to the relative position coding information of the natural language question, the blank of the absolute position coding information of the natural language question is filled to realize the slot filling of the natural language question;
[0153] The device entity element at least includes a name and / or a position and / or a quantifier and / or a time and / or a device subcomponent and / or device status information;
[0154] The fourth step is to determine whether there is a relationship between the device entity elements in the third step and what kind of relationship exists, and obtain the relationship information between the device entity elements;
[0155] Step 5: Based on the relationship information in step 4, the hidden relationships between the device entity elements are mined to obtain a question sentence graph;
[0156] Step 6: Build a knowledge graph model to perform subgraph matching on the question-sentence graph in step 5 to obtain knowledge graph triples.
[0157] In the seventh step, the similarity between the knowledge graph triple in the sixth step and the question-sentence graph is calculated, and the knowledge graph triple with the highest similarity is selected as the answer to the question and answer, thus realizing intelligent question and answer for outdoor substations.
[0158] The first specific embodiment of the intelligent question-answering robot used in the present invention for outdoor substations:
[0159] An intelligent question-answering robot for outdoor substations, comprising a knowledge base module and an intelligent question-answering module;
[0160] The knowledge base module is used to process and analyze unstructured substation equipment failure case text information, and after a series of unit processing in the module, a structured, easy-to-operate, and easy-to-use knowledge set is formed, which includes a substation equipment failure case text information input unit, an entity extraction unit, a relationship extraction unit, a knowledge storage unit, a relationship prediction unit, and a knowledge update unit;
[0161] A substation equipment failure case text information input unit, used to receive a substation equipment failure text data set input from outside the system;
[0162] The entity extraction unit is used to extract information elements such as equipment name, location, quantifier, time, equipment sub-components, equipment status, etc. from the substation equipment fault text data set;
[0163] The relation extraction unit is used to determine whether there is a relationship between entities in the text and what kind of relationship exists, that is, to extract relational feature words such as belong to, located in, discovered, produced, and triggered between entities;
[0164] Knowledge storage unit, used to store entity and relationship information extracted from text to form a knowledge graph;
[0165] The relationship prediction unit is used to mine hidden relationships between entities and further improve the existing knowledge graph;
[0166] The knowledge update unit is used to subsequently update new nodes: entities and edges: relationships on the existing knowledge graph;
[0167] The intelligent question-answering module is used to process and analyze the questions input by grassroots personnel and feedback relevant answers, and includes a question input unit, a question intention recognition unit, a slot filling unit, a subgraph matching unit, an answer sorting unit, and an answer encapsulation output unit;
[0168] A question input unit, used for receiving natural language questions input from outside the system;
[0169] The intention recognition unit is used to analyze the inquiry intentions of grassroots personnel and classify them according to inquiries about fault locations, fault causes, and recommended measures;
[0170] The slot filling unit is used to extract the device name, location, quantifier, time, device sub-component, and device status information elements in the question sentence;
[0171] The subgraph matching unit is used to perform subgraph matching between the constructed question-sentence graph and the knowledge graph in the knowledge base, and return the matching knowledge graph triples;
[0172] The answer ranking unit is used to select the subgraph matching results with the highest similarity;
[0173] The answer encapsulation output unit is used to encapsulate the subgraph matching results with similarity.
[0174] like Figure 1 As shown, a specific embodiment of the knowledge base module of the present invention:
[0175] The knowledge base module is used to process and analyze unstructured text information of substation equipment failure cases. After a series of unit processing within the module, a structured, easy-to-operate and easy-to-use knowledge set is formed.
[0176] The substation equipment failure case text information input unit is used to receive the substation equipment failure text data set entered from the outside of the system. The entity extraction unit is used to extract information elements such as equipment name, location, quantifier, time, equipment subcomponents, equipment status, etc. in the substation equipment failure text data set. The relationship extraction unit is used to determine whether there is a relationship between entities in the text and what kind of relationship exists, that is, to extract relationship feature words such as belong, located, discovered, generated, and triggered between entities. The knowledge storage unit is used to store the entity and relationship information extracted from the text to form a knowledge graph. The relationship prediction unit is used to mine the hidden relationships between entities and further improve the existing knowledge graph. The knowledge update unit is used to subsequently update new nodes (entities) and edges (relationships) on the existing knowledge graph.
