A real-time adaptive setting value checking method and system for a relay protection device

Abstract

The application provides a real-time self-adaptive setting value checking method and system for a relay protection device, which comprises the following steps: firstly, through information entropy and geometric topology analysis, an effective setting value table is automatically identified and analyzed from real-time data of the relay protection device to obtain a device end setting value item; a mixed feature vector containing text semantics, numerical values and context is constructed for the device end setting value item and a preset standard setting value item; through calculation of semantic similarity between the vectors, intelligent and accurate matching between setting value items with different sources and different naming modes is realized; finally, unit normalization and numerical value comparison are performed on the setting value item pairs that are successfully matched, and a final checking report is automatically generated. The application realizes automation of the relay protection setting value checking process and solves the problems of low efficiency and error-prone of manual checking.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of relay protection device detection, and particularly relates to a real-time self-adaptive setting value checking method and system for a relay protection device. BACKGROUND

[0002] A relay protection device is the nerve center for ensuring the safe and stable operation of a power system, and whether the protection setting values in the relay protection device are correct directly relates to whether the fault can be accurately and quickly removed. Therefore, before the device is put into operation and during regular maintenance, a special detection device must be used to read the actual setting values in the device, and the actual setting values are checked item by item with a standard setting value sheet issued by a design institute to ensure that the actual setting values are completely consistent with the standard setting value sheet. The checking work is a key link for ensuring the safety of a power grid, and the accuracy requirement is extremely high.

[0003] However, in the current practice, the checking work, which is crucial, still largely depends on the manual operation of an on-site engineer. The engineer needs to compare the setting values obtained from the relay protection device with a standard setting value sheet in paper or electronic form item by item. This process faces great challenges: the naming method, display order and unit representation of the setting values in the relay protection devices of different manufacturers and different models are significantly different. For example, the same "overcurrent I section setting value" may be displayed as "OC-1 Set" on one device and as "I> Set Value" on another device. The engineer must rely on his own professional knowledge and rich experience to understand and associate the setting value items with different names and structures, which is a huge workload and is extremely easy to cause missing checking or incorrect checking due to negligence or fatigue. SUMMARY

[0004] The application provides a real-time self-adaptive setting value checking method and system for a relay protection device to solve the problem that the manual workload of setting value checking is huge and missing checking or incorrect checking is easy to occur.

[0005] In a first aspect, the application provides a real-time self-adaptive setting value checking method for a relay protection device, which includes the following steps:

[0006] Data information entropy is calculated for real-time printing data obtained from the relay protection device, and a candidate data stream with a significant structured feature is identified from the real-time printing data based on a preset entropy value threshold;

[0007] Frame line structure feature points in the candidate data stream are extracted, and the candidate data stream is confirmed as a valid setting value table through a global grid topology model constructed by a geometric consistency verification algorithm;

[0008] The table boundary of the effective setting value table is defined according to the global grid topology model, and the content of the effective setting value table is parsed based on the table boundary to have the local abnormal row fault tolerance capability, and a device end setting value item of the relay protection device is parsed;

[0009] The device end setting value item and the standard setting value item in the preset standard setting value table are respectively fused in features to obtain a hybrid feature vector containing text semantics, numerical attributes and context information;

[0010] The semantic similarity between the hybrid feature vectors of the device end setting value item and the standard setting value item is calculated, and the cross-source matching of the device end setting value item and the standard setting value item is completed according to the semantic similarity;

[0011] The device end setting value item and the standard setting value item matched successfully in the cross-source are subjected to unit normalization processing and numerical comparison, and a setting value checking report is generated according to the numerical comparison result.

[0012] Optionally, the data information entropy calculation on the real-time printing data obtained from the relay protection device, and the identification of the candidate data stream with a structured feature from the real-time printing data based on a preset entropy value threshold value include the following steps:

[0013] The window size and the sliding step of the sliding window are set, the sliding window is moved along the real-time printing data, and the data byte frequency distribution in the window position of each sliding window is calculated;

[0014] The information entropy value is calculated for each window position according to the data byte frequency distribution and by applying the Shannon entropy formula, and an entropy value sequence curve corresponding to the data stream position is generated;

[0015] When the entropy value sequence curve crosses the preset first entropy value threshold value from high to low, the current data stream position is marked as the starting point of the candidate data stream, and when the entropy value sequence curve crosses the preset second entropy value threshold value from low to high, the current data stream position is marked as the end point of the candidate data stream, and the second entropy value threshold value is greater than the first entropy value threshold value.

[0016] Optionally, the frame line structure feature points in the candidate data stream are extracted, and the global grid topology model constructed by the geometric consistency verification algorithm confirms the candidate data stream as the effective setting value table, including the following steps:

[0017] A character set containing a plurality of table drawing characters is defined in advance;

[0018] All data in the candidate data stream are traversed, all target characters belonging to the character set are identified, and the two-dimensional coordinates of the target characters in the candidate data stream are extracted;

[0019] All two-dimensional coordinates are collected as two-dimensional point clouds;

[0020] The global grid topology model constructed based on the two-dimensional point cloud and through the geometric consistency checking algorithm confirms the candidate data stream as an effective fixed value table.

[0021] Optionally, the global grid topology model constructed based on the two-dimensional point cloud and through the geometric consistency checking algorithm confirms the candidate data stream as an effective fixed value table includes the following steps:

[0022] Randomly sampling a minimum point set from the two-dimensional point cloud, and assuming a global grid topology model based on the minimum point set;

[0023] Traversing all points in the two-dimensional point cloud according to the global grid topology model, and determining points that are geometrically consistent with the global grid topology model as inliers;

[0024] Repeating the iteration process of random sampling, model assumption, and inlier statistics until a preset iteration number is reached;

[0025] In all iteration processes, the global grid topology model that obtains the most inlier support is selected as the optimal model;

[0026] If the inlier number ratio of the optimal model exceeds a preset inlier ratio threshold, the candidate data stream is confirmed as an effective fixed value table.

[0027] Optionally, the feature fusion of the device-side fixed value item and the standard fixed value item in the preset standard fixed value single respectively to obtain a hybrid feature vector containing text semantics, numerical attribute, and context information includes the following steps:

[0028] For the device-side fixed value item or the standard fixed value item in the preset standard fixed value single, extracting fixed value item information of the fixed value item and relative position information of the fixed value item in the corresponding customization table or customization single, the fixed value item information including fixed value item description text, fixed value item value, and fixed value item unit;

[0029] Generating a text semantic vector based on the fixed value item description text, generating a numerical attribute vector by combining the fixed value item value and the fixed value item unit and through encoding, and generating a context feature vector based on the relative position information and through encoding;

[0030] Weighted fusion of the text semantic vector, the numerical attribute vector, and the context feature vector to obtain a hybrid feature vector.

