A method for recognizing handwritten information of a borehole formation associated with a spatial location

Abstract

This invention discloses a method for recognizing handwritten borehole formation records, comprising: S1, constructing a basic sample set of handwritten borehole formation records; S2, training a basic model for recognizing handwritten borehole formation records; S3, constructing a project-level sample set of handwritten borehole formation records; S4, training a project-level recognition model for handwritten borehole formation records; S5, constructing a spatially nested obfuscated text dictionary for formation records; S6, recognizing and automatically correcting handwritten record photos of borehole formation information; S7, manually verifying the recognition results; S8, adding the corrected recognized text and corresponding handwritten record photos from S7 to the project-level sample set from S3 and training the project-level recognition model, and adding the incorrectly recognized text and corrected text to the obfuscated text dictionary; S9, for the next recognition task, executing S6 and S7 to obtain the final recognition result. This method significantly improves the recognition accuracy and work efficiency of handwritten borehole formation records.

Description

Technical Field

[0001] This invention relates to the field of railway engineering geological exploration, specifically to a method for recognizing handwritten records of borehole strata information associated with spatial location. Background Technology

[0002] Drilling is a geological exploration method that uses mechanical rock-breaking to penetrate into the strata. It is one of the most widely used basic exploration methods in railway engineering geological exploration, and can directly obtain physical samples of rock, soil, and water from the strata. During railway drilling, borehole strata information needs to be recorded as required. Due to factors such as site conditions and personnel skill levels, borehole strata records are currently still mainly kept on paper, with information manually written down. When using borehole strata data, exploration staff need to repeatedly digitize these handwritten records, which is time-consuming and prone to errors. Therefore, quickly and accurately recognizing handwritten borehole strata records is crucial for improving the efficiency of railway engineering geological exploration.

[0003] In recent years, with the rapid development of artificial intelligence technology, intelligent text recognition technology has made significant progress. Intelligent recognition of printed text is relatively mature, achieving an accuracy rate of over 90%. However, due to the vast differences in handwritten styles, intelligent recognition of handwritten text still faces considerable challenges and is difficult to achieve mature application. Currently, large companies such as Tencent, Baidu, and Hanvon have developed intelligent handwritten text recognition products, but testing has shown that the recognition accuracy is still unsatisfactory. When used for recognizing handwritten records in borehole formations, the accuracy rate is only around 60%, making it unsuitable for direct application in engineering projects.

[0004] Borehole stratigraphic records from railway survey projects are characterized by patterned content and spatial correlation. Current handwritten text intelligent recognition methods mainly target general handwritten text recognition and have not yet considered the spatial correlation of handwritten information in borehole stratigraphic records, thus leaving considerable room for improvement. Summary of the Invention

[0005] To address the issues of the extensive manual and repetitive digitization required for handwritten borehole formation records, and the fact that conventional handwriting recognition methods do not consider the spatial characteristics of these records and suffer from low accuracy, this invention provides a method that takes into account the spatial location of handwritten borehole formation records and achieves high-accuracy recognition of these records through basic model training, project-level model training, and automatic correction.

[0006] Therefore, the technical solution of the present invention is as follows:

[0007] A method for recognizing handwritten borehole formation records includes the following steps:

[0008] S1. Construct a basic sample set of handwritten borehole formation records: Obtain handwritten record photos of borehole coordinates, recorders, and borehole formation information of completed projects, along with corresponding text data. Annotate the handwritten record photos and establish a relationship with the text data to form sample set X1. Obtain conventional formation description text data, generate handwritten style photos, and combine them with the conventional formation description text data to form sample set X2. Merge sample set X1 and sample set X2 to form the basic sample set X of handwritten borehole formation records.

