Speech recognition hybrid model construction method and system for power grid equipment monitoring

Abstract

The invention discloses a voice recognition hybrid model construction method and system for power grid equipment monitoring. Power grid equipment related document information is collected to form a corpus set in the power grid equipment monitoring field, the corpus set is filled into an instruction format template to form a monitoring professional voice recognition corpus, the monitoring professional voice recognition corpus is loaded into a model training algorithm, and incremental training is performed on a power grid equipment monitoring professional model and a hot word model; exporting the power grid equipment monitoring professional model and the universal recognition model as an acoustic model and a language model, respectively inputting a to-be-recognized voice signal segment into the acoustic model and the language model, calculating a word sequence score difference value, if the score difference value is higher than a threshold value, decoding and outputting the model corresponding to the highest score, and outputting the speech signal segment; and otherwise, judging whether the phoneme fragment contains the power grid professional word based on the hot word model, if so, decoding and outputting the phoneme fragment by a monitoring professional model, and if not, decoding and outputting the phoneme fragment by a general model. According to the invention, the recognition accuracy of voice recognition of monitoring service instructions and non-instruction services is realized.

Description

Technical Field

[0001] The present invention belongs to the field of power grid equipment monitoring, and in particular relates to a method and system for constructing a speech recognition hybrid model for power grid equipment monitoring. Background Art

[0002] With the advancement of the construction of new power systems, the scale of power grids continues to expand, information construction continues to be optimized, and the amount of command information is increasing. Monitoring personnel on duty have to send and receive a large number of instructions every day. Power monitoring speech recognition is a key technology in monitoring and is the basis for the correct understanding of information on the actual site. The acoustic model, pronunciation dictionary, language model and decoder constitute the entire speech recognition system.

[0003] With the rapid development of science and technology in recent years, speech recognition has also developed rapidly, especially with the deepening application of deep learning in recent years, the development speed is particularly obvious. However, speech recognition technology in the power field is still under development. Most of the existing speech recognition technologies are based on general recognition models or professional models. Only a few models combine general recognition models with professional models for power grid equipment monitoring, and do not use general recognition models with millions of parameters. Since the professional model for power grid equipment monitoring contains the professional name of the power grid and the names of equipment such as plants, stations, and lines, it often needs to be updated to ensure the recognition accuracy. Fine-tuning the professional monitoring model on the general recognition model requires a lot of computing resources, which has too high requirements on equipment, high hardware costs, and long training time each time. Summary of the invention

[0004] In order to solve the deficiencies in the prior art, the present invention provides a method and system for constructing a speech recognition hybrid model for power grid equipment monitoring, so as to solve the technical problems of long speech recognition model update time, large resource occupation and high equipment requirements.

[0005] In order to solve the above technical problems, the present invention adopts the following technical solutions.

[0006] The present invention first discloses a method for constructing a speech recognition hybrid model for power grid equipment monitoring, the method comprising the following steps:

[0007] Collect relevant document information of power grid equipment and perform text preprocessing to form a corpus in the field of power grid equipment monitoring;

[0008] Fill the power grid equipment monitoring domain corpus set into the instruction format template to form monitoring professional speech recognition corpus, use the corpus incremental training tool to load the monitoring professional speech recognition corpus into the model training algorithm, and perform incremental training on the power grid equipment monitoring professional model and the hot word model respectively;

[0009] The trained professional model for power grid equipment monitoring is exported as a professional acoustic model and a professional language model, and the universal recognition model is exported as a universal acoustic model and a universal language model. The speech signal segments to be recognized are respectively input into the professional acoustic model and the professional language model, as well as the universal acoustic model and the universal language model. The difference in word sequence scores of the two models is calculated based on the WFST-based Viterbi decoding algorithm. For phoneme segments whose score difference is higher than a preset threshold, the model corresponding to the highest score is used for decoding and output. For phoneme segments whose score difference is not higher than a preset threshold, it is determined based on the hot word model whether the phoneme segments contain professional power grid words. If so, the professional model for power grid equipment monitoring is used for decoding and output. If not, the universal recognition model is used for decoding and output.

