An information knowledge extraction method based on a power system, a controller and a storage medium

Abstract

The application discloses an information knowledge extraction method based on a power system, a controller and a storage medium, and the method comprises the following steps: reading non-structured data of an information document of the power system; performing data cleaning on the non-structured data to obtain power system data elements; performing preprocessing on the power system data elements to obtain words, and performing correction processing on the words by using a reverse maximum matching algorithm; based on a standard power system element database, establishing rules of data element information and entity extraction according to the construction direction of the power system, and realizing extraction of equipment entity category candidate words; performing recognition and filtering processing on the equipment entity category candidate words to obtain high-frequency nouns; performing entity joint extraction on equipment entity model candidate words to obtain high-frequency words with high confidence; generating equipment entity category and model labels according to the high-frequency nouns and the words with the highest confidence, and storing the labels in an index; and outputting equipment entity knowledge according to the index. The application improves the accuracy of naming equipment entity recognition of power system data elements.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power systems, and in particular to an information knowledge extraction method based on a power system, a controller and a storage medium. BACKGROUND

[0002] With the construction and development of the energy internet, the smart grid and the ubiquitous power internet of things, the long-term production and operation of the power system field has accumulated a large amount of power multi-source heterogeneous big data, including data of various links such as power production, transmission and sales. The big data can provide support for the construction of the smart grid. However, most of the accumulated power multi-source heterogeneous big data is non-structured data. The non-structured data has various formats and implicit data meanings, cannot be stored using a relational database, and can only be stored in different file forms. Massive multi-source heterogeneous data will bring challenges in storage, transmission and information processing, and therefore, how to automatically analyze and mine non-structured data in the power field has become one of the bottlenecks restricting the development of the smart grid at the present stage.

[0003] In related technologies, information extraction for non-structured data is mainly based on rules and dictionaries, conditional random fields (CRF) and classifiers. Such methods require a large amount of manually annotated templates and training data, have a relatively high requirement for domain knowledge, and have low scalability. In recent years, with the continuous development of deep learning models, entity recognition and relationship extraction methods based on deep learning have been improved in performance. However, due to the complex and diverse medical entity formation in the Chinese text of the power system, medical entities often appear in the form of declarative phrases, such as equipment entities. Therefore, the Chinese word segmentation and entity recognition task often faces a large number of nested entities, which causes the frequent occurrence of entity nesting in the Chinese text of the power system. However, the current model recognition entity method often ignores the unique entity nesting rules of the Chinese text of the power system and directly uses the sequence labeling method, which has problems such as low accuracy of domain entity recognition and inaccurate recognition of entities, and cannot accurately recognize and extract the content of the power grid data. SUMMARY

[0004] The present application aims to at least solve one of the technical problems existing in the prior art. To this end, the embodiments of the present application provide an information knowledge extraction method based on a power system, a controller and a storage medium, which realizes accurate recognition and extraction of power system data elements, and is beneficial to improving the accuracy of power system data element naming equipment entity recognition.

[0005] In a first aspect, the embodiments of the present application provide an information knowledge extraction method based on a power system, comprising:

[0006] An information document of a power system is acquired, unstructured data of the information document of the power system is read, and the unstructured data is converted into a txt format;

[0007] The unstructured data is subjected to data cleaning, classification, sorting, cataloging, and standardization processing to obtain power system data elements;

[0008] The power system data elements are preprocessed by a character embedding layer based on a RoBERTa model, a BiLSTM model, and a token speculation CRF layer based on an MFM model, combined with a power system text segmentation and part-of-speech tagging joint method to obtain segmented words, the segmented words are reorganized and vectorized, and a reverse maximum matching algorithm is used to correct the segmented words to obtain optimized segmentation results;

[0009] The optimized segmentation results are trained based on a power system special text library to obtain a standard power system element database, rules for data element information and entity extraction are established according to the construction direction of the power system to form a mapping rule library of data element information and equipment entities to realize equipment entity category candidate word extraction;

[0010] The equipment entity category candidate words are recognized and filtered to obtain high-frequency nouns;

[0011] The equipment entity model candidate words are subjected to entity joint extraction to obtain high-frequency words with high confidence;

[0012] The high-confidence high-frequency words are recognized to obtain the highest-confidence words;

[0013] The high-frequency nouns and the highest-confidence words are used to generate equipment entity categories and model labels corresponding to the power system data elements and store them in an index;

[0014] Equipment entity knowledge is output according to the index.

[0015] According to some embodiments of the present application, the recognition and filtering of the equipment entity category candidate words to obtain high-frequency nouns include:

[0016] A basic rule knowledge base of equipment entity application scenarios and entity boundaries and semantics is formed based on power system standards and specification knowledge;

[0017] Based on the basic rule knowledge base of equipment entity application scenarios and entity boundaries and semantics, a Chinese entity recognition method is fused with equipment entity nesting rules to convert the equipment entity recognition task into a joint training task of equipment entity boundary recognition and boundary head-tail relationship recognition;

[0018] In the decoding process of the joint training task, the decoding result is identified and filtered in combination with the entity nesting rules of the power system, so that the identification result conforms to the composition rules of the inner and outer entity nesting combination in the text.

[0019] According to the term standard of the power system, the device entity category candidate words are termized and standardized to obtain high-frequency nouns.

[0020] According to some embodiments of the present application, the power system data elements are preprocessed by combining the power system text segmentation and part-of-speech tagging joint method based on the character embedding layer of the RoBERTa model, the BiLSTM model, and the token speculation CRF layer of the MFM model to obtain the segmentation, including:

[0021] The power system data elements are input into the character embedding layer of the RoBERTa model to obtain word vectors reflecting context semantics and output the word vectors;

[0022] The word vectors are input into the BiLSTM model to capture the long-distance dependency relationship and context sequence information of the word vectors through the BiLSTM model, discard useless word vector sequences, and extract useful word vector sequences;

[0023] The useful word vector sequences are input into the token speculation CRF layer of the MFM model to jointly model the token sequence using the token speculation CRF layer of the MFM model, calculate the conditional probability of the input useful word vector sequences output in the token sequence;

[0024] The highest conditional probability is determined, and the useful word vector sequence corresponding to the highest conditional probability is selected as the segmentation.

[0025] According to some embodiments of the present application, the power system data elements are input into the character embedding layer of the RoBERTa model to obtain word vectors reflecting context semantics and output the word vectors, including:

[0026] The word vectors of the power system data elements are calculated with the first parameter matrix, the second parameter matrix, and the third parameter matrix respectively to obtain the first vector, the second vector, and the third vector with the same dimension as the word vector;

[0027] The attention matrix is constructed according to the first vector, the second vector, and the third vector;

[0028] The first vector, the second vector, and the third vector are mapped by different linear transformations, and are spliced through the attention matrix to obtain a multi-head attention matrix;

[0029] The multi-head attention matrix connection feedforward neural network encodes the power system data elements through linear transformation and a Relu activation function, and performs preliminary segmentation on the power system data elements using a Chinese word segmentation algorithm to obtain word vectors reflecting context semantics and output the word vectors.