[0009] S2, Training the basic model for information recognition of handwritten records of borehole formations: Divide the basic sample set X into a training set and a test set according to the proportion, construct the basic model for information recognition based on CRNN, and carry out model training and optimization;

[0010] S3, Construct a project-level sample set of borehole formation handwritten records: Obtain the borehole coordinates, mileage, recorder, borehole formation handwritten record photos, and corresponding text data for the ongoing project A. Annotate the borehole formation handwritten record photos and establish correlations with the text data to form a project-level sample set X. p ;

[0011] S4, Train the project-level recognition model for handwritten borehole formation records: For the project-level sample set X... p Perform random sampling to form the sub-task sample set X s Based on the basic model for information recognition in step S2, use the sub-task sample set to train the model and update the basic model parameters; repeat the sub-task sample set construction and model training until the model recognition test accuracy reaches the target requirement or the model training has converged.

[0012] S5, Construct a spatially nested obfuscated text dictionary for the stratigraphic record: Based on the text data in steps S1 and S3, construct a spatially nested obfuscated text dictionary, which includes obfuscated text, correct text, and coordinate positions;

[0013] S6, Recognition and Automatic Correction of Handwritten Record Photos of Borehole Formation Information: Preprocess the handwritten record photos of borehole formation information to be identified in Project A which is currently under construction; use the project-level recognition model trained in step S4 to recognize the handwritten record photos of borehole formation information to be identified; combine the spatial location of the borehole and use the obfuscated text dictionary constructed in step S5 to automatically correct the recognition results.

[0014] S7. Manually verify the borehole formation identification results. If there are no incorrect identification texts, the current identification result is taken as the final identification result, and S9 is executed. If there are incorrect identification texts, the incorrect identification texts are manually corrected to obtain the final identification result, and S8 is executed.

[0015] S8. Add the corrected recognition text from S7 and the corresponding handwritten record photos to the project-level sample set formed in S3. Repeat step S4 to train the borehole strata handwritten record project-level recognition model, and add the incorrectly recognized text and the corrected text to the confused text dictionary constructed in S5.

[0016] S9: For the next task to be identified in project A, execute S6 and S7 to obtain the final identification result.

[0017] In step S1 above, the steps for generating a handwritten-style photo are as follows:

[0018] (1) Randomly extract typical handwritten record photos of each recorder from the sample set X1, and integrate the handwritten record photos of other business personnel in the project team to form a target handwritten style photo dataset;

[0019] (2) Based on Generative Adversarial Network (GAN), a handwritten style photo generation model is constructed, and the model is trained and optimized using the target handwritten style photo dataset;

[0020] (3) Combine the conventional stratigraphic description text data obtained in S12 and the handwritten record photos obtained in S11, and use the handwritten style photo generation model trained and optimized in step (2) to generate handwritten style photos.

[0021] Preferably, in step (1) above, handwritten record photos from the open dataset are also integrated to form a target handwritten style photo dataset.

[0022] Preferably, the sample sets X1 and X2 in step S1 are merged in a ratio of ≥3:2.

[0023] The conventional stratigraphic description text data mentioned in step S1 includes: rock and soil name, color, structure and texture, density, moisture content, plasticity, weathering degree, joints and fissures, particle composition, particle size and inclusion characteristics.

[0024] In one embodiment of the present invention, the CRNN handwritten record recognition model described in step S2 includes the following structure:

[0025] The convolutional layers employ an improved MobileNet architecture;

[0026] The recurrent layer uses a general bidirectional long short-term memory network Bi-LSTM to enhance the contextual relationships between characters for sequence prediction;

[0027] The transcription layer was classified using a common CTC connection timing tool.

[0028] In step S4 above, the following random sampling method is used to construct the subtask sample set:

[0029] S411, the project-level sample set is divided into two groups according to drilling mileage and recorder. The first grouping attribute is the mileage range and the second grouping attribute is the recorder.

[0030] S412, from the project-level sample set X p The handwritten records of the borehole formations by each recorder were randomly selected from each mileage interval to form a sub-task sample set.

