A power semantic knowledge graph system and method

Abstract

A power semantic knowledge graph system and method are disclosed. A smart question-answering system is constructed by using natural language processing technology based on deep learning and combining it with the knowledge graph in the power field to optimize the power audit workflow and enhance audit efficiency.

Description

Technical Field

[0001] The present application relates to the field of electric power, and more specifically, to an electric power semantic knowledge graph system and method. Background Art

[0002] The auditing process of the power system is complicated and involves many links. Audit data has the complex characteristics of non-vertical fields and large horizontal spans. Auditors need to understand various laws, regulations, rules and regulations, historical issues, etc. in real time. Manual methods generally use paper data query, Internet search, etc. These methods are difficult to quickly obtain relevant information, which greatly affects the efficiency of audit operations.

[0003] With the rapid development of artificial intelligence technology in recent years, big data-driven artificial intelligence technology is rapidly integrating with the development of power and related industries with unprecedented breadth and depth, becoming an urgent need to assist audit work. The intelligent question-and-answer system is designed to automatically respond to questions raised by users to the system. However, the existing question-and-answer system is more focused on casual conversations. The research on intelligent question-and-answer in other professional fields generally involves medical, education, e-commerce and other fields, and there is less related work in the audit field. For auditing complex business processes, the existing question-and-answer system cannot well meet the needs of power audit question-and-answer.

[0004] Therefore, an optimized solution is expected. Summary of the invention

[0005] In order to solve the above technical problems, this application is proposed. The embodiment of this application provides a power semantic knowledge graph system and method, which uses natural language processing technology based on deep learning and combines the power field knowledge graph to build an intelligent question-answering system to optimize the power audit workflow and enhance audit efficiency.

[0006] According to one aspect of the present application, a power semantic knowledge graph system is provided, which includes:

[0007] The problem acquisition module is used to obtain power audit problems;

[0008] An alternative entity extraction module is used to extract the semantic embedding representation of the first candidate entity from the power semantic knowledge graph;

[0009] A matching module, configured to analyze and process the semantic embedding representation of the power audit problem and the first candidate entity based on a deep convolutional neural network model to obtain an optimized semantic matching feature matrix; and

[0010] An output result generation module is used to determine whether to output the first candidate entity based on the optimized semantic matching feature matrix.

[0011] According to another aspect of the present application, a smart floor scrubber control method is provided, which includes:

[0012] Get power audit questions;

[0013] Extract the semantic embedding representation of the first candidate entity from the power semantic knowledge graph;

[0014] Analyzing and processing the semantic embedding representation of the power audit problem and the first candidate entity based on a deep convolutional neural network model to obtain an optimized semantic matching feature matrix; and

[0015] Based on the optimized semantic matching feature matrix, it is determined whether to output the first candidate entity.

[0016] According to an embodiment of the present disclosure, a natural language processing technology based on deep learning is used in combination with a knowledge graph in the power field to construct an intelligent question-answering system to optimize the power audit workflow and enhance audit efficiency.

[0017] Further features and aspects of the present disclosure will become apparent from the following detailed description of exemplary embodiments with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0019] Figure 1 is a block diagram of a power semantic knowledge graph system according to an embodiment of the present application;

[0020] Figure 2 A system architecture diagram of a power semantic knowledge graph system according to an embodiment of the present application;

[0021] Figure 3 A block diagram of a matching module in a power semantic knowledge graph system according to an embodiment of the present application;

[0022] Figure 4 A block diagram of a semantic analysis unit in a power semantic knowledge graph system according to an embodiment of the present application;

[0023] Figure 5 A block diagram of an output result generation module in a power semantic knowledge graph system according to an embodiment of the present application;

[0024] Figure 6 It is a flowchart of the power semantic knowledge graph method according to an embodiment of the present application. DETAILED DESCRIPTION

[0025] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0026] As shown in this application and claims, unless the context clearly indicates an exception, the words "a", "an", "an" and / or "the" do not refer to the singular and may also include the plural. Generally speaking, the terms "include" and "comprise" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.

[0027] Flowcharts are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed accurately in order. On the contrary, various steps may be processed in reverse order or simultaneously as required. Meanwhile, other operations may also be added to these processes, or a certain step or several steps of operations may be removed from these processes.

[0028] Various exemplary embodiments, features and aspects of the present disclosure will be described in detail below with reference to the accompanying drawings. The same reference numerals in the accompanying drawings represent elements with the same or similar functions. Although various aspects of the embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless otherwise specified.

[0029] The word “exemplary” is used exclusively herein to mean “serving as an example, example, or illustration.” Any embodiment described herein as “exemplary” is not necessarily to be construed as preferred or advantageous over other embodiments.

[0030] In addition, in order to better illustrate the present disclosure, numerous specific details are given in the following specific embodiments. It should be understood by those skilled in the art that the present disclosure can also be implemented without certain specific details. In some examples, methods, means, components and circuits well known to those skilled in the art are not described in detail in order to highlight the main purpose of the present disclosure.