[0177] The method for constructing a knowledge base module specifically includes the following steps:
[0178] S1: First, the robot is fed with the existing historical failure case texts of the substation equipment through the failure case text information input unit of the substation equipment, and after the text is input, it is transmitted to the entity extraction unit.
[0179] S2: If knowledge extraction is done based on the Chinese word segmentation tool, word segmentation errors will occur, which will directly affect the prediction of entity boundaries and cause knowledge extraction to not match. If knowledge extraction is done based on characters, inter-vocabulary information will be lost. Therefore, in the present invention, the robot adopts the entity recognition BERT+FLAT model, and entity extraction is done based on the method of integrating character into vocabulary information, which will not affect the extraction result due to word segmentation errors, and can comprehensively utilize the information of characters and words, and finally obtain preliminary entity extraction results. When the text is manually input, mainly through pinyin and voice input, a large number of homophonic typos will appear. Therefore, according to the preliminary entity extraction results, it is constructed after de-duplication and manual screening of the entity pinyin homonym library, and the preliminary entity extraction results are all converted into Chinese character pinyin forms with tones using KTestpinyin4.8, and finally the final entity extraction results are obtained according to the entity pinyin homonym matching.
[0180] like Figure 2 As shown in the figure, the entity extraction unit of the robot first vectorizes the text transmitted through the input unit. In order to extract the vocabulary information in the sentence, it is necessary to generate candidate words in the training sentence, and use the BERT pre-trained word vector to generate a word tree. Starting from the head of each sentence, all words under the corresponding characters are matched according to the maximum matching principle, and the word information is directly placed at the end of the sentence. In BERT, the maximum text length that can be processed is 512, and each token is represented as a 768-dimensional vector. Using a single character and word as a token, the identity token-id conversion is performed, and a (512,768) vector matrix representation is obtained through id mapping.
[0181] Then enter the position encoding layer, which is divided into absolute position encoding and relative position encoding. The absolute position encoding is obtained according to the order of the text sequence. The relative position has 4 types of position encoding. Each character has two identical head and tail position encodings, and each word has two different head and tail position encodings. Four relative distances can be calculated between every two nodes:
[0182]
[0183]
[0184]
[0185]
[0186] In the formula head [ i ]、 tail [ i ] represent the head and tail codes of a character or word respectively, Representative i The first character or word and the j The distance between the beginning of the word or word, and the rest , , The meaning is similar.
[0187] The calculation results are fused through the fully connected layer to obtain the final relative position encoding:
[0188]
[0189]
[0190]
[0191] In the formula represents the learnable parameters, represents the concatenation operator, d yes , , , and k express ij Dimension index for position encoding.
[0192] Then use the multi-head self-attention mechanism to extract text information features:
[0193]
[0194]
[0195]
[0196]
[0197] In the formula , , , and u, These are all learnable parameters in the pre-trained model BERT.
[0198] Finally, the feature information extracted above X ={ X 1 ,X 2 ,X 3 ,...,X n} Input to the CRF layer, predict and output text sequence with entity labels Y ={ Y 1 ,Y 2 ,Y 3 ,...,Y n}, get the preliminary entity information of the substation equipment failure case, and then build an entity pinyin homonym library based on the preliminary entity extraction results after deduplication and manual screening. Use KTestpinyin4.8 to convert all the preliminary entity extraction results into Chinese pinyin forms with tones. Finally, get the final entity extraction results based on entity pinyin homonym matching and correction feedback.
[0199] S3: Figure 3 As shown in the figure, the robot inputs the text sequence with entity tags into the relation extraction unit. The present invention adopts the BiGRU-Attention model for relation extraction. First, the text is vectorized. Each text is converted into an index vector of the same length after vectorization. Each index corresponds to a word vector. c ijt The text vector is then transferred to the Bi-GRU neural network layer for semantic encoding to obtain the semantic encoding vector h ijt , the calculation formula is as follows:
[0200]
[0201] In the formula c ijt Represents the word vector of the t-th word in the j-th sentence input at time i. Bi-GRU includes forward and backward GRU, which can fully learn the contextual relationship.
[0202] Different word vectors contribute differently to text features. Some words are critical to determining the relationship between entities in the text, while others are not. This paper introduces the Attention mechanism to assign different weights to each word vector. Finally, the relationship classification result is obtained through the softmax classifier.