[0031] In step S4, the model training includes the following steps:

[0032] (1) Initialize the model parameters θ and set the number of iterations K;

[0033] (2) The model was trained using the stochastic gradient descent method on the subtask dataset, and the parameters after training were obtained as follows:

[0034] (3) Update model parameters ∈ is the step size coefficient, which can be between 0.01 and 0.2 depending on the convergence. Repeat step (2) until the number of iterations is reached.

[0035] In step S5, the spatially nested confusion dictionary uses the structure [T F ,(T R [,(X,Y))], where T F To obfuscate the text; T R This is the correct text; (X,Y) are the borehole space coordinates, T F By T R It generates text based on combinations of similar-looking characters and easily confused characters; when the text lacks corresponding spatial coordinate information, the coordinates are set to (0,0).

[0036] For position coordinates (X) b ,Y b The automatic correction in step S6 of the handwritten borehole formation record includes the following steps:

[0037] (1) Perform Chinese word segmentation on the recognition results to obtain a series of word groups W1, W2, ... W i …、W n ;

[0038] (2) For each phrase W i Search for obfuscated text dictionary records and retrieve all dictionary matching results (T) R,1 ,(X1,Y1)),(T R,2 ,(X2,Y2)), …(T R,j ,(X j ,Y j ))…、(T R,m,(X m ,Y m ));

[0039] (3) Based on the obtained matching results, calculate the distance between the spatial location of the borehole and the spatial locations of each dictionary matching record. Text correction is performed based on the shortest matching distance; if no matching result is found, no text correction is performed.

[0040] Compared with the prior art, the present invention has the following advantages and positive effects:

[0041] 1. This method considers the spatial location information of the samples when constructing the sample set, which enhances the text feature dimension of the handwritten records of borehole formations and is beneficial to feature mining of handwritten record information.

[0042] 2. This method employs a two-stage training approach, consisting of a basic model and a project-level model, which better adapts to the handwriting style characteristics of on-site recorders in exploration projects and improves the recognition accuracy of handwritten records from borehole strata.

[0043] 3. This method constructs a confused text dictionary that includes spatial location, which can automatically correct the recognition results by combining the spatial location of the borehole, further improving the recognition accuracy of handwritten records of borehole formations.

[0044] 4. Because this method achieves high accuracy in identifying handwritten borehole formation records, it significantly reduces the workload of manual digitization and improves work efficiency. Attached Figure Description

[0045] Figure 1 This is a flowchart of the identification method of the present invention. Detailed Implementation

[0046] The method for recognizing handwritten borehole formation records of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0047] See Figure 1 The borehole formation handwritten record information recognition method of the present invention includes: S1, constructing a basic sample set of borehole formation handwritten records; S2, training a basic model for borehole formation handwritten record recognition; S3, constructing a project-level sample set of borehole formation handwritten records; S4, training a project-level recognition model for borehole formation handwritten records; S5, constructing a dictionary of obfuscated text in formation records; S6, borehole formation handwritten record recognition and automatic correction; S7, manual verification of borehole formation record recognition results. Specifically, as follows:

[0048] S1, the basic sample set for constructing handwritten borehole formation records, including:

[0049] S11, obtain handwritten record photos of borehole coordinates, recorders, borehole formation information and corresponding text data of the completed project, label the handwritten record photos, establish a correlation with the text data, and form a sample set X1;

[0050] S12, Obtain conventional stratigraphic description text data, including rock and soil name, color, structure and texture, density, moisture content, plasticity, weathering degree, joints and fissures, particle composition, particle size, inclusion characteristics, etc., and generate handwritten-style photographs according to the following steps, and combine them with the conventional stratigraphic description text data to form sample set X2:

[0051] (1) Randomly extract typical handwritten record photos of each recorder from the sample set X1, and integrate the handwritten record photos of other business personnel in the project team to form a target handwritten style photo dataset;

[0052] Preferably, this step also integrates handwritten record photos from open datasets to form a target handwritten style photo dataset, so as to improve the generalization ability of the recognition method of the present invention;

[0053] (2) Based on Generative Adversarial Network (GAN), a handwritten style photo generation model is constructed, and the model is trained and optimized using the target handwritten style photo dataset;

[0054] (3) Combine the conventional stratigraphic description text data obtained in S12 and the handwritten record photos obtained in S11, and use the handwritten style photo generation model trained and optimized in step (2) to generate handwritten style photos.