[0031] The business process of power system auditing is complicated and involves many links. Among them, audit data has complex characteristics of non-vertical fields and large horizontal spans. Auditors need to understand various laws, regulations, rules and regulations, historical issues, etc. in real time. Manual methods generally use paper data query, Internet search, etc. These methods are difficult to quickly obtain relevant information, which greatly affects the efficiency of audit operations. Therefore, an optimized solution is expected.

[0032] Figure 1It is a block diagram of the power semantic knowledge graph system according to an embodiment of the present application. Figure 2 FIG. 1 is a system architecture diagram of the electric power semantic knowledge graph system according to an embodiment of the present application. Figure 1 and Figure 2 As shown, according to the electric power semantic knowledge graph system 300 of the embodiment of the present application, it includes: a question acquisition module 310, which is used to obtain electric power audit questions; an alternative entity extraction module 320, which is used to extract the semantic embedding representation of the first alternative entity from the electric power semantic knowledge graph; a matching module 330, which is used to analyze and process the electric power audit questions and the semantic embedding representation of the first alternative entity based on a deep convolutional neural network model to obtain an optimized semantic matching feature matrix; and an output result generation module 340, which is used to determine whether to output the first alternative entity based on the optimized semantic matching feature matrix.

[0033] In particular, the problem acquisition module 310 is used to acquire power audit problems. Power audit refers to the evaluation and analysis of the power usage of a building or equipment to determine whether energy consumption and costs can be reduced by improving energy usage and equipment.

[0034] In particular, in an example of the present application, the method for obtaining power audit questions can be implemented through a variety of ways, including the following: 1. User input: The user can directly input power audit questions into the intelligent question and answer system, and the system receives and obtains the power audit questions; 2. Text recognition: The intelligent question and answer system can use OCR technology to identify and extract text related to power audits, thereby obtaining power audit questions; 3. Voice recognition: The intelligent question and answer system can use voice recognition technology to recognize and convert user voice input, thereby obtaining power audit questions.

[0035] It is worth mentioning that in other specific examples of this application, electricity audit issues can also be obtained through other means, such as: finding a professional electricity audit company or agency that can provide professional electricity audit services and help you identify potential energy waste and improvement opportunities; looking for information about electricity audits on the website of the electricity audit company or agency to understand their service scope, charging standards, customer reviews and other aspects; contacting the customer service staff of the electricity audit company or agency to consult on electricity audit-related issues, such as how to conduct an electricity audit, what data and information needs to be provided, the cycle and cost of the electricity audit, etc.; preparing relevant data and information according to the requirements provided by the electricity audit company or agency, such as the electricity usage, energy consumption and cost of buildings or equipment; arranging electricity auditors to conduct on-site inspections of buildings or equipment, collect more detailed data and information, and provide improvement suggestions and plans.

[0036] In particular, the candidate entity extraction module 320 is used to extract the semantic embedding representation of the first candidate entity from the power semantic knowledge graph. Here, the knowledge graph is a method for representing and organizing knowledge based on a graph data structure.

[0037] Knowledge graph is a semi-structured knowledge representation method that uses entity-relationship diagrams to describe entities in the real world and the relationships between them. After extracting knowledge from text and databases, the knowledge graph performs processes such as concept classification, relationship extraction, and semantic analysis, and then transforms this knowledge into entities, attributes, and relationships, and organizes it into a large-scale graph database. Knowledge graphs have the following characteristics: 1. Easy to understand and expand: Knowledge graphs use edges and points to represent concepts and relationships, which is intuitive and easy to understand and expand; 2. Semantic richness: Knowledge graphs can improve machine understanding and application of data by standardizing and semanticizing entities and relationships; 3. Efficient reasoning mechanism: Knowledge graphs are built on graph databases, and graph-based algorithms can be used for reasoning and analysis.

[0038] Accordingly, in a possible implementation, the semantic embedding representation of the first candidate entity can be extracted from the electric power semantic knowledge graph by the following steps: determine the name or identifier of the first candidate entity. For example, assume that the name of the first candidate entity is "electric vehicle"; search for the entity node of "electric vehicle" in the electric power semantic knowledge graph. This can be achieved by using a graph query language (such as SPARQL) or a graph search engine (such as Neo4j); obtain all relevant attributes and relationships of "electric vehicle". These attributes and relationships may include the manufacturer, model, battery capacity, charging method, horsepower, etc. of "electric vehicle"; convert these attributes and relationships into vector representations. Some pre-trained natural language processing models (such as BERT, GPT, etc.) or graph neural network models (such as GCN, GAT, etc.) can be used to convert attributes and relationships into vector representations; combine all vector representations into a vector representing "electric vehicle". This can be achieved by simply concatenating all vectors together or using some aggregation functions (such as average, maximum, weighted average, etc.); the final vector obtained is the semantic embedding representation of "electric vehicle". It is important to note that the quality and accuracy of the semantic embedding representation depends on the model and algorithm used, as well as the completeness and accuracy of the extracted attributes and relations.