[0031] Optionally, the generation of the text semantic vector based on the fixed value item description text includes the following steps:

[0032] Carrying out word segmentation processing on the fixed value item description text to obtain a word sequence;

[0033] Loading a pre-trained word vector model, and mapping each word in the word sequence to a high-dimensional word vector using the word vector model.

[0034] The weighted average of all high-dimensional word vectors corresponding to the description text of the fixed value item is calculated, and the calculated weighted average vector is taken as the text semantic vector of the fixed value item.

[0035] Optionally, the semantic similarity between the mixed feature vector of the device-side fixed value item and the standard fixed value item is calculated, and the cross-source matching of the device-side fixed value item and the standard fixed value item is completed according to the semantic similarity, including the following steps:

[0036] For any device-side fixed value item, the cosine similarity between the first mixed feature vector of the device-side fixed value item and the second mixed feature vector of all standard fixed value items is calculated respectively;

[0037] The highest cosine similarity among all cosine similarities is selected, and if the highest cosine similarity exceeds a preset confidence threshold, it is determined that the device-side fixed value item and the standard fixed value item corresponding to the highest cosine similarity are successfully cross-source matched, forming a set of fixed value item pairs.

[0038] The above steps are repeated until the cosine similarity between the first mixed feature vector of all device-side fixed values and the second mixed feature vector is calculated.

[0039] Optionally, the successfully cross-source matched device-side fixed value item and the standard fixed value item are subjected to unit normalization processing and numerical comparison, and a fixed value checking report is generated according to the numerical comparison result, including the following steps:

[0040] All fixed value item pairs are traversed, and after unit normalization processing of each fixed value item pair, numerical comparison is performed, and the comparison result is classified as consistent items or inconsistent items;

[0041] All standard fixed value items that exist in the standard fixed value list and are not successfully cross-source matched are identified and classified as missing items;

[0042] All device-side fixed value items that exist in the device-side fixed value item and are not successfully cross-source matched are identified and classified as unplanned items;

[0043] The information of all consistent items, inconsistent items, missing items and unplanned items is summarized according to a preset format, and a final fixed value checking report is generated.

[0044] In a second aspect, the present application also provides a real-time adaptive fixed value checking system for a relay protection device, which comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor implements the real-time adaptive fixed value checking method for the relay protection device as described in the first aspect when executing the computer program.

[0045] In a third aspect, the present application also provides a computer readable storage medium, which stores instructions configured to cause a processor to perform the buried metal pipeline-based dual-source magnetic field comprehensive detection analysis method according to the first aspect when the instructions are executed by the processor.

[0046] The present application has the following beneficial effects:

[0047] The present application solves the core problems of complex manual comparison, low efficiency and error-prone in the traditional relay protection setting value checking work caused by the diversity of equipment sources and the inconsistency of naming specifications. By introducing information entropy and geometric topology analysis, the present application realizes automatic identification and analysis of setting value tables from complex real-time data streams, and ensures accurate extraction of content even in the case of local defects in the table. Moreover, a hybrid feature vector containing text semantics, numerical attributes and context information is constructed, and semantic similarity calculation is used to realize intelligent matching of device-side setting value items and standard setting value items. This enables the automatic identification process to understand the same technical connotation under different expressions, thereby overcoming the barriers between manufacturers, models and versions, and ensuring the uniqueness and accuracy of matching. The present application changes the entire checking process from manual operation relying on human experience to automated intelligent data processing and comparison process, greatly improving the efficiency and reliability of setting value checking. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 FIG. 1 is a flowchart of a relay protection device real-time adaptive setting value checking method according to one embodiment of the present application.

[0049] Figure 2 FIG. 4 is a device connection diagram for real-time acquisition of device-side setting value items according to one embodiment of the present application.

[0050] Figure 3 FIG. 6 is a setting value comparison diagram according to one embodiment of the present application.

[0051] Figure 4 FIG. 9 is a diagram of frame line structure feature points in a candidate data stream according to one embodiment of the present application.

[0052] Figure 5 FIG. 11 is a preview diagram of a setting value checking report according to one embodiment of the present application. DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the present application will be described clearly below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art belong to the scope of protection of the present application.

[0054] The terms "first", "second", etc. in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a particular order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second", etc. are generally a class and do not limit the number of objects, for example, the first object can be one or more. In addition, "and / or" in the specification and claims indicates at least one of the connected objects, and the character " / ", generally indicates that the objects before and after are in a "or" relationship.

[0055] Figure 1 The flowchart of the real-time adaptive setting value checking method of the relay protection device in one embodiment is shown. It should be understood that, although Figure 1 The steps in the flowchart are displayed in sequence according to the direction of the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps in the flowchart can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily sequential, but can be executed in rotation or alternation with at least part of other steps or sub-steps or stages of other steps. As Figure 1 The real-time adaptive setting value checking method of the relay protection device disclosed in the present application specifically includes the following steps:

[0056] S101. Calculate the data information entropy of the real-time print data obtained from the relay protection device, and identify the candidate data stream with significant structured features from the real-time print data based on the preset entropy value threshold.

[0057] Wherein, referring to Figure 2The message acquisition module is installed in the substation screen cabinet equipped with a relay protection device and is connected to the printing port of the protection device. The acquisition module is operated for a long time, and the mobile terminal and the printing data collector are temporarily connected (such as using an RJ11 quick plug interface) as needed, and the real-time printing data collected therein is synchronized. When calculating the information entropy of the real-time printing data obtained from the relay protection device, first, the size and step parameters of the sliding window are set, and then the window is moved along the data stream and the frequency distribution of the bytes in each window is counted. The information entropy value of each window position is calculated using the Shannon entropy formula. When the data stream contains structured content such as tables, the character distribution is relatively regular, and the information entropy is low; when random text is included, the character distribution is chaotic, and the information entropy is high. By setting two entropy threshold values, when the entropy value curve crosses the first threshold value from high to low, it is marked as the starting point of the candidate data stream, and when it crosses the second threshold value from low to high, it is marked as the end point. This method can automatically identify data paragraphs with obvious structured features and effectively filter out irrelevant text information.

[0058] S102. Extract the frame line structure feature points in the candidate data stream, and construct a global grid topology model through a geometric consistency verification algorithm to confirm the candidate data stream as an effective fixed value table.