[0030] According to some embodiments of the present application, the BiLSTM model includes a forget gate, an input gate, and an output gate. The word vectors are input into the BiLSTM model to capture long-distance dependency relationships and context sequence information of the word vectors, discard useless word vector sequences, and extract useful word vector sequences, including:

[0031] The word vectors are input into the forget gate of the BiLSTM model, which determines the information that needs to be discarded in the previous neuron through the forget gate. The input is the current word vector and the previous word vector, and the state of the previous neuron cell is mapped to 0-1, where 0 represents complete deletion and 1 represents complete retention. The result of the forget gate is obtained.

[0032] The input gate determines whether to record new word vectors in the neurons of the BiLSTM model. The input layer Sigmoid activation function determines the word vectors that need to be updated to update the neuron state, so as to determine the memory information that needs to be recorded. The word vector sequence that needs to be remembered is obtained.

[0033] The Sigmoid activation function of the output gate determines the part of the output neuron state. The neuron state is processed by the tanh function and multiplied by the output of the Sigmoid activation function to obtain a hidden layer state sequence with the same length as the word vector sequence. The long-distance dependency relationships and context sequence information of the word vectors are captured to discard useless word vector sequences and extract useful word vector sequences.

[0034] According to some embodiments of the present application, the reverse maximum matching algorithm is used to correct the segmentation to obtain an optimized segmentation result, including:

[0035] The characters composed of several consecutive word groups in the string to be re-segmented after the first segmentation are matched with the domain dictionary from right to left.

[0036] If the matching is successful, a new word is segmented to obtain an optimized segmentation result.

[0037] If the matching is unsuccessful, the leftmost word group in the first segmentation result is removed and matched with the domain dictionary. If the matching is successful, a new word is segmented to obtain an optimized segmentation result. If the matching is unsuccessful, the iteration is performed until only one word group is left in the first segmentation result, and an optimized segmentation result is obtained.

[0038] According to some embodiments of the present application, the entity joint extraction on the device entity model candidate words obtains high-frequency words with high confidence, including:

[0039] The device entity model candidate words are described by data elements and data cores to form a fixed feature expression form of power information.

[0040] According to the preset data element analysis and device entity extraction experience, the power system field factor is determined in combination with the standard text number, system classification number and detection report type, the power system field factor carries the data source and classification of the device entity model candidate words and the device entity extraction rule library serial number.

[0041] The feature vector of the device entity model candidate word is determined by bidirectional calculation of the long short-term memory network and the power system field factor, and the key features of the power system device entity model candidate word recognition are extracted by using the BiLSTM model.

[0042] The key features of the feature vector and the device entity model candidate word are combined and calculated to identify the device entity name and weight corresponding to the device entity model candidate word.

[0043] The global optimal sequence is labeled by the MFM model CRF layer, and the hidden state sequence is converted into the best label sequence, so as to realize the entity joint extraction on the device entity model candidate word and obtain high-frequency words with high confidence.

[0044] According to some embodiments of the present application, the confidence high-frequency words are identified to obtain the highest confidence word, including:

[0045] The distance between the confidence high-frequency words and the device model category is determined.

[0046] The distances are arranged from small to large to obtain a high-frequency word sequence table with high to low confidence.

[0047] Negative words are cleaned up in the high-frequency word sequence table, and the highest confidence high-frequency word in the high-frequency word sequence table is output to obtain the highest confidence word.

[0048] In a second aspect, the embodiments of the present application provide a controller, including a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor executes the computer program to perform the method in the technical solution of the first aspect.

[0049] In a third aspect, the embodiments of the present application provide a computer readable storage medium, characterized in that the computer readable storage medium stores computer executable instructions for causing a computer to execute the method in the technical solution of the first aspect.

[0050] The information knowledge extraction method based on the power system, the controller and the storage medium provided by the embodiments of the present application at least have one of the following advantages or beneficial effects: the information document of the power system is obtained, and the non-structured data of the information document of the power system is uniformly converted into a txt format, and then spaces, line breaks, invalid symbols and the like are cleaned through data cleaning, thereby improving the uniformity of the non-structured data; the non-structured data is classified, classified, layered, point-by-point collated, cataloged and standardized, so that the classification collation and standardization processing are realized, and power system data elements are obtained; the power system data elements are preprocessed through a character embedding layer based on a RoBERTa model, a BiLSTM model, a label inference CRF layer based on a MFM model, and a power system text segmentation and part-of-speech tagging combined method, so that segmentation is obtained, the segmentation is reorganized and vectorized, the power system data elements are vectorized after being divided into half sentences, the segmentation is corrected by using a reverse maximum matching algorithm, and an optimized segmentation result is obtained, thereby reducing the influence of an out-of-vocabulary word in the power system field on the segmentation effect. Then, the optimized segmentation result is trained based on a power system special text library, a standard power system element database is obtained, a standard element database with uniformity, fine granularity and identification can be established, rules of data element information and entity extraction are established according to the construction direction of the power system, a mapping rule library of data element information and equipment entities is formed, a good data foundation is provided for knowledge extraction of the power system, the recognition result of the equipment entity can conform to the composition rule of the inner and outer entity nesting combination in the actual text, the accuracy and efficiency of multi-modal power system data element naming equipment entity recognition are improved, and the low efficiency and inaccuracy of the current manual sampling are also compensated. The high-frequency nouns are obtained by recognizing and filtering the equipment entity category candidate words, the equipment entity model candidate words are subjected to entity joint extraction, the entity joint extraction is performed according to the upper entity and relationship concept extraction and the lower entity and relationship concept extraction, the high-frequency words with high confidence are obtained, the words with the highest confidence are obtained by recognizing the high-frequency words with high confidence, and the accuracy of power system data element naming equipment entity recognition is further improved. Finally, the equipment entity category and model label corresponding to the power system data element are generated according to the high-frequency nouns and the words with the highest confidence, and are stored in an index, the equipment entity knowledge is output according to the index, the accuracy of power system data element naming equipment entity recognition is improved, and thus the accurate recognition and extraction of the power system data are realized.

[0051] Other features and advantages of the present application will be set forth in the description that follows, and in part will be apparent from the description, or can be learned by practice of the application. The purposes and other advantages of the present application will be realized and attained by the structure particularly pointed out in the written description and claims hereof as well as the appended drawings. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 is a flowchart of a method for extracting information knowledge based on a power system according to an embodiment of the present application;

[0053] Figure 2 is a flowchart of a method for identifying and filtering device entity category candidate words to obtain high-frequency nouns according to an embodiment of the present application;

[0054] Figure 3 is a flowchart of a method for preprocessing power system data elements to obtain word segmentation according to an embodiment of the present application;

[0055] Figure 4 is a flowchart of a method for inputting power system data elements into a character embedding layer of a RoBERTa model to obtain word vectors reflecting context semantics and outputting the word vectors according to an embodiment of the present application;

[0056] Figure 5 is a flowchart of a method for capturing long-distance dependency relationships and context sequence information of word vectors by a BiLSTM model according to an embodiment of the present application;

[0057] Figure 6 is a flowchart of a method for correcting word segmentation by a reverse maximum matching algorithm to obtain an optimized word segmentation result according to an embodiment of the present application;

[0058] Figure 7 is a flowchart of a method for performing entity joint extraction on device entity model candidate words to obtain high-frequency words with high confidence according to an embodiment of the present application;

[0059] Figure 8 is a flowchart of a method for identifying high-confidence words with the highest confidence according to an embodiment of the present application;

[0060] Figure 9 is a structural schematic diagram of a controller according to an embodiment of the present application. DETAILED DESCRIPTION

[0061] The detailed description of the application will be described in detail in this part, the preferred embodiments of the application are shown in the drawings, the role of the drawings is to supplement the description of the text part with graphics, so that people can intuitively and visually understand each technical feature and the overall technical scheme of the application, but it cannot be understood as a limitation on the protection scope of the application.