[0039] In particular, in a specific example of the present application, Figure 3As shown, the matching module 330 includes: a semantic analysis unit 331, which is used to perform semantic analysis on the power audit problem to obtain a multi-scale power audit problem semantic encoding feature vector; an association matching unit 332, which is used to perform association encoding on the multi-scale power audit problem semantic encoding feature vector and the semantic embedding representation of the first candidate entity to obtain a semantic matching feature matrix; and an expression effect optimization unit 333, which is used to optimize the expression effect of the text semantic association feature of the semantic matching feature matrix to obtain the optimized semantic matching feature matrix.

[0040] Specifically, the semantic analysis unit 331 is used to perform semantic analysis on the power audit problem to obtain a multi-scale power audit problem semantic encoding feature vector. In a specific example of the present application, Figure 4 As shown, the semantic analysis unit 331 includes: a word segmentation subunit 3311, used to perform word segmentation processing on the power audit problem to obtain a sequence of power audit problem description words; a word embedding subunit 3312, used to pass the sequence of power audit problem description words through a word embedding layer to obtain a sequence of power audit problem description word embedding vectors; and a natural language processing subunit 3313, used to pass the sequence of power audit problem description word embedding vectors through a dual-pipeline model including a first natural language processing model and a second natural language processing model to obtain the multi-scale power audit problem semantic encoding feature vector.

[0041] More specifically, the word segmentation subunit 3311 is used to perform word segmentation processing on the power audit problem to obtain a sequence of power audit problem description words. In the technical solution of the present application, the word segmentation processing on the power audit problem can decompose the power audit problem into words one by one, so that the problem becomes a sequence that is easier to handle.

[0042] Word segmentation is the process of dividing a text into individual words according to certain rules. Word segmentation can help computers better understand human language, because computers need to convert human language into a form that computers can understand and process, and word segmentation is an important step.

[0043] Accordingly, in a possible implementation, the power audit problem can be segmented to obtain a sequence of power audit problem description words through the following steps: Data preprocessing: Preprocess the power audit problem, including operations such as removing stop words and punctuation marks, to obtain pure text; Word segmentation processing: Use a word segmentation tool to perform word segmentation processing on the pure text, splitting the text into individual words; Part-of-speech tagging: Perform part-of-speech tagging on each word to determine its grammatical role in the sentence, such as noun, verb, adjective, etc.; Removing stop words: Remove stop words, such as words like "de", "le", "shi", etc., which have no practical meaning for the description of the power audit problem; Stemming processing: For some words, such as different tenses of verbs and different forms of nouns, stemming processing is required to restore them to their original forms, such as restoring "running" to "run"; Obtaining a sequence of power audit problem description words: Arrange the processed words in the order in which they appear in the sentence to obtain a sequence of power audit problem description words.

[0044] More specifically, the word embedding subunit 3312 is used to obtain a sequence of power audit problem description word embedding vectors by passing the sequence of power audit problem description words through a word embedding layer. It should be understood that the word embedding layer can transform the sequence of power audit problem description words into a sequence of power audit problem description word embedding vectors. Generally, this process is also called word vectorization. That is, the word embedding layer can map words into a high-dimensional vector space and transform them into vector form.

[0045] Accordingly, in a possible implementation, the sequence of power audit problem description words can be passed through a word embedding layer to obtain a sequence of power audit problem description word embedding vectors through the following steps: Prepare a word embedding matrix, where each row corresponds to the embedding vector of a word. This matrix can be pre-trained or trained on the sequence of power audit problem description words; Convert each word in the sequence of power audit problem description words into its corresponding embedding vector. This can be achieved by looking up the embedding vector of each word in the word embedding matrix; Concatenate all the embedding vectors in sequence order to obtain a sequence of power audit problem description word embedding vectors; Optionally, some preprocessing steps, such as normalization or truncation, can be applied to the embedding vector sequence to ensure that the embedding vectors have the same length and range; Finally, provide the sequence of power audit problem description word embedding vectors to a downstream task, such as a classification or regression model, for further analysis and prediction.

[0046] More specifically, the natural language processing subunit 3313 is used to pass the sequence of the electric power audit problem description word embedding vectors through a dual-pipeline model including a first natural language processing model and a second natural language processing model to obtain the multi-scale electric power audit problem semantic encoding feature vector. The sequence of the electric power audit problem description word embedding vectors is passed through a dual-pipeline model including a first natural language processing model and a second natural language processing model to obtain the multi-scale electric power audit problem semantic encoding feature vector. That is, the dual-pipeline model processing including the first natural language processing model and the second natural language processing model is used to more comprehensively and accurately express the semantic information of the electric power audit problem. The first natural language processing model and the second natural language processing model respectively perform different degrees of semantic understanding on the sequence of the electric power audit problem description word embedding vectors.