[0055] S13, merge the sample set X1 and sample set X2 in a certain ratio (not less than 3:2) to form the basic sample set X of the borehole formation handwritten record.

[0056] S2, Training the basic model for information recognition of handwritten records of borehole formations: Divide the basic sample set X into a training set and a test set according to the proportion; construct the basic model for information recognition based on CRNN and carry out model training and optimization.

[0057] In one embodiment of the present invention, the model convolutional layers employ an improved MobileNet structure. As shown in Table 1, the recurrent layers use a general bidirectional long short-term memory network (Bi-LSTM) to enhance the contextual relationships between characters for sequence prediction; the transcription layers employ a general CTC connection-based temporal classification tool.

[0058] Table 1 Improved MobileNet Structure

[0059] 3ⅹ3 16 16 no Relu 1 3ⅹ3 64 24 no Relu (2,1) 3ⅹ3 72 24 no Relu 1 5ⅹ5 72 40 yes Relu (2,1) 5ⅹ5 120 40 yes Relu 1 5ⅹ5 120 40 yes Relu 1 3ⅹ3 240 80 no Hardswish 1 3ⅹ3 200 80 no Hardswish 1 3ⅹ3 184 80 no Hardswish 1 3ⅹ3 184 80 no Hardswish 1 3ⅹ3 480 112 yes Hardswish 1 3ⅹ3 672 112 yes Hardswish 1 5ⅹ5 672 160 yes Hardswish (2,1) 5ⅹ5 960 160 yes Hardswish 1 5ⅹ5 960 160 yes Hardswish 1

[0060] S3, Construct a project-level sample set of borehole formation handwritten records: Obtain the borehole coordinates, mileage, recorder, borehole formation handwritten record photos, and corresponding text data for the ongoing project A. Annotate the borehole formation handwritten record photos and establish correlations with the text data to form a project-level sample set X. p .

[0061] S4, training a project-level recognition model for handwritten borehole formation records. Includes:

[0062] S41, for the project-level sample set X p Perform random sampling to construct the subtask sample set X s The construction steps are as follows:

[0063] S411, the project-level sample set is divided into two groups according to drilling mileage and recorder. The first grouping attribute is the mileage range and the second grouping attribute is the recorder.

[0064] S412, from the project-level sample set X p Randomly select handwritten record photos of borehole formations from each recorder within each mileage interval to form a sub-task sample set;

[0065] S42, Based on the information recognition model from step S2, model training is performed using the sub-task sample set to update the basic model parameters. Model training includes the following steps:

[0066] (1) Initialize the model parameters θ and set the number of iterations K;

[0067] (2) Using the aforementioned subtask dataset, the basic model for borehole formation handwritten record recognition is trained using the stochastic gradient descent method, and the model parameters are obtained after training.

[0068] (3) Update model parameters Repeat step (2) until the number of iterations is reached. Wherein, ∈ is the step size coefficient, which takes a value ranging from 0.01 to 0.2 depending on the convergence.

[0069] S43, repeat S412-S42 to construct the sub-task sample set and train the model until the model recognition test accuracy reaches the target requirement or the model training has converged.

[0070] S5, Construct a spatially nested obfuscated text dictionary for the stratigraphic record: Based on the text data from steps S1 and S3, construct a spatially nested obfuscated text dictionary. The dictionary includes obfuscated text, correct text, and coordinate positions. The obfuscated dictionary adopts the structure [T F ,(T R [,(X,Y))], where T F To obfuscate the text; TR This is the correct text; (X,Y) are the borehole space coordinates, T F By T R Generates text based on combinations of similar-looking and easily confused characters. When the text lacks corresponding spatial coordinate information, the coordinates are set to (0,0).

