A speech recognition method for power grid dispatching
Through voice recognition technology and machine learning, the automation of power grid scheduling is achieved, the problems of low efficiency and high cost of manual scheduling are solved, and the automation efficiency and cost-effectiveness of power grid scheduling are improved.
Patent Information
- Application Number
- CN202110690248.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-06-22
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2041-06-22
AI Technical Summary
The existing grid scheduling relies on manual operations, is inefficient and costly, and cannot meet the increasing workload needs.
Using speech recognition technology, voice is collected through voice input interface, converted into text, and semantics are used to understand semantics using machine learning and power semantic database, scheduling commands are generated, SMS is automatically sent to realize automatic scheduling.
It improves the automation efficiency of power grid scheduling, reduces the number of dispatchers, reduces the operating costs, and meets the efficient workload needs.
Smart Images

Figure CN113869040B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of speech recognition methods, and in particular to a speech recognition method for power grid dispatching. Background Art
[0002] With the continuous development of the economy, the society's demand for electricity is constantly increasing, the scale of the power grid is also increasing, the requirements for lean control and operation management are constantly improving, the daily work undertaken by control personnel is also increasing, the workload capacity remains high, and the existing work model has gradually failed to adapt to the requirements of the new situation.
[0003] As one of the power production and operation organizations, the current power dispatching and control in the power dispatching and control center relies on manpower. The daily workload of dispatchers is increasing, while the number of regional dispatchers has not increased accordingly. The existing manual dispatching method is inefficient and costly, and cannot meet the increasing workload. Summary of the invention
[0004] In view of this, an object of the present invention is to provide a voice recognition method for power grid dispatching, which can perform automatic dispatching more accurately and efficiently based on voice.
[0005] A speech recognition method for power grid dispatching of the present invention comprises the steps of:
[0006] (1) Configure a voice input interface in the log system to collect input voice, convert voice recognition into text, and then recognize the text. At the same time, establish a power semantic library for the dispatching operation log, and use machine learning to automatically understand the semantics of the recorded log;
[0007] (2) After understanding the semantics of the log, different voice commands are generated according to different scenarios. SMS messages of operation events are automatically sent according to the voice sending commands, and the recipient number is pre-established in the database;
[0008] (3) Use machine learning combined with the knowledge graph in the power semantic library to post-process operation events, generate dispatch commands and publish them externally.
[0009] Furthermore, in the step (1), the speech recognition needs to recognize text, and the recognition process includes text segmentation, part-of-speech tagging, entity recognition and text error correction, and the recognition process is based on the pre-established power semantic library.
[0010] Furthermore, the text segmentation adopts N shortest path segmentation, CRF segmentation or fast dictionary segmentation;
[0011] The part-of-speech tagging adopts hidden Markov model part-of-speech tagging, perceptual machine part-of-speech tagging, entity naming tagging, a method based on feature templates or a method based on neural network.
[0012] Furthermore, the semantic understanding in step (1) includes scenario classification, sentiment analysis, content summary, word vector, dependency syntax and sensitive word processing.
[0013] Furthermore, the power semantic library includes a term library, a synonym forest and a corpus
[0014] Furthermore, the scenarios in step (2) include compliance verification, shift handover, customer service scheduling and intelligent recommendation.
[0015] Furthermore, the establishment of the knowledge graph in the power semantic library requires knowledge representation, knowledge extraction modeling, knowledge reasoning, semantic search and machine learning of power knowledge.
[0016] Furthermore, the post-processing in step (3) includes fusion decision, selection of response template and result processing. After the result processing, a scheduling command is generated and released to the outside, and the release structure is returned and deep learning is performed.