[0047] In an embodiment of the present application, the first natural language processing model is a recurrent neural network model (RNN), and the second natural language processing model is a long short-term memory network model (LSTM). Recurrent neural network (RNN) is mainly used for processing sequence data, such as speech, text, time series, etc. It has the ability to process variable-length sequence data and can learn correlations and patterns from sequences. There is a feedback connection between neurons in the recurrent neural network model, which can pass previous information to the current moment to form a dynamic neural network structure. The core of the RNN model is the recurrent unit (RU), which uses the input at the current moment, the output at the previous moment, and the state information at the previous moment to calculate the output and state at the current moment. The long short-term memory network (LSTM) is different from the traditional recurrent neural network model. Each neuron in the LSTM network has three gates: input gate, forget gate, and output gate. These gates allow the LSTM neural network to decide what information to input, what information needs to be retained, and when to output information. The LSTM network is usually composed of multiple LSTM cells, each of which has an internal state unit and three gates. By recursively connecting multiple LSTM cells, a deep LSTM network can be formed to learn and model the correlation and pattern in multi-time series data. Since the recursive neural network model and the long short-term memory network model learn the semantic association information under different receptive fields from the sequence of the word embedding vectors describing the power audit problem, that is, the sequence of the word embedding vectors describing the power audit problem is processed through a dual-pipeline model composed of the first natural language processing model and the second natural language processing model, and the sequence of the word embedding vectors describing the power audit problem can be converted into a multi-scale semantic encoding feature vector of the power audit problem, which can improve the model's understanding and classification capabilities of the power audit problem.

[0048] In a specific example of the present application, the natural language processing subunit 3313 includes: a first-scale natural language processing secondary subunit, which is used to pass the sequence of word embedding vectors describing the power audit problem through the first natural language processing model of the dual-pipeline model to obtain a first-scale power audit problem semantic coding feature vector; a first-scale natural language processing secondary subunit, which is used to pass the sequence of word embedding vectors describing the power audit problem through the second natural language processing model of the dual-pipeline model to obtain a second-scale power audit problem semantic coding feature vector; a multi-scale fusion secondary subunit, which is used to fuse the first-scale power audit problem semantic coding feature vector and the second-scale power audit problem semantic coding feature vector to obtain the multi-scale power audit problem semantic coding feature vector.

[0049] It is worth mentioning that in other specific examples of the present application, the sequence of word embedding vectors describing the power audit problem can also be passed through a dual-pipeline model including a first natural language processing model and a second natural language processing model to obtain the multi-scale power audit problem semantic encoding feature vector, for example: first, the power audit problem description is converted into a sequence of word embedding vectors, and a pre-trained word vector model, such as Word2Vec or GloVe model, can be used to map each word to a vector representation of a fixed length; then, the first natural language processing model, such as a convolutional neural network (CNN) or a long short-term memory network (LSTM), is used to process the word embedding vector sequence to extract local features. These local features can be features at the word, phrase or sentence level; then, a second natural language processing model, such as a recurrent neural network (RNN) or an attention mechanism network (Attention), is used to process the local features extracted by the first model to extract global features. These global features can be the semantic information of the entire power audit problem description; and then the semantic encoding feature vector of the audit problem. This multi-scale feature vector can be used for classification, clustering or other related tasks.

[0050] It is worth mentioning that in other specific examples of the present application, the power audit problem can also be semantically analyzed by other methods to obtain a multi-scale semantic encoding feature vector of the power audit problem, for example: data preprocessing: preprocess the power audit problem, including removing stop words, word segmentation, stemming and other operations to obtain a text sequence; semantic embedding representation: convert the text sequence into a vector representation, you can use a pre-trained word vector model (such as Word2Vec, GloVe, etc.) or input the text sequence into a deep learning model for training to obtain a semantic embedding representation of each word; multi-scale semantic analysis: use multiple different convolution kernels to perform convolution operations on the semantic embedding representation to obtain semantic feature maps of different scales. Convolution kernels of different sizes can be used to achieve multi-scale analysis; pooling operation: pooling operation is performed on the semantic feature map of each scale to obtain a feature vector of fixed size; feature fusion: feature vectors of different scales can be fused using simple weighted average or more complex attention mechanisms; normalization operation: normalize the fused feature vectors to make them have the same scale and range; obtain the semantic encoding feature vector of multi-scale power audit issues: concatenate the normalized feature vectors into a vector to obtain the semantic encoding feature vector of multi-scale power audit issues.