[0203] S4: The robot transmits the extraction results of the entity extraction unit and the relationship extraction unit to the knowledge storage unit. Neo4j is a graph database with flexible storage, index-free adjacency attributes, ACID support, and can provide integrated query and display. It can guarantee zero latency for data of different quantities and depths. The present invention uses the Neo4j graph database to store the extracted entities and relationships as knowledge, forming a knowledge graph of substation equipment failures composed of nodes (entities) and edges (relationships), such as Figure 4 As shown, the entire atlas data is stored in the hard disk of the robot system.
[0204] S5: Each entity pair extracted in the entity extraction unit and the relationship between them can form a triple< h , r , t >, respectively representing <head entity, relationship, tail entity>. The relationship type is manually defined and annotated. Faced with a huge amount of data, the coverage of the defined relationship type is likely to be incomplete. For example: (main transformer, including, cooler system) and (cooler system, including, fan). In fact, there is also an inclusion relationship between the main transformer and the fan. In the present invention, the robot uses the relationship prediction unit to complete the path between the main transformer and the fan, and finds potential directed edges (relationships) for different nodes (entities) in the existing knowledge graph.
[0205] like Figure 5 As shown, the relationship prediction unit of the intelligent question-answering robot designed and developed by the present invention adopts the TransH model, see Figure 6 , for each relationship r , both give a hyperplane W r ,exist W r Define the relationship vector on the hyperplane d r , and then the original head entity h and tail entity t Mapped onto the hyperplane is h r , t r .
[0206] The triples are required to satisfy hr + d r = t r , for triples with the same head entity and relation ( h , r , t 1 )and( h , r , t 2 ) Through the relationship r The hyperplane mappings are:
[0207] h+r=t 1
[0208] h+r=t 2
[0209] In this way, t 1 and t 2 The problems that cannot be represented simultaneously on the TransE model are solved, that is, the one-to-many, many-to-one, and reflexive problems are solved.
[0210] S6: There are various types of equipment failures at the substation site, and the knowledge base needs to be continuously updated. Therefore, the present invention adds a knowledge update unit, and there is no need to retrain all text data. When there are new failure cases or new knowledge to be supplemented, the inspection personnel only need to input new knowledge texts, and in the knowledge update unit, after the knowledge extraction process described in S2 and S3, new points (entities) and edges (relationships) can be added to the original graph database to form a new knowledge base, so as to realize the automatic update of the knowledge base when the system is in service, continuously expand the knowledge of the system, and improve the system's question-answering ability.
[0211] like Figure 7 As shown, a specific embodiment of the intelligent question-answering module of the present invention is:
[0212] The intelligent question-answering module is used to process and analyze questions input by grassroots personnel and feedback the most relevant answers. The question input unit is used to receive natural language questions input from outside the system. The intention recognition unit is used to analyze the inquiry intentions of grassroots personnel. In the present invention, the robot is classified according to the inquiry intentions such as fault location inquiry, fault cause inquiry, and recommended measures inquiry. The slot filling unit is used to extract information elements such as equipment name, location, quantifier, time, equipment sub-components, equipment status, etc. in the question. The subgraph matching unit is used to perform subgraph matching on the constructed question sentence graph with the knowledge graph in the knowledge base, and return the knowledge graph triples that meet the requirements. The answer sorting unit is used to screen the subgraph matching results with the highest similarity. The answer encapsulation output unit is used to encapsulate the subgraph matching results with the highest similarity and output them to the display interface of the robot.
[0213] The construction method of the intelligent question-answering module is as follows:
[0214] S1: First, the grassroots staff inputs the question to the robot through the robot's question input unit, and after the question is input, it is transmitted to the question intention recognition unit.
[0215] S2: Figure 8 As shown in the figure, the intention recognition unit first vectorizes the question text, then loads the BERT pre-trained model parameters, optimizes the word vector through the BERT pre-trained model, and then performs semantic encoding through the Bi-LSTM neural network layer to obtain the semantic encoding vector. The weight of each vector is calculated through the Attention layer, and then all vectors are weighted summed to obtain the feature vector. Finally, the inquiry intention recognition result of the grassroots personnel is obtained through softmax classification.
[0216] S3: The slot filling process is essentially entity extraction, so the operation process of the slot filling unit of the robot in the present invention is the same as that described in the entity extraction unit of the knowledge base module.