[0010] The present invention further includes the following preferred embodiments:

[0011] The collecting of power grid equipment related document information and performing text preprocessing to form a corpus in the field of power grid equipment monitoring further includes:

[0012] Read the power grid equipment model, run the real-time library and the historical library, organize the power grid equipment model dictionary, including the latest rules and regulations, procedures and specifications, work instructions, office documents, power grid dictionary, equipment model, interactive instructions, and convert special symbols in the power grid equipment model into nouns;

[0013] The natural language processing algorithm is used to perform segmentation and sentence processing, and the processed sentences are analyzed by word segmentation according to the power grid equipment model dictionary. The equipment parameters and monitoring vocabulary in the sentences are analyzed to build a professional monitoring dictionary.

[0014] The corpus incremental training tool comprises:

[0015] Monitoring vocabulary expansion tool, used to expand the power grid monitoring vocabulary, by importing files, adding new interfaces, obtaining all the words that need to be expanded, adding new words, maintaining the original words, updating the power grid monitoring vocabulary, and enriching and expanding the power grid monitoring vocabulary;

[0016] Model loading tool, used to load different models, load old models for incremental training of new models, and parse the relationship between intent, corpus, and entities in the model;

[0017] Corpus loading tool, used to load corpus during model training, extract corpus in text file format, generate different JSON format content according to the format requirements of model training, load it into the model training algorithm, and provide data for model training;

[0018] Model correction tool, used to use model correction algorithm to provide a variety of different corpora for multiple training of similar corpora, and obtain a new recognition accuracy after a preset number of trainings, and obtain the final training model after meeting the requirements;

[0019] Corpus training tools, which are used to optimize the corpus training process by using model correction algorithms and corpus training algorithms, screen the corpus, and extract key parts;

[0020] Model generation tool, used to export trained model files.

[0021] The loading of corpus further comprises:

[0022] According to the power grid equipment model dictionary and the monitoring professional dictionary, the corpus is spliced ​​in combination with the instruction format, and each template is simplified according to the instruction format. The dictionary name is used in the template to replace the specific value in the instruction. The power grid equipment model dictionary and the monitoring professional dictionary are read and filled into the instruction format template to complete the automatic generation of monitoring professional speech recognition corpus. Then the generated corpus is saved in the database, and a corpus file in JSON format or TXT format is generated.

[0023] For a corpus file in JSON format, the corpus file includes the corpus text content text and its entities and intents intent, wherein each entity includes the entity starting position start, the entity ending position end, the entity content value, and the entity name entity.

[0024] The present invention also discloses a speech recognition hybrid model construction system for power grid equipment monitoring using the aforementioned speech recognition hybrid model construction method for power grid equipment monitoring, comprising:

[0025] Corpus collection module, used to collect document information related to power grid equipment and perform text preprocessing to form a corpus set in the field of power grid equipment monitoring;

[0026] A model training module is used to fill the power grid equipment monitoring field corpus set into the instruction format template to form a monitoring professional speech recognition corpus, use the corpus incremental training tool to load the monitoring professional speech recognition corpus into the model training algorithm, and perform incremental training on the power grid equipment monitoring professional model and the hot word model respectively;

[0027] A speech recognition module is used to export the trained professional model for power grid equipment monitoring into a professional acoustic model and a professional language model, and to export the general recognition model into a general acoustic model and a general language model, and to input the speech signal segments to be recognized into the professional acoustic model and the professional language model, as well as the general acoustic model and the general language model, respectively, and to calculate the difference in word sequence scores of the two models based on the WFST Viterbi decoding algorithm. For phoneme segments whose score difference is higher than a preset threshold, the model corresponding to the highest score is used for decoding and outputting. For phoneme segments whose score difference is not higher than a preset threshold, it is determined based on the hot word model whether the phoneme segments contain professional power grid words. If so, the professional model for power grid equipment monitoring is used for decoding and outputting. If not, the general recognition model is used for decoding and outputting.

[0028] Accordingly, the present application also discloses a terminal, including a processor and a storage medium;

[0029] The storage medium is used to store instructions;

[0030] The processor is used to operate according to the instructions to execute the steps of the aforementioned method for constructing a speech recognition hybrid model for power grid equipment monitoring.

[0031] Correspondingly, the present application also discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the aforementioned method for constructing a speech recognition hybrid model for monitoring power grid equipment.