[0051] Specifically, the association matching unit 332 is used to perform association coding on the semantic encoding feature vector of the multi-scale power audit problem and the semantic embedding representation of the first candidate entity to obtain a semantic matching feature matrix. In the technical solution of the present application, after obtaining the vectorized semantic expression of the problem and the first candidate entity, in the technical solution of the present application, it is expected to calculate the matching degree and the correlation degree between the semantic representation of the problem and the semantic representation of the first candidate entity. That is, the mapping relationship between the semantic representation of the problem and the semantic representation of the first candidate entity is established by using association coding. In this way, the semantic matching feature matrix obtained by association coding can fully consider the mutual relationship between the semantic expression of the problem and the semantic expression of the first candidate entity, rather than being limited to the separate semantic expressions of each. More specifically, the semantic encoding feature vector of the multi-scale power audit problem and the semantic embedding representation of the first candidate entity are associated coded using the following association formula to obtain the semantic matching feature matrix; wherein, the formula is: Where V m represents the semantic encoding feature vector of the multi-scale power audit problem, represents the transposed vector of the semantic encoding feature vector of the multi-scale power audit problem, V n represents the semantic embedding representation of the first candidate entity, M represents the semantic matching feature matrix, Represents vector multiplication.

[0052] Accordingly, in a possible implementation, the semantic encoding feature vector of the multi-scale power audit problem and the semantic embedding representation of the first candidate entity can be associated and encoded to obtain a semantic matching feature matrix by the following steps: semantically encode the multi-scale power audit problem into a feature vector. This can be done by using natural language processing techniques, such as a bag-of-words model or a word embedding model, to convert the problem into a numerical vector; represent the semantic embedding of the first candidate entity as a feature vector. This can be done by using a word embedding model, such as Word2Vec or GloVe, to convert the entity into a numerical vector; associate the problem feature vector with the entity feature vector to obtain a semantic matching feature matrix. This can be done by using some methods, such as cosine similarity or dot product calculation; input the semantic matching feature matrix into a machine learning algorithm for classification or clustering to solve the multi-scale power audit problem.

[0053] Specifically, the expression effect optimization unit 333 is used to optimize the expression effect of the text semantic association features of the semantic matching feature matrix to obtain the optimized semantic matching feature matrix. In a specific example of the present application, the expression effect optimization unit 333 includes: a multi-source information fusion pre-test distribution evaluation optimization sub-unit, which is used to perform multi-source information fusion pre-test distribution evaluation optimization on each row feature vector of the semantic matching feature matrix to obtain multiple optimized row feature vectors; and an arrangement sub-unit, which is used to arrange the multiple optimized row feature vectors into the optimized semantic matching feature matrix.

[0054] More specifically, the multi-source information fusion pre-test distribution evaluation optimization subunit is used to perform multi-source information fusion pre-test distribution evaluation optimization on each row feature vector of the semantic matching feature matrix to obtain multiple optimized row feature vectors. In the technical solution of the present application, when the multi-scale power audit problem semantic encoding feature vector and the semantic embedding representation of the first candidate entity are associated and encoded to obtain the semantic matching feature matrix, the multi-scale power audit problem semantic encoding feature vector and the semantic embedding representation of the first candidate entity are associated position by position, so each row feature vector of the semantic matching feature matrix can be regarded as the associated feature vector of each eigenvalue of the multi-scale power audit problem semantic encoding feature vector and the semantic embedding representation of the first candidate entity as a whole, so the semantic matching feature matrix is ​​equivalent to the combined feature set of the local feature sets corresponding to each row feature vector. In addition, since the feature distribution of each row feature vector has the text semantic association feature distribution of the multi-scale power audit problem expressed by the multi-scale power audit problem semantic encoding feature vector, each row feature vector has a multi-source information association relationship corresponding to the graph semantic embedding association distribution information of the candidate entity in addition to the mutually associated neighborhood distribution relationship. Therefore, in order to improve the expression effect of the semantic matching feature matrix as a whole on the text semantic association features of different scales of the power audit problem, the applicant of this application performs a feature vector matching operation on each row, for example, V i Perform multi-source information fusion prior distribution evaluation optimization to obtain the optimized row feature vector V ′ i , specifically expressed as:

[0055]

[0056]

[0057] Where V i is the i-th row feature vector of the semantic matching feature matrix, V j is the j-th row feature vector of the semantic matching feature matrix, is the mean eigenvector, n is the neighborhood setting hyperparameter, log represents the logarithmic function value with base 2, Represents positional subtraction, V ′ iis the i-th optimized row feature vector of the optimized semantic matching feature matrix. Here, the optimization of the multi-source information fusion prior distribution evaluation can achieve the effective conversion of the prior information of each feature vector on the local synthetic distribution for a local feature set composed of multiple interrelated neighborhood parts based on the quasi-maximum likelihood estimation of the robustness of feature distribution fusion, and obtain the optimization paradigm that can be used to evaluate the standard expected fusion information between the internal correlation within the set and the change relationship between sets through the prior distribution construction in the multi-source case, so as to improve the information expression effect of the feature vector based on the multi-source information correlation fusion. Therefore, the optimized row feature vector V ′ i Arranging them into the semantic matching feature matrix can improve the expression effect of the semantic matching feature matrix as a whole on the text semantic association features of different scales of power audit issues.