[0071] S6, Handwritten Record Photo Recognition and Automatic Correction of Borehole Formation Information: Preprocess the handwritten record photos of the borehole formation information to be identified in Project A (including conventional methods such as scaling, normalization, and histogram equalization); use the project-level recognition model trained in step S4 to recognize the handwritten record photos of the borehole formation information; and automatically correct the recognition results using the obfuscated text dictionary constructed in step S5, combined with the borehole spatial location. For location coordinates (X... b ,Y b The automatic correction includes the following steps: (The text appears to be incomplete and contains several errors. A more accurate translation would require the full context.)

[0072] 1) Perform Chinese word segmentation on the recognition results to obtain a series of word groups W1, W2, ... W i …、W n ;

[0073] 2) For each phrase W i Search for obfuscated text dictionary records and retrieve all dictionary matching results (T) R,1 ,(X1,Y1)),(T R,2 ,(X2,Y2)), …(T R,j ,(X j ,Y j ))…、(T R,m ,(X m ,Y m ));

[0074] 3) Based on the obtained matching results, calculate the distance between the spatial location of the borehole and the spatial locations of each dictionary matching record. Text correction is performed based on the shortest matching distance; if no matching result is found, no text correction is performed.

[0075] S7: Manually verify the borehole formation identification results. If there are no incorrectly identified texts, the current identification result is taken as the final identification result, and S9 is executed; if there are incorrectly identified texts, S8 is executed.

[0076] S8, manually correct the incorrectly identified text to obtain the final identification result, add the corrected identification text from S7 and the corresponding handwritten record photo to the project-level sample set formed in S3, repeat step S4, train the borehole strata handwritten record project-level identification model, and add the incorrectly identified text and the corrected text to the confused text dictionary constructed in S5.

[0077] S9: For the next task to be identified in project A, execute S6 and S7 to obtain the final identification result.

[0078] Example

[0079] A method for recognizing handwritten borehole formation records includes the following steps:

[0080] S1. Obtain handwritten records of borehole strata from completed railway geological survey projects. Obtain photos of handwritten records containing borehole coordinates, recorders, and borehole strata information, as well as corresponding text data (hereinafter referred to as "handwritten record text"), to form a sample set X1. Some data examples are shown in Table 2.

[0081] Table 2. Examples of Stratigraphic Handwritten Record Sample Set X1

[0082]

[0083]

[0084] Obtain conventional stratigraphic description text data and generate a handwritten style sample set X2. Some data examples are shown in Table 3.

[0085] Table 3 shows examples of the generated handwriting style sample set X2.

[0086]

[0087] The sample sets X1 and X2 were merged in a 3:2 ratio to form the basic sample set X of the borehole formation handwritten records.

[0088] S2, the sample set X is divided into a training set and a test set in an 8:2 ratio to carry out basic model training and optimization for information recognition.

[0089] S3, for a survey project that is currently under construction, construct a project-level sample set X. p Some data examples are shown in Table 4.

[0090] Table 4 Sample Set of Project-Level Stratigraphic Handwritten Records X p Example

[0091]

[0092]

[0093] S4, grouped by mileage intervals of 44500-45000, 45000-45500, and 45500-46000, from the project-level sample set X. pThe labeled data of each recorder is randomly sampled to form a sub-task sample set. Based on the basic model for identifying handwritten records of borehole formations, the model is trained until the accuracy of the model recognition test reaches the target requirement.

[0094] S5. Based on the combination and replacement of similar-looking and easily confused characters, a spatially nested confusion text dictionary of the stratigraphic record is constructed. Some data are shown in Table 5.