[0217] S4: The subgraph matching unit of the robot in the present invention transmits the query intent and entity according to the intent recognition unit and the slot filling unit to construct the question sentence graph G 1 , using the VF2 search tree model and the knowledge graph in the knowledge base G 2 Perform subgraph matching. G 1 =( E 1 , R 1 ), G 2 =( E 2 , R 2 ), ERepresents a collection of nodes (entities), R Represents a set of edges (relationships). Each node (state) in the VF2 model is s , s Contains a series of G 1 and G 2 Mapping of midpoints (entities)<key,value> , key represents G 1 The node in the value represents G 2 As the height of the search tree increases, the number of key-value pairs in s will gradually increase. s All node pairs in the state satisfy the following constraints, which is called this state s To a consistent state.
[0218]
[0219]
[0220]
[0221]
[0222]
[0223]
[0224]
[0225] M E1XE2 represents the corresponding relationship set of nodes in the two graphs. If the node ,but Denotes G corresponding to node u 2 Nodes in ,but Denotes G corresponding to node v 1 Nodes in .
[0226] S5: In the answer sorting unit of the robot, the similarity between all triple sets returned in the subgraph matching unit and the question sentence graph is calculated based on the similarity calculation formula, and the triple set with the highest similarity is selected and sent to the answer encapsulation unit. The similarity calculation is as follows:
[0227]
[0228] In the formulaA and B Represent two subgraph sets respectively. Jaccard The output range of the similarity value is [0,1]. The larger the value, the higher the similarity between the two subgraphs.
[0229] S6: The output result of the answer sorting unit is packaged by the answer packaging unit and output to the display interface of the robot.
[0230] like Fig. 9 As shown, the second specific embodiment of the intelligent question-answering robot of the present invention is placed in an outdoor substation:
[0231] The hardware configuration of the system of the present invention includes an embedded keyboard, a 12.9-inch touch display screen with a resolution of 2388x1668, and mobile robot wheels. The internal system of the robot includes an iFlytek speech recognition module, a knowledge base module, an intelligent question and answer module, an ultrasonic ranging sensor, an Arduino MEGA128 controller, a dual H-bridge motor driver, a DC motor, and a power module.
[0232] When grassroots personnel are dealing with equipment failure incidents at outdoor substations, they can choose manual input or voice input to describe the fault phenomenon according to the actual situation. After the internal system of the intelligent question-answering robot operates, a picture of the faulty equipment is returned on the display screen and the specific parts are marked and circled. A list of similar historical failure cases is displayed on the right side of the picture. You can click to view the details of similar failure cases for grassroots personnel to make comprehensive judgments. iFlytek's speech recognition module is used for speech recognition, the ultrasonic ranging sensor is used to prevent the robot from hitting obstacles when moving, the DC motor powers the mobile robot, the dual H-bridge motor driver is used to drive the DC motor and controller, and the power module is used to power the entire robot. The knowledge base module includes six units: substation equipment failure case text information input unit, entity extraction unit, relationship extraction unit, knowledge storage unit, relationship prediction unit, and knowledge update unit.
[0233] The intelligent question-answering module consists of six units: question input unit, question intention recognition unit, slot filling unit, subgraph matching unit, answer sorting unit, and answer encapsulation output unit.
[0234] At the same time, considering the situation of placing the intelligent question-answering robot in an outdoor substation, the anti-electromagnetic interference performance is also increased. The following will briefly describe the intelligent question-answering robot in three parts: the knowledge base module, the intelligent question-answering module, and the external performance of the robot.
[0235] A specific embodiment of the external structure of the intelligent question-answering robot of the present invention:
[0236] Considering that the robot is placed in a special scenario such as an outdoor substation, the present invention adds anti-electromagnetic interference and waterproof capabilities to the robot to ensure the normal operation of the robot in the station.
[0237] Anti-electromagnetic interference
[0238] The robot designed and studied by the present invention is powered by batteries, and the batteries are sealed with metal parts. In addition, each circuit module inside the robot is encapsulated with metal parts with electromagnetic shielding function, and the cables connecting each module are also inserted into a metal shielding sleeve for protection. Silver-copper conductive paint paper with an inert metal plated on the surface is used to cover the connections and gaps of each structural part to strictly ensure the electromagnetic shielding effect.