[0058] More specifically, the arrangement subunit is used to arrange the multiple optimized row feature vectors into the optimized semantic matching feature matrix. That is, after obtaining the multiple optimized row feature vectors, they are further arranged in two dimensions to obtain the optimized semantic matching feature matrix.

[0059] Accordingly, in a possible implementation, the multiple optimized row feature vectors can be arranged into the optimized semantic matching feature matrix by the following steps: Arrange each optimized row feature vector in a certain order to form a matrix. This order can be determined according to factors such as feature importance and dimension after dimensionality reduction; Standardize each optimized row feature vector according to certain rules to ensure that they are on the same scale; Arrange all standardized optimized row feature vectors together to form an optimized semantic matching feature matrix.

[0060] It is worth mentioning that in other specific examples of the present application, the semantic matching feature matrix can also be optimized by other methods to obtain the optimized semantic matching feature matrix, for example: first, the semantic matching feature matrix is ​​preprocessed, including removing stop words, stem extraction, part-of-speech tagging and other operations. These preprocessing steps can improve the expressiveness and discrimination of the feature vector; then, the semantic matching feature matrix is ​​optimized using some expressions of text semantic association features, for example, using the TF-IDF method to weight word vectors, or using some text similarity calculation methods, such as edit distance, Jaccard similarity, etc.; then, some dimensionality reduction methods, such as principal component analysis (PCA) or linear discriminant analysis (LDA), are used to reduce the dimension of the feature vector to improve the training speed and generalization ability of the model; and then, some model selection and parameter adjustment methods, such as cross-validation, grid search, etc., are used to select the optimal machine learning model and parameters to improve the performance and robustness of the model.

[0061] It is worth mentioning that in other specific examples of the present application, the semantic embedding representation of the power audit problem and the first candidate entity can also be analyzed and processed based on the deep convolutional neural network model in other ways to obtain an optimized semantic matching feature matrix, for example: data preprocessing: preprocessing the text data of the power audit problem and the first candidate entity, including removing stop words, word segmentation, stemming and other operations to obtain a text sequence; semantic embedding representation: converting the text sequence into a vector representation, you can use a pre-trained word vector model (such as Word2Vec, GloVe, etc.) or input the text sequence into a deep learning model for training to obtain the semantic embedding representation of each word; constructing a deep convolutional neural network model: construct an appropriate deep convolutional neural network model according to task requirements. The convolutional neural network model, including convolutional layers, pooling layers, fully connected layers, etc., is used to extract features from the semantic embedding representation; the deep convolutional neural network model is trained using a labeled data set, and the model parameters are adjusted to enable it to better extract semantic features; feature extraction: the trained deep convolutional neural network model is used to extract features from the semantic embedding representation of the power audit problem and the first candidate entity to obtain an optimized semantic matching feature matrix; feature matrix classification: the optimized semantic matching feature matrix is ​​input into the machine learning algorithm for classification to determine the degree of semantic matching between the power audit problem and the first candidate entity; result analysis: the result analysis is performed based on the classification results to obtain the degree of semantic matching between the power audit problem and the first candidate entity, providing a reference for subsequent decision-making.

[0062] In particular, the output result generation module 340 is used to determine whether to output the first candidate entity based on the optimized semantic matching feature matrix. Figure 5As shown, the output result generation module 340 includes: a classification unit 341, used to pass the optimized semantic matching feature matrix through a classifier to obtain a classification result, and the classification result is used to indicate whether the probability that the first candidate entity is the entity most relevant to the power audit issue exceeds a predetermined threshold; and an output unit 342, used to determine whether to output the first candidate entity based on the classification result.

[0063] Specifically, the classification unit 341 is used to pass the optimized semantic matching feature matrix through a classifier to obtain a classification result, and the classification result is used to indicate whether the probability that the first candidate entity is the entity most relevant to the power audit issue exceeds a predetermined threshold. That is, the semantic matching feature matrix is ​​passed through a classifier to obtain a classification result, and the classification result is used to indicate whether the probability that the first candidate entity is the entity most relevant to the power audit issue exceeds a predetermined threshold. In a specific example of the present application, the classification unit 341 includes: an expansion subunit, used to expand the optimized semantic matching feature matrix into a classification feature vector based on a row vector or a column vector; a fully connected encoding subunit, used to use multiple fully connected layers of the classifier to perform fully connected encoding on the classification feature vector to obtain an encoded classification feature vector; and a classification result generation subunit, used to pass the encoded classification feature vector through the Softmax classification function of the classifier to obtain the classification result.

