A Deep Learning-Based Method for Archiving Geotechnical Engineering Investigation Data

By using spatial segment indexing and an improved Perceiver IO model and RIME optimization algorithm, the problems of low efficiency and association errors in the archiving of geotechnical engineering investigation data were solved, achieving accurate archiving of investigation data and traceability of evidence, and improving the efficiency of investigation data organization and retrieval.

CN123115141APending Publication Date: 2026-11-17JINAN SURVEYING & MAPPING RES INST
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611352560.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-02
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing methods for archiving geotechnical engineering investigation data are inefficient, making it difficult to establish a continuous correspondence between investigation data and engineering spatial sections, result versions, and interpretation content. Furthermore, the lack of continuous recording of changes in investigation and interpretation processes leads to data omissions, duplications, and errors in location correlation, affecting verification and long-term utilization.

Method used

By combining spatial segment indexing, an improved Perceiver IO model, and the RIME optimization algorithm, heterogeneous exploration data are standardized, spatially segmented, and versioned. Evidence associations are established through secondary cross-attention reading of the continuous potential array and residual directional write-back, forming interpretation branches and version relationships that bind supporting or excluding evidence pointers.

Benefits of technology

It has enabled accurate archiving and clear version evolution of exploration data, complete evidence retrospection, improved the accuracy of exploration data organization and archiving and the convenience of retrieval and verification, and enhanced the traceability and long-term utilization value of exploration data.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115141A_ABST
    Figure CN123115141A_ABST
Patent Text Reader

Abstract

This invention discloses a deep learning-based method for archiving geotechnical engineering investigation data, relating to the field of data storage and management technology. The method includes the following steps: S1, collecting and normalizing investigation data; S2, mapping to engineering space, segmenting continuous ranges, and establishing segment indexes; S3, organizing data fragments and output content, and encoding them according to version order; S4, integrating with an improved Perceiver IO model, performing two cross-attention reads using a persistent latent array and residual write-back to form evidence association states; S5, optimizing residuals using the RIME optimization algorithm and re-decoding; S6, merging consistent interpretation items, separating conflicting items, and binding evidence pointers; S7, dividing interpretation branch boundaries according to the official output version; S8, establishing structured archive records and retrieval indexes to form archive results that allow for the retrospective understanding of original evidence and the reasons for changes. This invention improves the accuracy and traceability of archiving.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of data storage and management technology, and in particular to a method for archiving geotechnical engineering investigation data based on deep learning. Background Technology

[0002] Existing geotechnical engineering investigation data usually come from drilling records, in-situ tests, laboratory tests, and investigation results documents. Different data differ in file format, field names, units of measurement, spatial reference, and recording methods. Before archiving, data often relies on manual sorting and format conversion, which is inefficient and can easily lead to data omissions, duplications, or errors in spatial location correspondence.

[0003] Existing archiving methods typically categorize data by project, file type, or creation time, making it difficult to establish a continuous correspondence between survey data and engineering spatial sections, result versions, and interpretation content. When different survey conclusions appear within the same spatial area, existing systems often employ overwrite updates or parallel storage methods, making it difficult to distinguish between supporting and conflicting relationships between interpretation content.

[0004] Furthermore, existing methods lack continuous recording of changes in exploration interpretation. When new data or official results are updated, the connection between historical interpretations and original evidence is easily interrupted, making it difficult to accurately trace the exploration data, results content, and version boundaries on which the interpretation changes are based. This affects the verification, retrieval, and long-term use of geotechnical engineering exploration data.

[0005] Therefore, how to provide a deep learning-based method for archiving geotechnical engineering investigation data is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a deep learning-based method for archiving geotechnical engineering investigation data. This invention employs a combination of spatial segment indexing, an improved Perceiver IO model, and the RIME optimization algorithm to standardize, spatially segment, and version-encode heterogeneous investigation data. Evidence associations are established through secondary cross-attention reading of the continuous latent array and residual-oriented write-back, forming interpretation branches and version relationships that bind support or exclude evidence pointers. This enables structured archiving and change tracing of investigation data, interpretation results, and original evidence, offering advantages such as accurate archiving associations, clear version evolution, complete evidence backtracking, and convenient retrieval and verification.