[0239] water proof
[0240] The whole body of the robot designed and studied by the present invention is covered with a Weirs waterproof pressure relief membrane, which is breathable and waterproof. When it rains outdoors, rainwater can be prevented from entering the body, so that the internal circuit is not damaged and the normal use of the robot is not affected.
[0241] Explanation of terms in this application:
[0242] Intelligent question answering system
[0243] The intelligent question-and-answer system uses a question-and-answer format to accurately locate the knowledge that website users need, and provides personalized information services to website users by interacting with them.
[0244] entity
[0245] An entity is something that exists objectively and can be distinguished from each other. In terms of databases, an entity often refers to a collection of a certain type of thing, and the individual of each type of data object is called an entity.
[0246] relation
[0247] The state of interaction and mutual influence between entities.
[0248] Deep Neural Networks
[0249] Deep neural networks are a technique in the field of machine learning.
[0250] BERT
[0251] The full name is Bidirectional Encoder Representation from Transformers, which refers to the bidirectional encoding representation based on transformer. It is a pre-training model proposed by Google AI Research Institute in October 2018.
[0252] BiLSTM
[0253] BiLSTM stands for Bidirectional Long Short Term Memory, which refers to a bidirectional long short-term memory artificial neural network with an attention mechanism. It is suitable for scenarios with higher complexity and can better capture long-term dependencies in text.
[0254] CRF
[0255] The full name is conditional random field, which is a discriminant probability model that is often used to label or analyze sequence data.
[0256] Elasticsearch
[0257] Elasticsearch is a popular enterprise-level search engine developed in Java and released as open source under the terms of the Apache license.
[0258] Knowledge Graph
[0259] Knowledge graph is called knowledge domain visualization or knowledge domain mapping map in the library and information industry. It is a series of various graphics that show the development process and structural relationship of knowledge. It uses visualization technology to describe knowledge resources and their carriers, and to mine, analyze, construct, draw and display knowledge and their interrelationships.
[0260] Triple
[0261] Usually consists of entity-relationship-entity.
[0262] Slot
[0263] Well-defined properties of an entity.
[0264] Slot Filling
[0265] Extract values of well-defined attributes of a given entity from a large corpus.
[0266] ACID
[0267] Refers to the four characteristics that a database management system (DBMS) must have in order to ensure that transactions are correct and reliable during the process of writing or updating data: atomicity (also known as indivisibility), consistency (consistency), isolation (also known as independence), and durability.
[0268] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.
[0269] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0270] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.
Claims
1. An intelligent question-answering method for outdoor substations, It is characterized in that The following steps are involved: The first step is to receive natural language questions from external input; The second step is to identify the intent of the natural language questions in the first step, analyze the questioner's inquiry intention, and classify them into inquiries about fault location, fault cause, and recommended measures; The third step is to build an entity recognition BERT+FLAT model based on the intent recognition results in the second step, extract device entity elements in natural language questions, and fill in the slots of natural language questions. The entity recognition BERT+FLAT model performs preliminary extraction of device entity elements based on a method of integrating characters into vocabulary information to obtain preliminary entity extraction results; After removing duplicates and filtering the preliminary entity extraction results, an entity pinyin homonym database is constructed; Then convert the entity pinyin homonym database into Chinese character pinyin data with tones; The final entity extraction result is obtained by matching the Chinese character pinyin data; At the same time, according to the relative position coding information of the natural language question, the blank of the absolute position coding information of the natural language question is filled to realize the slot filling of the natural language question; The device entity element at least includes a name and / or a position and / or a quantifier and / or a time and / or a device subcomponent and / or device status information; The fourth step is to determine whether there is a relationship between the device entity elements in the third step and what kind of relationship exists, and obtain the relationship information between the device entity elements; Step 5: Based on the relationship information in step 4, the hidden relationships between the device entity elements are mined to obtain a question sentence graph; Step 6: Build a knowledge graph model to perform subgraph matching on the question-sentence graph in step 5 to obtain knowledge graph triples. Step 7: Calculate the similarity between the knowledge graph triples in step 6 and the question-sentence graph, select the knowledge graph triples with the highest similarity as the answer to the question, and realize intelligent question-answering for outdoor substations. In the third step, the method for constructing the entity recognition BERT+FLAT model is as