Power grid dispatching operation ticket structured analysis method

Abstract

The invention belongs to the technical field of power distribution network operation ticket scheduling management, and particularly relates to a power grid scheduling operation ticket structured analysis method. Comprising the steps of determining power distribution network topology elements based on operation ticket content, and creating an operation ticket text element database; establishing a three-layer expansion storage structure character string matching dictionary; carrying out optimization processing on the operation ticket image; establishing an operation order table feature extraction model based on a multilayer convolutional neural network, and restoring a relative position relationship between an actual table and a text sequence in the table; based on an operation ticket text element recognition extraction result, carrying out structured analysis on the generated operation ticket character string; the method is mainly used for better obtaining the structural data in the operation ticket in the operation ticket scheduling management process, improving the accuracy of the data, and analyzing the specific element content of the operation ticket.

Description

Technical Field

[0001] The present invention belongs to the technical field of dispatching management of distribution network operation tickets, and in particular relates to a structured parsing method for a power grid dispatching operation ticket. Background Art

[0002] The dispatching operation ticket is a text record required for equipment inspection, maintenance and dispatching operations in the power system. It contains detailed operation steps, safety precautions, equipment information, etc. Through the various types of information recorded in the operation ticket, the distribution network status is predicted, and the corresponding operation steps, equipment name, time, etc. are obtained. The actual operation content during the operation of the distribution network is traced and analyzed to confirm the distribution network control process. Among them, using NLP technology to identify the information in the operation ticket and extract valid data is an effective solution in the process of operation ticket dispatching management. However, on the one hand, the operation ticket file information has the characteristics of unstructured, no unified standard, expression differences, etc. At the same time, the large number and high processing difficulty limit the further application and development of the corresponding solution. Summary of the invention

[0003] The purpose of the present invention is to provide a structured parsing method for power grid dispatching operation tickets based on actual needs, which can reduce the difficulty of identifying and extracting operation ticket text information elements in the operation ticket dispatching management process and improve the efficiency of extracting large amounts of operation ticket data information.

[0004] To achieve the above purpose, the present invention adopts the following technical solution.

[0005] A structured parsing method for a power grid dispatching operation ticket of the present invention comprises the following steps:

[0006] Step 1: Create a database of operation ticket text elements

[0007] It refers to determining the distribution network topology elements involved in various operations based on the specific operation content of the operation ticket, and creating an operation ticket text element database according to the text expression of the corresponding elements in the operation ticket issuance rules or marking methods, and the distribution network topology topology elements and topology nodes include: distribution line elements, distribution port elements, distribution equipment elements and operation content elements;

[0008] Step 2: Create a dictionary for matching operation ticket strings;

[0009] Establishing a string matching dictionary based on the aforementioned operation ticket text element database;

[0010] The string matching dictionary is a three-layer expansion storage structure, and the strings in each layer are expanded downward in sequence;

[0011] The first layer element is all the single strings A in the string data pool. iSet, represented by W1 = {A1, A2...A i ...A I}, any single string A in the string set W1 i Expand downward to form the second layer of double string set W2, I represents the total number of single strings, A i Represents the i-th single string;

[0012] The second layer double string set W2 stores the single string A i Double string starting with A i B j and contains double string A i B j The remaining length of the longest string is L ij , and a single string A i The first two characters of a multi-string that begins with A i C k and contains double string A i C k The remaining length of the longest string is L i ' j , we get W2={A i B1L i1 ,A i B2L i2 ...A i B j L ij ...A i B J L iJ ,A i C1L′ i1 ...A i C k L′ ik ....A i C K L′ iK .}, where L ij =L maxij -2; L′ ik =L′ maxik -2;L maxij It means that it contains double string A i B j The length of the longest string, L′ maxik It means that it contains double string A i C k The length of the longest string; any double string A in the string set W2 i B j or A i C k Expand downward to form the third-level multi-string set W3;

