Error correction method and system for paper equipment operation record recognition result

Abstract

The invention discloses an error correction method and system for a paper equipment operation record recognition result, and the method comprises the steps: firstly scanning a paper equipment operation record, carrying out the character recognition of the scanned record through an OCR technology, and evaluating the confidence; performing hierarchical error detection on an identification result: identifying characters with low confidence and characters which do not conform to a column mode, and identifying and marking abnormal characters in all columns by adopting an abnormal value detection method; and finally, selecting characters which are most similar to the wrong characters and conform to the column mode from the candidate character set for correction, and generating an accurate equipment operation record table. According to the method disclosed by the invention, errors in the identification result are positioned and corrected by combining the content and the organization mode of the equipment operation record data aiming at a paper equipment operation record document digital scene, so that the correction workload in the equipment operation record sheet digital process is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and in particular to an error correction method and system for paper-based equipment operation record recognition results. Background Art

[0002] In business operations, paper documents, especially equipment operation records, play a vital role. These records detail the key parameters of equipment operation, such as pressure, temperature, and speed, which are essential for equipment maintenance, fault diagnosis, and safety prevention. With the development of technology, optical character recognition (OCR) and form recognition technology have become effective means to digitize these paper documents. In particular, the application of deep learning technology has greatly improved the accuracy of recognition, laying the foundation for the digital transformation of paper documents.

[0003] Existing OCR technologies use deep learning models to recognize and convert text information in documents. With predefined character sets and patterns, paper documents are scanned and converted into digital formats using OCR technology for further analysis and processing. These models are able to handle a variety of character types including Chinese characters, numbers, uppercase and lowercase English letters, and special symbols. However, these technologies still face some challenges when processing handwritten text and text in specific formats.

[0004] Although existing OCR technology has made significant progress in recognition accuracy, there are still some problems and limitations in practical applications. First, the fixed recognition range of the OCR model limits its flexibility in specific application scenarios, which may lead to incorrect recognition of unexpected characters. Secondly, the OCR model is not accurate enough in recognizing handwritten text. Differences in writing habits among individuals and inconsistent document quality, such as unclear scans, all increase the difficulty of recognition. These problems not only limit the overall effectiveness of the OCR system, but also increase the workload of subsequent proofreading work, affecting the efficiency of the digitization process. Summary of the invention

[0005] The present invention proposes a method and system for correcting the results of paper-based equipment operation record recognition, which solves the problem that the existing optical character recognition method has insufficient accuracy when recognizing handwritten text and is not targeted enough in equipment recording scenarios.

[0006] In order to solve the above technical problems, the present invention provides a method for correcting errors in paper equipment operation record recognition results, comprising the following steps:

[0007] Step S1: Scan and identify the paper equipment operation record to obtain an equipment operation record table and the confidence level corresponding to each character in the equipment operation record table;

[0008] Step S2: Perform hierarchical error detection and marking on the recognized device operation record form:

[0009] Mark the characters with a confidence level less than the set confidence threshold in the device operation record form as error characters;

[0010] Mark the characters that do not conform to the preset column pattern in each column of the device operation record form as error characters;

[0011] Take each column of the device operation record form as a time series, calculate the features of each time series, regard the time series with features greater than the set feature threshold as an abnormal series, calculate the relative distance between characters in the abnormal series, and mark the characters with a relative distance greater than the set distance threshold as error characters;

[0012] Step S3: Select the character with the highest similarity to the error character and conforming to the column pattern of the column where the error character is located from the preset candidate character set as the corrected character, and obtain the corrected device operation record form.

[0013] Preferably, in step S2, the calculating the features of each time series and regarding the time series with features greater than the set feature threshold as an abnormal series includes the following steps:

[0014] Step S201: Convert a column of characters with length n into a time series:

[0015] X(t) = (x(t1), x(t2), x(t3),..., x(t n ));

[0016] In the formula, is the character at time t i moment;

[0017] Step S202: Use a sliding window with length w and w < n to equally divide the time series, with a sliding step of r, and continuously slide (n - w) / r times to obtain (n - w) / r + 1 sub-time series with length w;

[0018] Step S203: Calculate the confidence interval radius of the slope of each sub-time series as the feature of the sub-time series, and regard the sub-time series with a confidence interval radius greater than the set confidence interval threshold as an abnormal series. x(t i )

[0019] Preferably, the calculating the confidence interval radius of the slope of each sub-time series in step S203 includes the following steps:

[0020] Step S2031: Calculate the slope of each character in the sub-time series to obtain w - 1 slope values;

[0021] Step S2032: Calculate the mean and mean square error of the slope of each sub-time series;

[0022] Step S2033: Calculate the confidence interval radius of the slope of each sub-time series. The expression for calculating the confidence interval radius is:

[0023]

[0024] Where, d j is the radius of the confidence interval; θ j↑ is the upper confidence limit; θ j↓ is the lower confidence limit; s j is the mean of the slope; δ j is the mean square error of the slope; Z is a random variable that satisfies the normal distribution; α is the confidence level.