[0032] The beneficial effect of the present invention is that, compared with the prior art, the present invention provides a method and system for constructing a speech recognition hybrid model for power grid equipment monitoring, which reduces the model training cost and can complete the model training on the existing server; reduces the model training time, and each training does not involve the general language model part, and performs incremental training on the original monitoring professional model, which is convenient and fast. In terms of social and economic benefits, the recognition accuracy of monitoring professional vocabulary and daily general vocabulary is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flow chart of the method for constructing a speech recognition hybrid model for power grid equipment monitoring in the present invention.

[0034] Figure 2 It is a schematic diagram of the power grid equipment model dictionary in the present invention.

[0035] Figure 3 It is a schematic diagram of the monitoring professional speech recognition corpus format in the present invention.

[0036] Figure 4 It is a schematic diagram of the speech recognition adaptation network framework in the present invention.

[0037] Figure 5 It is a schematic diagram of the acoustic model based on DNN-HMM in the present invention.

[0038] Figure 6 This is a flowchart of the acoustic model construction based on DNN-HMM in the present invention. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solution and advantages of the present invention more clear, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0040] The embodiments described in this application are only some embodiments of the present invention, not all embodiments. Based on the spirit of the present invention, other embodiments obtained by ordinary technicians in this field without creative work are all within the protection scope of the present invention.

[0041] Since professional documents such as power grid monitoring duty regulations and operation instructions are updated regularly, and power grid plants, lines, switches and other equipment will also increase, in order to meet the accuracy of speech recognition in the field of power grid equipment monitoring, the professional model of power grid equipment monitoring needs to be updated in time after the knowledge in the field of power grid equipment monitoring is updated. In view of the shortcomings of the prior art, based on the existing speech recognition service of the monitoring duty station, the present invention proposes a speech recognition hybrid model construction method and system for power grid equipment monitoring, constructs a general recognition model, adds a monitoring field corpus set on the basis of the general recognition model for fine-tuning, trains the general recognition model and the professional model separately, and combines and decodes, collects and organizes the monitoring corpus, trains the monitoring professional model, and models hot words based on the nouns of primary and secondary equipment in the power grid and the power grid dictionary for subsequent wake-up words and professional words judgment, so as to improve the accuracy of speech recognition of professional vocabulary for power grid equipment monitoring; integrates and fine-tunes the monitoring professional model with the general recognition model, calculates the difference in word sequence scores of the two models based on the Viterbi decoding algorithm of WFST for confidence judgment, and the judge takes the decoding output of the high-scoring model; implements an incremental training tool for professional models and hot word models, which is conducive to the organization of monitoring professional vocabulary corpus.

[0042] The model construction process of the speech recognition system in the field of power grid equipment monitoring is as follows: Figure 1 As shown, the method for constructing a speech recognition hybrid model for power grid equipment monitoring disclosed in the present invention includes the following steps:

[0043] Step 1: Collect document information related to power grid equipment and perform text preprocessing to form a corpus in the field of power grid equipment monitoring.

[0044] First, we collect monitoring professional vocabulary and process the corpus. We organize the grid equipment model dictionary by reading the grid equipment model, running the real-time library and the historical library, such as Figure 2 As shown, it includes the latest rules and regulations, procedures and specifications, work instructions, office documents, power grid dictionaries, equipment models, interactive instructions, etc. of this year, and converts special symbols in the power grid equipment model into nouns. For example, "#1 machine" and "I section bus" are converted into "No. 1 machine" and "a section bus".

[0045] The natural language processing algorithm is used to perform segmentation and sentence processing, and the processed sentences are analyzed by word segmentation according to the power grid equipment model dictionary. The equipment parameters, monitoring vocabulary, etc. in the sentences are analyzed to build a professional monitoring dictionary.

[0046] Step 2: Fill the power grid equipment monitoring field corpus set into the instruction format template to form the monitoring professional speech recognition corpus, use the corpus incremental training tool to load the monitoring professional speech recognition corpus into the model training algorithm, and perform incremental training on the power grid equipment monitoring professional model and hot word model respectively.