[0013] The third-layer multi-string set W3 stores the double string A in the second layer. i B j or A i C k A multi-character string with the first character;

[0014] Step 3: Optimizing the image of the operation ticket, including the following steps:

[0015] Step 3a, preprocessing of the operation ticket image, graying the acquired original operation ticket image to obtain a gray image to remove unnecessary chromaticity information, binarizing the gray image to extract its binary contour, and then filtering the image to filter out pixels below a preset threshold to reduce noise and remove impurities; then filtering the original operation ticket image according to the position of the pixels that have not been filtered out in the image to intercept valid data in the original operation ticket image;

[0016] Step 3b, image correction of the operation ticket: for the aforementioned binarized image, straight line recognition and extraction are performed, straight line lengths are counted, a length threshold for distinguishing text lines and boundary lines is set, text lines are removed, non-text lines are retained, horizontal and vertical coordinate values ​​of non-text lines are counted, straight lines with similar coordinate values ​​but not linked are extended to make them intersect, the inclination angle of the straight lines in the image is calculated, and reverse adjustment and correction are performed based on the inclination angle result, so that the straight lines in the operation ticket are in a relatively horizontal direction, and the operation ticket is restored and corrected as a whole based on the relative position relationship between each text element and the straight line;

[0017] Step 4: Identification and extraction of elements in the operation ticket image refers to determining the table element information while recognizing the text in the operation ticket; specifically:

[0018] 4a. Identify and extract elements of the operation ticket form;

[0019] A model for extracting operation ticket table features is established based on a multi-layer convolutional neural network, wherein the input of the model is the operation ticket image after optimization processing in step 3, and the output is the table graph structure and the adjacency relationship of the text sequences in each table graph structure; the operation ticket image is input into the table structure recognition model to extract the table structure in the operation ticket, and the actual table and the relative position relationship of the text sequence in the table are restored according to the adjacency relationship of the text sequence in the table structure;

[0020] 4c. Identification and extraction of text elements of the operation ticket, including the following steps:

[0021] Text element feature extraction: Based on the recognition and extraction of table elements, the relative position relationship of each text sequence can be confirmed, but the text elements in the text sequence cannot be directly obtained. At the same time, the text element information of the operation ticket often includes different types of text elements such as Chinese characters, letters, numbers, symbols, etc. Therefore, in order to achieve effective recognition and extraction of text elements, this application adopts the DSSD network model to extract text element features. The DSSD network model is obtained by replacing the convolution kernel network of the SSD network with a fast training residual network, introducing a residual module before classification and regression, and adding a deconvolution layer to the auxiliary convolution layer to form an improved model;

[0022] Text sequence corner point detection: Text corner point detection is used to locate the border of the text sequence to facilitate the extraction of text content in the operation ticket image, and to remove data outside the text border to improve recognition efficiency. At the same time, the coordinates and relative position relationship of the text box are used to restore the text order in the text sequence to ensure the effective acquisition of text information. Because the text in the operation ticket is relatively standardized and the size is basically the same, the diagonal box and its center position obtained by corner point detection can easily locate the corner point. Corner point detection converts the detection of the text boundary box into the detection process of the corner point through the convolutional neural network.

[0023] Step 5: Based on the results of the operation ticket text element recognition and extraction in step 4, the generated operation ticket string is structurally parsed, including the following steps:

[0024] 5a. For an operation ticket string, for example, D1D2D3...D x ...D X , in the string matching dictionary, whether the first double string just queried exists in the second-layer double string set W2, that is, whether the double string D1D2 is an element in the second-layer double string set W2;

[0025] 5b. If the first double character string does not exist in the second-layer double character string set W2, separate the first and last single character strings D1, end the word segmentation, separate the first single character string D1 from the operation ticket character string, and return to step 5a to continue analyzing the remaining character strings;

[0026] 5c. If the first double character string exists in the second-layer double character string set W2, verify whether the double character string exists in the corresponding third layer; if not, end the word segmentation; if so, count the length n of the longest character string in the third layer; record x=2, and go to step 5d;

[0027] 5d. Let x = x + 1. If x = n + 3, go to step A6. Otherwise, go to step 5e.