[0059] Among them, referring to Figure 3 When extracting the frame line structure features in the candidate data stream, a character set containing various table drawing characters is defined in advance, such as horizontal lines, vertical lines, intersection points, and the like. All characters in the candidate data stream are traversed, target characters belonging to the character set are identified, and their two-dimensional coordinate positions in the data stream are recorded to form two-dimensional point cloud data. The RANSAC geometric consistency verification algorithm is used to analyze the point cloud, randomly sample the minimum point set to assume a global grid topology model, and then count the number of inliers that conform to the geometric position of the model. After repeated iterations, the model with the most inlier support is selected as the optimal solution. When the inlier proportion of the optimal model exceeds a preset threshold, the candidate data stream is confirmed as an effective fixed value table. This method effectively distinguishes between real tables and pseudo table structures through geometric constraints, ensuring that only data with complete grid topology is identified as a fixed value table, significantly improving the accuracy and robustness of table recognition.

[0060] S103. Define the table boundary of the effective fixed value table according to the global grid topology model, and perform content analysis on the effective fixed value table based on the table boundary with local abnormal row fault tolerance capability to obtain the device end fixed value item of the relay protection device.

[0061] Wherein, after defining the table boundary according to the global grid topology model, a content parsing algorithm based on grid structure is used to extract data from the fixed value table. This algorithm has the ability to tolerate local abnormal rows and can handle problems such as irregular format, missing or misaligned characters in the table. The boundary range of each cell is determined by the grid topology model, and then the text content is extracted within the boundary. For detected abnormal rows, the algorithm will infer and correct according to the structural characteristics of adjacent normal rows to ensure the integrity of data extraction. During the parsing process, the name, value, unit and other key information of the fixed value item are identified, and their relative position relationship in the table is preserved. This fault-tolerant mechanism enables the algorithm to cope with poor print quality, format changes and other situations in actual applications, ensuring that the key relay protection device fixed value information can be accurately extracted.

[0062] S104. The device-end fixed value item and the standard fixed value item in the preset standard fixed value table are respectively fused in feature, to obtain a mixed feature vector containing text semantics, numerical attribute and context information.

[0063] Wherein, the complete information of each fixed value item is extracted, including the fixed value item description text, value, unit and relative position in the table. After the fixed value item description text is processed by word segmentation, the word is mapped to a high-dimensional vector using a pre-trained word vector model, and then a weighted average is taken to obtain a text semantic vector. Combined with the value and unit information of the fixed value item, a numerical attribute vector is generated through encoding technology, which can capture the magnitude feature of the value and the unit type. Based on the row and column position information of the fixed value item in the table, a context feature vector is encoded to reflect the structured position relationship of the fixed value item. Finally, the three types of feature vectors are fused according to the preset weight to form a mixed feature vector containing text semantics, numerical attribute and context information.

[0064] S105. The semantic similarity between the mixed feature vectors of the device-end fixed value item and the standard fixed value item is calculated, and the cross-source matching of the device-end fixed value item and the standard fixed value item is completed according to the semantic similarity.

[0065] Wherein, when calculating the semantic similarity between the device-end fixed value item and the standard fixed value item, the cosine similarity algorithm is used to measure the similarity of the mixed feature vectors. For each device-end fixed value item, the cosine similarity with all standard fixed value items is calculated, and the highest similarity value is selected for judgment. When the highest similarity exceeds the preset confidence threshold, it is determined that the device-end fixed value item and the corresponding standard fixed value item are successfully matched, forming a fixed value item pair. This cross-source matching method based on semantic similarity can handle expression differences between different data sources, such as synonyms, abbreviations, changes in description order, etc., and realizes intelligent fixed value item association. By setting an appropriate confidence threshold, the accuracy of the matching can be ensured while avoiding false matching.

[0066] S106. The device-side fixed value items that match the standard fixed value items successfully are subjected to unit normalization processing and numerical comparison, and a fixed value checking report is generated according to the numerical comparison result.

[0067] wherein, referring to Figure 4 , first, the numerical values under different units are unified, such as converting kilovolts to volts and converting milliamperes to amperes, to ensure the accuracy of the numerical comparison. After completing the unit normalization, the numerical values of the two fixed values are directly compared, and it is judged whether they are consistent according to the preset error tolerance. The comparison result is classified into consistent items and inconsistent items, and missing items that exist in the standard fixed value table but are not matched successfully and unexpected items that exist in the device side but are not matched are identified. According to the preset format, all categories of fixed value information are summarized, including fixed value name, standard value, actual value, deviation degree and other detailed information, and a structured fixed value checking report is generated. The fixed value checking report is shown in Figure 5 , and the "uncertain" in the check result column in the report is "missing item" or "unexpected item". The report can clearly show the accuracy of the fixed value configuration of the relay protection device, help the operation and maintenance personnel quickly identify the configuration errors, omissions or redundant fixed value items, and ensure the reliable operation of the power system protection device.

[0068] In one of the embodiments, the real-time print data obtained from the relay protection device is subjected to data information entropy calculation, and based on a preset entropy value threshold, a candidate data stream with a significant structured feature is identified from the real-time print data, including the following steps:

[0069] The window size and sliding step of the sliding window are set, the sliding window is moved along the real-time print data, and the data byte frequency distribution in each sliding window position is calculated;

[0070] According to the data byte frequency distribution and applying the Shannon entropy formula, the information entropy value of each window position is calculated to generate an entropy value sequence curve corresponding to the data stream position;

[0071] When the entropy value sequence curve crosses the preset first entropy value threshold from high to low, the current data stream position is marked as the starting point of the candidate data stream, and when the entropy value sequence curve crosses the preset second entropy value threshold from low to high, the current data stream position is marked as the end point of the candidate data stream, and the second entropy value threshold is greater than the first entropy value threshold.

[0072] In this embodiment, two key parameters, window size and sliding step, need to be determined first. Window size determines the data range for each analysis, usually set to a length that can contain enough character samples, such as 64 or 128 bytes, to ensure the reliability of statistical results. Sliding step controls the distance of window movement, smaller step can provide finer analysis granularity but increase the amount of calculation, larger step is on the contrary. Move the sliding window along the real-time print data stream by the set step, perform statistical analysis on all data bytes in each window position. Record the number of occurrences of each byte value during the statistical process, calculate its frequency distribution in the current window. This frequency distribution reflects the character composition characteristics of the data: structured data such as tables have relatively regular character distribution, while random text presents more uniform or chaotic distribution pattern. According to the data byte frequency distribution of each window position, apply the Shannon entropy formula to calculate the information entropy value. Then arrange the information entropy values corresponding to each window position in the order of the position of the data stream to form a continuous entropy value sequence curve. This curve can intuitively show the trend of information complexity change in the entire data stream, providing a quantitative index for identifying structured data segments.