[0062] In the description of the application, the meaning of one or more is one or more, the meaning of multiple is more than two, greater than, less than, more than, etc. are understood as not including the number, above, below, within, etc. are understood as including the number, "any one" means one or more, "at least one of the following" and the like means any combination of these items, including any combination of single or multiple items. If the first, second is described for the purpose of distinguishing technical features, it cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features or implicitly indicating the order of indicated technical features.

[0063] It should be noted that the words such as setting, installing and connecting in the embodiments of the application should be understood broadly, and those skilled in the art can reasonably determine the specific meaning of the above words in the embodiments of the application in combination with the specific content of the technical scheme. For example, the term "connection" can be mechanical connection, electrical connection or can communicate with each other; it can be directly connected, or indirectly connected through an intermediate medium.

[0064] It should be noted that the technical features involved in each embodiment of the application described below can be combined with each other as long as there is no conflict between them.

[0065] Referring to Figure 1 , the Figure 1 is a flowchart of an information knowledge extraction method based on a power system provided by the embodiments of the application. The information knowledge extraction method based on the power system includes but is not limited to steps S100 to S900, specifically,

[0066] Step S100: obtaining the information document of the power system, reading the non-structured data of the information document of the power system, and converting the non-structured data into txt format;

[0067] Step S200: data cleaning, classification, sorting, cataloging and standardization processing are performed on the non-structured data to obtain power system data elements;

[0068] Step S300: The power system data elements are preprocessed by combining the power system text segmentation and part-of-speech tagging joint method through the character embedding layer based on the RoBERTa model, the BiLSTM model, and the token speculation CRF layer based on the MFM model, to obtain the segmentation, reorganize and vectorize the segmentation, and correct the segmentation by using the reverse maximum matching algorithm to obtain the optimized segmentation result;

[0069] Step S400: The optimized segmentation result is trained based on the power system special text word library to obtain a standard power system element database, rules of data element information and entity extraction are established according to the construction direction of the power system, a mapping rule library of data element information and equipment entities is formed to realize the extraction of equipment entity category candidate words;

[0070] Step S500: The equipment entity category candidate words are recognized and filtered to obtain high-frequency nouns;

[0071] Step S600: The equipment entity model candidate words are subjected to entity joint extraction to obtain high-confidence frequency words;

[0072] Step S700: The high-confidence frequency words are recognized to obtain the highest-confidence words;

[0073] Step S800: The equipment entity category and model labels corresponding to the power system data elements are generated according to the high-frequency nouns and the highest-confidence words and are stored in the index;

[0074] Step S900: The equipment entity knowledge is output according to the index.

[0075] In some embodiments of the present application, the information knowledge extraction method based on the power system includes: obtaining the information document of the power system, reading the non-structured data of the information document of the power system, and converting the word, pdf, txt, image and other non-structured data into txt format through document reading; then through data cleaning, the space, line feed, invalid symbol and the like are cleaned, the data cleaning of the non-structured data is realized, the uniformity of the non-structured data is improved, and the standardization of the data, the reduction of the sharing cost, the improvement of the sharing efficiency, the support of the efficient processing and the deep application of the data play an important role; the non-structured data in the power system field is classified, classified, layered, point combed and standardized, the classification, combed and standardized processing is realized, and the power system data element is obtained; the power system data element is pretreated through the character embedding layer based on the RoBERTa model, the BiLSTM model and the label speculation CRF layer based on the MFM model, the power system text segmentation and the word class tagging combined method is combined, the segmentation is obtained, the segmentation is reorganized and vectorized, the power system data element is vectorized after being divided into half sentences, the reverse maximum matching algorithm is used to correct the segmentation, the optimized segmentation result is obtained, and the influence of the out-of-vocabulary word in the power system field on the segmentation effect is reduced. Then, the optimized segmentation result is trained based on the power system special text library, the standard power system element database is obtained, the standard element database with uniformity, fine granularity and identification can be established, the rules of data element information and entity extraction are established according to the construction direction of the power system, the mapping rule library of data element information and equipment entity is formed, a good data foundation is provided for the knowledge extraction of the power system, the recognition result of the equipment entity can meet the composition rule of the inner and outer entity nesting combination in the actual text, the accuracy and efficiency of the multi-modal power system data element naming equipment entity recognition are improved, and the low efficiency and inaccuracy of the current manual sampling are also compensated. The high-frequency nouns are obtained by recognizing and filtering the equipment entity category candidate words, the equipment entity model candidate words are subjected to entity joint extraction, the entity joint extraction is carried out according to the upper entity and relationship concept extraction, the lower entity and relationship concept extraction, the confidence high-frequency words are obtained, the confidence highest words are obtained by recognizing the confidence high-frequency words, and the accuracy of the power system data element naming equipment entity recognition is further improved. Finally, the equipment entity class and model label corresponding to the power system data element are generated according to the high-frequency nouns and the confidence highest words and stored in the index, the equipment entity knowledge is output according to the index, the accuracy of the power system data element naming equipment entity recognition is improved, and the accurate recognition and extraction of the power system data are realized.

[0076] In some embodiments of this application, a power industry lexicon is obtained from industry experts, national standards, Baidu Input Method, Sogou Input Method, etc., and filtered using artificial intelligence technology. The optimized word segmentation results are trained based on a power system-specific text lexicon to obtain a standard power system metadata database. According to the construction direction of the power system, rules for data metadata and entity extraction are established. Then, the power system data metadata is split into half sentences and vectorized, and compared with the power industry lexicon to obtain words whose word vectors have a cosine value > 0. This forms a mapping rule library between data metadata and equipment entities, which makes up for the inefficiency and inaccuracy of current manual screening.

[0077] This application's standard power system metadata database includes, but is not limited to, the following information: Power system information resources: power system information resource classification, information resource name, information resource provider, and information resource summary; Power system standard documents: release date, issuing organization, standard name, uploaded attachments, and standard number; Power system data elements: belonging information resource, definition description, Chinese name of the data element, English name of the data element, code set name, remarks, number, data element domain, data element provider, data format, data type, and original data type; Power system value domain: code set name, description, encoding rules, and code name. A unified, fine-grained, and interconnected standard power system metadata database will be established. Based on the power system's construction direction, rules for data element information and entity extraction will be established, forming a mapping rule base between data element information and equipment entities, providing a sound data foundation for knowledge extraction from the power system.

[0078] Facing massive, unstructured power system data, complex entities in power system data lack obvious features and exhibit complex and diverse syntactic structures and parts of speech. Furthermore, the complex structure of the Chinese language itself makes it difficult to identify word boundaries, which are generally also the boundaries of named entities. Existing models widely employ span-based methods for identifying nested entities exhibit fuzziness in entity boundary detection, affecting recognition performance. Therefore, many problems remain in recognizing named device entities in long texts. This application provides a method that can improve the accuracy and efficiency of named device entity recognition in multi-power system data.