[0028] 5e. Determine whether there is a D3D4...D xIf the string headed by exists, go to step 5f, otherwise go to step 5d;

[0029] 5f, extraction can be combined with D3D4...D x The matched string is separated from the D1D2 component string, and the process returns to step 5a to continue analyzing the remaining string.

[0030] Further improvement or specific implementation steps of the aforementioned structured parsing method for power grid dispatching operation tickets, distribution line elements are regional attribute elements in the topological structure of the distribution network, which are usually isolated by various switch nodes and used to supply specific areas or achieve specific purposes to form relatively independent distribution areas; including busbar bypass, bus tie line, wiring line, feeder line, segmented line, etc.;

[0031] The distribution port element refers to the topological connection node between the topological lines or within the topological structure, mainly including the busbar and each bypass connection port, the line and the distribution equipment or the port that meets the connection;

[0032] The power distribution equipment element refers to the power distribution equipment corresponding to the operation object or execution object involved in the operation ticket, including various electrical equipment such as power generation equipment and transformer equipment;

[0033] The operation content element refers to the specific operation content or status information involved in the operation ticket, among which the status elements include operation, hot standby, cold standby, etc.; the action elements include maintenance, input, output, grounding, etc.

[0034] Further improvement or specific implementation steps of the aforementioned structured parsing method for power grid dispatching operation tickets, the table structure is established based on a graph structure composed of a set of vertex elements and edge elements, in the table graph structure, the vertex elements are the text sequences of each cell in the table, the edge elements are the row and column position relationship between text rows, and the position relationship includes: in the same row and column, adjacent rows, adjacent columns, and non-adjacent columns; wherein in the same row and column means that two text sequences are in the same cell in the table, adjacent rows mean that two text sequences are respectively located in two adjacent rows, and adjacent columns mean that two text sequences are respectively located in two adjacent columns;

[0035] During the training and learning process, the adjacency matrix is ​​used to express the adjacency relationship of the aforementioned edge elements. A table database is established and divided into a training data set and a verification data set. The tables in the training data set are used to train and optimize the table structure recognition model, and the verification data set is used to verify and feedback the model training results until the accuracy of the table structure recognition model meets the requirements.

[0036] Further improvement or specific implementation steps of the aforementioned structured parsing method for power grid dispatching operation tickets. In the process of constructing a priori box for corner point detection, since the text size and table width in the operation ticket are relatively fixed, the pixel height of the text determines the length of the left and right borders of the text box, and the maximum width of the cell determines the maximum length of the upper and lower borders of the text box. Therefore, based on this, the limit size of the priori box can be further optimized and restricted, thereby reducing the amount of calculation.

[0037] The further improvement or specific implementation steps of the aforementioned structured parsing method for power grid dispatching operation tickets include a step for preprocessing function words in the structured parsing step of the operation ticket string, specifically: determining the character string that may be a function word according to the operation ticket text, establishing a function word library, and for the character string D to be analyzed x , determine whether it is a string element in the virtual word library or the first character of a string element. If so, respectively form a double string with the previous string and the next string, and search whether the double string belongs to an element in the string matching dictionary. If so, retain the string to be analyzed D x , otherwise delete it. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a basic flow chart of the structured parsing method of power grid dispatching operation tickets. DETAILED DESCRIPTION

[0039] The present invention is described in detail below in conjunction with specific embodiments.

[0040] The present application relates to a structured parsing method for power grid dispatching operation tickets, which is mainly used to better obtain the structured data in the operation ticket during the operation ticket dispatching management process, improve the accuracy of the data, and parse the specific elements of the operation ticket.