follows: Step 31, vectorizing the inputted substation equipment historical fault case text to obtain a case word vector; Step 32, using the pre-trained model BERT training case word vector to generate a word tree, starting from the head of each sentence, matching all words under the corresponding characters according to the maximum matching principle, and placing the word information directly at the end of the sentence to obtain the training text; In the pre-trained model BERT, the maximum text length that can be processed is ZD, and each token is represented as a BS-dimensional vector; The identification token is a single character or word, which is converted into an identification token-id and mapped to an address id to obtain a vector matrix representation of (ZD, BS); Step 33, inputting the training text in step 32 into the position encoding layer for encoding, the encoding is divided into absolute position encoding and relative position encoding; The absolute position encoding is obtained according to the order of the training text sequence; Relative position coding has four types of position coding. Each character has two identical head and tail position codings, and each word has two different head and tail position codings. Four relative distances are calculated between every two nodes. The specific calculation formula is as follows: In the formula head [ i ]、 tail [ i ] represent the head and tail codes of a character or word respectively, , , , Respectively represent i The first character or word and the j The distance between the heads of individual characters or words; The calculation results are fused through the fully connected layer to obtain the final relative position encoding, and the calculation formula is as follows: In the formula represents the learnable parameters, represents the concatenation operator, d yes , , , and k express ij Dimension index of position encoding; Step 34, using the multi-head self-attention mechanism, combined with the relative position encoding in step 33, extracts text information features. The calculation formula of text information features is as follows: In the formula , , , and u, These are all learnable parameters in the pre-trained model BERT; Step 35: Input the text information features in step 34 into the discriminative probability model CRF layer to predict and output the text sequence with entity labels. Y ={ Y 1 ,Y 2 ,Y 3 ,...,Y n }, and obtain the preliminary entity extraction results of substation equipment failure cases; In the fourth step, the method for obtaining the relationship information is as follows: Step 41, vectorizing the device entity elements to convert them into index vectors of the same length; Step 42: The word vector corresponding to the index vector in step 41 c ijt Transmitted to the neural network layer Bi-GRU for semantic encoding to obtain the semantic encoding vector h ijt ; Step 43: convert the semantic encoding vector in step 42 h ijt Input into the logistic regression softmax classifier to obtain the relationship classification result; Semantic Encoding Vector h ijt The calculation formula is as follows: In the formula c ijt Represents the word vector of the tth word in the jth sentence input at time i. Bi-GRU is a neural network function, which includes forward and backward neural network GRU, which can fully learn the context relationship; In the sixth step, the method for constructing the knowledge graph model is as follows: Step 61, vectorizing the device entity elements to obtain index vectors with the same length; Step 62: The word vector corresponding to the index vector in step 61 c ijt Transmitted to the neural network layer Bi-GRU for semantic encoding to obtain the semantic encoding vector h ijt ; Step 63: convert the semantic encoding vector in step 62 h ijt Input into the logistic regression softmax classifier to obtain the relationship classification result; Step 64: Based on the relationship classification result in step 63, the extracted equipment entity elements and relationships are stored as knowledge using the graph database Neo4j to form a knowledge graph about substation equipment failures consisting of nodes: equipment entity elements and edges: relationships; Each device entity element pair and the relationship between them in the knowledge graph form a triple< h , r , t >,<h,r,t> Respectively represent <head entity, relationship, tail entity>; Step 65, using the relationship prediction model to complete the paths of different device entity elements, and find potential directed edges for different nodes in the knowledge graph in step 64; in the sixth step, the subgraph matching method is as follows: The question-sentence graph uses the search tree model in the graph-speaking subgraph isomorphism algorithm VF2 to match the subgraph with the knowledge graph. The matching relationship is as follows: G1=(E1,R1),G2=(E2,R2),where G1 is the question sentence graph, G2 is the knowledge graph, E is the set of nodes, i.e., equipment entity elements, and R is the set of edges, i.e., relationships; Each node in the subgraph isomorphism algorithm VF2 is s. s contains a series of mappings of nodes in G1 and G2<key,value> , Among them, key represents the node in G1, and value represents the node in G2; As the height of the search tree increases, the number of key-value pairs in s will gradually increase; When all node pairs in s satisfy the constraint formula, the state s is called a consistent state, and subgraph matching is completed; The constraint formula is as follows: M E1XE2 represents the corresponding relationship set of nodes in the question sentence graph and knowledge graph. ,but represents the node in G2 corresponding to node u; if node ,but Denotes the node in G1 corresponding to node v.