[0073] The boundary positions of the candidate data stream are identified based on the entropy value sequence curve and a preset double threshold mechanism. Set two entropy value thresholds: a lower first threshold for detecting the start of structured data, and a higher second threshold for detecting the end of structured data. When the entropy value sequence curve changes from a high entropy value region to a low entropy value region and crosses the first threshold, it indicates that the data stream changes from random text to structured content, and the current position is marked as the starting point of the candidate data stream. Conversely, when the curve changes from a low entropy value region to a high entropy value region and crosses the second threshold, it indicates that the structured content ends, and the current position is marked as the end point of the candidate data stream. The double threshold design avoids the frequent switching problem that may be caused by a single threshold, and improves the stability of boundary detection by introducing a hysteresis effect. This method can automatically identify data paragraphs containing tables, lists and other structured features, effectively filtering out irrelevant text information.

[0074] In one embodiment, calculating the information entropy value for each window position according to the data byte frequency distribution and applying the Shannon entropy formula includes the following steps:

[0075] Based on the structural role of characters in the fixed value table, all possible data bytes are pre-divided into multiple character categories, including at least: table frame line category, digit and decimal point category, and letter and space symbol category;

[0076] For each window position, the frequency distribution of data bytes falling into each preset character category is counted respectively;

[0077] The Shannon entropy formula is independently applied to the frequency distribution of each character category, and a set of classification information entropy values corresponding to each character category is calculated for the window position.

[0078] A preset weight coefficient is set for the classification information entropy values of each character category, wherein the weight coefficient of the table frame line category is set to be the highest.

[0079] The set of classification information entropy values of the window position and its corresponding weight coefficients are weighted and summed, thereby calculating the composite information entropy value of the enhanced structured feature perception.

[0080] In the present embodiment, the characters are classified based on their structural role in the value table, and a classification system is established according to the functional characteristics of different character types in table construction. The table frame line category includes characters used to draw table boundaries and dividing lines, such as Figure 4 As shown, including but not limited to the top-left corner border symbol "┏", the top-right corner border symbol "┓", the bottom-left corner border symbol "┗", the bottom-right corner border symbol "┛", the cell separator "┯" and "┠", etc., these characters are the core elements of the visual structure of the table. The number and decimal point category covers all numerical characters and decimal points, mainly carrying specific numerical information of the value. The letter and blank symbol category includes various letter characters, spaces, tabs, etc., which are usually used to represent value item names, unit identifiers and format adjustments. Through this functional-based character classification method, the originally mixed byte data can be orderly classified according to their actual role in the table. Each character category has a specific information expression function, and its distribution characteristics can reflect the structured degree of the data content. In the character category frequency statistics process for each sliding window position, all data bytes in the current window need to be traversed, and each byte is classified into the corresponding category according to the preset character classification rule.

[0081] The statistical process records the occurrence frequency of each specific character in each character category by counting, and then calculates the frequency distribution in the category. For the table frame line category, the distribution of various line characters is counted; for the number and decimal point category, the occurrence frequency of different numbers and decimal points is recorded; for the letter and space symbol category, the distribution pattern of letters and space characters is analyzed. This classification and statistical method can reveal the composition ratio and distribution law of different functional characters in the current data segment. When the data contains table structure, the frame line characters will show a relatively concentrated distribution characteristic, the number characters will appear intensively in the numerical area, and the letter characters will mainly distribute in the title and description area. When calculating the category information entropy by independently applying the Shannon entropy formula to each character category, the character frequency distribution within the category is used as the calculation basis. In the calculation process, first, the occurrence frequency of each character in each category is divided by the total number of characters in the category to obtain the probability distribution within the category, and then the formula is substituted to obtain the category information entropy. This classification entropy calculation method can independently evaluate the complexity and randomness within each character category. The low entropy value of the table frame line category indicates the distribution law of the line characters, the entropy value of the number category reflects the diversity of the numerical value, and the entropy value of the letter category reflects the complexity of the text content.

[0082] Next, the weight distribution is determined according to the importance of different category characters for table structure recognition. The table frame line characters are the key elements that constitute the visual structure of the table, and their distribution law directly determines whether the data has table characteristics, so they are given the highest weight coefficient. The number and decimal point characters carry the core information of the numerical value and play an important role in table recognition, so the weight coefficient is set to a medium level. The letter and space symbol characters mainly serve as auxiliary and formatting, and have relatively small influence on table structure judgment, so the weight coefficient is set to a lower level. When calculating the composite information entropy by weighting the category information entropy values of each character category and the corresponding weight coefficients, a linear combination method is used to calculate the composite information entropy.

[0083] The calculation formula of the composite information entropy value is as follows:

[0084]

[0085] wherein, is the composite information entropy value, m represents the total number of character categories, k represents the index of the character category, represents the weight coefficient of the kth character category, represents the set of the kth character category, represents the range of summation is all bytes belonging to the category , represents that under the premise that the byte belongs to the category , the byte is exactly ​The probability is actually calculated as follows: within the data window, the number of bytes... The number of times it appears, divided by the number of all items in the window belonging to that category. The total number of bytes. The part in parentheses calculates the internal entropy of the k-th character category itself, considering the variation within a single character category. For example, if table borders use only one type of character, their classification entropy will be low; if multiple types are used, the classification entropy will be higher. During the calculation, the classification entropy of each category is multiplied by its corresponding weight coefficient, and then all weighted results are summed to obtain the final composite entropy value. This weighted fusion method comprehensively considers the contributions of different character categories, forming a more comprehensive and accurate measure of structured features. Composite entropy inherits the advantages of each classification entropy while highlighting the role of key features through a weighting mechanism. When the data contains a regular table structure, the low entropy value of border characters will significantly reduce the composite entropy value under high weighting; when the data is random text, the entropy values ​​of each category are relatively high, and the composite entropy value also increases accordingly.

[0086] In one implementation, extracting the frame structure feature points within the candidate data stream and confirming the candidate data stream as a valid value table using a global mesh topology model constructed through a geometric consistency check algorithm includes the following steps:

[0087] A predefined character set containing various table drawing characters is used.

[0088] Traverse all data within the candidate data stream, identify all target characters belonging to the character set, and extract the two-dimensional coordinates of the target characters within the candidate data stream;

[0089] All two-dimensional coordinates are aggregated into a two-dimensional point cloud;

[0090] The global grid topology model, constructed based on two-dimensional point clouds and using a geometric consistency check algorithm, confirms the candidate data stream as a valid value table.