[0041] The specific steps are as follows Figure 1 As shown, including the following:

[0042] Step 1: Create a database of operation ticket text elements

[0043] It refers to determining the distribution network topology elements involved in various operations based on the specific operation content of the operation ticket, and creating an operation ticket text element database based on the text expression of the corresponding elements in the operation ticket issuance rules or marking methods;

[0044] The distribution network topology topology elements and topology nodes include: distribution line elements, distribution port elements, distribution equipment elements and operation content elements;

[0045] Distribution line elements are regional attribute elements in the topology of the distribution network. They are usually isolated by various switch nodes and used to supply specific areas or achieve specific purposes to form relatively independent distribution areas; including busbar bypass, bus tie line, wiring line, feeder line, segmented line, etc.

[0046] The distribution port element refers to the topological connection node between the topological lines or within the topological structure, mainly including the busbar and each bypass connection port, the line and the distribution equipment or the port that meets the connection;

[0047] The power distribution equipment element refers to the power distribution equipment corresponding to the operation object or execution object involved in the operation ticket, including various electrical equipment such as power generation equipment and transformer equipment;

[0048] The operation content element refers to the specific operation content or status information involved in the operation ticket, among which the status elements include operation, hot standby, cold standby, etc.; the action elements include maintenance, input, output, grounding, etc.;

[0049] Since the text information of the operation ticket involves a large number of homogeneous elements, and similar elements usually involve the same or similar operation content or attributes, by analyzing and extracting the text elements of the operation ticket content, we can better establish the text information classification, recognition and extraction process, provide an optimized reference for the recognition of text elements, and thus enhance the effect of text information extraction and recognition and improve recognition efficiency.

[0050] Step 2: Create a dictionary for matching operation ticket strings

[0051] Establishing a string matching dictionary based on the aforementioned operation ticket text element database;

[0052] The string matching dictionary is a three-layer expansion storage structure, and the strings in each layer are expanded downward in sequence;

[0053] The first layer element is all the single strings A in the string data pool. i Set, represented by W1 = {A1, A2...A i ...A I}, any single string A in the string set W1 i Expand downward to form the second layer of double string set W2, I represents the total number of single strings, A i Represents the i-th single string;

[0054] The second layer double string set W2 stores the single string A i Double string starting with A i B j and contains double string A i B j The remaining length of the longest string is L ij, and a single string A i The first two characters of a multi-string that begins with A i C k and contains double string A i C k The remaining length of the longest string is L i ' j , we get W2={A i B1L i1 ,A i B2L i2 ...A i B j L ij ...A i B J L iJ ,A i C1L i1 '...A i C k L i ' k ....A i C K L i ' K .}, where L ij =L maxij -2;L i ' k =L′ maxik -2;L maxij It means that it contains double string A i B j The length of the longest string, L′ maxik It means that it contains double string A i C k The length of the longest string; any double string A in the string set W2 i B j or A i C k Expand downward to form the third-level multi-string set W3;

[0055] The third-layer multi-string set W3 stores the double string A in the second layer. i B j or A i C k A multi-character string with the first character;

[0056] The content of the operation ticket is usually concise and does not contain too much meaningless vocabulary and text. In addition, a small number of consecutive character strings in the text information have clear attribute meanings. By constructing a three-layer structure of single-character and double-character first-character strings, better operating conditions can be created for the recognition and extraction of subsequent text information, reducing unnecessary operation content, further reducing the consumption of virtual resources such as system storage and computing power, and simplifying the solution processing flow.

[0057] Step 3: Optimize the operation ticket image

[0058] Since the operation ticket may be deformed losslessly during use, and the image may be tilted due to factors such as shooting angle and lighting during the image acquisition of the operation ticket, it is necessary to pre-process it to improve the recognizability of the information elements in the operation ticket and reduce the difficulty of processing. Specifically, it can be divided into the following steps:

[0059] Step 3a: Preprocessing of operation ticket images

[0060] In order to reduce the difficulty of data processing, the original image of the operation ticket is grayed to obtain a gray image to remove unnecessary chromaticity information and improve text recognizability. The gray image is binarized to extract its binary contour, and then the image is filtered to filter out pixels below a preset threshold to reduce noise and remove impurities. Then, the original operation ticket image is filtered according to the position of the pixels that have not been filtered out in the image to intercept the valid data in the original operation ticket image.