[0017] The beneficial effects of the present invention are as follows: a voice recognition method for power grid dispatching of the present invention automatically understands the intention of the incoming call by collecting the incoming call voice, converting it into text and understanding its semantics, and utilizing machine learning in combination with a power semantic library, thereby automatically generating dispatching commands according to different scenarios and in combination with different templates, and sending them via text messages; the present invention can automatically perform dispatching and solve the problem of insufficient dispatching personnel. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present application and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative work:
[0019] Figure 1 It is a schematic diagram of the overall process of the present invention;
[0020] Figure 2 This is a schematic diagram of the N shortest path word segmentation of the present invention;
[0021] Figure 3 Schematic diagram for modeling knowledge extraction of the present invention. DETAILED DESCRIPTION
[0022] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations.
[0023] Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the present application for which protection is sought, but merely represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present application.
[0024] like Figure 1-Figure 3 As shown: A speech recognition method for power grid dispatching in this embodiment, the overall steps include:
[0025] (1) Configure a voice input interface in the log system to collect input voice, convert voice recognition into text, and then recognize the text. At the same time, establish a power semantic library for the dispatching operation log, and use machine learning to automatically understand the semantics of the recorded log;
[0026] (2) After understanding the semantics of the log, different voice commands are generated according to different scenarios. SMS messages of operation events are automatically sent according to the voice sending commands, and the recipient number is pre-established in the database;
[0027] (3) Use machine learning combined with the knowledge graph in the power semantic library to post-process operation events, generate dispatch commands and publish them externally.
[0028] Specifically, in the present embodiment, the speech recognition in step (1) requires the recognition of text, and the recognition process includes text segmentation, part-of-speech tagging, entity recognition and text error correction, and the recognition process is based on a pre-established power semantic library;
[0029] Speech recognition technology is an important prerequisite for improving friendly interaction. A good speech service engine directly affects the processing of other modules. Based on the cloud speech service voiceprint recognition, speech recognition, speech synthesis and other technologies, combined with the power dispatch resource library, a unified speech recognition, synthesis and voiceprint recognition intelligent speech service platform is built to provide technical support and capability support for upper-level services;
[0030] Speech synthesis converts text into speech streams, allowing machines to speak like humans. If the above speech recognition enables machines to learn to "listen", then speech synthesis enables machines to learn to "speak". By calling third-party interfaces, the intelligent voice service platform has different languages, dialects, and timbres.
[0031] Specifically, in this embodiment, semantic understanding of the speech-converted text needs to be preprocessed, and the preprocessing includes text segmentation and part-of-speech tagging, wherein the text segmentation adopts N shortest path segmentation, CRF segmentation or fast dictionary segmentation;
[0032] N Shortest Path Participle
[0033] N shortest paths are actually a variation of K shortest paths (KSP), which refers to the first K shortest paths among the single-source paths in the DAG. The algorithms for solving KSP include the deletion algorithm, the improved deletion algorithm (MS Algorithm), and the recursive enumeration algorithm (REA, Recursive Enumeration Algorithm).
[0034] The general graph representation uses an adjacency matrix to record the edges with a weight of cost from vertex from to vertex to, and a one-dimensional array to record the subscripts of the vertices, such as Figure 2 As shown;
[0035] CRF word segmentation
[0036] CRF (Conditional Random Field) is a machine learning technology (model). CRF represents a new generation of machine learning technology for word segmentation. Its basic idea is to annotate Chinese characters, that is, to form words (group words) from characters. It not only considers the frequency information of the occurrence of words, but also considers the context. It has good learning ability, so it has a good effect on the recognition of ambiguous words and unregistered words; its disadvantages are that the training cycle is long, the amount of calculation is large during operation, and the performance is not as good as dictionary word segmentation. Its main principles:
[0037] CRF treats word segmentation as a word position classification problem, and usually defines the word position information of a word as follows:
[0038] The beginning of a word, usually represented by B
[0039] In words, M is often used to represent
[0040] End of a word, usually represented by E
[0041] Monad, usually represented by S
[0042] The process of CRF word segmentation is to mark the word position and then form word segments with the words between B and E and the single word S.
[0043] CRF word segmentation example:
[0044] Original example: I love Beijing.
[0045] CRF labeling results: I / Sai / SBeijing / BJing / E
[0046] Word segmentation result: I / love / Beijing.