[0007] A method for archiving geotechnical engineering investigation data based on deep learning according to an embodiment of the present invention includes the following steps: S1. Collect geotechnical engineering investigation data, and perform normalization processing on data from different sources to form standardized investigation data; S2. Map the standardized survey data to the engineering space, divide it into continuous spatial ranges and write it into the spatial segment index to form survey data fragments; S3. Organize the exploration data segments and exploration results according to the spatial segment index, retain the version order and encode them as the archived input sequence; S4. The archived input sequence is connected to the improved Perceiver IO model. The persistent latent array, which is retained in version order, is read through cross-attention after two encodings. The residual formed by the relative temporary latent state after the second reading is written back to the persistent latent array and the write-back position is recorded. The archived evidence association state is then decoded. The S5 and RIME optimization algorithms use the recorded write-back positions as the updatable dimension of the hard frost puncture mechanism, optimize the corresponding residuals and re-decode them to form optimized evidence association results. S6. Based on the results of optimized evidence association, merge exploration interpretation items with consistent attributes within the same spatial segment and separate conflicting items to form interpretation branches that bind supporting or excluding evidence pointers. S7. Correct the interpretation branches with the official survey results, divide the version boundaries along the evidence pointers that caused the interpretation changes, and form the interpretation branch version relationship; S8. Establish structured archive records according to the interpretation of branch version relationships, connect evidence pointers to the retrieval index, and form geotechnical engineering investigation data archive results that can trace back the original evidence and the reasons for changes.

[0008] Optionally, S1 specifically includes: S11. Analyze the data format of geotechnical engineering investigation data, and extract data fields, field content, and source information; S12. Match synonymous fields from different sources based on the meaning of the data fields, map the matched fields to a unified field, and retain the original field name and source information for unmatched fields. S13. Perform conversion on the unit of measurement, time format and text encoding in the unified field and the unmatched field. For fields that cannot be converted, retain the original value and source information. S14. Integrate the converted field content according to the unified field name and the original field name, merge duplicate records with the same content, and retain records with conflicting content and source information in parallel to form standardized survey data.

[0009] Optionally, S2 specifically includes: S21. Read the spatial coordinates and depth range of standardized survey data, and convert the location data using different spatial benchmarks to the same engineering space; S22. Arrange the standardized survey data according to the converted location data, compare the spatial coverage boundaries of adjacent data, classify data with adjacent or overlapping boundaries into the same continuous spatial range, and cut at the boundary separation point to form a continuous spatial range. S23. Write the start and end boundaries and arrangement order of the continuous spatial range into the spatial segment index, and associate the spatial segment index with the standardized exploration data within the range to form exploration data fragments.

[0010] Optionally, S3 specifically includes: S31. Read the spatial range recorded in the survey results, retrieve the spatial segment index, and associate the survey results that intersect with the spatial segment with the corresponding survey data segment. S32. After parsing the version identifiers of the associated exploration results, arrange the exploration results within the same spatial segment according to the order of the version identifier records to form a version sequence. S33. Encode the exploration data fragments and associated exploration results into a feature vector of a unified dimension, and superimpose the spatial segment index code and version sequence code to form an archived input sequence.

[0011] Optionally, the improved Perceiver IO model includes a persistent latent array, a first encoding cross-attention layer, a latent processing module, a second encoding cross-attention layer, a residual write-back connection, an output query array, and a decoding cross-attention layer. The latent processing module includes a latent self-attention layer and a multilayer perceptron.

[0012] Optionally, the formation of the archived evidence association status specifically includes: The archived input sequence is read in version order. The first version initializes the persistent latent array with the latent vectors obtained from model training, and the remaining versions read the persistent latent arrays from the previous version after the residuals have been written back. The persistent latent array is copied, and the first encoding cross-attention layer projects the copied latent vectors into query vectors, and projects the current version of the archived input sequence into key vectors and value vectors; Calculate the attention weight of each potential index for each sequence position and aggregate the value vector. After projecting the aggregated result onto the output, add it to the copied potential vector to form the first read state. The continuous potential array keeps the potential vector before the read. The latent self-attention layer projects the first read state into a query vector, a key vector, and a value vector, aggregates latent features along the latent index, and adds them to the first read state; The multilayer perceptron performs linear transformation, nonlinear activation, and linear recovery along the feature channels, and adds the recovery result to the output of the latent self-attention layer to form a temporary latent state; The second encoding cross-attention layer projects the temporary latent state into a query vector and reprojects the current version of the archived input sequence into a key vector and a value vector, while masking the padding positions caused by length padding of the archived input sequence. Aggregate the value vector along the unmasked sequence position, project the aggregation result onto the output, and add it to the temporary latent state to form the second read state; Based on the same potential index and feature channel, the temporary potential state is subtracted from the second read state to form the residual write-back parameters; The residual write-back connection bypasses the first read state and the temporary latent state, adds the residual write-back parameters to the persistent latent array before the read, keeps the feature channels that do not generate residual components unchanged, and determines the latent index and feature channel that generate residual components as the valid write-back position; The spatial segment index encoding and version sequence encoding are arranged in the order of the archived input sequence to form the output query array. The decoding cross-attention layer uses the output query array as the query and the persistent potential array that has completed the residual write-back as the key and value to perform attention reading. Record the exploration interpretation items obtained from decoding and their association results with exploration data fragments and exploration results, and attach the valid write-back location and corresponding residual write-back parameters to form the associated status of archived evidence.