[0061] Step 3b: Correct the image of the operation ticket

[0062] Since the content of the operation ticket involves strict operation steps and corresponding operation objects, it is reflected in the top-down operation content sequence and the specific object mark of the operation ticket on the operation ticket. The above content is strictly constrained in the form of a table or a table-like form combined with text. In order to ensure that the line boundary and other information in the operation ticket are retained, simplify the difficulty of subsequent text recognition processing, and solve the problem of skewed boundaries of the operation ticket image due to angle changes and other reasons, it is necessary to correct the operation ticket image. Specifically:

[0063] For the aforementioned binarized image, straight line recognition and extraction are performed, straight line lengths are counted, a length threshold for distinguishing text lines from boundary lines is set, text lines are removed, non-text lines are retained, horizontal and vertical coordinate values ​​of non-text lines are counted, straight lines with similar coordinate values ​​but not linked are extended to make them intersect, the inclination angle of the straight lines in the image is calculated, and reverse adjustment and correction are performed based on the inclination angle result so that the straight lines in the operation ticket are in a relatively horizontal direction, and the operation ticket is restored and corrected as a whole based on the relative position relationship between each text element and the straight line;

[0064] Step 4: Identify and extract elements from the operation ticket image

[0065] From the above, we can see that in addition to the basic text elements, the information of the operation ticket image elements also needs to extract table or table-like information elements, while retaining the relative position relationship between each text element block and the table element, so as to confirm important information such as the specific operation steps of the operation ticket. In order to ensure that the structure of the operation ticket text element recognition can still retain these element information, it is necessary to determine the table element information while recognizing the text in the operation ticket; specifically:

[0066] 4a. Identify and extract elements of the operation ticket form;

[0067] A model for extracting operation ticket table features is established based on a multi-layer convolutional neural network, where the input of the model is the operation ticket image after optimization processing in step 3, and the output is a table graph structure and the adjacency relationship of text sequences in each table graph structure;

[0068] The table structure is established based on a graph structure composed of a set of vertex elements and edge elements. In the table graph structure, the vertex elements are text sequences of each cell in the table, and the edge elements are row and column position relationships between text rows. The position relationships include: in the same row and column, adjacent rows, adjacent columns, and non-adjacent columns. In the same row and column, two text sequences are in the same cell in the table, adjacent rows mean that two text sequences are located in two adjacent rows, and adjacent columns mean that two text sequences are located in two adjacent columns.

[0069] During the training and learning process, the adjacency matrix is ​​used to express the adjacency relationship of the aforementioned edge elements, a table database is established, and it is divided into a training data set and a verification data set. The tables in the training data set are used to train and optimize the table structure recognition model, and the verification data set is used to verify and feedback the model training results until the accuracy of the table structure recognition model meets the requirements;

[0070] Input the operation ticket graphic into the table structure recognition model to extract the table structure in the operation ticket, and restore the actual table and the relative position relationship of the text sequence in the table according to the adjacency relationship of the text sequence in the table structure;

[0071] 4c. Identification and extraction of text elements of the operation ticket, including the following steps:

[0072] Text element feature extraction: Based on the recognition and extraction of table elements, the relative position relationship of each text sequence can be confirmed, but the text elements in the text sequence cannot be directly obtained. At the same time, the text element information of the operation ticket often includes different types of text elements such as Chinese characters, letters, numbers, symbols, etc. Therefore, in order to achieve effective recognition and extraction of text elements, this application adopts the DSSD network model to extract text element features. The DSSD network model is obtained by replacing the convolution kernel network of the SSD network with a fast training residual network, introducing a residual module before classification and regression, and adding a deconvolution layer to the auxiliary convolution layer to form an improved model;