[0079] Reference Figure 2 As shown, Figure 2 This application provides a flowchart of a method for identifying and filtering candidate words for device entity categories to obtain high-frequency nouns. The method includes, but is not limited to, steps S510 to S520. Specifically,

[0080] Step S510: Based on power system standards and specifications, form a basic rule knowledge base for equipment entity application scenarios, entity boundaries, and semantics;

[0081] Step S520: Based on the device entity application scene and the basic rule knowledge base of entity boundary and semantics, the device entity nested rule is fused into the Chinese entity recognition method, and the recognition task of the device entity is converted into a joint training task of boundary recognition and boundary head-tail relationship recognition of the device entity.

[0082] Step S530: In the decoding process of the joint training task, the decoding result is identified and filtered in combination with the entity nesting rules of the power system, so that the recognition result conforms to the composition rule of the inner and outer entity nesting combination in the text.

[0083] Step S540: According to the terminology standard of the power system, the device entity class candidate word is terminologized and standardized, and high-frequency nouns are obtained.

[0084] In some embodiments of the present application, the device entity in the power system field is mostly nested entity, and has the characteristics of long characters and strong context correlation. The original power system data elements collected and collected are cleaned and processed, the multi-source heterogeneous data fusion for power system text entity recognition is faced, the basic rule knowledge base of device entity application scene and entity boundary and semantics is formed based on power system standards, expert consensus, power system specifications and other knowledge; and the Chinese entity recognition method is designed by fusing the device entity nesting rules, the recognition task of the device entity is converted into a joint training task of boundary recognition and boundary head-tail relationship recognition of the device entity, and the decoding result is filtered in combination with the device entity nesting rules summarized from the actual power system text in the decoding process, so that the recognition result can conform to the composition rule of the inner and outer entity nesting combination in the actual text, and the accuracy and efficiency of the power system data element naming device entity recognition are improved. At the same time, the important fields related to the power system are terminologized and standardized according to the related terminology standards in the industry, and high-frequency nouns are obtained, so as to further achieve the purpose of data structuring, improve the accuracy and efficiency of multi-modal power system data element naming device entity recognition, and form a structured power system data knowledge base suitable for key features in multiple dimensions of power system operation and evaluation.

[0085] In the development of Chinese word segmentation and part-of-speech tagging for the power system field, two major challenges have remained unresolved: word segmentation ambiguity and out-of-vocabulary (OV) word identification. 1) Word segmentation ambiguity: In real-world text environments, word segmentation ambiguity is common, stemming from the flexibility of Chinese characters. Characters can freely appear in any position within Chinese text, forming diverse word combinations, which greatly increases the complexity of defining lexical boundaries. If a natural language processing system lacks sufficient ambiguity resolution capabilities, it will struggle to handle large-scale corpora, leading to frequent word segmentation errors. These errors directly impact the accuracy of automatic Chinese word segmentation and even the accuracy of the entire syntactic analysis. 2) Out-of-vocabulary (OV) word problem: OV words refer to words that appear in the text to be processed but are not included in existing dictionaries. In practical research, we collectively refer to all words that appear in the test corpus but are not included in the training corpus as OV words. These words broadly cover various equipment names, place names, organization names, and other proper nouns, making their processing difficult.

[0086] To address the challenges of insufficient domain-specific corpora and difficulty in correctly segmenting out-of-vocabulary terms and phrases in power system text segmentation, this application proposes a method for preprocessing power system data elements to obtain segmented words.

[0087] Reference Figure 3 As shown, Figure 3 This is a flowchart illustrating a method for preprocessing power system data elements to obtain word segments, as provided in an embodiment of this application. The method includes, but is not limited to, steps S310 to S340. Specifically,

[0088] Step S310: Input the power system data elements into the character embedding layer of the RoBERTa model to obtain character vectors that reflect the contextual semantics and output the character vectors;

[0089] Step S320: Input the word vectors into the BiLSTM model. The BiLSTM model captures the long-distance dependencies and contextual sequence information of the word vectors to discard useless word vector sequences and extract useful word vector sequences.

[0090] Step S330: Input the useful word vector sequence into the label inference CRF layer of the MFM model, use the label inference CRF layer of the MFM model to jointly model the label sequence, and calculate the conditional probability of the input useful word vector sequence output in the label sequence;

[0091] Step S340: Determine the highest conditional probability of the output and select the useful word vector sequence corresponding to the highest conditional probability as the word segmentation.

[0092] The embodiment of the application provides a power system text segmentation and part-of-speech tagging joint method based on a RoBERTa model, a BiLSTM model and an MFM model, which is used for preprocessing power system data elements to obtain a segmented method, including: taking the RoBERTa model with the ability of transfer learning and text feature representation as a feature representation layer; inputting the power system data elements into a character embedding layer of the RoBERTa model to obtain word vectors reflecting context semantics and outputting the word vectors; combining the BiLSTM model to extract global and local features of the power system data elements for text segmentation; capturing long-distance dependency and context sequence information of the word vectors through the BiLSTM model to discard useless word vector sequences and extract useful word vector sequences; then inputting the useful word vector sequences into a label inference CRF layer of the MFM model; the label inference CRF layer of the MFM model has the ability to process overlapping features and long-distance dependencies, can well process the label bias problem, uses the label inference CRF layer of the MFM model to jointly model the label sequence, calculates the conditional probability of the input useful word vector sequence in the output label sequence, determines the highest conditional probability, and selects the useful word vector sequence corresponding to the highest conditional probability as the segmentation, thereby improving the accuracy and efficiency of power system data element named device entity recognition.

[0093] Referring to Figure 4 as shown, Figure 4 is a flow chart of a method for inputting power system data elements into a character embedding layer of a RoBERTa model to obtain word vectors reflecting context semantics and outputting the word vectors, the method for inputting power system data elements into a character embedding layer of a RoBERTa model to obtain word vectors reflecting context semantics and outputting the word vectors includes but is not limited to steps S311 to S314, specifically,

[0094] Step S311: calculating the word vectors of the power system data elements with a first parameter matrix, a second parameter matrix and a third parameter matrix respectively to obtain a first vector, a second vector and a third vector with the same dimension as the word vectors;

[0095] Step S312: constructing an attention matrix according to the first vector, the second vector and the third vector;

[0096] Step S313: mapping the first vector, the second vector and the third vector through different linear transformations, and splicing through the attention matrix to obtain a multi-head attention matrix;

[0097] Step S314: connecting the multi-head attention matrix to a feedforward neural network, performing encoding processing on the power system data elements through linear transformation and a Relu activation function, and performing preliminary segmentation processing on the power system data elements using a Chinese segmentation algorithm to obtain word vectors reflecting context semantics and output the word vectors.