2. An intelligent question-answering method for outdoor substations according to claim 1, characterized in that: In the second step, the method of intent recognition is as follows: Step 21, vectorize the natural language question to obtain a word vector; Step 22, construct a pre-trained model BERT to optimize the word vector in step 21 to obtain an optimized word vector; Step 23, constructing a neural network layer Bi-LSTM to semantically encode the optimized word vector in step 22 to obtain a semantic encoding vector; Step 24, calculating the weight of the semantic encoding vector in step 23 through the attention layer Attention; Step 25, multiply all semantic coding vectors by corresponding weights and sum them up to obtain a feature vector; Step 26, input the feature vector in step 25 into the logistic regression function softmax to obtain the questioning intention recognition result of the questioner.
3. The intelligent question-answering method for outdoor substation according to claim 1, characterized in that: The completion method of the relationship prediction model is as follows: Step 651, using the vector translation algorithm TransH, for each relationship r , giving a hyperplane W r ; Step 652, calculate the relationship vector on the hyperplane Wr in step 651 d r ; Step 653, according to the relationship vector in step 652 d r , the original head entity h and tail entity t Mapped onto the hyperplane is h r , t r ; Step 654, requiring the triple to satisfy h r + d r = t r , for triples with the same head entity and relation ( h , r , t 1) and ( h , r , t 2) Through relationships r The hyperplane mappings are: h+r=t 1 h+r=t 2; Implement path completion for different device entity elements.
4. An intelligent question-answering method for an outdoor substation according to any one of claims 1 to 3, characterized in that: In the seventh step, the similarity calculation formula is as follows: In the formula A and B They represent the knowledge graph triple and question sentence graph respectively. The output interval of the similarity value is [0,1]. The larger the value, the higher the similarity between the knowledge graph triple and the question sentence graph.
5. An intelligent question-answering robot placed in an outdoor substation, characterized in that: Applying an intelligent question-answering method for an outdoor substation as described in any one of claims 1 to 4; It includes a knowledge base module, an intelligent question-answering module, a battery, several circuit modules, and a body; The knowledge base module is used to process and analyze unstructured substation equipment failure case text information, and after a series of unit processing in the module, a structured, easy-to-operate, and easy-to-use knowledge set is formed, which includes a substation equipment failure case text information input unit, an entity extraction unit, a relationship extraction unit, a knowledge storage unit, a relationship prediction unit, and a knowledge update unit; A substation equipment failure case text information input unit, used to receive a substation equipment failure text data set input from outside the system; The entity extraction unit is built based on the entity recognition BERT+FLAT model, which is used to extract the equipment name, location, quantifier, time, equipment sub-components, and equipment status information elements in the substation equipment fault text dataset; The relation extraction unit is used to determine whether there is a relationship between entities in the text and what kind of relationship exists; Knowledge storage unit, used to store entity and relationship information extracted from text to form a knowledge graph; The relationship prediction unit is used to mine hidden relationships between entities and improve the existing knowledge graph; The knowledge update unit is used to subsequently update new nodes: entities and edges: relationships on the existing knowledge graph; The intelligent question-answering module is used to process and analyze the questions input by grassroots personnel and feedback relevant answers, and includes a question input unit, a question intention recognition unit, a slot filling unit, a subgraph matching unit, an answer sorting unit, and an answer encapsulation output unit; A question input unit, used for receiving natural language questions input from outside the system; The intention recognition unit is used to analyze the inquiry intentions of grassroots personnel and classify them according to inquiries about fault locations, fault causes, and recommended measures; The slot filling unit is built based on the entity recognition BERT+FLAT model, which is used to extract the device name, location, quantifier, time, device sub-components, and device status information elements in the question; The subgraph matching unit is used to perform subgraph matching between the constructed question-sentence graph and the knowledge graph in the knowledge base, and return the matching knowledge graph triples; The answer ranking unit is used to select the subgraph matching results with the highest similarity; The answer encapsulation output unit is used to encapsulate the subgraph matching results with similarity; The battery is sealed with a metal member; The circuit module is encapsulated with metal parts with electromagnetic shielding function, and the cables connecting each circuit module are also inserted into a metal shielding sleeve for protection; and the connections and gaps of each component are covered with silver-copper conductive paint paper with an inert metal plated on the surface; the entire outer surface of the fuselage is covered with a Micros waterproof pressure relief film.