[0073] Text sequence corner point detection: Text corner point detection is used to locate the border of the text sequence to facilitate the extraction of text content in the operation ticket image, and to remove data outside the text border to improve recognition efficiency. At the same time, the coordinates and relative position relationship of the text box are used to restore the text order in the text sequence to ensure the effective acquisition of text information. Because the text in the operation ticket is relatively standardized and the size is basically the same, the diagonal box and its center position obtained by corner point detection can easily locate the corner point. Corner point detection converts the detection of the text boundary box into the detection process of the corner point through the convolutional neural network.

[0074] In particular, in the process of constructing the prior box for corner detection, since the text size and table width in the operation ticket are relatively fixed, the pixel height of the text determines the length of the left and right borders of the text box, and the maximum width of the cell determines the maximum length of the upper and lower borders of the text box. Therefore, based on this, the limit size of the prior box can be further optimized and restricted, thereby reducing the amount of calculation.

[0075] Step 5: Based on the results of the operation ticket text element recognition and extraction in step 4, the generated operation ticket string is structurally parsed, including the following steps:

[0076] 5a. For an operation ticket string, for example, D1D2D3...D x ...D X , in the string matching dictionary, whether the first double string just queried exists in the second-layer double string set W2, that is, whether the double string D1D2 is an element in the second-layer double string set W2;

[0077] 5b. If the first double character string does not exist in the second-layer double character string set W2, separate the first and last single character strings D1, end the word segmentation, separate the first single character string D1 from the operation ticket character string, and return to step 5a to continue analyzing the remaining character strings;

[0078] 5c. If the first double character string exists in the second-layer double character string set W2, verify whether the double character string exists in the corresponding third layer; if not, end the word segmentation; if so, count the length n of the longest character string in the third layer; record x=2, and go to step 5d;

[0079] 5d. Let x=x+1. If x=n+3, go to step A6, otherwise go to step 5e.

[0080] 5e. Determine whether there is a D3D4...D x If the string headed by exists, go to step 5f, otherwise go to step 5d;

[0081] 5f, extraction can be combined with D3D4...D x The matched string is separated from the D1D2 component string, and the process returns to step 5a to continue analyzing the remaining string.

[0082] In particular, in the aforementioned step of structural analysis of the operation ticket string, a step for preprocessing function words is also included, specifically, determining the character string that may be a function word according to the operation ticket text, establishing a function word library, and for the character string D to be analyzed x , determine whether it is a string element in the virtual word library or the first character of a string element. If so, respectively form a double string with the previous string and the next string, and search whether the double string belongs to an element in the string matching dictionary. If so, retain the string to be analyzed D x , otherwise delete it.

[0083] Through function word analysis, the function words in the operation ticket string text and other content that does not directly affect the meaning of the text information can be streamlined and screened out, reducing the total amount of data that needs to be processed and analyzed, improving the processing capacity of the solution, and reducing the consumption of system computing resources.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention, rather than to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the essence and scope of the technical solution of the present invention.

