CN112215012A - Power distribution network maintenance list safety measure semantic analysis method based on conditional random field - Google Patents

Abstract

The invention relates to a semantic analysis method for a power distribution network maintenance list safety measure based on a conditional random field. The method comprises the steps of establishing a power regulation and control corpus on the basis of realizing a natural semantic analysis basic algorithm engine by adopting a machine learning mode based on a conditional random field, simultaneously refining a historical text of application form safety measures, forming analysis samples and training, using a trained model for text analysis of the application form safety measures, and if semantic rule changes occur, only adding the analysis samples and training to realize the text analysis of the application form safety measures.

Description

Power distribution network maintenance list safety measure semantic analysis method based on conditional random field

Technical Field

The invention belongs to the technical field of electric power, and particularly relates to a semantic analysis method for a power distribution network maintenance order safety measure based on a conditional random field.

Background

The essence of intelligent analysis of the characters of the safety measures of the maintenance application form in the power distribution network is an analysis process of natural language, and people mainly conduct basic research work of natural language processing before 1956. Shannon, 1948, applied probabilistic models of discrete markov processes to automata describing languages, while incorporating the concept of "entropy" into language processing. And Kleene studied finite automata and regular expressions in the same phase. In 1956, Chomsky proposed a context free grammar. These works lead to the emergence of two different natural language processing methods based on rules and based on probabilities, so that research in the field is divided into two major camps of symbolic dispatching by using a rule method and random dispatching by using a probability method, and further, disputes about whether the two methods are good or bad are triggered for decades. In 1956, after artificial intelligence emerged, natural language processing was rapidly merged into the research of artificial intelligence. Cognitive psychology was proposed by Neisser, American psychologist in 1967, to link natural language processing with human cognition. In the early 70 s, since some problems in the natural language processing research could not be solved in a short time and new problems were continuously emerging, the natural language processing research entered the valley period. Hidden Markov Model (HMM) based statistical methods and speech analysis have made significant progress during this period. In the 80 s, the recovery began with finite state models and empirical studies. After 90 s, with the great increase of the speed and the storage capacity of the computer, the material basis of natural language processing is greatly improved, and the commercialized development of voice and language processing becomes possible; meanwhile, the development of network technology and Internet commercialization in 1994 have made the demand for natural language-based information retrieval and information extraction more prominent. From the end of the 90 s to the beginning of the 21 st century, it became increasingly recognized that natural language processing was unsuccessful using either rule-based methods alone or statistical-based methods alone. Statistical, instance-based, and rule-based corpus technologies have been developed vigorously during this time, various processing technologies have been fused, and research on natural language processing has been flourishing.

At present, with the continuous development of computer technology, the text analysis of the safety measures of the power distribution network overhaul application form mostly adopts an accurate word matching method of a vector space model, namely, words input by a user are accurately matched with words existing in a vector space, so that the analysis of the safety measure text is realized, but the following problems still exist:

1. the analytical model realized by pure codes is poor in universality, and the related technology is difficult to popularize;

2. the technical requirements on the customizers of the analytical model are high, and after the rule changes, the original functions cannot be influenced when the program is modified, so that the stability of the system is poor

3. The utilization rate of the historical data is low, and the analytic samples cannot be directly extracted from the historical data.

4. The change of the local rule also needs to be modified by a program, and the change of the local rule cannot be adapted through simple sample customization or configuration modification.

Disclosure of Invention

The invention aims to provide a power distribution network maintenance list safety measure semantic analysis method based on a conditional random field, which is characterized in that a machine learning mode based on the conditional random field is adopted, a power regulation and control corpus is created on the basis of realizing a natural semantic analysis basic algorithm engine, meanwhile, a historical text of application list safety measures is refined, analysis samples are formed and trained, a trained model is used for text analysis of the application list safety measures, and if the semantic rule changes, the text analysis of the application list safety measures can be realized only by adding the analysis samples and training.

In order to achieve the purpose, the technical scheme of the invention is as follows: a power distribution network maintenance list safety measure semantic analysis method based on a conditional random field comprises the following steps:

step S1, constructing an electric power regulation corpus:

the construction of the power regulation language database comprises the construction of a maintenance request sheet sample database, a physical environment database and a term database; wherein, the overhaul application form sample library is an overhaul application form safety measure; the physical environment library theme is a keyword sample library, and the keyword sample library is divided into a CIM type library, an action type library, a state type library and a function type library; the term library comprises units, fixed collocation and inherent phrases which cannot be classified into the physical environment library;

step S2, training sample establishment:

extracting historical texts of safety measures of the overhaul application form, performing word segmentation standard design, and further constructing part-of-speech tagging training samples, wherein the part-of-speech tagging training samples adopt a form that each character and part-of-speech tags thereof occupy one line, and sentences are separated by spaces;

step S3, materialized participle:

according to the word segmentation standard of the step S2, carrying out word segmentation operation on the safety measures of the maintenance request form to be analyzed in the maintenance request form sample base by applying a Viterbi algorithm based on dynamic programming, and then determining a unique entity corresponding to the word after word segmentation by using a physical environment base;

step S4, analysis result data format:

and outputting a text analysis result of the safety measures of the overhaul application form to be analyzed in the overhaul application form sample library by adopting XML.