[0098] The character embedding layer of the RoBERTa model is the core of the RoBERTa model, and the main role of this layer is to calculate the correlation between different words in a sentence. The overall framework of the character embedding layer is stacked by multiple encoder structures of the Transformer, and each encoder is composed of a multi-head attention and a feed-forward neural network, wherein the multi-head attention expands the ability of the model to focus on a subspace, and each multi-head attention structure is composed of h self-attention structures. In the embodiment of the present application, the power system data element is input into the character embedding layer of the RoBERTa model, and the word vector of the power system data element is calculated with the first parameter matrix WQ, the second parameter matrix WK and the third parameter matrix Wv respectively to obtain the first vector Q, the second vector K and the third vector V which have the same dimension as the word vector; the attention matrix is constructed according to the first vector Q, the second vector K and the third vector V, and then the multi-head attention matrix is connected to the feed-forward neural network through the mapping of h different linear transformations on the first vector Q, the second vector K and the third vector V, which can effectively increase the nonlinear fitting ability of the model. The power system data element is encoded and processed through linear transformation and Relu activation function, and the power system data element is preliminarily segmented by using the Chinese segmentation algorithm to obtain the word vector reflecting the context semantics and output the word vector. The word vector is taken as the input information of the BiLSTM model for subsequent semantic encoding.

[0099] In some embodiments of the present application, the attention matrix constructed according to the first vector Q, the second vector K and the third vector V is realized by the following formula:

[0100]

[0101] Wherein, Attention(Q,K,V) is the attention matrix, Q is the first vector, K is the second vector, V is the third vector, and d is the dimension of the word vector of the input power system data element.

[0102] Referring to Figure 5 , the Figure 5 is a flowchart of a method for capturing long-distance dependency relationships and context sequence information of a word vector by a BiLSTM model provided by the embodiment of the present application. The method for capturing long-distance dependency relationships and context sequence information of a word vector by a BiLSTM model includes but is not limited to steps S321 to S323, specifically,

[0103] Step S321: input the word vector into the forget gate of the BiLSTM model to determine the information to be discarded in the previous neuron through the forget gate, the input being the current word vector and the previous word vector, and the previous neuron cell state being mapped to 0-1, wherein 0 represents complete deletion and 1 represents complete retention, to obtain the result of the forget gate;

[0104] Step S322: Determine whether to record a new word vector into the neuron of the BiLSTM model through the input gate, determine the word vector that needs to be updated through the input layer Sigmoid activation function, update the neuron state, determine the memory information that needs to be recorded, and obtain the word vector sequence that needs to be remembered;

[0105] Step S323: Determine the part of the output neuron state through the Sigmoid activation function of the output gate, process the neuron state through the tanh function, and multiply it with the output of the Sigmoid activation function, to obtain a hidden layer state sequence with the same length as the word vector sequence, so as to capture the long distance dependency relationship and context sequence information of the word vector, discard the useless word vector sequence, and extract the useful word vector sequence.

[0106] The BiLSTM model is an improved recurrent neural network (RNN) model, which can be used to capture long distance dependency relationship and context sequence information. The BiLSTM model includes a forgetting gate, an input gate, and an output gate. By introducing a gating mechanism, the problem of gradient disappearance or gradient dispersion caused by long text sequence in traditional RNN is overcome. The forgetting gate determines which information needs to be discarded, the input gate determines which information is retained by the memory unit, and the output gate determines which information is output and enters the next cycle iteration. The algorithm implementation process is as follows:

[0107] First step: Calculate the forgetting gate and select the information to be forgotten: input the word vector into the forgetting gate of the BiLSTM model, determine the information that needs to be discarded in the previous neuron through the forgetting gate, the input is the current word vector and the previous word vector, and the previous neuron cell state is mapped to 0-1, wherein 0 represents complete deletion and 1 represents complete retention, to obtain the result of the forgetting gate.

[0108] The BiLSTM model determines the information that needs to be discarded in the previous neuron through the forgetting gate, and the input is the output of the current word vector and the previous word vector. The previous neuron cell state is mapped to 0-1, wherein 0 represents complete deletion and 1 represents complete retention. The result of the forgetting gate at time t is:

[0109] f t =σ(W fx x t +W fh h t-1 +W fc c t-1 +b f );

[0110] In the formula, W fx ,W fh ,Wfc forget gate and input x t , intermediate output h t-1 and neuron state c of BiLSTM model at t-1 moment t-1 matrix weight multiplied by, b f bias term of forget gate, and σ is sigmoid activation function. x t input data at t moment. h t-1 output data at t-1 moment.

[0111] Second step: calculate the input gate and select the information to be remembered. Determine whether to record the new word vector into the neuron of the BiLSTM model through the input gate, and determine the word vector to be updated through the input layer sigmoid activation function to update the neuron state to determine the memory information to be recorded, and obtain the word vector sequence to be remembered.

[0112] The input gate determines whether to record new information into the neuron of the BiLSTM model, and the input is the output of the current word vector and the previous word vector, and the value i t to be updated is determined through the input layer sigmoid activation function to update the neuron state to obtain c t , which is realized by the following formula:

[0113] i t = σ(W ix x t +W ih h t-1 +W ic c t-1 +b i );

[0114] c t =f t c t-1 +i t tanh(W cx x t +W ch h t-1 +b c );

[0115] In the formula, W ix , W ih , and W ic are matrix weights multiplied by input gate and input x t and intermediate output h t-1 and neuron state c of LSTM at t-1 moment t-1 , b f is the bias term of the forget gate. W cx , W ch are matrix weights multiplied by input node and input xt and the intermediate output h t-1 the matrix weight multiplied by the input x c is the bias term of the input node. c t is the neuron state of the BiLSTM model at time t. tanh is the activation function. i t is the n-dimensional vector at time t in the input gate, taking a number between 0 and 1.

[0116] Step 3: Calculate the output gate and the current time hidden layer state, select the output value. Determine the part of the output neuron state through the Sigmoid activation function of the output gate, process the neuron state through the tanh function, and multiply it with the output of the Sigmoid activation function to obtain a hidden layer state sequence with the same length as the word vector sequence, to capture the long-distance dependence of the word vector and the context sequence information, discard the useless word vector sequence, and extract the useful word vector sequence.

[0117] The output gate is used to determine the output information, and the Sigmoid activation function of the output layer is used to determine the part of the output neuron state o t , process the neuron state through the tanh function, and multiply it with the output of the Sigmoid gate to obtain the final output h t , and finally, a hidden layer state sequence with the same length as the word vector can be obtained {h1, h2, h3, …, h n}.

[0118] o t = σ(W ox x t +W oh h t-1 +W oc c t +b o );

[0119] h t = o t tanh(c t );

[0120] In the formula, W ox , W oh , and W oc are the matrix weights multiplied by the input x t , the intermediate output h t-1 , and the neuron state c t-1 of the BiLSTM model at time t-1. o is the bias matrix of the output gate. o t is the generation matrix of the logistic regression after the full connection layer. h t is the output content of the current state at time t.

[0121] From the working principle of the BiLSTM model, it can be seen that the BiLSTM model uses memory gate, forget gate and output gate to process information, which can discard some useless information, enhance the memory of neurons, and is suitable for processing long-term time relationship data and solving long-term dependency problems.

[0122] Word segmentation tasks are often viewed as sequence labeling tasks, where there is a strong dependency between the labels of consecutive characters. In character-based part-of-speech tagging and Chinese word segmentation, the collocation relationships between adjacent tags must be considered. Therefore, tagging decisions cannot be made solely by the segmentation tool; instead, an MFM model is used to jointly model the tag sequence. The MFM model is an undirected graphical model used to compute the conditional probability of the output of a random variable given an input random variable. It combines the characteristics of Hidden Markov Models and Maximum Entropy Models, possessing the ability to handle overlapping features and long-distance dependencies, and can effectively address the tag bias problem.