Claims

1. A structured parsing method for power grid dispatching operation tickets, characterized in that: The steps include: Step 1: Create a database of operation ticket text elements It refers to determining the distribution network topology elements involved in various operations based on the specific operation content of the operation ticket, and creating an operation ticket text element database according to the text expression of the corresponding elements in the operation ticket issuance rules or marking methods, and the distribution network topology topology elements and topology nodes include: distribution line elements, distribution port elements, distribution equipment elements and operation content elements; Step 2: Create a dictionary for matching operation ticket strings Establishing a string matching dictionary based on the aforementioned operation ticket text element database; The string matching dictionary is a three-layer expansion storage structure, and the strings in each layer are expanded downward in sequence; The first layer element is all the single strings A in the string data pool. i Set, represented by W1 = {A1, A2...A i ...A I }, any single string A in the string set W1 i Expand downward to form the second layer of double string set W2, I represents the total number of single strings, A i Represents the i-th single string; The second layer double string set W2 stores the single string A i Double string starting with A i B j and contains double string A i B j The remaining length of the longest string is L ij , and a single string A i The first two characters of a multi-string that begins with A i C k and contains double string A i C k The remaining length of the longest string is L i ' j , we get W2 = {AiB1Li1,AiB2Li2...AiBjLij...AiB J Li J ,AiC1Li′1...AiCkLi′k....AiC K Li′ K .}, where L ij =L maxij -2;L i ' k =L′ maxik -2;L maxij It means that it contains double string A i B j The length of the longest string, L′ maxik It means that it contains double string A i C k The length of the longest string; any double string A in the string set W2 i B j or A i C k Expand downward to form the third-level multi-string set W3; The third-layer multi-string set W3 stores the double string A in the second layer. i B j or A i C k A multi-character string with the first character; Step 3: Optimizing the image of the operation ticket, including the following steps: Step 3a, preprocessing of the operation ticket image, graying the acquired original operation ticket image to obtain a gray image to remove unnecessary chromaticity information, binarizing the gray image to extract its binary contour, and then filtering the image to filter out pixels below a preset threshold to reduce noise and remove impurities; then filtering the original operation ticket image according to the position of the pixels that have not been filtered out in the image to intercept valid data in the original operation ticket image; Step 3b, image correction of the operation ticket: for the aforementioned binarized image, straight line recognition and extraction are performed, straight line lengths are counted, a length threshold for distinguishing text lines and boundary lines is set, text lines are removed, non-text lines are retained, horizontal and vertical coordinate values ​​of non-text lines are counted, straight lines with similar coordinate values ​​but not linked are extended to make them intersect, the inclination angle of the straight lines in the image is calculated, and reverse adjustment and correction are performed based on the inclination angle result, so that the straight lines in the operation ticket are in a relatively horizontal direction, and the operation ticket is restored and corrected as a whole based on the relative position relationship between each text element and the straight line; Step 4: Identification and extraction of elements in the operation ticket image refers to determining the table element information while recognizing the text in the operation ticket; specifically: 4a. Identify and extract elements of the operation ticket form; A model for extracting operation ticket table features is established based on a multi-layer convolutional neural network, wherein the input of the model is the operation ticket image after optimization processing in step 3, and the output is the table graph structure and the adjacency relationship of the text sequences in each table graph structure; the operation ticket image is input into the table structure recognition model to extract the table structure in the operation ticket, and the actual table and the relative position relationship of the text sequence in the table are restored according to the adjacency relationship of the text sequence in the table structure; 4c. Identification and extraction of text elements of the operation ticket, including the following steps: Text element feature extraction: Based on the recognition and extraction of table elements, the relative position relationship of each text sequence can be confirmed, but the text elements in the text sequence cannot be directly obtained. At the same time, the text element information of the operation ticket often includes different types of text elements such as Chinese characters, letters, numbers, symbols, etc. Therefore, in order to achieve effective recognition and extraction of text elements, this application adopts the DSSD network model to extract text element features. The DSSD network model is obtained by replacing the convolution kernel network of the SSD network with a fast training residual network, introducing a residual module before classification and regression, and adding a deconvolution layer to the auxiliary convolution layer to form an improved model; Text sequence corner point detection: Text corner point detection is used to locate the border of the text sequence to facilitate the extraction of text content in the operation ticket image, and to remove data outside the text border to improve recognition efficiency. At the same time, the coordinates and relative position relationship of the text box are used to restore the text order in the text sequence to ensure the effective acquisition of text information. Because the text in the operation ticket is relatively standardized and the size is basically the same, the diagonal box and its center position obtained by corner point detection can easily locate the corner point. Corner point detection converts the detection of the text boundary box into the detection process of the corner point through the convolutional neural network. Step 5: Based on the results of the operation ticket text element recognition and extraction in step 4, the generated operation ticket string is structurally parsed, including the following steps: 5a. For an operation ticket string, for example, D1D2D3...D x ...D X , in the string matching dictionary, whether the first double string just queried exists in the second-layer double string set W2, that is, whether the double string D1D2 is an element in the second-layer double string set W2; 5b. If the first double character string does not exist in the second-layer double character string set W2, separate the first and last single character strings D1, end the word segmentation, separate the first single character string D1 from the operation ticket character string, and return to step 5a to continue analyzing the remaining character strings; 5c. If the first double character string exists in the second-layer double character string set W2, verify whether the double character string exists in the corresponding third layer; if not, end the word segmentation; if so, count the length n of the longest character string in the third layer; record x=2, and go to step 5d; 5d. Let x=x+1. If x=n+3, go to step A6, otherwise go to step 5e. 5e. Determine whether there is a D3D4...D x If the string headed by exists, go to step 5f, otherwise go to step 5d; 5f, extraction can be combined with D3D4...D x The matched string is separated from the D1D2 component string, and the process returns to step 5a to continue analyzing the remaining string.