[0047] First, incremental training of monitoring professional models and hot word models is carried out. The corpus incremental training tool of professional models and hot word models is used. The corpus incremental training tool is based on machine learning and deep learning. By adding corpus for learning and training, a language model that meets the needs of power grid monitoring business is established. Specifically, the corpus incremental training tool specifically includes:

[0048] Monitoring vocabulary expansion tool, used to expand the power grid monitoring vocabulary, obtain all the words that need to be expanded by importing files, adding new words to the interface, etc., add new words, keep the original words, update the power grid monitoring vocabulary, and enrich and expand the power grid monitoring vocabulary;

[0049] Model loading tool, used to load different models, load old models for incremental training of new models, and parse the relationship between intent, corpus, and entities in the model;

[0050] The corpus loading tool is used to load the corpus first during model training, extract the corpus in the form of text files, and generate different JSON format contents according to the format requirements of model training, and load them into the model training algorithm to provide data for model training.

[0051] The model correction tool is used to expand similar corpora using a model correction algorithm to provide a large number of different corpora, conduct training multiple times, obtain a new recognition accuracy after a preset number of trainings, and obtain the final training model after meeting the requirements. The model correction algorithm compares similar corpora, obtains more comprehensive parameters such as equipment and time by querying database tables, etc., parses the previous corpus structure for filling, thereby increasing the number of corpora, retraining, and improving recognition accuracy.

[0052] The corpus training tool is used to optimize the corpus training process by using the model correction algorithm and the corpus training algorithm, screen a large amount of corpus, extract the key parts, and only configure part of the corpus, while all similar corpora can be identified and modeled. The corpus training algorithm mainly includes parsing the corpus structure, parsing out the parameters such as equipment, time, and location, and parsing the intent of the corpus, screening out corpora with the same intent, performing comparative analysis, and extracting keywords or key words.

[0053] Model generation tool, used to export trained model files.

[0054] First, new corpus is added to a pattern file and a dictionary file, and a regular expression is used to parse the pattern file, in which the instruction syntax is stored.

[0055] The dictionary file is trained to obtain a hot word model, and then sentences are parsed from the paradigm file and the dictionary file, and trained together to obtain a professional model.

[0056] The organized monitoring corpus is trained separately using the general recognition model training method to obtain a professional model.

[0057] Hot word modeling is performed based on the nouns of primary and secondary equipment in the power grid and the power grid dictionary for subsequent wake-up word and professional word judgment.

[0058] In the corpus model training phase, the corpus is spliced ​​according to the power grid equipment model dictionary and the monitoring professional dictionary in combination with the instruction format, and each template is simplified according to the instruction format. The dictionary name is used in the template to replace the specific value in the instruction. For example, query the voltage of $(st_name)$(line_name), where $(st_name) and $(line_name) correspond to the plant dictionary and the line dictionary respectively. The power grid equipment model dictionary and the monitoring professional dictionary are read and filled into the instruction format template to complete the automatic generation of monitoring professional speech recognition corpus. The corpus format is as follows Figure 3 As shown, the generated corpus is saved in the database and a corpus file in JSON or TXT format is generated.

[0059] The generated JSON corpus file consists of specific corpus text and its entities and intents. The intent field is the intent name, that is, the function name corresponding to the corpus sentence, the text field is the text content of the corpus, and the entities field is a collection of entities in the corpus text, where each entity contains start (entity starting position), end (entity ending position), value (entity content), and entity (entity name).

[0060] The newly generated corpus is input into the corpus incremental training tool for incremental training to obtain the latest professional monitoring model.

[0061] Step 3: Export the trained professional model for power grid equipment monitoring into a professional acoustic model and a professional language model, and export the general recognition model into a general acoustic model and a general language model. Input the speech signal segments to be recognized into the professional acoustic model and the professional language model, as well as the general acoustic model and the general language model, respectively. Calculate the difference in word sequence scores of the two models based on the WFST-based Viterbi decoding algorithm. For phoneme segments whose score difference is higher than a preset threshold, use the model corresponding to the highest score for decoding and output. For phoneme segments whose score difference is not higher than a preset threshold, determine whether the phoneme segments contain professional power grid words based on the hot word model. If so, use the professional model for power grid equipment monitoring for decoding and output. If not, use the general recognition model for decoding and output.