[0095] Table 5 Examples of Spatial Nested Obfuscated Text Dictionaries

[0096] 1 Dry phyllite Phyllite 601079,4499280 2 Fen Tu silt 611417,4499911 3 Dividing soil silt 611417,4499911 4 fine sand fine sand 601280,4499773 5 silica conglomerate 601280,4499773 6 silica sandstone 610983,4499903 7 Grains Particle size 610983,4499903 8 Siphon gland fissure 411401,4467963 9 leaning columnar 411401,4467963 10 Weakness Weak weathering 601079,4499280

[0097] S6. For the handwritten records of the stratigraphic information of the borehole to be identified in the current project (number: 20-ZD-4-232, spatial location: 608047, 4499851), the project-level recognition model trained in S4 is used for recognition, and the confused text dictionary constructed in S5 is used for correction. The recognition results are shown in Table 6.

[0098] Table 6. Examples of Comparison of Borehole Formation Handwritten Record Recognition Results

[0099]

[0100]

[0101] As can be seen from the examples, by implementing the borehole strata handwritten record information recognition method of the present invention with associated spatial location, constructing basic models and project-level models, and automatically correcting the recognition results in combination with the borehole spatial location, it can better adapt to the handwriting style characteristics of on-site recorders in exploration projects, improve the recognition accuracy of borehole strata handwritten records, and significantly reduce the workload of manual digital data entry.

Claims

1. A method for recognizing handwritten borehole formation records associated with spatial location, characterized in that, Includes the following steps: S1. Construct a basic sample set of handwritten borehole formation records: Obtain handwritten record photos of borehole coordinates, recorders, and borehole formation information of completed projects, along with corresponding text data. Annotate the handwritten record photos and establish a relationship with the text data to form sample set X1. Obtain conventional formation description text data, generate handwritten style photos, and combine them with the conventional formation description text data to form sample set X2. Merge sample set X1 and sample set X2 to form the basic sample set X of handwritten borehole formation records. S2, Training the basic model for information recognition of handwritten records of borehole formations: Divide the basic sample set X into a training set and a test set according to the proportion, construct the basic model for information recognition based on CRNN, and carry out model training and optimization; S3, Construct a project-level sample set of borehole formation handwritten records: Obtain the borehole coordinates, mileage, recorder, borehole formation handwritten record photos, and corresponding text data for the ongoing project A. Annotate the borehole formation handwritten record photos and establish correlations with the text data to form a project-level sample set X. p ; S4, Train the project-level recognition model for handwritten borehole formation records: For the project-level sample set X... p Perform random sampling to form the sub-task sample set X s Based on the information recognition model in step S2, model training is carried out using the sub-task sample set to update the basic model parameters. Repeat the subtask sample set construction and model training until the model recognition test accuracy reaches the target requirement or the model training has converged. S5, Construct a spatially nested obfuscated text dictionary for the stratigraphic record: Based on the text data in steps S1 and S3, construct a spatially nested obfuscated text dictionary, which includes obfuscated text, correct text, and coordinate positions; S6, Recognition and Automatic Correction of Handwritten Record Photos of Borehole Formation Information: Preprocess the handwritten record photos of borehole formation information to be identified in Project A which is currently under construction; use the project-level recognition model trained in step S4 to recognize the handwritten record photos of borehole formation information to be identified; combine the spatial location of the borehole and use the obfuscated text dictionary constructed in step S5 to automatically correct the recognition results. S7. Manually verify the borehole formation identification results. If there are no incorrect identification texts, the current identification result is taken as the final identification result, and S9 is executed. If there are incorrect identification texts, the incorrect identification texts are manually corrected to obtain the final identification result, and S8 is executed. S8. Add the corrected recognition text from S7 and the corresponding handwritten record photos to the project-level sample set formed in S3. Repeat step S4 to train the borehole strata handwritten record project-level recognition model, and add the incorrectly recognized text and the corrected text to the confused text dictionary constructed in S5. S9: For the next task to be identified in project A, execute S6 and S7 to obtain the final identification result.