[0047] Fast Dictionary Segmentation
[0048] The word segmentation of the fast dictionary focuses on speed, with high throughput and low resource consumption, but its disadvantage is that the accuracy is not high. Its main algorithm is to implement an extremely high-performance Aho Corasick automaton in combination with Double Array Trie. When applied to word segmentation, it can achieve 14 million words per second, which is about 27MB / s. The basic principle is to attach additional information to each state (reflected as a subscript) of a double array Trie tree. The basis of the AC automaton (success table) is the Trie tree, but it has an output table and a fail table more than the Trie tree. Then the construction principle of the Aho Corasick Double Array Trie is to build output[i] and fail[i] for each state (base[i] and check[i]). When creating base[i] for a Trie tree State, let the State remember its own i, so that a mapping between State and subscript is established.
[0049] Part-of-Speech tagging (POS tagging) refers to the process of assigning a suitable part of speech to each word in a sentence, that is, to determine whether each word is a noun, verb, adjective or other part of speech. It is also called part-of-speech tagging or simply tagging. Part-of-speech tagging is a basic task in natural language processing and plays an important role in speech recognition, information retrieval and many other fields of natural language processing.
[0050] Part-of-speech tagging uses hidden Markov model part-of-speech tagging, perceptual machine part-of-speech tagging, entity naming tagging, feature template-based methods, or neural network-based methods;
[0051] HMM (Hidden Markov Model) Part of Speech Tagging
[0052] The hidden Markov model is a mathematical model represented by a bunch of mathematical symbols and parameters, including a set of hidden states, a set of observed states, an initial probability vector, a state transfer matrix A, and a confusion matrix B.
[0053] The observable sequence in HMM is the word segmentation of a given sentence in part-of-speech tagging, and the hidden state is the different parts of speech. The prior probability of the part of speech is the parameter. Therefore, in order to achieve part-of-speech tagging of sentences, we need to first train an HMM using the corpus, and then segment and tag the sentences.
[0054] The number of states (parts of speech) in the model is the number of part-of-speech symbols N; the number of different symbols (words) that may be output from each state is the number of words M. Assume that in a statistical sense, the probability distribution of each part of speech is only related to the part of speech of the previous word (i.e., the bigram of the part of speech), and the probability distribution of each word is only related to its part of speech. Then, we can calculate the parameter λ of the HMM by counting the training corpus that has been segmented and tagged with part of speech. Then, based on the known words, the part of speech corresponding to each word can be calculated through the Viterbi algorithm, that is, the part-of-speech tagging is completed.
[0055] Perceptron Part-of-Speech Tagging
[0056] The perceptron can be considered as a classifier with built-in weights. When a feature is input, the perceptron outputs whether the input belongs to this category. However, perceptrons are generally used for binary classification. For multi-classification problems such as Chinese part-of-speech tagging, multiple perceptrons can be used to form a perceptron network to output the category of the neural network. This is the main core idea of the perceptron part-of-speech tagging algorithm.
[0057] Entity naming annotation
[0058] Named entity recognition is a very basic task in NLP, also known as "proper name recognition", which refers to the recognition of entities with specific meanings in text, mainly including names of people, places, institutions, proper nouns, etc., mainly paving the way for tasks such as relationship extraction. Of course, in a specific field, various entity types in the field will be defined accordingly.
[0059]
[0060] Names of people, organizations, and places are all marked here.
[0061] Feature template based method
[0062] The statistical machine learning method regards NER as a sequence labeling task, and uses large-scale corpus to learn the labeling model, so as to label each position of the sentence. Commonly used models applied to NER tasks include generative models HMM and discriminative models CRF. The more popular method is the feature template + CRF scheme: the feature template is usually some manually defined binary feature functions, which attempts to explore the composition characteristics of the named entity and the context. For a given position in a sentence, the position where the feature is extracted is a window, that is, the context position. Moreover, different feature templates can be combined to form a new feature template. The advantage of CRF is that it can use the previously labeled information in the process of labeling a position, and use Viterbi decoding to obtain the optimal sequence. When extracting features for each position in the sentence, the feature that meets the conditions is taken as 1, and the feature that does not meet the conditions is taken as 0; then the features are fed to CRF, the transfer of model labels is modeled in the training stage, and then the various positions of the test sentence are labeled in the inference stage.