[0062] In the speech recognition process, professional models are integrated and fine-tuned with general recognition models.

[0063] The speech signal segments to be recognized are input into the acoustic model and language model of the professional model and the general recognition model respectively to obtain the corresponding acoustic score and language score. The acoustic model is a knowledge representation of acoustic, phonetic, environmental variables, as well as differences in speaker gender and accent. The language model is a knowledge representation of a set of character sequences.

[0064] The decoder uses the Viterbi decoding algorithm based on WFST (weighted finite state transducer) to calculate the difference in word sequence scores between the professional model and the general model for confidence judgment. For phoneme segments with a score difference of more than 20%, the judge takes the model decoding output corresponding to the highest score. If the difference does not exceed 20%, the hot word model is used to further determine whether the phoneme segment contains professional power grid words. If so, the professional model is used for decoding and output; if not, the general recognition model is used for decoding and output.

[0065] like Figure 4As shown in the figure, the process of speech recognition from speech signal input to recognition result output shows that the decoder gives the final recognition result through arbitration of the decoder module according to the acoustic model, pronunciation dictionary, and language model. The decoder converts the input speech feature vector sequence into a character sequence according to the scores given by the acoustic model and language model.

[0066] The acoustic model is constructed using statistical acoustics, which is suitable for power grid monitoring terminology and meets the needs of monitoring professionals. It processes the vocabulary set and pronunciation of monitoring professional dictionaries and provides mapping between acoustic modeling and language modeling.

[0067] The acoustic model is to link the observed features of the speech signal with the speech modeling unit of the sentence, that is, to calculate the acoustic model P(X|W). The Hidden Markov Model (HMM) is usually used to solve the indefinite length relationship between speech and text.

[0068] P(X|W)=P(w1)P(x1|w1)P(w2|w1)P(x2|w2)P(w2|w2)P(x3|w2)

[0069] Among them, the initial state probability P(w1) and the state transition probability P(w2|w1), P(w2|w2) are calculated by conventional statistical methods, and the emission probabilities P(x1|w1), P(x2|w2), and P(x3|w2) are solved by mixed Gaussian models GMM or deep neural networks DNN. The acoustic model is constructed using DNN-HMM (deep neural network-hidden Markov model). The strong learning ability of DNN can improve the recognition performance. The schematic diagram of the acoustic model is shown below. Figure 5 shown.

[0070] The acoustic model construction process is as follows Figure 6 As shown, the speech data comes from the accumulated audio. In audio standardization, different formats are unified into wav or pcm formats; in the feature extraction stage, the speech signal is converted into a speech feature vector that can be processed by the computer. The extraction methods include Mel cepstral coefficients (mfcc), Fbank features (fbank) based on filter banks, etc.; after feature extraction, the acoustic model is trained, and the temporal connection classification model (CTC) is used to enumerate all possible alignment methods, and then these probabilities are accumulated to calculate the probability distribution of different outputs at each moment to predict possible results. At the same time, a blank character is introduced to distinguish consecutive identical characters, thereby distinguishing some similar words and improving the search speed of speech recognition. The acoustic model integrates the knowledge of acoustics and phonetics, takes the features extracted by the feature extraction module as input, and generates an acoustic model score.

[0071] In view of the large number of professional terms and high complexity of equipment naming in the power industry, a language model suitable for power grid monitoring terms is constructed. For the situations of digital numbers, special nouns, special pronunciations, etc. in the power industry, a language model that is mixed and adapted to the power grid monitoring profession is constructed. A language model that supports the construction of a Chinese-English mixed corpus based on the power grid monitoring context and data augmentation is provided. Language modeling can use various language models including regular languages and context-free grammars, and mainly adopts statistical-based N-gram and its variants.

[0072] The text data contains data from different fields. The normalization of the text mainly converts special symbols, punctuation marks, numbers, etc. contained in the text into Chinese characters for convenient processing. After normalization, language model training is carried out. The training method is mainly based on the algorithms of the N-gram model and neural network for training. And the language models in different fields can enhance the generalization ability through interpolation. The purpose of the language model is to, according to the results output by the acoustic model, rank the candidate words according to the scores of the language model in descending order and give the text sequence with the highest probability as the speech recognition result.