[0013] Optionally, S5 specifically includes: S51. Arrange the residual write-back parameters corresponding to the effective write-back positions in the order of potential index and feature channel, and use the arranged parameters as the candidate solution dimension to form a candidate solution group; S52. Write each candidate solution back to the persistent potential array and decode it again. Check whether the interrelated exploration data fragments and exploration results in the decoding results correspond to the same spatial segment. Determine the fitness of the candidate solution by subtracting the number of different spatial segments from the number of identical spatial segments. S53. The soft-frost search strategy uses the same-dimensional parameters of the candidate solution with the highest fitness as a benchmark, and combines the iteration progress and random direction to update the dimension corresponding to the effective write-back position. S54. Map the fitness of candidate solutions to a uniform numerical range. The hard frost puncture mechanism generates random numbers dimension by dimension. When the random number is lower than the mapped value, it is replaced with the same dimension parameter of the candidate solution with the highest fitness. The search results of soft frost are retained for the missing dimension. S55. Re-decode the updated candidate solution and calculate the fitness. The positive greedy selection mechanism replaces the original candidate solution when the fitness after the update is greater than the fitness before the update. Write the candidate solution with the maximum fitness back to the valid write-back position to form the optimized evidence association result.

[0014] Optionally, S6 specifically includes: S61. Read the exploration interpretation items in the optimized evidence association results according to the spatial segment index, as well as the exploration data fragments and exploration results associated with each exploration interpretation item; S62. Compare the interpretation content of the same interpretation attribute in the same spatial segment item by item, and merge the exploration interpretation items with consistent interpretation content and their associated exploration data fragments and exploration results. S63. The merged exploration interpretation items are identified as the same interpretation branch. Exploration interpretation items with conflicting interpretation content are established as interpretation branches according to different interpretation contents, and their associated exploration data fragments and exploration results are retained. S64. Establish supporting evidence pointers for exploration data fragments and exploration results associated with the same interpretation content as the interpretation branch, establish exclusion evidence pointers for exploration data fragments and exploration results associated with conflicting interpretation content, and bind them to the interpretation branch.

[0015] Optionally, S7 specifically includes: S71. According to the spatial segment index and interpretation attributes, match the official exploration results version with the interpretation branch item by item, and arrange them in the order of version. S72. Compare the content of the formal survey results with the interpretation content of the interpretation branch version by version. If the content is consistent, continue the interpretation branch. If the content is conflicting, copy the interpretation branch, replace the conflicting interpretation content with the content of the formal survey results, and bind the evidence pointer pointing to the corresponding survey results content. S73. Evidence pointers that are newly added after copying and point to the content of the formal survey results shall be identified as evidence pointers that cause changes in interpretation. S74. Divide the version boundaries by the version of the survey results indicated by the evidence pointer that causes the change in interpretation, connect the interpretation branches before and after the boundary, and form the interpretation branch version relationship.

[0016] Optionally, S8 specifically includes: S81. Arrange the interpretation branches according to the spatial segment index and version order, and connect the interpretation branches of adjacent versions according to the version boundary; S82. Write the interpretation branches, version boundaries, and interpretation content before and after the boundaries into the structured archive record, and associate the evidence pointers that cause interpretation changes with the corresponding version boundaries; S83. Analyze supporting evidence pointers and excluding evidence pointers, and locate the exploration data fragments and exploration results content pointed to by the pointers; S84. Incorporate spatial segment indexes, version boundaries, and evidence pointers into the retrieval index, and connect the structured archive records with the content pointed to by the pointers to form the archived results of geotechnical engineering investigation data.