[0063] Neural network based methods
[0064] At present, the NN / CNN / RNN-CRF model that combines the neural network with the CRF model has become the mainstream model of NER. For CNN and RNN, neither has an absolute advantage, and each has its own advantages. Since RNN has a natural sequence structure, RNN-CRF is more widely used. The NER method based on the neural network structure inherits the advantages of the deep learning method and does not require a large number of artificial features. Only word vectors and character vectors are needed to reach the mainstream level, and adding high-quality dictionary features can further improve the effect.
[0065] Specifically, in this embodiment, the semantic understanding in step (1) includes scenario classification, sentiment analysis, content summary, word vector, dependency syntax and sensitive word processing;
[0066] Scenario classification associates different keywords with different dispatch scenarios based on different scenarios in the power dispatch process. After obtaining the corresponding keywords, it can help the machine to quickly understand the semantics and improve the generation efficiency when making decisions to generate dispatch commands.
[0067] Sentiment Analysis
[0068] The tendency calculation based on semantic sentiment dictionary is different from the machine learning algorithm that requires a large number of training data sets. It mainly uses sentiment dictionary and sentence word library to analyze the special structure of text sentences and sentiment tendency words, and adopts weight algorithm to replace traditional manual judgment or simple statistical method for sentiment classification. Different weights are assigned to sentiment words with different sentiment intensity, and then weighted sum is performed. Using weighted average algorithm (1) for calculation can effectively improve the efficiency and accuracy of sentiment classification. The calculation formula is as follows:
[0069]
[0070] Among them, Np and Nn represent the number of words expressing positive emotions and negative emotions respectively, and wpi and wpj represent the weights of positive emotion words and negative emotion words respectively. By setting the threshold and comparing the threshold with the calculated value of E, the relative positive emotion result or relative negative emotion result of the semantics is obtained to help the machine make judgments on the semantics.
[0071] Summary
[0072] Content summarization obtains a certain paragraph in the text through word segmentation, compares it with previous cases, and helps the machine understand the semantics based on the similarity. Content summarization requires a lot of template training to be more accurate.
[0073] Word vectors
[0074] To hand over natural language to machine learning algorithms for processing, it is usually necessary to first mathematize the language. Word vectors are a way to mathematize the words in the language. Word vectors also require a large number of statistical templates. In the power semantic library, a large number of words are mapped and associated based on a large number of template training to help the machine understand the semantics.
[0075] Dependency Syntax
[0076] Help machines understand semantics through fixed collocations between the meanings and parts of speech of different words;
[0077] Sensitive word processing
[0078] A sensitive word library needs to be established in the power semantic library. Once a sensitive word appears, the corresponding instruction will be triggered.
[0079] Specifically, in this embodiment, the power semantic library includes a term library, a synonym forest and a corpus. The power semantic library is the key to machine learning in this solution. It uses natural language processing technology to perform word segmentation, part-of-speech tagging, named entity recognition and text error correction on the input operation instructions. These preprocessing are used as input for semantic understanding. This system is a speech recognition in the context of scheduling, which will inevitably involve a lot of operation professional terms and equipment professional terms, so it is also necessary to establish a power-related semantic library. It is based on the term library, and a synonym forest and a corpus are established on this basis.
[0080] Specifically, in this embodiment, the scenarios in step (2) include compliance verification, shift handover, customer service scheduling and intelligent recommendation.