[0073] The language model mainly expands the text corpus provided by the user and expands the expressions according to the common power grid word list, thus generating a large amount of text corpus. In addition, a crawler script is used to collect a large amount of power grid-related text corpus from the website to enhance the generalization of the text and improve the recognition effect.

[0074] For the special pronunciations (including Chinese and English) in the power grid field, such as "#1" is pronounced as "No. 1", "0" is pronounced as "dong", etc., a dedicated pronunciation dictionary is customized to retain the user's usage habits and improve the user experience.

[0075] The beneficial effects of the present invention are as follows. Compared with the prior art, the present invention provides a method and system for constructing a speech recognition hybrid model for power grid equipment monitoring, which reduces the model training cost and can complete the model training on the existing server; reduces the model training time. Each training does not involve the general language model part and performs incremental training on the original monitoring professional model. The training duration is only about 10 min / time to 30 min / time, which is convenient and fast. In terms of social and economic benefits, the recognition accuracy of monitoring professional terms reaches 100%, and the recognition accuracy of daily general terms is higher than 98%.

[0076] The present invention can be a system, a method and / or a computer program product. The present invention also discloses a speech recognition hybrid model construction system for power grid equipment monitoring based on the aforementioned method for constructing a speech recognition hybrid model for power grid equipment monitoring, including:

[0077] Corpus collection module, used to collect document information related to power grid equipment and perform text preprocessing to form a corpus set in the field of power grid equipment monitoring;

[0078] A model training module is used to fill the power grid equipment monitoring field corpus set into the instruction format template to form a monitoring professional speech recognition corpus, use the corpus incremental training tool to load the monitoring professional speech recognition corpus into the model training algorithm, and perform incremental training on the power grid equipment monitoring professional model and the hot word model respectively;

[0079] A speech recognition module is used to export the trained professional model for power grid equipment monitoring into a professional acoustic model and a professional language model, and to export the general recognition model into a general acoustic model and a general language model, and to input the speech signal segments to be recognized into the professional acoustic model and the professional language model, as well as the general acoustic model and the general language model, respectively, and to calculate the difference in word sequence scores of the two models based on the WFST Viterbi decoding algorithm. For phoneme segments whose score difference is higher than a preset threshold, the model corresponding to the highest score is used for decoding and outputting. For phoneme segments whose score difference is not higher than a preset threshold, it is determined based on the hot word model whether the phoneme segments contain professional power grid words. If so, the professional model for power grid equipment monitoring is used for decoding and outputting. If not, the general recognition model is used for decoding and outputting.

[0080] Based on the spirit of the present invention, those skilled in the art can easily think that a computer program product can be obtained based on the aforementioned method for building a speech recognition hybrid model for power grid equipment monitoring. The computer program product may include a computer-readable storage medium, which carries computer-readable program instructions for enabling a processor to implement various aspects of the present disclosure. That is, the present application also includes a terminal, including a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps according to the aforementioned method for building a speech recognition hybrid model for power grid equipment monitoring.

[0081] Computer readable storage medium can be a tangible device that can keep and store the instructions used by the instruction execution device. Computer readable storage medium can be, for example, - but not limited to - electrical storage device, magnetic storage device, optical storage device, electromagnetic storage device, semiconductor storage device or any suitable combination of the above. More specific examples (non-exhaustive list) of computer readable storage medium include: portable computer disk, hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disk read-only memory (CD-ROM), digital versatile disk (DVD), memory stick, floppy disk, mechanical encoding device, for example, punch card or groove protrusion structure with instructions stored thereon and any suitable combination of the above. Computer readable storage medium used here is not interpreted as instantaneous signal itself, such as radio wave or other free propagating electromagnetic wave, electromagnetic wave propagated by waveguide or other transmission medium (for example, light pulse by optical fiber cable) or electrical signal transmitted by wire.

[0082] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing / processing device.