Compared with the prior art, the invention has the following beneficial effects:

the method adopts a machine learning mode based on a conditional random field, creates a power regulation and control corpus on the basis of realizing a natural semantic analysis basic algorithm engine, simultaneously refines a historical text of application form safety measures, forms analysis samples and trains, uses the trained model for text analysis of the application form safety measures, and can realize text analysis of the application form safety measures only by adding the analysis samples and training if the semantic rule changes.

Drawings

FIG. 1 is a schematic flow chart of the method of the present invention.

FIG. 2 is an example of a service application form sample library according to the present invention.

FIG. 3 is an exemplary TXT format of the physical environment library of the present invention.

FIG. 4 is an example of a part-of-speech tagging corpus sample according to the present invention.

FIG. 5 is an exemplary materialization model of the present invention.

FIG. 6 is an example XML parsing form of overhaul application form security.

FIG. 7 is a semantic parsing function interface.

Detailed Description

The technical scheme of the invention is specifically explained below with reference to the accompanying drawings.

As shown in fig. 1, the invention provides a semantic analysis method for a power distribution network maintenance order safety measure based on a conditional random field, which comprises the following steps:

step S1, constructing an electric power regulation corpus:

the construction of the power regulation language database comprises the construction of a maintenance request sheet sample database, a physical environment database and a term database; wherein, the overhaul application form sample library is an overhaul application form safety measure; the physical environment library theme is a keyword sample library, and the keyword sample library is divided into a CIM type library, an action type library, a state type library and a function type library; the term library comprises units, fixed collocation and inherent phrases which cannot be classified into the physical environment library;

step S2, training sample establishment:

extracting historical texts of safety measures of the overhaul application form, performing word segmentation standard design, and further constructing part-of-speech tagging training samples, wherein the part-of-speech tagging training samples adopt a form that each character and part-of-speech tags thereof occupy one line, and sentences are separated by spaces;

step S3, materialized participle:

according to the word segmentation standard of the step S2, carrying out word segmentation operation on the safety measures of the maintenance request form to be analyzed in the maintenance request form sample base by applying a Viterbi algorithm based on dynamic programming, and then determining a unique entity corresponding to the word after word segmentation by using a physical environment base;

step S4, analysis result data format:

and outputting a text analysis result of the safety measures of the overhaul application form to be analyzed in the overhaul application form sample library by adopting XML.

The following is a specific implementation of the present invention.

A semantic analysis method for a power distribution network maintenance list safety measure based on a conditional random field is realized as follows:

1. construction of electric power regulation corpus

The corpus is the basis of follow-up research, and the corpus which can support follow-up semantic analysis and can be used in detail is designed aiming at the safety measures of the power distribution network maintenance application form. Because the corpus is required to be established to balance the corpus range and the coverage type as much as possible, and the characteristics of the safety measure text object are combined, the whole corpus is divided into three parts, namely an overhaul application form sample library, a physical environment library and a term library in the corpus establishing process.

1.1 examination and repair application form sample library

And the overhaul application form sample library is used for taking overhaul application form safety measures. The sample library can be used for early-stage scheme verification, model training and knowledge body construction in the semantic analysis process. An example of a service application form sample library is shown in fig. 2.

1.2 physical Environment library

The physical environment library theme is a keyword sample library, and the keyword sample library is divided into a CIM type library, an action type library, a state type library and a function type library. The libraries contain most of physical meaning information under the power grid rule, and can be added to general natural language analysis in an auxiliary mode to form a specific analysis rule for power grid dispatching, so that a specific physical environment is further built. The physical environment library may be used for context recognition, semantic recognition, rule making, and the like.

Taking a CIM type keyword sample library as an example, the sample library table includes nominal information such as most device names in the tickets, for example: the key words with high occurrence rate such as transformer substations, voltage grades, synchronous motors, buses, transformers, alternating current lines, circuit breakers, grounding switches, disconnecting switches and towers and the key words with low occurrence rate such as series capacitors, series reactors, parallel capacitors, parallel reactors, PT, CT, fuses, equivalent loads, lightning arresters and arc suppression coils. According to actual requirements, each sample base can be expanded and modified. The sample library format is initially intended as a TXT text format (as shown in fig. 3). Table 1 is an example physical environment library illustration.