[0123] In this embodiment, the useful word vector sequence is input into the label prediction CRF layer of the MFM model, and the label prediction CRF layer of the MFM model is used to jointly model the label sequence and calculate the conditional probability of the input useful word vector sequence output in the label sequence.

[0124] The calculation process of the label inference CRF layer based on the MFM model is as follows: For the output sequence X = (x1, x2, ..., x...) of the BiLSTM layer of the model... t (This is also the input sequence of the CRF layer), assuming P∈R n×k Let P be the output score matrix of BiLSTM, where n is the number of words, k is the number of tags, and P is the matrix element. ij This is the output score for the i-th word under the j-th label. For a useful word vector sequence Y = (y1, y2, ..., yj), ... t If the total score of the label sequence is:

[0125]

[0126] In the formula, A is the transition score matrix, A i,j P represents the transition score from label i to label j. i,j The output score of the i-th word under the j-th label, n is the number of words, k is the number of labels, A is k+2, X=(x1,x2,…,x t Y = (y1, y2, ..., y3) is the input sequence of the CRF layer. t ) is a useful sequence of word vectors, and s(X,Y) represents the score for each labeled sequence.

[0127] The probability of generating a useful word vector sequence Y is expressed by the following formula:

[0128]

[0129] In the formula, X = (x1, x2, ..., x t Y = (y1, y2, ..., y3) is the input sequence of the CRF layer. t ) is a useful sequence of word vectors, where s(X,Y) represents the score for each labeled sequence. The actual labeled sequence; Y X Given all possible label sequences, P(Y|X) represents the probability of each label. The label sequence with the highest probability is selected as the conditional probability of the output in the label sequence, and the prediction result is obtained.

[0130] Reference Figure 6 As shown, Figure 6 This is a flowchart illustrating a method for obtaining optimized word segmentation results by correcting word segmentation using a reverse maximum matching algorithm, as provided in an embodiment of this application. The method includes, but is not limited to, steps S350 to S370. Specifically,

[0131] Step S350: Match the characters composed of several consecutive word groups in the string to be re-segmented after the initial word segmentation from right to left with the domain dictionary;

[0132] Step S360: If the match is successful, new words are segmented to obtain the optimized word segmentation result;

[0133] Step S370: If the match fails, remove the leftmost word from the first word segmentation result and match it with the domain dictionary. If the match succeeds, segment out new words and obtain the optimized word segmentation result. If the match fails, iterate until only a single word remains in the first word segmentation result and obtain the optimized word segmentation result.

[0134] This application employs a reverse maximum matching algorithm to correct word segmentation and obtain optimized word segmentation results, thereby reducing the impact of out-of-vocabulary words in the power system field on the word segmentation effect.

[0135] In some embodiments of this application, assuming the string obtained from the initial word segmentation is S1, the matching result is S2 (initially S2 is an empty string), and the maximum word length in the domain dictionary D is N, the algorithm steps for performing reverse maximum matching on S1 are as follows:

[0136] Step 1. Initialize the algorithm parameters and input the string S1 to be re-segmented;

[0137] Step 2. Set the maximum word length of the domain dictionary D to N;

[0138] Step 3. Start from the rightmost side of S1 and take the candidate string W.

[0139] Step 4. Match the candidate string W with the domain dictionary D. If the match is successful, proceed to step 6; otherwise, execute step 5.

[0140] Step 5. Delete the leftmost original phrase from the candidate string W and assign it to S2; determine whether the candidate string W is a single phrase. If it is, go to step 6; otherwise, return to step 4.

[0141] Step 6. S2 = "" + W + S2 ("" is 2 spaces), remove the original phrase contained in W from the far right of S2;

[0142] Step 7. Output the matching result S2.

[0143] Then, an automatic word segmentation optimization algorithm is used to merge and optimize the "scattered words." To address the bias and prevent its spread, transfer learning is introduced to enhance the system's adaptability. This application embodiment combines domain-specific domain-specific corpus annotation data for domain-specific word segmentation transfer learning. Considering the data similarity between the power system domain-specific corpus and the target domain corpus, different levels of network transfer are performed on the label inference CRF layer based on RoBERTa, BiLSTM, and MFM models to solve the domain adaptability problem of word segmentation. Finally, optimized word segmentation results are output.

[0144] Reference Figure 7 As shown, Figure 7 This is a flowchart illustrating a method for extracting high-confidence high-frequency words from candidate device entity model names, as provided in this application embodiment. The method includes, but is not limited to, steps S610 to S650. Specifically,

[0145] Step S610: Describe the candidate terms for equipment entity model using data elements to form a fixed feature expression form for power information;

[0146] Step S620: Based on the preset data element sorting and equipment entity extraction experience, and combined with the standard document number, system classification number and test report type, determine the power system domain factor. The power system domain factor carries the data source and classification of the equipment entity model candidate words and the equipment entity extraction rule base sequence number.

[0147] Step S630: Determine the feature vector of candidate words for equipment entity models through bidirectional calculation of long short-term memory network and power system domain factor, and extract key features for identification of candidate words for power system equipment entity models using BiLSTM model;

[0148] Step S640: the feature vector is combined with the key features of the device entity model candidate word to calculate, so as to identify the device entity name corresponding to the device entity model candidate word and the weight;

[0149] Step S650: the global optimal sequence is marked by the mark inference CRF layer of the MFM model, and the hidden state sequence is converted into the best mark sequence, so as to realize the entity joint extraction of the device entity model candidate word, and obtain the confidence high frequency word.

[0150] For massive and unstructured power system data, the multi-modal power system data complex entity lacks obvious features, and the syntactic structure and word composition are complex and diverse. The embodiment of the application provides a method for performing entity joint extraction on the device entity model candidate word to obtain the confidence high frequency word, thereby improving the accuracy of power system data element naming device entity recognition.

[0151] According to the data standard file, the information system database, the power detection report and the like, the keywords are recognized, at this time, the Chinese word segmentation technology for the power system field is adopted, and the stop words are removed, and finally the key equipment entity model candidate words of the data source are left, then, the equipment entity model candidate words are described by the data elements to form the fixed feature expression form of the power information of different types and multiple channels, and according to the data element analysis and the equipment entity extraction experience, the data element and the entity extraction rule library are continuously improved, which are used to guide the power system field factor generation. According to the preset data element analysis and equipment entity extraction experience, the standard number, the system classification number and the detection report type are combined to determine the power system field factor, the power system field factor carries the data source and the classification of the equipment entity model candidate words and the equipment entity extraction rule library sequence number, which is beneficial to the power data information knowledge expression, the efficient identification of abnormal values and the source, and provides the basis for the power data information knowledge expression; the power system data element and the equipment entity extraction rule library sequence number are used to guide the confirmation of the same field entity during the entity extraction, and the entity extraction efficiency is improved. The vector generated by the RoBERTa model in the character embedding layer is generated by superimposing the word vector, the sentence vector and the position vector. The word vector is converted by querying the word vector table; the sentence vector represents the text information of the sentence, which is used to distinguish different sentences; the position vector is obtained by coding the position information corresponding to each word, which can distinguish the semantic information of different positions. The output layer is the word vector that can reflect the context semantics after the coding layer processing, and these word vector sequences are used as the input information of the BiLSTM model for subsequent semantic coding. The feature vector of the equipment entity model candidate word is determined through the long short-term memory network and the power system field factor bidirectional calculation, the key features of the power system equipment entity model candidate word recognition are extracted by using the BiLSTM model, the feature vector and the key features of the equipment entity model candidate word are combined and calculated to identify the equipment entity name corresponding to the equipment entity model candidate word and the weight; the global optimal sequence is marked by the CRF layer through the marking of the MFM model, and the hidden state sequence is converted into the best marking sequence, the equipment entity model candidate word is entity joint extracted, and the high confidence degree word is obtained. The recognition result of the equipment entity can conform to the composition rule of the inner and outer entity nesting combination in the actual text, the accuracy and efficiency of the multi-modal power system data element naming equipment entity recognition are improved, so that the power system data is accurately recognized and extracted.