[0081] Specifically, in this embodiment, the knowledge graph is composed of some interconnected entities and their attributes. In the power dispatching system, a knowledge base is constructed based on the knowledge related to power. This knowledge base contains a lot of power knowledge, and there are many associations between these knowledge, which constitutes a knowledge graph related to power. The knowledge graph integrates many aspects. From the perspective of the Web, the knowledge graph establishes semantic links between data like hyperlinks between texts, and supports semantic search. From the perspective of NLP, it mainly focuses on how to extract semantic and structured data from text. From the perspective of knowledge representation, it is how to use computer symbols to represent and process knowledge. From the perspective of AI, it is how to use knowledge bases to assist in understanding human language. From the perspective of the database, it is to store knowledge in the form of graphs. Therefore, to do a good job of the knowledge graph, it is necessary to make comprehensive use of methods and technologies in many aspects such as knowledge representation, natural language understanding, network, machine learning, and database.
[0082] First of all, the knowledge graph is the foundation of knowledge. It uses technologies such as knowledge extraction, knowledge fusion, and knowledge crowdsourcing to obtain the required data from a large number of data sources such as text, structured databases, and multimedia files. Then, through knowledge representation, knowledge reasoning, and knowledge linking, the knowledge is organized and stored in a standardized and orderly manner. It is ultimately used in knowledge question and answer, semantic search, and visualization.
[0083] The establishment of knowledge graph in the power semantic database requires knowledge representation, knowledge extraction modeling, knowledge reasoning, semantic search and machine learning of power knowledge;
[0084] Knowledge Representation
[0085] The representation of knowledge is a description of knowledge, or a set of conventions for knowledge, a data structure that can be accepted by computers to describe knowledge. In a sense, representation can be regarded as a combination of data structure and its processing mechanism: representation = data structure + processing mechanism. Therefore, in ES, knowledge representation is a series of technical means in ES that can complete the computer processing of expert knowledge. Common ones include production rules, semantic web, framework method, etc.
[0086] Knowledge extraction modeling
[0087] Knowledge extraction modeling is to abstract knowledge in a formalized and structured way. Knowledge model is not knowledge, but an abstraction of knowledge. Its main process is to obtain text data and obtain clean text data after text preprocessing. Then, with the help of machine learning related programs, the text is segmented, part-of-speech tagged, lexical parsed, and dependency analyzed. At this time, the analysis of lexical and syntactic levels is completed. Next, the text is subjected to NER and entity linking to prepare for relation extraction and time extraction, and finally the knowledge graph consisting of triples, multi-relationships, modal knowledge, etc. for KR is formed.
[0088] like Figure 3 As shown in the figure, the specific operation is to first use natural language processing technology to extract unstructured data such as text into the database, completing the first step of data cleaning and data organization. Then, the table in the database is converted into RDF triples by writing heuristic rules.
[0089] Follow the following methods:
[0090] Convert table into class
[0091] Convert columns to attributes
[0092] Convert a row of data into a resource
[0093] Convert the value in the cell into a text value
[0094] There are two solutions. On the one hand, we can use the NTriples to TDB tool of Apache Jena to convert the triples in the data into RDF data for query service of SPARQL server Fuseki. On the other hand, we can transfer the data of relational database to NOSQL graph database Neo4j and store structured data on the network instead of in tables. So far, the knowledge modeling service is completed. Good knowledge modeling lays the foundation for knowledge reasoning.
[0095] Knowledge Reasoning
[0096] Knowledge reasoning is to use existing knowledge through various logical reasoning to obtain new knowledge or conclusions, and these knowledge and conclusions meet the semantics.
[0097] There are many ontology reasoning methods in technology and theory. The first one is based on Tableaux operation. This method checks the satisfiability of a certain ontology and instance detection. The idea is similar to the reductive refutation of first-order logic. The second one is based on logic programming rewriting. Rules can be customized according to specific scenarios to achieve user-defined reasoning processes. The third one is based on first-order query rewriting. This method can efficiently combine data sources in different data formats. The last one is based on production rules. The production system is a forward reasoning system that can execute rules according to a certain mechanism to achieve certain goals. It is similar to first-order logic, but there are also differences.