2. The method for recognizing handwritten borehole formation records according to claim 1, characterized in that: The steps for generating a handwritten-style photo in step S1 are as follows: (1) Randomly extract typical handwritten record photos of each recorder from the sample set X1, and integrate the handwritten record photos of other business personnel in the project team to form a target handwritten style photo dataset; (2) Based on Generative Adversarial Network (GAN), a handwritten style photo generation model is constructed, and the model is trained and optimized using the target handwritten style photo dataset; (3) Combine the conventional stratigraphic description text data obtained in S12 and the handwritten record photos obtained in S11, and use the handwritten style photo generation model trained and optimized in step (2) to generate handwritten style photos.

3. The method for recognizing handwritten borehole formation records according to claim 2, characterized in that, In step (1), handwritten record photos from the open dataset are also integrated to form a target handwritten style photo dataset.

4. The method for recognizing handwritten borehole formation records according to claim 1, characterized in that: In S1, sample sets X1 and X2 are merged in a ratio of ≥3:

2.

5. The method for recognizing handwritten borehole formation records according to any one of claims 1 to 4, characterized in that, The conventional stratigraphic description text data mentioned in step S1 includes: rock and soil name, color, structure and texture, density, moisture content, plasticity, weathering degree, joints and fissures, particle composition, particle size and inclusion characteristics.

6. The method for recognizing handwritten borehole formation records according to claim 5, characterized in that: The CRNN handwritten record recognition model described in step S2 includes the following structure: The convolutional layers employ an improved MobileNet architecture; The recurrent layer uses a general bidirectional long short-term memory network Bi-LSTM to enhance the contextual relationships between characters for sequence prediction; The transcription layer was classified using a common CTC connection timing tool.

7. The method for recognizing handwritten borehole formation records according to claim 6, characterized in that: In step S4, the following random sampling method is used to construct the subtask sample set: S411, the project-level sample set is grouped into two levels according to drilling mileage and recorder. The first level grouping attribute is the mileage range, and the second level grouping attribute is the recorder. S412, from the project-level sample set X p The handwritten records of the borehole formations by each recorder were randomly selected from each mileage interval to form a sub-task sample set.

8. The method for recognizing handwritten borehole formation records according to claim 7, characterized in that: The model training described in step S4 includes the following steps: (1) Initialize the model parameters θ and set the number of iterations K; (2) The model was trained using the stochastic gradient descent method on the subtask dataset, and the parameters after training were obtained as follows: (3) Update model parameters ∈ is the step size coefficient, which can be between 0.01 and 0.2 depending on the convergence. Repeat step (2) until the number of iterations is reached.

9. The method for recognizing handwritten borehole formation records according to claim 8, characterized in that: In step S5, the spatially nested confusion dictionary uses the structure [T F ,(T R [,(X,Y))], where T F To obfuscate the text; T R This is the correct text; (X,Y) are the borehole space coordinates, T F By T R It generates text based on combinations of similar-looking characters and easily confused characters; when the text lacks corresponding spatial coordinate information, the coordinates are set to (0,0).

10. The method for recognizing handwritten borehole formation records according to claim 9, characterized in that: For position coordinates (X) b ,Y b The automatic correction in step S6 of the handwritten borehole formation record includes the following steps: (1) Perform Chinese word segmentation on the recognition results to obtain a series of word groups W1, W2, ... W i …、W n ; (2) For each phrase W i Search for obfuscated text dictionary records and retrieve all dictionary matching results (T) R,1 ,(X1,Y1)),(T R,2 ,(X2,Y2)), …(T R,j ,(X j ,Y j ))…、(T R,m ,(X m ,Y m )); (3) Based on the obtained matching results, calculate the distance between the spatial location of the borehole and the spatial locations of each dictionary matching record. Text correction is performed based on the shortest matching distance; if no matching result is found, no text correction is performed.

Images