[0083] The computer program instructions for performing the operation of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages-such as Smalltalk, C++, etc., and conventional procedural programming languages-such as "C" language or similar programming languages. Computer-readable program instructions may be executed completely on a user's computer, partially on a user's computer, executed as an independent software package, partially on a user's computer, partially on a remote computer, or completely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any type of network-including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., using an Internet service provider to connect via the Internet). In some embodiments, an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA) may be personalized by utilizing the state information of a computer-readable program instruction, and the electronic circuit may execute a computer-readable program instruction, thereby realizing various aspects of the present disclosure.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for constructing a speech recognition hybrid model for power grid equipment monitoring, characterized in that: The following steps are involved: Collect relevant document information of power grid equipment and perform text preprocessing to form a corpus in the field of power grid equipment monitoring; Fill the power grid equipment monitoring domain corpus set into the instruction format template to form monitoring professional speech recognition corpus, use the corpus incremental training tool to load the monitoring professional speech recognition corpus into the model training algorithm, and perform incremental training on the power grid equipment monitoring professional model and the hot word model respectively; The trained professional model for power grid equipment monitoring is exported as a professional acoustic model and a professional language model, and the universal recognition model is exported as a universal acoustic model and a universal language model. The speech signal segments to be recognized are respectively input into the professional acoustic model and the professional language model, as well as the universal acoustic model and the universal language model. The difference in word sequence scores of the two models is calculated based on the WFST-based Viterbi decoding algorithm. For phoneme segments whose score difference is higher than a preset threshold, the model corresponding to the highest score is used for decoding and output. For phoneme segments whose score difference is not higher than a preset threshold, it is determined based on the hot word model whether the phoneme segments contain professional power grid words. If so, the professional model for power grid equipment monitoring is used for decoding and output. If not, the universal recognition model is used for decoding and output.

2. The method for constructing a speech recognition hybrid model for power grid equipment monitoring according to claim 1, characterized in that: The collecting of power grid equipment related document information and performing text preprocessing to form a corpus in the field of power grid equipment monitoring further includes: Read the power grid equipment model, run the real-time library and the historical library, organize the power grid equipment model dictionary, including the latest rules and regulations, procedures and specifications, work instructions, office documents, power grid dictionary, equipment model, interactive instructions, and convert special symbols in the power grid equipment model into nouns; The natural language processing algorithm is used to perform segmentation and sentence processing, and the processed sentences are analyzed by word segmentation according to the power grid equipment model dictionary. The equipment parameters and monitoring vocabulary in the sentences are analyzed to build a professional monitoring dictionary.

3. The method for constructing a speech recognition hybrid model for power grid equipment monitoring according to claim 2, characterized in that: The corpus incremental training tool comprises: Monitoring vocabulary expansion tool, used to expand the power grid monitoring vocabulary, by importing files, adding new interfaces, obtaining all the words that need to be expanded, adding new words, maintaining the original words, updating the power grid monitoring vocabulary, and enriching and expanding the power grid monitoring vocabulary; Model loading tool, used to load different models, load old models for incremental training of new models, and parse the relationship between intent, corpus, and entities in the model; Corpus loading tool, used to load corpus during model training, extract corpus in text file format, generate different JSON format content according to the format requirements of model training, load it into the model training algorithm, and provide data for model training; Model correction tool, used to use model correction algorithm to provide a variety of different corpora for multiple training of similar corpora, and obtain a new recognition accuracy after a preset number of trainings, and obtain the final training model after meeting the requirements; Corpus training tools, which are used to optimize the corpus training process by using model correction algorithms and corpus training algorithms, screen the corpus, and extract key parts; Model generation tool, used to export trained model files.

4. The method for constructing a speech recognition hybrid model for power grid equipment monitoring according to claim 3, characterized in that: The loading of corpus further comprises: According to the power grid equipment model dictionary and the monitoring professional dictionary, the corpus is spliced ​​in combination with the instruction format, and each template is simplified according to the instruction format. The dictionary name is used in the template to replace the specific value in the instruction. The power grid equipment model dictionary and the monitoring professional dictionary are read and filled into the instruction format template to complete the automatic generation of monitoring professional speech recognition corpus. Then the generated corpus is saved in the database, and a corpus file in JSON format or TXT format is generated.

5. The method for constructing a speech recognition hybrid model for power grid equipment monitoring according to claim 4, characterized in that: For a corpus file in JSON format, the corpus file includes the corpus text content text and its entities and intents intent, wherein each entity includes the entity starting position start, the entity ending position end, the entity content value, and the entity name entity.