[0025] Preferably, in step S2, by calculating the distance between each character in the abnormal sequence and other characters, characters whose distance is greater than a set distance threshold are regarded as erroneous characters, including the following steps:

[0026] Step S211: calculating the reachable distance of each character in the abnormal sequence, and calculating the local density of each character according to the reachable distance;

[0027] Step S212: calculating the local outlier factor of each character according to the local density;

[0028] Step S213: taking 1 / 2 of the maximum value of the local outlier factor as a distance threshold, and taking the characters whose local outlier factor is greater than the distance threshold as erroneous characters.

[0029] Preferably, the expression for calculating the reachable distance in step S211 is:

[0030] RD k (p,q)=max{k-distance(q),dist(p,q)};

[0031]

[0032] Where, RD k (p,q) is the reachable distance between character p and character q; k-distance(q) is the distance between character q and the kth character closest to character q; dist(p,q) is the Euclidean distance between character p and character q; N is the dimension of the character.

[0033] Preferably, the expression for calculating the local density in step S212 is:

[0034]

[0035] Where, LRD k (p) is the local density of character p; N k-distance (p) is the set whose distance to character p is less than k-distance(q).

[0036] Preferably, the expression for calculating the local outlier factor in step S213 is:

[0037]

[0038] Where LOF(p) is the local outlier factor of character p.

[0039] Preferably, calculating the similarity between the erroneous character and all characters in the candidate character set of the column to which the erroneous character belongs in step S3 includes the following steps: converting the erroneous character and all characters in the candidate character set into binary images, calculating the intersection-and-union ratio between the binary image of the erroneous character and the binary images of all characters in the candidate character set, and using the intersection-and-union ratio as the similarity between the erroneous character and the characters in the candidate character set.

[0040] Preferably, in step S2, a corresponding regular expression is designed for each column in the equipment operation record table according to the pattern of characters in each column, as the column pattern of each column.

[0041] A paper equipment operation record recognition result error correction system, applicable to the above-mentioned paper equipment operation record recognition result error correction method, including a character recognition module, an error location module and a character correction module;

[0042] The character recognition module: uses OCR technology to convert the text on the paper equipment operation record table into characters, and outputs the equipment operation record table and the confidence level corresponding to each character in the equipment operation record table;

[0043] The error location module includes a confidence error location module, a column mode error location module and a time series outlier location module;

[0044] The confidence error positioning module is used to mark characters in the equipment operation record table whose confidence is lower than a set confidence threshold as error characters;

[0045] The column mode error locating module is used to check whether the recognition result of each column character of the equipment operation record table conforms to a preset column mode, and mark the characters that do not conform to the column mode of the corresponding column as error characters;

[0046] The time series outlier location module: regards each column of characters in the equipment operation record table as a time series, and uses an outlier detection method to identify and mark all abnormal data in the time series as erroneous characters;

[0047] The character correction module calculates the similarity between the erroneous character and the characters in the candidate character set, selects the character with the highest similarity as the correction candidate, ensures that the selected character conforms to the column mode of the column to which the erroneous character belongs, and outputs the corrected equipment operation record table.

[0048] The benefits of the present invention include at least:

[0049] 1. Through the preset column mode, it can more accurately recognize text with specific formats or regularities, such as date, time, temperature, etc. in equipment operation records, thereby solving the problem that existing OCR technology is not targeted enough when processing text with specific formats;

[0050] 2. Combining confidence evaluation and time series analysis, it is possible to more effectively identify and correct errors in handwritten text, so that the method can still maintain a high recognition accuracy when processing low-quality scanned documents, reducing the dependence on high-definition scanned documents and improving the recognition accuracy of handwritten text;

[0051] 3. Through the multi-level error detection and correction process, the error detection and correction process is made more systematic and automated, which significantly improves the practicality of the OCR system, reduces the proofreading workload caused by recognition errors, and thus improves the overall work efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 The figure is a schematic diagram of a method flow of an embodiment of the present invention. DETAILED DESCRIPTION

[0053] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the protection scope of the present invention.