[0098] Semantic Search
[0099] Semantic search is a search method that searches for the meaning of sentences beyond the surface of words. It is a supplement to knowledge reasoning. Semantic search aims to improve search accuracy by understanding the searcher's intention through contextual meaning. Through concept matching, synonyms, and natural language algorithms, semantic search provides more interactive search results by converting structured and unstructured data into an intuitive and responsive database. Semantic search can enhance the understanding of searcher intent, the ability to extract answers, and provide more personalized results.
[0100] Previously, the commonly used search was based on document retrieval. Information retrieval (IR) supports the retrieval of documents. It expresses the user's search requirements and resource content through a lightweight grammatical model, such as AND OR. This is the currently dominant keyword model: the bag-of-words model. It works well for topic searches, but cannot cope with more complex information retrieval requirements.
[0101] Databases (DB) and knowledge-based expert systems can provide more accurate answers (data retrieval). They use more expressive models to represent user needs, utilize the inherent structure and semantic associations between data, allow complex queries, and return specific answers that accurately match the queries.
[0102] Machine Learning Techniques
[0103] At its core, machine learning is about “using an algorithm to parse data, learn from it, and then make a decision or prediction about something in the world.” This means that instead of explicitly writing a program to perform some task, you teach the computer how to develop an algorithm to do it. There are three main types of machine learning: supervised learning, unsupervised learning, and reinforcement learning, all of which have specific advantages and disadvantages.
[0104] Supervised learning involves a set of labeled data. The computer can use specific patterns to recognize new examples of each labeled type. The two main types of supervised learning are classification and regression. In classification, the machine is trained to classify a group into specific classes. A simple example of classification is a spam filter on an email account. The filter analyzes emails that you have previously marked as spam and compares them to new emails. If they match a certain percentage, these new emails are marked as spam and sent to the appropriate folder. Those emails that are less similar are classified as ham and sent to your mailbox.
[0105] The second type of supervised learning is regression. In regression, the machine uses previous (labeled) data to predict the future. Weather apps are a good example of regression. Using historical data on meteorological events (i.e. average temperature, humidity, and precipitation), your phone weather app can look at the current weather and make a prediction about what the weather will be like in the future.
[0106] In unsupervised learning, the data is unlabeled. Since most real-world data is unlabeled, these algorithms are particularly useful. Unsupervised learning is divided into clustering and dimensionality reduction. Clustering is used to group objects based on attributes and behaviors. This is different from classification because the groups are not provided by you. An example of clustering is dividing a group into different subgroups (for example, based on age and marital status) and then applying targeted marketing programs. Dimensionality reduction reduces the variables of a data set by finding commonalities. Most big data visualizations use dimensionality reduction to identify trends and rules.
[0107] Finally, reinforcement learning uses the machine's personal history and experience to make decisions. Unlike supervised and unsupervised learning, reinforcement learning does not involve providing the "correct" answer or output. Instead, it focuses solely on performance. This mirrors how humans learn based on positive and negative outcomes.
[0108] Artificial intelligence (AI) is a science, machine learning (ML) is currently the most mainstream method to implement artificial intelligence, and deep learning (DL) is a branch of machine learning (ML) and is also the most popular type of machine learning (ML).
[0109] Deep learning (DL) is a method in machine learning based on data representation and learning. It is a machine learning method that can simulate the neural structure of the human brain. The concept of deep learning originated from the study of artificial neural networks. The artificial neural network ANN (Artificial Neural Network) abstracts the human brain neural network from the perspective of information processing, establishes a simple model, and forms different networks according to different connection methods. It is referred to as a neural network or a quasi-neural network.
[0110] In this solution, we will use both traditional machine learning algorithms and deep learning algorithms, and the selection is mainly based on data scale, scenario and performance. For example, text classification uses traditional machine learning support vector machines, which have always had good performance and performance indicators in text classification.
[0111] Specifically, in this embodiment, the post-processing in step (3) includes fusion decision, selection of response template and result processing. After the result processing, a scheduling command is generated and released to the outside, and the release structure is returned and deep learning is performed.
[0112] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.