[0091] In this embodiment, when defining the table drawing character set, all possible character types for constructing the table structure need to be collected, and a complete character library is established. This character set includes horizontal line characters for drawing the horizontal boundaries and dividing lines of the table, vertical line characters for constructing the vertical boundaries and column dividing lines of the table, intersection point characters located at the intersection of horizontal lines and vertical lines to form grid nodes, corner point characters for identifying the four corner positions of the table, and other auxiliary drawing characters such as T-shaped connectors, L-shaped corner symbols, etc. The establishment of the character set needs to consider the differences of different printing devices and character coding standards, and ensure to cover various possible table drawing symbols. These characters have corresponding numerical representations in ASCII code, Unicode or other coding systems. By establishing the mapping relationship between character code values and character functions, a standardized character recognition library is formed. When traversing the candidate data stream to identify the target characters, each character in the data stream needs to be checked one by one to determine whether it belongs to the predefined table drawing character set. The identification process uses character code value matching to compare the encoding value of the current character with the standard encoding in the character set to determine whether it is the target character. For each identified target character, its two-dimensional coordinate position information in the candidate data stream needs to be accurately extracted. In the coordinate extraction process, a coordinate system is established with the starting position of the data stream as the origin, the horizontal coordinate represents the column position of the character in the current row, and the vertical coordinate represents the row number of the character.

[0092] All the extracted target character coordinates are collected as a two-dimensional point cloud, which is composed of a series of coordinate points, each point representing a table drawing character in the spatial position of the data stream, and the coordinate values of the points reflect the row and column position relationship of the characters. The point cloud data structure is usually stored in the form of array or list, each element contains two numerical values of horizontal and vertical coordinates. In the point cloud construction process, coordinate standardization needs to be performed to eliminate the coordinate offset and scaling differences that may exist in different data sources. The quality of the point cloud directly affects the accuracy of subsequent geometric analysis, therefore, abnormal coordinate points need to be detected and filtered to eliminate coordinate values that obviously deviate from the normal range. Next, based on the two-dimensional point cloud, a global grid topology model is constructed through a geometric consistency verification algorithm and the RANSAC random sample consensus algorithm is used to identify the regular grid structure in the point cloud. The algorithm first randomly selects a minimum point set from the point cloud, and based on these points, a candidate grid topology model is assumed, which defines the row and column spacing, angle and overall layout parameters of the grid. Then all other points in the point cloud are traversed, the geometric deviation of each point from the assumed model is calculated, and the points with a deviation less than a predetermined threshold are determined as inliers that support the model.

[0093] The iterative process of random sampling, model assumption, and inner point statistics is repeatedly performed, and each iteration produces a candidate model and a corresponding inner point set. After all iterations are completed, the model that obtains the most inner point support is selected as the optimal global grid topology model. When the proportion of the number of inner points of the optimal model to the total number of points exceeds a preset threshold, it is confirmed that the candidate data stream has a complete table grid structure, and it is identified as a valid constant table. This geometric consistency test method can effectively distinguish between real table structures and randomly distributed characters, ensuring that only data with regular grid characteristics are identified as constant tables.

[0094] In one embodiment, the global grid topology model constructed based on the two-dimensional point cloud and through the geometric consistency test algorithm confirms the candidate data stream as a valid constant table, including the following steps:

[0095] Randomly sample a minimum point set from the two-dimensional point cloud, and assume a global grid topology model based on the minimum point set;

[0096] According to the global grid topology model, traverse all points in the two-dimensional point cloud, and determine points that are geometrically consistent with the global grid topology model as inner points;

[0097] Repeat the iteration process of random sampling, model assumption, and inner point statistics until a preset number of iterations is reached;

[0098] In all iteration processes, select the global grid topology model that obtains the most inner point support as the optimal model;

[0099] If the proportion of the number of inner points of the optimal model exceeds a preset inner point proportion threshold, the candidate data stream is confirmed as a valid constant table.

[0100] In this embodiment, for a regular rectangular grid structure, the minimum point set usually contains 4 non-collinear points, which can define the basic geometric parameters of the grid. The random sampling process uses a uniform distribution random number generator to randomly select a specified number of point indexes from the total number of points in the point cloud, ensuring that each point has an equal probability of being selected. Based on the selected minimum point set, the key parameters of the global grid topology model are derived through geometric calculation, including the row spacing, column spacing, starting coordinate position, and overall tilt angle of the grid. The grid model assumes that data points are distributed in a regular rectangular array, with a fixed vertical spacing between adjacent rows and a fixed horizontal spacing between adjacent columns. The calculation of model parameters needs to consider the geometric relationship of the point set, and the best grid parameter combination is determined through the least squares method or other fitting algorithms. For any point in the point cloud, first calculate the theoretical grid node position corresponding to the point according to the grid model parameters, and then calculate the Euclidean distance between the actual point coordinates and the theoretical position. The distance calculation formula is wherein represents the actual point coordinates, represents the corresponding theoretical grid node coordinates. When the calculated distance is less than the preset geometric tolerance threshold, the point is determined as an inlier point supporting the current grid model; otherwise, it is determined as an outlier point.

[0101] The iterative process of random sampling, model assumption, and inlier point statistics is repeatedly performed, and the random sampling of the minimum point set is independently performed in each iteration to avoid the influence of the previous iteration result on the current sampling. The determination of the number of iterations is usually based on probability theory, considering the probability requirement that at least one iteration can sample a minimum point set of all inliers under a given proportion of outliers. More iterations can improve the probability of finding the optimal model, but will increase the computational overhead; fewer iterations may miss the best solution. In each iteration, the entire process from random sampling to inlier point statistics needs to be completely performed, and the grid model parameters and the corresponding number of inlier points obtained in the current iteration are recorded. After all iterations are completed, the number of inlier points of the grid model obtained in each iteration is compared, and the model with the most inliers is selected as the global optimal solution, because more inlier support means that the model can better explain the distribution rule of the actual data. In the case of the same number of inliers, auxiliary indicators such as geometric fitting error and model complexity can be further considered for optimization.