[0054] Paper equipment operation records are mostly organized in the form of tables. One page of the document consists of one or more tables, each of which records different equipment operation data, or each table records the operation data of different parts of the same equipment. The table consists of a header and a table content part. The header part is mostly printed text, describing the content of each column record; the table content is handwritten numbers, English letters, or special characters, etc.

[0055] Since everyone has different writing habits, the writing methods of the same character or number vary greatly, which makes it difficult to identify paper equipment operation records when using optical character recognition (OCR) technology to scan and identify them. In addition, the low quality of paper documents and unclear scanning further increase the difficulty of identification.

[0056] Without solving the above problems, the embodiments of the present invention provide Figure 1 A method for correcting errors in paper equipment operation record recognition results is shown, comprising the following steps:

[0057] Step S1: Scan and identify the paper equipment operation record to obtain a digitized equipment operation record sheet.

[0058] Specifically, after scanning the paper equipment operation record, OCR recognition and table structure recognition technology are used to save the paper equipment operation record table in Excel file format.

[0059] Step S2: performing layered error detection and marking on the identified equipment operation record table.

[0060] Specifically, the manifestations of OCR recognition errors are mainly in two categories: column pattern mismatch and column logic non-compliance. Column pattern refers to the same pattern that the data in the same column of the equipment operation record table conforms to. Column pattern mismatch refers to the situation where the recognition result does not meet the semantic specification due to the recognition error of the OCR algorithm. For example, if the column pattern to which a column of data belongs is equipment temperature, and the OCR algorithm recognizes numbers as English characters and decimal points as English letters, etc., the recognition result does not match the column pattern of the equipment temperature column. Column logic non-compliance means that the OCR recognition result meets the pattern of the column to which it belongs, but due to the OCR recognition error, the decimal point is lost in the recognition result, resulting in the recognition result being too large, one more number is recognized, resulting in the recognition result being too large, one less number is recognized, resulting in the recognition result being too small, etc., resulting in a situation that is obviously inconsistent with the business logic of the column.

[0061] According to the manifestation of recognition result errors, a hierarchical error detection and location scheme is proposed:

[0062] First layer: Error detection and location for column pattern mismatch.

[0063] The data of equipment operation records records the recording date, time, and parameter values ​​of equipment operation parameters. Equipment operation parameters can be physical parameters such as temperature and pressure of equipment components, and these parameters have obvious patterns. In view of this feature, corresponding patterns can be designed for the types of equipment operation parameters, including date patterns, time patterns, pressure patterns, temperature patterns, digital patterns, etc., and the column pattern set can be expanded according to actual application scenarios. Define the corresponding regular expression for each column of data to form a column pattern.

[0064] When performing error detection and location, first obtain a column of characters in the equipment operation record table, determine the mode of each cell in this column of characters, use a voting algorithm to get the column mode of the column, mark the characters that do not match the column mode as error characters, and complete the error location.

[0065] The second layer: confidence-based error detection and localization.

[0066] The last layer of the OCR recognition network is an N classification task, where N is the number of characters to be recognized. From the loss function of the classification network, the more accurate the prediction of a character, the higher the probability P of the output of this character in the softmax function of the classification task. The OCR recognition result will return the probability P, which is called confidence. The higher the confidence, the higher the accuracy of character recognition. Based on this principle, confidence is used as prior knowledge to detect and locate recognition errors. Set a suitable confidence threshold and mark characters with confidence lower than the confidence threshold as incorrect characters to attract the attention of proofreaders.

[0067] The third layer: outlier error location based on sliding window.

[0068] Even if the confidence of the character recognition result exceeds the set threshold and matches the column pattern, these characters may still be recognized incorrectly. For example, the temperature data "32.5" may be mistakenly recognized as "325". In the scenario of equipment operation records, each column of operation parameters is usually recorded in chronological order, so a column of data can be regarded as a time series. These time series usually follow specific change rules and show consistent characteristics.

[0069] Therefore, detecting and locating the errors can adopt the outlier detection method in time series. The present invention proposes a time series outlier detection algorithm based on sliding window, which is used to detect and locate the error data.