[0152] Events in power system have certain range and theme, therefore, the method for extracting entity relationship of the device is different from the traditional method, the bootstrapping method and the Word2vec model are used to complete the extraction of entity relationship. Word2vec can get the similarity between single word and other words by processing text content into vector space form through model training. Word2vec mainly has two models, CBOW model and improved Skip-gram model, CBOW model is to predict the generation probability of the current word according to the words around the current word, and Skip-gram model is to predict the surrounding words according to the current word, so as to extract relationship knowledge. Mainly divided into upper entity and relationship concept extraction, lower entity and relationship concept.

[0153] The upper entity concept extraction method based on mixed morphological rules is as follows:

[0154] When extracting the upper entity concept, a mixed morphological rule is proposed for the extraction of the relation extraction (RE) part. In the sentence containing the relationship feature word "is", the RE part structure is simple, so it is easy to obtain the upper entity concept from the RE, mainly according to the morphological analysis method to extract the upper concept. The specific extraction rules are as follows:

[0155] 1) If the RE does not contain the word "of", the noun after the relationship feature word is directly obtained through morphological analysis;

[0156] 2) If the RE contains the word "of", the noun after the "of" word with the largest position number in the RE is directly extracted through morphological analysis;

[0157] 3) The mixed morphological rule extracts the RE, if the RE contains "of", the noun after "of" is extracted, otherwise the noun after the feature word is extracted.

[0158] The lower entity concept extraction method based on mixed syntactic rules is as follows:

[0159] The extraction of lower entity concept is different from the extraction of upper entity concept. By matching the "is a" relationship feature sentence, the LE part containing the lower entity concept may contain multiple commas. After analyzing the sentence containing multiple commas, it is found that most of them contain multiple subject-predicate structures. The application analyzes the LE part from the perspective of syntax, thereby summarizing some syntactic relationship features to accurately extract the lower entity concept.

[0160] Referring to Figure 8 , Figure 8is a flowchart of a method for identifying the highest-confidence word from the high-frequency words with confidence provided by an embodiment of the present application. The method for identifying the highest-confidence word from the high-frequency words with confidence includes but is not limited to steps S710 to S730, specifically,

[0161] Step S710: Determine the distance between the high-frequency word with confidence and the device model category.

[0162] Step S720: Arrange the distances from small to large to obtain a high-frequency word sequence table with confidence from high to low.

[0163] Step S730: Clean up the negative words in the high-frequency word sequence table, and output the highest-confidence high-frequency word in the high-frequency word sequence table to obtain the highest-confidence word.

[0164] In some embodiments of the present application, the device entity model candidate words are extracted by entity joint extraction to obtain high-frequency words with confidence, and the high-frequency words with confidence are processed to obtain the highest-confidence word. The method for identifying the highest-confidence word from the high-frequency words with confidence includes: first, calculate the semantic distance, similarity score or other relatedness quantitative indicators between each high-frequency word with confidence and the device model category to evaluate the relevance of each high-frequency word with confidence and the device model. After calculating the distance between all high-frequency words with confidence and the device model, the high-frequency words with confidence are sorted according to the distance values, and the distances are arranged from small to large to obtain a high-frequency word sequence table with confidence from high to low. That is, the high-frequency word with confidence with the smallest distance value (i.e. the most relevant word to the device model) will be placed at the front of the high-frequency word sequence table, thereby forming a sequence table with confidence from high to low. Since negative words may reduce the confidence of the word, as they usually represent negative or opposite meanings. Therefore, in the present application, the negative words are cleaned and removed from the high-frequency word sequence table. After cleaning, the highest-confidence high-frequency word is selected from the remaining words as the final result, and the highest-confidence high-frequency word is output to obtain the highest-confidence word.

[0165] By this method, the key words in the power system high-frequency word query can be more accurately identified, thereby providing more relevant content or services, thereby improving the accuracy of power system data element naming device entity identification, thereby achieving accurate identification and extraction of power system data.

[0166] Referring to Figure 9 , as shown, Figure 9Fig. 1 is a structural schematic diagram of a controller 1000 provided by an embodiment of the present application, which comprises a processor 1001, which can be implemented in a manner of a general-purpose CPU (Central Processing Unit), a microprocessor, an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits, etc., and is used to execute a related program to implement the method provided by the embodiment of the present application; a memory 1002, which can be implemented in a manner of a ROM (Read Only Memory), a static storage device, a dynamic storage device, or a RAM (Random Access Memory) etc. The memory 1002 can store an operating system and other application programs, and when the technical solution provided by the embodiment of the present application is implemented by software or firmware, the related program code is saved in the memory 1002 and is called and executed by the processor 1001 to implement the embodiment of the present application; an input / output interface 1003, which is used to implement information input and output; a communication interface 1004, which is used to implement the communication interaction between the device and other devices, and can realize the communication in a wired manner (for example, a USB, a network cable, etc.) or in a wireless manner (for example, a mobile network, WIFI, Bluetooth, etc.); and a bus, which is used to transmit information between various components (for example, the processor 1001, the memory 1002, the input / output interface 1003, and the communication interface 1004) of the device. The processor 1001, the memory 1002, the input / output interface 1003, and the communication interface 1004 are connected to each other in the device through the bus.

[0167] As will be appreciated by one of ordinary skill in the art, all or some of the steps, systems, etc. in the methods disclosed above can be embodied in software, firmware, hardware, and / or suitable combinations thereof. Some or all of the physical components can be implemented with software executed by a processor, such as a central processing unit, a digital signal processor, or a microprocessor, or can be implemented as hardware, or as an integrated circuit, such as an application- specific integrated circuit. Such software can be distributed on computer readable media, which can comprise computer readable storage media (or non-transitory media), and communication media (or transitory media). As is known to those of ordinary skill in the art computing readable storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Computer readable storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to store the desired information and which can be accessed by a computer. Further, as will be appreciated by one skilled in the art, communication media typically embodies computer readable instructions, data structures, program modules, or other data in a modulated data signal, such as carrier waves or other transport mechanisms, and includes any information delivery media.

[0168] Additional features and advantages of the application will be set forth in the description that follows, and in part will be apparent from the description, or can be learned by practice of the application. The objectives and other advantages of the application will be realized and attained by the structure particularly pointed out in the written description and claims thereof.