[0102] Next, the proportion of the number of inlier points in the optimal model to the total number of points needs to be calculated and compared with the preset threshold. When the proportion of inlier points exceeds the preset threshold, it indicates that most of the table drawing characters conform to the regular grid distribution pattern, and the candidate data stream is confirmed as a valid fixed value table; when the proportion is lower than the threshold, it indicates that the data lacks obvious table structure characteristics and is not identified as a fixed value table. The setting of the threshold needs to balance between recognition accuracy and fault tolerance, and a too high threshold may lead to missed detection of real tables, and a too low threshold may misidentify non-table data. Through this quantitative evaluation method based on geometric consistency, it can objectively and accurately judge whether the data has a complete table structure.

[0103] In one embodiment, the device-side fixed value item and the standard fixed value item in the preset standard fixed value sheet are respectively fused to obtain a mixed feature vector containing text semantics, numerical attributes, and context information, including the following steps:

[0104] For the device-side fixed value item or the standard fixed value item in the preset standard fixed value sheet, the fixed value item information of the fixed value item and the relative position information of the fixed value item in the corresponding customized table or customized sheet are extracted, and the fixed value item information includes fixed value item description text, fixed value item value, and fixed value item unit;

[0105] The text semantic vector is generated based on the fixed-value item description text, the numerical attribute vector is generated by combining the fixed-value item value and fixed-value item unit and through coding, and the context feature vector is generated based on relative position information and through coding;

[0106] The text semantic vector, the numerical attribute vector and the context feature vector are weighted and fused to obtain a mixed feature vector.

[0107] In the embodiment, the fixed-value item information extraction includes three core components: the fixed-value item description text is usually located in the left column or the title row of the table, contains the name, function description and technical parameter description of the fixed value; the fixed-value item value is a specific quantitative value of the fixed value, which can include an integer, a decimal or a value represented by scientific notation; the fixed-value item unit represents the dimension of the value, such as the voltage unit volt, the current unit ampere, the time unit second, etc. The relative position information extraction needs to record the row number and column number of the fixed-value item in the table, and the spatial relationship with adjacent fixed-value items. The position information is represented by a standardized coordinate system, with the upper left corner of the table as the origin. Through regular expression and pattern matching technology, the three components of text, value and unit can be accurately identified and separated.

[0108] Next, the description text is preprocessed, including removing punctuation, unifying case, word segmentation processing and the like, to convert continuous text into a word sequence. A pre-trained word vector model is used to map each word to a high-dimensional numerical vector, and the word vector can capture the semantic information and context relationship of the word. For a description text containing multiple words, a weighted average method is used to calculate the overall text semantic vector, and the weight can be determined based on word frequency, importance or position information. The generation of the numerical attribute vector needs to jointly encode the value and unit information of the fixed-value item, the value part is converted into a fixed range of values through standardization processing, and the unit part is converted into a numerical representation through one-hot encoding or embedding vector. The context feature vector is generated based on the relative position information of the fixed-value item, including the row number, column number and distance relationship with the key reference point, and the spatial position information is converted into a vector form through position coding technology.

[0109] The mixed feature vector is constructed in a linear combination manner, and the fusion formula is wherein represents the mixed feature vector, represents the text semantic vector, represents the numerical attribute vector, represents the context feature vector, , , are corresponding weight coefficients respectively and satisfy The setting of the weight coefficient needs to adjust the importance of the fixed value item matching according to different feature types, and the text semantics usually carries the most important identification information and is given a higher weight; the numerical attribute reflects the quantification characteristics of the fixed value, and has a medium importance; the position information provides a structured context, and the weight is relatively low. Through weighted fusion, the mixed feature vector can comprehensively express the multi-dimensional information such as semantic content, numerical characteristics and structural position of the fixed value item.

[0110] In one of the embodiments, generating the text semantic vector based on the fixed value item description text includes the following steps:

[0111] Carrying out word segmentation processing on the fixed value item description text to obtain a word sequence;

[0112] Loading a pre-trained word vector model, and mapping each word in the word sequence into a high-dimensional word vector by using the word vector model;

[0113] Carrying out weighted average calculation on all high-dimensional word vectors corresponding to the fixed value item description text, and taking the calculated weighted average vector as the text semantic vector of the fixed value item.

[0114] In the embodiment, the word segmentation process first carries out text preprocessing, including removing redundant white spaces, punctuation marks and special symbols, unifying the case format of the text, and ensuring the standardization of the input text. For Chinese text, a word segmentation algorithm based on the combination of a dictionary and a statistical model is adopted, such as the jieba word segmenter or other mature Chinese word segmentation tools, which can accurately identify word boundaries and handle ambiguous segmentation problems. For English text, word segmentation is relatively simple, mainly based on spaces and punctuation marks for word segmentation. Special attention needs to be paid to professional terms in the field of relay protection during word segmentation, and a special field dictionary needs to be established to ensure that technical terms can be correctly identified as complete word units without being incorrectly segmented. The word segmentation result generates an ordered word sequence, maintaining the relative position relationship of the words in the original text. The word vector model is usually trained on a large amount of text data using algorithms such as Word2Vec, GloVe or FastText, which can map words into fixed-dimensional dense vectors, and each dimension in the vector contains specific semantic information. The model loading process includes reading the pre-trained model file, initializing the vocabulary and vector matrix, etc. For each word obtained by word segmentation, the corresponding high-dimensional word vector is obtained by querying the vocabulary of the word vector model. The dimension of the word vector is usually set to 100 to 300, which can effectively capture the semantic features and context relationships of the words.

[0115] Next, weighted average calculation needs to be carried out on all high-dimensional word vectors to generate a vector representation that can represent the entire text semantics. The weighted average formula is:

[0116]

[0117] wherein denotes the weighted average vector, denotes the word vector of the i-th word, denotes the weight of the i-th word, and L denotes the total number of words. The determination of the weight can be based on word frequency statistics, TF-IDF value, or the position of the word in the sentence, etc. For professional terms in the field of relay protection, a higher weight can be given to highlight their importance. During the calculation, it is necessary to ensure that all word vectors have the same dimension, and vector alignment processing is needed for inconsistent dimensions.

[0118] In one embodiment, the semantic similarity between the hybrid feature vector of the device-side setting value item and the standard setting value item is calculated, and the cross-source matching of the device-side setting value item and the standard setting value item is completed according to the semantic similarity, including the following steps:

[0119] For any device-side setting value item, the cosine similarity between the first hybrid feature vector of the device-side setting value item and the second hybrid feature vector of all standard setting value items is calculated respectively;

[0120] The highest cosine similarity among all cosine similarities is selected, and if the highest cosine similarity exceeds a preset confidence threshold, it is determined that the device-side setting value item and the standard setting value item corresponding to the highest cosine similarity are successfully matched cross-source to form a set of setting value item pairs.