[0064] The classifier can learn a classification rule based on the relationship between the feature matrix and the classification label in the training data, which is used to classify and predict the semantic matching feature matrix input during inference, thereby obtaining the classification result. It is worth mentioning that the classifier obtains a probability, that is, the classifier calculates the maximum value of the probability value of "the probability that the first candidate entity is the entity most relevant to the power audit problem exceeds the predetermined threshold" and the probability value of "the probability that the first candidate entity is the entity most relevant to the power audit problem does not exceed the predetermined threshold".

[0065] A classifier is a machine learning model that is used to classify input data into different categories or labels. It can be a simple set of rules or a complex mathematical model. Classifiers usually use a known training data set to learn how to classify new data into known categories. In practical applications, classifiers can be used in areas such as image recognition, speech recognition, and natural language processing.

[0066] It is worth mentioning that in other specific examples of the present application, the optimized semantic matching feature matrix can also be passed through a classifier in other ways to obtain a classification result, and the classification result is used to indicate whether the probability that the first candidate entity is the entity most relevant to the power audit problem exceeds a predetermined threshold, for example: using the optimized semantic matching feature matrix as input, and using a machine learning algorithm to train a classifier. Commonly used classifiers include support vector machines (SVM), decision trees, random forests, etc.; the trained classifier is applied to the test data to obtain the classification result of the test data. Test data refers to entities to be classified, such as fitness equipment, power equipment, etc.; determine whether the classification result exceeds a predetermined threshold. The predetermined threshold is set according to actual needs and performance indicators, and is used to determine whether the classification result meets the expected goal.

[0067] Specifically, the output unit 342 is used to determine whether to output the first candidate entity based on the classification result. That is, if the probability of the classification result exceeds a predetermined threshold, the entity is considered to be most relevant to the power audit issue; otherwise, the entity is considered to be irrelevant to the power audit issue; the classification result is output to indicate whether the probability that the first candidate entity is the entity most relevant to the power audit issue exceeds a predetermined threshold. If the classification result exceeds the predetermined threshold, the entity is considered to be the entity most relevant to the power audit issue, and the first candidate entity is output; otherwise, other candidate entities need to be considered.

[0068] It is worth mentioning that in other specific examples of the present application, other methods can be used to determine whether to output the first candidate entity based on the optimized semantic matching feature matrix. For example: collect text data and perform preprocessing, including word segmentation, stop word removal, stem extraction, etc.; convert the preprocessed text data into a numerical vector, and use methods such as bag of words model, TF-IDF, Word2Vec, etc.; perform semantic matching on text data, and use methods such as cosine similarity, Jaccard similarity, and Euclidean distance; construct a semantic matching feature matrix based on the semantic matching results, in which each row represents a text data and each column represents a semantic feature. PCA, LDA and other methods can be used to reduce the dimension of semantic features to reduce feature dimensions; feature matrix can be optimized by using feature selection, feature weighting, feature combination and other methods to improve the expression effect and generalization ability of the feature matrix; the machine learning model can be trained on the optimized feature matrix using training data, and classification, regression, clustering and other methods can be used to solve multi-scale power audit problems; test data can be predicted by using the trained machine learning model to perform classification, regression, clustering and other operations on the test data to output the first candidate entity.

[0069] As described above, the power semantic knowledge graph system 300 according to the embodiment of the present application can be implemented in various wireless terminals, such as a server with a power semantic knowledge graph algorithm. In one possible implementation, the power semantic knowledge graph system 300 according to the embodiment of the present application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the power semantic knowledge graph system 300 can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the power semantic knowledge graph system 300 can also be one of the many hardware modules of the wireless terminal.

[0070] Alternatively, in another example, the power semantic knowledge graph system 300 and the wireless terminal may also be separate devices, and the power semantic knowledge graph system 300 may be connected to the wireless terminal via a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.

[0071] Furthermore, a power semantic knowledge graph method is also provided.

[0072] Figure 6 Flowchart of the power semantic knowledge graph method according to an embodiment of the present application. Figure 6 As shown, in the power semantic knowledge graph method, it includes: S110, obtaining power audit problems; S120, extracting the semantic embedding representation of the first candidate entity from the power semantic knowledge graph; S130, analyzing and processing the power audit problems and the semantic embedding representation of the first candidate entity based on a deep convolutional neural network model to obtain an optimized semantic matching feature matrix; and, S140, determining whether to output the first candidate entity based on the optimized semantic matching feature matrix.

[0073] In summary, based on the power semantic knowledge graph method of the embodiment of the present application, a smart question-answering system is constructed by utilizing natural language processing technology based on deep learning and combining it with the power field knowledge graph to optimize the power audit workflow and enhance audit efficiency.