[0070] The algorithm is divided into two stages:

[0071] The first stage: Detection of abnormal sub-time series. In this stage, the entire time series is segmented into multiple shorter sub-time series by using the sliding window technique. Then, features are extracted from each sub-time series, and it is detected whether these features exceed the set feature threshold. If the features of a certain sub-time series exceed the threshold, then it will be marked as an abnormal sub-time series. This stage specifically includes the following steps:

[0072] S1: Convert a column of characters with length n into a time series:

[0073] X(t) = (x(t1), x(t2), x(t3),..., x(t n ));

[0074] In the formula, x(t i ) is the character at time t i , 1 ≤ i ≤ n, and the acquisition times are strictly increasing.

[0075] S2: Use a sliding window with length w, where w < n, to equally segment the time series with a sliding step of r, and slide continuously (n - w) / r times to obtain (n - w) / r + 1 sub-time series with length w.

[0076] S3: Calculate the confidence interval radius of the slope of each sub-time series as the feature of the sub-time series, including the following steps:

[0077] S31: Let a sub-time series of the time series X(t) be represented as X j (t), and calculate the slope between any two adjacent data points in this sub-time series according to the following formula:

[0078]

[0079] In the embodiment of the present invention, the sliding step is set to 1, that is, t i -t i-1 = 1.

[0080] Calculate the slope of each character in the sub-time series X j (t) to obtain w - 1 slope values;

[0081] S32: Calculate the mean s j of the slope of each sub-time series and the mean square deviation δ j :

[0082]

[0083] S33: Calculate the confidence interval radius of the slope of each sub-time series. The expression for calculating the confidence interval radius is:

[0084]

[0085]

[0086] Where, d j is the radius of the confidence interval; θ j↑ is the upper confidence limit; θ j↓ is the lower confidence limit; s j is the mean of the slope; δ j is the mean square error of the slope; Z is a random variable that satisfies the normal distribution; α is the confidence level.

[0087] S4: The sub-time series whose confidence interval radius is larger than the set confidence interval threshold is regarded as an abnormal series. The confidence interval threshold is generally set based on empirical values ​​or after statistics of normal data.

[0088] Phase 2: Outlier Detection. After detecting the abnormal sub-time series, the local outlier factor (LOF) algorithm is further used to identify outliers in these sub-time series. The LOF algorithm identifies outliers by evaluating the local density deviation of data points, thereby accurately locating those data points that do not conform to the expected pattern. This phase specifically includes the following steps:

[0089] S1: Calculate the reachable distance of each data point in the anomaly sequence:

[0090] RD k (p,q)=max{k-distance(q),dist(p,q)};

[0091]

[0092] Where, RD k (p,q) is the reachable distance between point p and point q; k-distance(q) is the distance between the kth data point closest to the data point q to be tested and the data point q to be tested; dist(p,q) is the Euclidean distance between point p and point q; N is the dimension of the data point.

[0093] S2: Calculate the local density of each data point based on the reachable distance:

[0094]

[0095] Where, LRD k (p) is the local density at point p; N k-distance (p) is the set of data points whose distance to point p is less than k-distance(q).

[0096] S3: Calculate the local outlier factor of each data point based on the local density. If the local outlier factor of a data point is greater than 1, it means that the density of the data point is smaller than the average density of the data points in its neighborhood, and the data point is more likely to be an outlier point. Otherwise, it means that the data point is a normal point. The expression for calculating the local outlier factor is:

[0097]

[0098] S4: Take 1 / 2 of the maximum value of the local outlier factor as the distance threshold, and regard the characters whose local outlier factor is greater than the distance threshold as error characters.

[0099] Step S3: For each erroneous character: select a character from a preset candidate character set that has the highest similarity to the erroneous character and conforms to the column pattern of the column to which it belongs as a correction character, and obtain a corrected equipment operation record table.

[0100] Specifically, the error types of the recognition result include column pattern mismatch, low confidence, and column logic failure. Among them, low confidence is likely to appear in the error type of column pattern mismatch or logic failure. The embodiment of the present invention uses the error type of column pattern mismatch to illustrate the error correction process.

[0101] Column pattern mismatches are mostly cases where numbers are recognized as characters, characters are recognized as numbers, punctuation marks are recognized as letters, or letters and numbers are recognized as punctuation marks. After locating the wrong characters according to the column pattern, the candidate for the correction character is first determined by the text dot matrix similarity, and then the candidates that do not conform to the column pattern are filtered out, and finally the correction character candidate with the greatest similarity and conforming to the column pattern is selected as the correction character. The following steps are included:

[0102] S1: Design candidate character sets based on device operation record scenarios.