Claims

1. A method for extracting information and knowledge based on power systems, characterized in that, include: Obtain information documents of the power system, read the unstructured data of the information documents of the power system, and convert the unstructured data into txt format; The unstructured data is cleaned, classified, cataloged, and standardized to obtain power system data elements; The power system data elements are preprocessed by a character embedding layer based on the RoBERTa model, a BiLSTM model, and a tag-inference CRF layer based on the MFM model, combined with a joint method of power system text segmentation and part-of-speech tagging to obtain segmented words. The segmented words are then reorganized and vectorized, and the inverse maximum matching algorithm is used to correct the segmented words to obtain optimized segmentation results. The optimized word segmentation results are trained based on a dedicated text lexicon for power systems to obtain a standard power system data element library. According to the construction direction of the power system, rules for data element information and entity extraction are established to form a mapping rule library between data element information and equipment entities, so as to realize the extraction of candidate words for equipment entity categories. The candidate words for the device entity categories are identified and filtered to obtain high-frequency nouns; Keyword identification is performed on data standard documents, information system databases, and power testing reports to obtain candidate words for equipment entity models; Candidate terms for equipment entity models are described using data elements to form a fixed feature expression form for power information; Based on the preset data element sorting and equipment entity extraction experience, and combined with the standard document number, system classification number and test report type, the power system domain factor is determined. The power system domain factor carries the data source and classification of the equipment entity model candidate words and the equipment entity extraction rule base sequence number. The feature vectors of candidate words for equipment entity models are determined by bidirectional calculation of long short-term memory network and power system domain factors, and the key features for identifying candidate words for power system equipment entity models are extracted by BiLSTM model. Feature vectors are introduced and key features of candidate words for device entity models are combined for calculation to identify the device entity name and weight corresponding to the candidate words for device entity models; By inferring the globally optimal sequence of CRF layer labels through the labeling of the MFM model, and transforming the hidden state sequence into the best label sequence, entity joint extraction is achieved for the candidate words of the device entity model to obtain high-frequency words with confidence. The high-frequency words with the highest confidence level are identified and processed to obtain the words with the highest confidence level; Based on the high-frequency nouns and the words with the highest confidence, generate equipment entity category and model tags corresponding to the power system data elements and store them in the index; Based on the entity knowledge of the aforementioned lead output device.

2. The information knowledge extraction method based on power systems according to claim 1, characterized in that, The process of identifying and filtering candidate words for the device entity category yields high-frequency nouns, including: A basic rule knowledge base for the application scenarios, entity boundaries, and semantics of equipment entities is formed based on power system standards and specifications. Based on the knowledge base of basic rules of device entity application scenarios, entity boundaries, and semantics, the Chinese entity recognition method of device entity nesting rules is integrated to transform the device entity recognition task into a joint training task of device entity boundary recognition and boundary beginning and end relationship recognition. In the decoding process of the joint training task, the entity nesting rules of the power system are combined to identify and filter the decoding results so that the identification results conform to the composition rules of the inner and outer layer entity nesting combination in the text. Based on the terminology standards of the power system, the candidate terms for the equipment entity categories are terminologically and standardized to obtain high-frequency nouns.

3. The information knowledge extraction method based on power systems according to claim 1, characterized in that, The process involves preprocessing the power system data elements using a RoBERTa-based character embedding layer, a BiLSTM model, and an MFM-based tokenized CRF layer, combined with a power system text segmentation and part-of-speech tagging method, to obtain segmented words, including: The power system data elements are input into the character embedding layer of the RoBERTa model to obtain character vectors that reflect contextual semantics and output the character vectors. The word vectors are input into the BiLSTM model, which captures the long-distance dependencies and contextual sequence information of the word vectors, thereby discarding useless word vector sequences and extracting useful word vector sequences. The useful word vector sequence is input into the label inference CRF layer of the MFM model, and the label inference CRF layer of the MFM model is used to jointly model the label sequence to calculate the conditional probability of the input useful word vector sequence output in the label sequence. Determine the highest conditional probability of the output, and select the useful word vector sequence corresponding to the highest conditional probability as the word segmentation.

4. The information knowledge extraction method based on power systems according to claim 3, characterized in that, The step of inputting the power system data elements into the character embedding layer of the RoBERTa model to obtain character vectors that reflect contextual semantics and then outputting the character vectors includes: The word vectors of the power system data elements are calculated with the first parameter matrix, the second parameter matrix, and the third parameter matrix respectively to obtain the first vector, the second vector, and the third vector with the same dimension as the word vectors; Construct an attention matrix based on the first vector, the second vector, and the third vector; The first vector, the second vector, and the third vector are mapped by different linear transformations and then concatenated using the attention matrix to obtain a multi-head attention matrix; The multi-head attention matrix is ​​connected to a feedforward neural network, and the power system data elements are encoded using linear transformation and ReLU activation function. The power system data elements are then preliminarily segmented using a Chinese word segmentation algorithm to obtain character vectors that reflect the contextual semantics, and the character vectors are output.

5. The information knowledge extraction method based on power systems according to claim 3, characterized in that, The BiLSTM model includes a forget gate, an input gate, and an output gate. The word vectors are input into the BiLSTM model, which captures the long-range dependencies and contextual sequence information of the word vectors to discard useless word vector sequences and extract useful ones. This includes: The word vector is input into the forget gate of the BiLSTM model. The forget gate determines the information that needs to be discarded in the previous neuron. The input is the current word vector and the previous word vector. The cell state of the previous neuron is mapped to 0~1, where 0 represents complete deletion and 1 represents complete retention. The result of the forget gate is obtained. The input gate determines whether to record new word vectors into the neurons of the BiLSTM model. The input layer sigmoid activation function determines the word vectors that need to be updated and updates the neuron state to determine the memory information that needs to be recorded, thus obtaining the sequence of word vectors that need to be memorized. The output neuron state is determined by the sigmoid activation function of the output gate. The neuron state is then processed by the tanh function and multiplied by the output of the sigmoid activation function to obtain a hidden layer state sequence with the same length as the word vector sequence. This process captures the long-distance dependencies and contextual sequence information of the word vectors, discarding useless word vector sequences and extracting useful ones.

6. The information knowledge extraction method based on power systems according to claim 3, characterized in that, The step of using the inverse maximum matching algorithm to correct the word segmentation and obtain optimized word segmentation results includes: From right to left, match the characters composed of several consecutive word groups in the string to be re-segmented after the initial word segmentation with the domain dictionary; If the match is successful, new words are segmented to obtain the optimized word segmentation result; If the match fails, the leftmost word group in the initial word segmentation result is removed and then matched with the domain dictionary. If the match is successful, a new word is segmented to obtain the optimized word segmentation result. If the match fails, the process is iterated until only a single word group remains in the initial word segmentation result to obtain the optimized word segmentation result.

7. The information knowledge extraction method based on power systems according to claim 1, characterized in that, The process of identifying and processing the high-frequency words with the highest confidence levels to obtain the words with the highest confidence levels includes: Determine the distance between the high-frequency confidence words and the device model category; Arrange the distances from smallest to largest to obtain a high-frequency word order list with confidence levels from highest to lowest; Negative words are removed from the high-frequency word order list, and the high-frequency word with the highest confidence in the high-frequency word order list is output to obtain the word with the highest confidence.

8. A controller, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, performs the method as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions for causing a computer to perform the method as described in any one of claims 1 to 7.

Images