2. The structured parsing method for power grid dispatching operation tickets according to claim 1, characterized in that: The distribution line element is a regional attribute element in the distribution network topology, which is usually isolated by various switch nodes and used to supply specific areas or achieve specific purposes to form a relatively independent distribution area; Including mother bypass, mother tie line, connection line, feeder line, section line, etc.; The distribution port element refers to the topological connection node between the topological lines or within the topological structure, mainly including the busbar and each bypass connection port, the line and the distribution equipment or the port that meets the connection; The power distribution equipment element refers to the power distribution equipment corresponding to the operation object or execution object involved in the operation ticket, including various electrical equipment such as power generation equipment and transformer equipment; The operation content element refers to the specific operation content or status information involved in the operation ticket, among which the status elements include operation, hot standby, cold standby, etc.; the action elements include maintenance, input, output, grounding, etc.

3. The structured parsing method for power grid dispatching operation tickets according to claim 1, characterized in that: The table structure is established based on a graph structure composed of a set of vertex elements and edge elements. In the table graph structure, the vertex elements are text sequences of each cell in the table, and the edge elements are row and column position relationships between text rows. The position relationships include: in the same row and column, adjacent rows, adjacent columns, and non-adjacent columns. In the same row and column, two text sequences are in the same cell in the table, adjacent rows mean that two text sequences are located in two adjacent rows, and adjacent columns mean that two text sequences are located in two adjacent columns. During the training and learning process, the adjacency matrix is ​​used to express the adjacency relationship of the aforementioned edge elements. A table database is established and divided into a training data set and a verification data set. The tables in the training data set are used to train and optimize the table structure recognition model, and the verification data set is used to verify and feedback the model training results until the accuracy of the table structure recognition model meets the requirements.

4. The structured parsing method for power grid dispatching operation tickets according to claim 1, characterized in that: In the process of constructing the prior box for corner detection, since the text size and table width in the operation ticket are relatively fixed, the pixel height of the text determines the length of the left and right borders of the text box, and the maximum width of the cell determines the maximum length of the upper and lower borders of the text box. Therefore, based on this, the limit size of the prior box can be further optimized and restricted, thereby reducing the amount of calculation.

5. The structured parsing method for power grid dispatching operation tickets according to claim 1, characterized in that: The step of structural analysis of the operation ticket character string also includes a step for preprocessing function words, specifically, determining the character string that may be a function word according to the operation ticket text, establishing a function word library, and for the character string to be analyzed, x , determine whether it is a string element in the virtual word library or the first character of a string element. If so, respectively form a double string with the previous string and the next string, and search whether the double string belongs to an element in the string matching dictionary. If so, retain the string to be analyzed D x , otherwise delete it.

Images