[0103] First, a candidate character set is designed according to the actual scenario. The candidate character set is defined as characters that may appear in the OCR recognition results that need to be corrected. The present invention is oriented to the equipment operation record data, so the candidate character set includes numbers, English characters, punctuation marks, and physical parameter symbols such as temperature and pressure. For example, if the scenario is the recognition of equipment pressure, temperature and other operating parameters, the candidate character set includes numbers, pressure symbols, temperature symbols, etc.

[0104] S2: Calculate the text dot matrix similarity between the wrong character and the candidate character set.

[0105] Each character in the candidate character set is represented by a dot matrix of a fixed size, and the dot matrix of each character is a binary image. The foreground of the binary image is black, and the background is white. The values ​​of the pixels corresponding to the foreground of the binary image are set to 1 or 255, and the values ​​of the pixels corresponding to the background are set to 0, and the segmentation result of an image is obtained. The similarity measurement of two characters is to use the dot matrix image of one of the characters as the "gold standard" of image segmentation, and the dot matrix image of the other character as the segmentation result. The similarity of the two characters is measured by calculating the intersection and union ratio between the segmentation result and the "gold standard", that is, A similarity threshold is set, and characters with similarity greater than the set similarity threshold are used as a candidate set of correction characters.

[0106] S3: Filter the correction character candidate set by column pattern: delete the characters in the correction character candidate set that do not conform to the column pattern to which the incorrect character belongs.

[0107] S4: sorting the characters in the filtered correction character candidate set according to their similarity to the wrong character, and taking the character with the highest similarity as the correction character.

[0108] Embodiment 1

[0109] The method of the present invention is used to identify and correct time series, including the following steps:

[0110] S1: Set a candidate character set according to the time recognition scenario, the candidate character set includes numbers 0-9 and the ":" symbol, and generate dot matrix images of these 11 characters at the same time.

[0111] S2: Design a column pattern for the time column. The regular expression of the column pattern of the time column is: ^[0-2][0-9]:[0-5][0-9]:[0-5][0-9]$.

[0112] S3: Layered error detection and localization:

[0113] First layer: Set the text recognition confidence threshold to 0.90. The confidence returned by a cell "12:75:49" in the time column is 0.95, which meets the confidence requirement and there is no recognition confidence error.

[0114] The second layer: Get all the data in the column where "12:75:49" is located, determine the column model to which all the cell data belongs, and use a voting algorithm to determine that the column mode of the column is a time column mode.

[0115] The third layer: The interactive window outlier detection method based on time series locates the error. According to the empirical value, the slope confidence radius threshold is set to 0.03, and the "12:75:49" in the time column is located as an abnormal sequence, of which "7" is an abnormal point and needs to be corrected.

[0116] S4: Similarity correction based on text dot matrix:

[0117] Set the similarity threshold to 0.9, measure the similarity between the dot matrix image of "7" and the dot matrix images of all characters in the candidate character set, select the characters with a similarity greater than 0.9 as candidate correction characters, and select "1" as the candidate correction character with the highest similarity. Correct "12:75:49" to "12:15:49", and judge that it meets the time column pattern, and the correction is completed.

[0118] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. Only the preferred embodiments of the present invention are expressed. The description is more specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. As long as there is no contradiction in the combination of these technical features, they should be considered as the scope recorded in this specification.

[0119] It should be noted that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these modifications and improvements all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A method for correcting errors in paper equipment operation record recognition results, characterized in that: It includes the following steps: Step S1: Scan and identify the paper-based device operation records to obtain the device operation record form and the confidence level corresponding to each character in the device operation record form; Step S2: Perform hierarchical error detection and marking on the identified device operation record form: Mark the characters with a confidence level less than the set confidence level threshold in the device operation record form as error characters; Mark the characters that do not conform to the preset column pattern in each column of the device operation record form as error characters; Take each column of the device operation record form as a time series, calculate the features of each time series, take the time series with features greater than the set feature threshold as an abnormal series, calculate the relative distance between characters in the abnormal series, and mark the characters with a relative distance greater than the set distance threshold as error characters; Step S3: Select, from the preset candidate character set, the character with the highest similarity to the error character and conforming to the column pattern of the column where the error character is located as the corrected character to obtain the corrected device operation record form.