[0121] The above steps are repeated until the cosine similarity between the first hybrid feature vector of all device-side setting values and the second hybrid feature vector is calculated.

[0122] In this embodiment, the cosine similarity algorithm measures the similarity of two vectors by calculating the cosine of the included angle between them. During the calculation, the dot product of the two vectors is first calculated, that is, the sum of the products of corresponding dimension elements; then the modulus of the two vectors is calculated, that is, the square root of the sum of the squares of each dimension element; finally, the dot product is divided by the product of the two modulus to obtain the cosine similarity value. The cosine similarity value ranges from -1 to 1, the closer to 1, the more similar the two vectors, the closer to -1, the less similar, and the value of 0 means that the two vectors are orthogonal. For any device-side setting value item, similarity calculation needs to be performed with each standard setting value item in the standard setting value item, forming a similarity array. Then compare all similarity values and record the maximum value and its corresponding standard setting value item index. The confidence threshold needs to balance between matching accuracy and coverage, and the appropriate threshold range is usually determined according to historical data statistics and cross-validation results. When the highest cosine similarity exceeds the preset threshold, it is confirmed that the device-side setting value item and the corresponding standard setting value item are successfully matched, and the association relationship of the setting value item pair is established.

[0123] The iterative process of similarity calculation and matching judgment is repeated, and the iterative process is processed in a loop, and each item in the device-end fixed value item list is processed in turn. For each device-end fixed value item, the complete process of similarity calculation, highest similarity selection and confidence judgment with all standard fixed value items needs to be completed. During the iteration process, the matching state record needs to be maintained to track which device-end fixed value items have been successfully matched and which standard fixed value items have been occupied to avoid repeated matching and conflict allocation. After the iteration is completed, the complete matching result statistics can be obtained, including the number of successfully matched fixed value item pairs, the number of unmatched device-end fixed value items and the number of unmatched standard fixed value items.

[0124] In one embodiment, the device-end fixed value items and the standard fixed value items successfully matched across sources are subjected to unit normalization processing and numerical comparison, and a fixed value check report is generated according to the numerical comparison result, including the following steps:

[0125] All fixed value item pairs are traversed, and after unit normalization processing of each fixed value item pair, numerical comparison is performed, and the comparison result is classified as consistent items or inconsistent items;

[0126] All standard fixed value items that exist in the standard fixed value list and are not successfully matched across sources are identified and classified as missing items;

[0127] All device-end fixed value items that exist in the device-end fixed value item and are not successfully matched across sources are identified and classified as unexpected items;

[0128] The information of all consistent items, inconsistent items, missing items and unexpected items is summarized according to the preset format, and the final fixed value check report is generated.

[0129] In this embodiment, the unit normalization process establishes a standard unit conversion table, including the unit conversion relationship of various physical quantities such as voltage, current, time, power, etc., such as converting kilovolts to volts, milliamperes to amperes, milliseconds to seconds, etc. For each fixed value pair, first identify the unit type of the device end fixed value and the standard fixed value, and then apply the corresponding conversion coefficient to unify the two values to the same reference unit. The numerical comparison uses the relative error calculation method, and when the relative error is less than the preset tolerance threshold, it is determined as a consistent item; when the error exceeds the threshold, it is determined as an inconsistent item. The setting of the tolerance threshold needs to consider factors such as measurement accuracy, device error and engineering practice, and is usually set to a range of 1% to 5%. The missing item identification process traverses all fixed value items in the standard fixed value table to check if each standard fixed value item appears in the successfully matched fixed value item pair. The planned item identification traverses the device end fixed value item list to check if each device end fixed value item participates in the successful cross-source matching. For device end fixed value items that do not participate in matching, the reason needs to be analyzed, which may include non-standard fixed value items defined by the device manufacturer, newly added fixed value items after device upgrade, redundant fixed value items caused by configuration errors, etc. The existence of planned items may indicate that the device configuration exceeds the standard requirements, and the impact on system operation needs to be evaluated. For each planned item, the complete fixed value information needs to be recorded, including the fixed value name, actual value, unit and location in the device end fixed value table.

[0130] The final fixed value check report is generated according to the preset format, and the report format usually adopts a table form, which lists the detailed information of consistent items, inconsistent items, missing items and planned items. The report also needs to include statistical summary information, such as the total number of fixed value items, the number of items in each category, the overall consistency ratio and other key indicators. The report generation adopts a templated manner to ensure the consistency and readability of the format. The final check report provides a comprehensive and accurate fixed value configuration state evaluation for the operation and maintenance personnel, helping to quickly identify configuration problems and develop corresponding processing schemes to ensure the reliable operation of the relay protection device.

[0131] The application also discloses a relay protection device real-time adaptive fixed value checking system, which comprises 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 realize the relay protection device real-time adaptive fixed value checking method as described in any one of the above embodiments.

[0132] The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), programmable logic devices (PLD), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., and the general-purpose processor can be a microprocessor or any conventional processor, etc., and the present application is not limited thereto.

[0133] The memory can be an internal storage unit of the computer device, for example, a hard disk or a memory of the computer device, or an external storage device of the computer device, for example, a plug-in hard disk, a smart memory card (SMC), a secure digital card (SD), or a flash memory card (FC) equipped on the computer device, and the memory can also be a combination of the internal storage unit and the external storage device of the computer device, and the memory is used to store a computer program and other programs and data required by the computer device, and the memory can also be used to temporarily store data that has been output or will be output, and the present application is not limited thereto.

[0134] The present application also discloses a computer readable storage medium, which stores instructions, and the instructions make the processor be configured to execute the real-time adaptive setting value checking method of the relay protection device described in any one of the embodiments when executed by the processor.

[0135] The computer program can be stored in a machine readable medium, and the computer program includes computer program code, which can be in the form of source code, object code, executable files, or some intermediate forms, etc., and the machine readable medium includes any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc., and it should be noted that the machine readable medium includes but is not limited to the above-mentioned components.

[0136] The computer readable storage medium stores the real-time adaptive setting value checking method of the relay protection device in the computer readable storage medium, and is loaded and executed on the processor to facilitate the storage and application of the method.

[0137] Those skilled in the art should understand that the above discussion of any embodiment is only intended to be illustrative and is not intended to be limiting to the scope of the present application; the above embodiments or technical features among different embodiments can also be combined, the steps can be implemented in any order, and there are many other changes such as the different aspects of one or more embodiments of the present application as described above, which are not provided in details for the sake of brevity.