Table 1 physical environment library illustration:

serial number	Name of field	Meaning of a field	Type (B)	Default value	Remarks for note
						1．	CIM type	Name of the device	varchar	null
2.	Status of state	Device status	varchar	null
						3.	Type of action	Device action name	varchar	null
4.	Type of function	Device function name	varchar	null

1.3 term library

The term library comprises units, fixed collocation, inherent phrases which cannot be classified in the physical environment library and the like, namely the term library is used as a supplement of the physical environment library and can play a role in auxiliary recognition during matching, recognition and escape. See table 2 for an example of a portion of the library of device names for a certain area.

TABLE 2 partial equipment name term library for a certain area

Line name	Line name (Default)	Transformer substation	Name of switch
				New 193 line	Car city 214	Garden transformer	Red tower 162 switch
Huafang
	116 line	Wuxin 206	Transverse port transformer	Spring 169 switch
Xinming 330 line					Black copper 222	Change from Wentang to Tan	Clockwise 489 switch
	Garden 307 line	Lion mountain 203	Landscape transformer	Phoenix 432 switch
Garden 305 line					To be used 213	Big universal transformer	16 switch for river crossing
	Taixin 3A4 circuit	Wuhai 212	City change of thank-xifrage	Tianyang 822 switch
Copper conductor 159 circuit					Longtan 211	Phoenix transformer	Black copper 222 switch
	Zhongfu 1G1 circuit	Weft one 209	Change of uterus	Zhu Dong 519 switch

2. Training and testing sample design

The common nouns are analyzed and classified by extracting the nouns of the power dispatching system in the sample, and an available noun word list is formed and serves as one of bases of word segmentation. The marking rule is formed through the summary analysis of the analysis object, so that the addition of a new sample is facilitated, and the expandability of the system is improved.

2.1 Power dispatching System noun (see Table 3)

TABLE 3 Power dispatching system nomenclature table

Serial number	Station name	Device name
				1.	Transverse port transformer	Huafang	116 line
2.	Change from Wentang to Tan	Xinming 330 line
			3.	Landscape transformer	Garden 307 line
4.	Big universal transformer	Garden 305 line
			5.	Ancient water change	Segmented 00 switch
6.	Phoenix transformer	Red tower 162 switch
			7.	Change of uterus	Spring 169 switch
8.	Change of village	Clockwise 489 switch
			9.	Change of Licun	Phoenix 432 switch
10.	Xie Cheng	New 193 line

2.2 participle Standard design

The following rules are segmentation rules for the first time segmentation: the method is subdivided into the most basic semantic units as much as possible, and partial rules are shown in tables 4-6:

2.3 part-of-speech tagging training sample establishment

The model adopts a method of joint labeling with Chinese text, namely, words and the parts of speech to which the words belong are identified at the same time. Generating a label set suitable for joint labeling by adopting a mode of connecting Chinese word segmentation labels and part-of-speech labels, such as: "B _ D" represents the beginning character of the number.

The part-of-speech tagging training sample adopts a form that each character and a label thereof occupy one line, and sentences are separated by spaces. The word segmentation text is derived from sentences actually used in the actual power system scheduling control process, 1015 sentences are commonly used in the experiment process, and 8943 characters are counted. Fig. 4 shows a part of the part-of-speech tagging corpus sample.

3. Materialized participle based on corpus

And (3) solving the optimal state sequence value for a given hidden Markov model and an observation sequence value by utilizing the word segmentation standard of 2.2 and applying a Viterbi algorithm based on dynamic programming, wherein the solving process of the Viterbi algorithm.

The present invention separates words by using the Viterbi algorithm, and corresponds specific words to a unique entity by using the physical environment library, which needs to be converted into an materialization model as shown in fig. 5.

As shown in fig. 5, there are multiple entities corresponding to "5032 switch" in the physical environment of the power grid, but the substation entities corresponding to "zhongxing change" and "shun-ploughing change" are unique in the physical environment of the whole power grid. The materialization sub-model determines the entities corresponding to the switches of 500kV and 5032 by determining the substation entities 'crowd-going transformation' and 'shun-ploughing transformation'. The attribute values in the entity word dictionary include five attribute values such as "entity ID", "container ID", "abstract word", "entity word", "unique name correspondence container", and the like. When the materialization sub-model is established, the materialization sub-model is established according to the hierarchical relationship embodied by the entity ID and the container ID. The meaning of the attribute value of "unique name corresponding to container" is that under a certain container, the entity corresponding to the word is unique. In fig. 5, the "container ID" of the entity corresponding to the "5032 switch" is the "entity ID" of the entity corresponding to the "500 kV". The entity corresponding to "500 kV" can be uniquely determined under a container such as "substation". The abstract word corresponding to the 'popular transformation' is 'transformer substation'.