Claims
1. A speech recognition method for power grid dispatching, characterized in that: Included steps (1) Configure a voice input interface in the log system to collect input voice, convert voice recognition into text, and then recognize the text. At the same time, establish a power semantic library for the dispatching operation log, and use machine learning to automatically understand the semantics of the recorded log; (2) After understanding the semantics of the log, different voice commands are generated according to different scenarios. SMS messages of operation events are automatically sent according to the voice sending commands, and the recipient number is pre-established in the database; (3) Use machine learning combined with the knowledge graph in the power semantic library to post-process the operation events, generate dispatch commands and publish them externally; In the step (1), the speech recognition needs to recognize text, and the recognition process includes text segmentation, part-of-speech tagging, entity recognition and text error correction, and the recognition process is based on the pre-established power semantic library; The text segmentation adopts N shortest path segmentation, CRF segmentation or fast dictionary segmentation; The part-of-speech tagging adopts hidden Markov model part-of-speech tagging, perceptual machine part-of-speech tagging, entity naming tagging, a method based on feature templates or a method based on neural network; The semantic understanding in step (1) includes scenario classification, sentiment analysis, content summary, word vector, dependency syntax and sensitive word processing; Scenario classification is based on different scenarios in the power dispatch process, and different keywords are associated with different dispatch scenarios; Sentiment analysis uses sentiment dictionaries and sentence word libraries to analyze the special structure and sentiment tendency words of text sentences, and adopts weighted algorithms to replace traditional manual judgment or simple statistical methods for sentiment classification; different weights are assigned to sentiment words with different sentiment intensities, and then weighted sums are taken and calculated using weighted average algorithm (1), which can effectively improve the efficiency and accuracy of sentiment classification. The calculation formula is as follows: Among them, N p and N n Represents the number of words expressing positive and negative emotions, wp i and wp j Represent the weights of positive sentiment words and negative sentiment words respectively; by setting the threshold and comparing the threshold with The calculated value of can obtain the relative positive emotion result or the relative negative emotion result of the semantics, which helps the machine to judge the semantics; Content summary obtains a certain paragraph in the text through word segmentation, compares it with previous cases, and helps the machine understand the semantics based on the similarity; Word vectors hand over natural language to machine learning algorithms for processing, which requires mathematizing the language. Dependency syntax helps machines understand semantics through fixed collocations between the meanings and parts of speech of different words; Sensitive word processing: A sensitive word library is established in the power semantic library. Once a sensitive word appears, the corresponding instruction is triggered; The electric power semantic library includes a term library, a synonym forest and a corpus; The scenarios in step (2) include compliance verification, shift handover, customer service scheduling, and intelligent recommendation; The establishment of the knowledge graph in the power semantic library requires knowledge representation, knowledge extraction modeling, knowledge reasoning, semantic search and machine learning of power knowledge; Knowledge representation is used to describe the data structure of knowledge. Representation can be considered as a combination of data structure and its processing mechanism. Knowledge extraction modeling is used to abstract knowledge in a formalized and structured manner; Detailed steps of knowledge extraction modeling: obtain text data, and get clean text data after text preprocessing; then use machine learning related programs to perform word segmentation, part-of-speech tagging, lexical parsing and dependency analysis on the text, and the analysis of lexical and syntactic levels is completed; then perform NER and entity linking on the text to prepare for relationship extraction and time extraction; finally form a knowledge graph consisting of triples, multi-relationships and modal knowledge used for KR; Knowledge reasoning uses existing knowledge to obtain new knowledge or conclusions through various logical reasoning. The knowledge and conclusions satisfy semantics. Semantic Search Through concept matching, synonyms and natural language algorithms, semantic search provides more interactive search results by transforming structured and unstructured data into an intuitive and responsive database; Machine learning uses algorithms to analyze data, including supervised learning, unsupervised learning, and reinforcement learning; The post-processing in step (3) includes fusion decision, selection of response template and result processing. After the result processing, a scheduling command is generated and released to the outside, and the release structure is returned and deep learning is performed.