6. A speech recognition hybrid model construction system for power grid equipment monitoring, characterized in that: include: Corpus collection module, used to collect document information related to power grid equipment and perform text preprocessing to form a corpus set in the field of power grid equipment monitoring; A model training module is used to fill the power grid equipment monitoring field corpus set into the instruction format template to form a monitoring professional speech recognition corpus, use the corpus incremental training tool to load the monitoring professional speech recognition corpus into the model training algorithm, and perform incremental training on the power grid equipment monitoring professional model and the hot word model respectively; A speech recognition module is used to export the trained professional model for power grid equipment monitoring into a professional acoustic model and a professional language model, and to export the general recognition model into a general acoustic model and a general language model, and to input the speech signal segments to be recognized into the professional acoustic model and the professional language model, as well as the general acoustic model and the general language model, respectively, and to calculate the difference in word sequence scores of the two models based on the WFST Viterbi decoding algorithm. For phoneme segments whose score difference is higher than a preset threshold, the model corresponding to the highest score is used for decoding and outputting. For phoneme segments whose score difference is not higher than a preset threshold, it is determined based on the hot word model whether the phoneme segments contain professional power grid words. If so, the professional model for power grid equipment monitoring is used for decoding and outputting. If not, the general recognition model is used for decoding and outputting.

7. The speech recognition hybrid model construction system for power grid equipment monitoring according to claim 6, characterized in that: The corpus collection module is further used to: Read the power grid equipment model, run the real-time library and the historical library, organize the power grid equipment model dictionary, including the latest rules and regulations, procedures and specifications, work instructions, office documents, power grid dictionary, equipment model, interactive instructions, and convert special symbols in the power grid equipment model into nouns; The natural language processing algorithm is used to perform segmentation and sentence processing, and the processed sentences are analyzed by word segmentation according to the power grid equipment model dictionary. The equipment parameters and monitoring vocabulary in the sentences are analyzed to build a professional monitoring dictionary.

8. The speech recognition hybrid model construction system for power grid equipment monitoring according to claim 7, characterized in that: The corpus incremental training tool comprises: Monitoring vocabulary expansion tool, used to expand the power grid monitoring vocabulary, by importing files, adding new interfaces, obtaining all the words that need to be expanded, adding new words, maintaining the original words, updating the power grid monitoring vocabulary, and enriching and expanding the power grid monitoring vocabulary; Model loading tool, used to load different models, load old models for incremental training of new models, and parse the relationship between intent, corpus, and entities in the model; Corpus loading tool, used to load corpus during model training, extract corpus in text file format, generate different JSON format content according to the format requirements of model training, load it into the model training algorithm, and provide data for model training; Model correction tool, used to use model correction algorithm to provide a variety of different corpora for multiple training of similar corpora, and obtain a new recognition accuracy after a preset number of trainings, and obtain the final training model after meeting the requirements; Corpus training tools, which are used to optimize the corpus training process by using model correction algorithms and corpus training algorithms, screen the corpus, and extract key parts; Model generation tool, used to export trained model files.

9. The speech recognition hybrid model construction system for power grid equipment monitoring according to claim 8, characterized in that: The corpus loading tool is further used to: According to the power grid equipment model dictionary and the monitoring professional dictionary, the corpus is spliced ​​in combination with the instruction format, and each template is simplified according to the instruction format. The dictionary name is used in the template to replace the specific value in the instruction. The power grid equipment model dictionary and the monitoring professional dictionary are read and filled into the instruction format template to complete the automatic generation of monitoring professional speech recognition corpus. Then the generated corpus is saved in the database, and a corpus file in JSON format or TXT format is generated.

10. The speech recognition hybrid model construction system for power grid equipment monitoring according to claim 9, characterized in that: For a corpus file in JSON format, the corpus file includes the corpus text content text and its entities and intents intent, wherein each entity includes the entity starting position start, the entity ending position end, the entity content value, and the entity name entity.

11. A terminal comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is used to operate according to the instructions to execute the steps of the method for constructing a speech recognition hybrid model for power grid equipment monitoring according to any one of claims 1-5.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method for constructing a speech recognition hybrid model for power grid equipment monitoring described in any one of claims 1 to 5 are implemented.

Images