Taking the text "the crow's change 500kV crow's line/the dragon wire 5032 switch first and second groups of control power supply disappear" as an example, when the entity ID of an entity is uniquely determined, the entity is considered to be uniquely determined, as shown in the following table, before the materialization, only the word "the crow's change" has the uniquely corresponding entity ID, i.e., the entity corresponding to the word is considered to be uniquely determined, other words correspond to a plurality of entity IDs, i.e., the entity IDs corresponding to other words are considered not uniquely determined, and limited by the influence of the space of the table, in the actual environment, the entity IDs corresponding to the "500 kV" and "5032" switch are more than three listed in the table. After the materialization submodel is carried out, the unique corresponding entity can be found in the whole physical environment of the power grid by the words. Table 7 shows comparison results before and after materialization.

There are multiple entities in the physical environment of the power grid, corresponding to "500 kV" and "5032 switch", but the substation entities corresponding to "crowd change" and "shun change" are unique in the physical environment of the whole power grid. The materialization sub-model determines the entities corresponding to the switches of 500kV and 5032 by determining the substation entities 'crowd-going transformation' and 'shun-ploughing transformation'.

4. Analysis result data format design

The markup language, which uses XML for marking electronic documents to have a structure, can be used for marking data and defining data types, and is a source language that allows a user to define his own markup language. XML is a subset of the Standard Generalized Markup Language (SGML) and is well suited for Web transport. XML provides a unified way to describe and exchange structured data that is independent of the application or vendor.

The currently used semantic parsing result form is an XML form, taking an instruction "the 10kV huafang 116 line is changed from maintenance to operation" as an example, and the XML parsing form is shown in fig. 6.

In the above example, the node labels are xml4nlp, para, send, word, arg, and the meanings are shown in Table 8:

XML is an information model for communication between modules, is a modular, universal, objectified, regularized information representation with definite physical meaning, and formally is an XML result containing specific nouns and their physical environment tags. Extensibility of model parsing (escape) may be based on additions and deletions of physical environment library statements. Node tags can be set up and changed (number, meaning) to instantiate (objectify) the physical environment. Compatibility can be achieved by node tagging to apply data results of different attributes (forms) of the neural network.

The knowledge body in the XML form is a standard form containing semantic meanings, is formed according to preset standards, and particularly, what contents can be set through the standards, and is read by a user to form a data structure.

5. Implementation of semantic parsing function

The semantic parsing function interface is shown in fig. 7, and the semantic parsing interface includes a parsing object, a parsing method, and a log record.

The analysis object part provides a pull-down selection menu of the physical environment to which the operation unit and the operation plant station wait for analysis instruction service belong, and can be used for selecting a corresponding regulation center so as to correspond to a model obtained by training of corpus of different fields, the operation content is a text to be analyzed, the analysis method part comprises the steps of analyzing application form security errors, instruction ticket tasks, instruction ticket steps, operation ticket tasks, operation ticket steps and the like, and the characters can be analyzed according to actual requirements. The log recording part is used for recording operation and debugging information in the parsing process.

The above are preferred embodiments of the present invention, and all changes made according to the technical scheme of the present invention that produce functional effects do not exceed the scope of the technical scheme of the present invention belong to the protection scope of the present invention.

Claims (1)

1. A power distribution network maintenance list safety measure semantic analysis method based on a conditional random field is characterized by comprising the following steps:

step S1, constructing an electric power regulation corpus:

the construction of the power regulation language database comprises the construction of a maintenance request sheet sample database, a physical environment database and a term database; wherein, the overhaul application form sample library is an overhaul application form safety measure; the physical environment library theme is a keyword sample library, and the keyword sample library is divided into a CIM type library, an action type library, a state type library and a function type library; the term library comprises units, fixed collocation and inherent phrases which cannot be classified into the physical environment library;

step S2, training sample establishment:

extracting historical texts of safety measures of the overhaul application form, performing word segmentation standard design, and further constructing part-of-speech tagging training samples, wherein the part-of-speech tagging training samples adopt a form that each character and part-of-speech tags thereof occupy one line, and sentences are separated by spaces;

step S3, materialized participle:

according to the word segmentation standard of the step S2, carrying out word segmentation operation on the safety measures of the maintenance request form to be analyzed in the maintenance request form sample base by applying a Viterbi algorithm based on dynamic programming, and then determining a unique entity corresponding to the word after word segmentation by using a physical environment base;

step S4, analysis result data format:

and outputting a text analysis result of the safety measures of the overhaul application form to be analyzed in the overhaul application form sample library by adopting XML.

Images