[0074] The embodiments of the present disclosure have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A power semantic knowledge graph system, characterized in that: include: The problem acquisition module is used to obtain power audit problems; An alternative entity extraction module is used to extract the semantic embedding representation of the first candidate entity from the power semantic knowledge graph; A matching module, used for analyzing and processing the semantic embedding representation of the power audit problem and the first candidate entity based on a deep convolutional neural network model to obtain an optimized semantic matching feature matrix; as well as An output result generating module, used for determining whether to output the first candidate entity based on the optimized semantic matching feature matrix; Wherein, the matching module includes: A semantic analysis unit, used for performing semantic analysis on the power audit problem to obtain a multi-scale power audit problem semantic encoding feature vector; an association matching unit, configured to perform association encoding on the multi-scale power audit question semantic encoding feature vector and the semantic embedding representation of the first candidate entity to obtain a semantic matching feature matrix; and An expression effect optimization unit, used for optimizing the expression effect of text semantic association features on the semantic matching feature matrix to obtain the optimized semantic matching feature matrix; The association matching unit is used to: perform association encoding on the multi-scale power audit question semantic encoding feature vector and the semantic embedding representation of the first candidate entity using the following association formula to obtain the semantic matching feature matrix; Wherein, the formula is: Where V m represents the semantic encoding feature vector of the multi-scale power audit problem, represents the transposed vector of the semantic encoding feature vector of the multi-scale power audit problem, V n represents the semantic embedding representation of the first candidate entity, M represents the semantic matching feature matrix, Represents vector multiplication; Wherein, the expression effect optimization unit includes: A multi-source information fusion pre-test distribution evaluation optimization subunit is used to perform multi-source information fusion pre-test distribution evaluation optimization on each row feature vector of the semantic matching feature matrix to obtain a plurality of optimized row feature vectors; and An arranging subunit, used for arranging the plurality of optimized row feature vectors into the optimized semantic matching feature matrix; The multi-source information fusion pre-test distribution evaluation optimization subunit is used to: perform multi-source information fusion pre-test distribution evaluation optimization on each row feature vector of the semantic matching feature matrix according to the following optimization formula to obtain multiple optimized row feature vectors; Wherein, the formula is: Where V i is the i-th row feature vector of the semantic matching feature matrix, V j is the j-th row feature vector of the semantic matching feature matrix, is the mean eigenvector, n is the neighborhood setting hyperparameter, log represents the logarithmic function value with base 2, Represents positional subtraction, V ′ i is the i-th optimized row feature vector of the optimized semantic matching feature matrix.

2. The electric power semantic knowledge graph system according to claim 1, characterized in that: The semantic analysis unit comprises: A word segmentation subunit, used for performing word segmentation processing on the power audit problem to obtain a sequence of power audit problem description words; A word embedding subunit, used for passing the sequence of power audit problem description words through a word embedding layer to obtain a sequence of power audit problem description word embedding vectors; and The natural language processing subunit is used to pass the sequence of word embedding vectors describing the power audit problem through a dual-pipeline model including a first natural language processing model and a second natural language processing model to obtain the multi-scale power audit problem semantic encoding feature vector.

3. The electric power semantic knowledge graph system according to claim 2 is characterized in that: The natural language processing subunit comprises: A first-scale natural language processing secondary subunit, used for passing the sequence of embedding vectors of the electric power audit problem description words through the first natural language processing model of the dual-pipeline model to obtain a first-scale electric power audit problem semantic encoding feature vector; A first-scale natural language processing secondary subunit is used to pass the sequence of the electric power audit problem description word embedding vectors through the second natural language processing model of the dual-pipeline model to obtain a second-scale electric power audit problem semantic encoding feature vector; The multi-scale fusion secondary sub-unit is used to fuse the first-scale power audit question semantic coding feature vector and the second-scale power audit question semantic coding feature vector to obtain the multi-scale power audit question semantic coding feature vector.

4. The electric power semantic knowledge graph system according to claim 3 is characterized in that: The output result generation module comprises: a classification unit, configured to pass the optimized semantic matching feature matrix through a classifier to obtain a classification result, wherein the classification result is used to indicate whether the probability that the first candidate entity is the entity most relevant to the power audit problem exceeds a predetermined threshold; and An output unit is used to determine whether to output the first candidate entity based on the classification result.

5. The electric power semantic knowledge graph system according to claim 4, characterized in that: The classification unit includes: An expansion subunit, used for expanding the optimized semantic matching feature matrix into a classification feature vector based on a row vector or a column vector; a fully connected encoding subunit, configured to perform fully connected encoding on the classification feature vector using a plurality of fully connected layers of the classifier to obtain an encoded classification feature vector; and The classification result generating subunit is used to pass the encoded classification feature vector through the Softmax classification function of the classifier to obtain the classification result.

6. A method for electric power semantic knowledge graph, using the electric power semantic knowledge graph system according to claim 1, characterized in that: include: Get electricity audit questions; Extracting semantic embedding representation of the first candidate entity from the power semantic knowledge graph; Analyzing and processing the semantic embedding representation of the power audit problem and the first candidate entity based on a deep convolutional neural network model to obtain an optimized semantic matching feature matrix; as well as Based on the optimized semantic matching feature matrix, it is determined whether to output the first candidate entity.

Images