[0138] One or more embodiments of the present application are intended to cover all such alternatives, modifications, and variations as fall within the broad scope of the present application. Accordingly, any one of the above-cited examples, or any other unrecited example, can be prepared by any method and materials equivalent to those described above, without departing from the spirit and scope of one or more embodiments of the present application.

Claims

1. A method for real-time adaptive setting verification of a relay protection device, characterized in that, Includes the following steps: Set the window size and sliding step of the sliding window, move the sliding window along the real-time printed data, and calculate the frequency distribution of data bytes within the window position of each sliding window; Based on the frequency distribution of data bytes and applying the Shannon entropy formula, the information entropy value is calculated for each window position, and an entropy value sequence curve corresponding to the information entropy value and the data stream position is generated. When the entropy value sequence curve crosses the preset first entropy value threshold from high to low, the current data stream position is marked as the starting point of the candidate data stream. When the entropy value sequence curve crosses the preset second entropy value threshold from low to high, the current data stream position is marked as the ending point of the candidate data stream. The second entropy value threshold is greater than the first entropy value threshold. Extract the frame structure feature points in the candidate data stream, and confirm the candidate data stream as a valid value table by constructing a global grid topology model through a geometric consistency test algorithm; The table boundaries of the effective setting table are defined based on the global grid topology model, and the effective setting table is parsed with local abnormal row fault tolerance capability based on the table boundaries to obtain the equipment-side setting items of the relay protection device. The device-side setpoints and the standard setpoints in the preset standard setpoint sheet are respectively fused to obtain a hybrid feature vector containing text semantics, numerical attributes and contextual information; Calculate the semantic similarity between the mixed feature vectors of the device-side fixed value and the standard fixed value, and perform cross-source matching between the device-side fixed value and the standard fixed value based on the semantic similarity; The device-side setpoints that are successfully matched across sources are normalized in units and compared with the standard setpoints. A setpoint verification report is then generated based on the comparison results.

2. The real-time adaptive setting verification method for relay protection devices according to claim 1, characterized in that, The step of extracting the frame structure feature points within the candidate data stream and confirming the candidate data stream as a valid value table using a global mesh topology model constructed through a geometric consistency check algorithm includes the following steps: A predefined character set containing various table drawing characters is used. Traverse all data within the candidate data stream, identify all target characters belonging to the character set, and extract the two-dimensional coordinates of the target characters within the candidate data stream; All two-dimensional coordinates are aggregated into a two-dimensional point cloud; The global grid topology model, constructed based on two-dimensional point clouds and using a geometric consistency check algorithm, confirms the candidate data stream as a valid value table.

3. The real-time adaptive setting verification method for relay protection devices according to claim 2, characterized in that, The process of confirming candidate data streams as valid value tables using a global grid topology model constructed based on two-dimensional point clouds and through a geometric consistency check algorithm includes the following steps: A minimum point set is randomly sampled from a 2D point cloud, and a global grid topology model is assumed based on the minimum point set. The minimum point set contains 4 non-collinear points. Traverse all points in the 2D point cloud according to the global grid topology model, and determine the points that match the geometric position of the global grid topology model as interior points; Repeat the iterative process of random sampling, model assumptions, and interior point statistics until the preset number of iterations is reached; In all iterations, the global grid topology model that receives the most interior point support is selected as the optimal model. If the proportion of interior points in the optimal model exceeds the preset threshold for the proportion of interior points, then the candidate data stream is confirmed as a valid setpoint table.

4. The real-time adaptive setting verification method for relay protection devices according to claim 1, characterized in that, The step of fusing features between the device-side setpoints and the standard setpoints in the preset standard setpoint sheet to obtain a hybrid feature vector containing text semantics, numerical attributes, and contextual information includes the following steps: For the standard setting items in the device-side setting items or the preset standard setting items, extract the setting item information and the relative position information of the setting item in the corresponding customization table or customization order. The setting item information includes the setting item description text, the setting item value and the setting item unit. Text semantic vectors are generated based on the description text of the fixed value terms. Numerical attribute vectors are generated by combining the numerical value and unit of the fixed value terms and encoding. Contextual feature vectors are generated based on relative position information and encoding. The text semantic vector, numerical attribute vector, and context feature vector are weighted and fused to obtain a hybrid feature vector.

5. The real-time adaptive setting verification method for relay protection devices according to claim 4, characterized in that, The process of generating a text semantic vector based on a fixed-value term description includes the following steps: The text describing the fixed-value items is segmented into words to obtain a word sequence; Load the pre-trained word vector model and use the word vector model to map each word in the word sequence into a high-dimensional word vector; A weighted average is calculated for all high-dimensional word vectors corresponding to the text describing the fixed-value item, and the calculated weighted average vector is used as the text semantic vector of the fixed-value item.

6. The real-time adaptive setting verification method for relay protection devices according to claim 1, characterized in that, The semantic similarity between the hybrid feature vectors of the device-side fixed value and the standard fixed value, and the cross-source matching between the device-side fixed value and the standard fixed value based on the semantic similarity, includes the following steps: For any device-side fixed value, calculate the cosine similarity between the first mixed feature vector of the device-side fixed value and the second mixed feature vector of all standard fixed value items; Select the highest cosine similarity among all cosine similarities. If the highest cosine similarity exceeds the preset confidence threshold, it is determined that the device-side fixed value item and the standard fixed value item corresponding to the highest cosine similarity have successfully matched across sources, forming a set of fixed value item pairs. Repeat the above steps until the cosine similarity between the first mixed feature vector and the second mixed feature vector has been calculated for all device-side values.

7. The real-time adaptive setting verification method for relay protection devices according to claim 6, characterized in that, The process of performing unit normalization and numerical comparison on the device-side setpoints that have successfully matched across sources and the standard setpoints, and generating a setpoint verification report based on the numerical comparison results, includes the following steps: Iterate through all pairs of fixed values, perform unit normalization on each pair of fixed values, compare the values, and classify the comparison results into consistent or inconsistent items. Identify all standard value items that exist in the standard value sheet but have not been successfully matched across sources, and classify them as missing items; Identify all device-side settings that exist in the device-side settings but have not been successfully matched across sources, and classify them as unplanned items; Summarize all consistent, inconsistent, missing, and unplanned items according to the preset format to generate the final value verification report.

8. A real-time adaptive setting verification system for a relay protection device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the real-time adaptive setting verification method for relay protection devices as described in any one of claims 1 to 7.

9. A computer-readable storage medium storing instructions thereon, characterized in that, When executed by a processor, this instruction causes the processor to be configured to perform the real-time adaptive setting verification method for a relay protection device according to any one of claims 1 to 7.

Images