2. The method for correcting errors in paper equipment operation record recognition results according to claim 1, characterized in that: The step of calculating the features of each time series in Step S2 and taking the time series with features greater than the set feature threshold as an abnormal series includes the following steps: Step S201: Convert a column of characters with a length of n into a time series: X(t)=(x(t1),x(t2),x(t3),...,x(t n )); Where x(ti) is t i Characters of the moment; Step S202: Use a sliding window with a length of w and w < n to equally divide the time series, with a sliding step of r, and continuously slide (n - w) / r times to obtain (n - w) / r + 1 sub-time series with a length of w; Step S203: Calculate the confidence interval radius of the slope of each sub-time series as the feature of the sub-time series, and take the sub-time series with a confidence interval radius greater than the set confidence interval threshold as an abnormal series.

3. The method for correcting errors in paper equipment operation record recognition results according to claim 2, characterized in that: The step of calculating the confidence interval radius of the slope of each sub-time series in Step S203 includes the following steps: Step S2031: Calculate the slope of each character in the sub-time series to obtain w - 1 slope values; Step S2032: Calculate the mean and variance of the slope of each sub-time series; Step S2033: Calculate the confidence interval radius of the slope of each sub-time series, and the expression for calculating the confidence interval radius is: Where, d j is the radius of the confidence interval; θ j↑ is the upper confidence limit; θ j↓ is the lower confidence limit; s j is the mean of the slope; δ j is the mean square error of the slope; Z is a random variable that satisfies the normal distribution; α is the confidence level.

4. The method for correcting errors in paper equipment operation record recognition results according to claim 3, characterized in that: The step of taking the characters with a distance greater than the set distance threshold as error characters by calculating the distance between each character and other characters in the abnormal series in Step S2 includes the following steps: Step S211: Calculate the reachable distance of each character in the abnormal series, and calculate the local density of each character according to the reachable distance; Step S212: Calculate the local outlier factor of each character according to the local density; Step S213: Take 1 / 2 of the maximum value of the local outlier factor as the distance threshold, and take the characters with a local outlier factor greater than the distance threshold as error characters.

5. The method for correcting errors in paper equipment operation record recognition results according to claim 4, characterized in that: The expression for calculating the reachable distance in Step S211 is: RD k (p,q)=max{k-distance(q),dist(p,q)}; Where, RD k (p,q) is the reachable distance between character p and character q; k-distance(q) is the distance between character q and the kth character closest to character q; dist(p,q) is the Euclidean distance between character p and character q; N is the dimension of the character.

6. The method for correcting errors in paper-based equipment operation record recognition results according to claim 5, characterized in that: The expression for calculating the local density in Step S212 is: Where, LRD k (p) is the local density of character p; N k-distance (p) is the set whose distance to character p is less than k-distance(q).

7. The method for correcting errors in paper equipment operation record recognition results according to claim 6, characterized in that: The expression for calculating the local outlier factor in Step S213 is: Where LOF(p) is the local outlier factor of character p.

8. The method for correcting errors in paper equipment operation record recognition results according to claim 1, characterized in that: In step S3, the similarity between the erroneous character and all the characters in the candidate character set of the column to which the erroneous character belongs is calculated, including the following steps: converting the erroneous character and all the characters in the candidate character set into binary images, calculating the intersection-and-union ratio between the binary image of the erroneous character and the binary images of all the characters in the candidate character set, and using the intersection-and-union ratio as the similarity between the erroneous character and the characters in the candidate character set.

9. The method for correcting errors in paper equipment operation record recognition results according to claim 1, characterized in that: In step S2, a corresponding regular expression is designed for each column in the equipment operation record table according to the pattern of each column character, as the column pattern of each column.

10. A paper-based device operation record recognition result error correction system, applicable to a paper-based device operation record recognition result error correction method as claimed in any one of claims 1 to 9, characterized in that: It includes a character recognition module, an error location module and a character correction module; The character recognition module: uses OCR technology to convert the text on the paper equipment operation record table into characters, and outputs the equipment operation record table and the confidence level corresponding to each character in the equipment operation record table; The error location module includes a confidence error location module, a column mode error location module and a time series outlier location module; The confidence error positioning module is used to mark characters in the equipment operation record table whose confidence is lower than a set confidence threshold as error characters; The column mode error locating module is used to check whether the recognition result of each column character of the equipment operation record table conforms to a preset column mode, and mark the characters that do not conform to the column mode of the corresponding column as error characters; The time series outlier location module: regards each column of characters in the equipment operation record table as a time series, and uses an outlier detection method to identify and mark all abnormal data in the time series as error characters; The character correction module calculates the similarity between the erroneous character and the characters in the candidate character set, selects the character with the highest similarity as the correction candidate, ensures that the selected character conforms to the column mode of the column to which the erroneous character belongs, and outputs the corrected equipment operation record table.

Images