[0017] The beneficial effects of this invention are: This invention normalizes geotechnical engineering survey data from different sources and maps the standardized survey data to a unified engineering space. It establishes a spatial segment index according to continuous spatial range, so that the survey data fragments, survey results content and their version order form a stable correspondence. This reduces data omissions, duplications and location association errors caused by inconsistencies in data format, spatial benchmarks and recording methods in the traditional archiving process, and improves the accuracy of survey data organization and archiving.

[0018] This invention employs an improved Perceiver IO model to process archived input sequences. It stores the latent states between different versions through a persistent latent array, uses cross-attention with two encodings to read archived content, and writes the residuals formed by the second reading back to the persistent latent array. This allows newly added exploration information to update the evidence association state while preserving historical associations. The RIME optimization algorithm further limits the residual write-back dimension, reduces the interference of irrelevant feature updates on the evidence association results, and improves the reliability of the association between exploration interpretation items and original data.

[0019] This invention forms an interpretation branch relationship with version boundaries by merging consistent interpretation items, separating conflicting interpretation items, and binding supporting evidence pointers and excluding evidence pointers. This enables the interpretation content, original evidence, and reasons for changes before and after the official exploration results are updated to be continuously saved and retrieved, thereby improving the traceability, verification efficiency, and long-term utilization value of geotechnical engineering exploration data. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a deep learning-based geotechnical engineering investigation data archiving method proposed in this invention; Figure 2 This is a schematic diagram of the structure of the improved Perceiver IO model for a deep learning-based geotechnical engineering investigation data archiving method proposed in this invention. Figure 3 This is a flowchart illustrating the cross-version residual directional write-back and optimization process of a deep learning-based geotechnical engineering investigation data archiving method proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figures 1-3A deep learning-based method for archiving geotechnical engineering investigation data includes the following steps: S1. Collect geotechnical engineering investigation data, and perform normalization processing on data from different sources to form standardized investigation data; S2. Map the standardized survey data to the engineering space, divide it into continuous spatial ranges and write it into the spatial segment index to form survey data fragments; S3. Organize the exploration data segments and exploration results according to the spatial segment index, retain the version order and encode them as the archived input sequence; S4. The archived input sequence is connected to the improved Perceiver IO model. The persistent latent array, which is retained in version order, is read through cross-attention after two encodings. The residual formed by the relative temporary latent state after the second reading is written back to the persistent latent array and the write-back position is recorded. The archived evidence association state is then decoded. The S5 and RIME optimization algorithms use the recorded write-back positions as the updatable dimension of the hard frost puncture mechanism, optimize the corresponding residuals and re-decode them to form optimized evidence association results. S6. Based on the results of optimized evidence association, merge exploration interpretation items with consistent attributes within the same spatial segment and separate conflicting items to form interpretation branches that bind supporting or excluding evidence pointers. S7. Correct the interpretation branches with the official survey results, divide the version boundaries along the evidence pointers that caused the interpretation changes, and form the interpretation branch version relationship; S8. Establish structured archive records according to the interpretation of branch version relationships, connect evidence pointers to the retrieval index, and form geotechnical engineering investigation data archive results that can trace back the original evidence and the reasons for changes.

[0023] In this embodiment, S1 specifically includes: Geotechnical engineering investigation data includes spreadsheets, text files, database exported files, word processing documents, searchable PDF files, and scanned image files. The archiving system reads the file extension, file header bytes, internal directory information, and text layer information. The file header bytes determine the actual data format. When the extension and file header are inconsistent, the parsing program corresponding to the file header is used, and the original file name and the actual format identification result are preserved.

[0024] Spreadsheets read cell content by worksheet, row, and column; text files are split into fields by character encoding and delimiters; word processing documents read content by paragraph and table area; searchable PDF files restore reading order by text coordinates; scanned image files first perform layout area detection, then perform character recognition on text and table areas, write blank values ​​for unrecognized characters and retain page numbers, image coordinates, and original image areas to prevent failed recognition content from deviating from the source material.

[0025] Data fields are extracted from the header text, field labels, and adjacent content, and the field content is saved synchronously with the original field name; the source information includes the source file check value, file name, page number or worksheet name, row and column position, and page coordinates. The source file check value is obtained by calculating the hash value of all bytes in the file. The same field content is always recorded in pairs with the corresponding source information. When the file is copied or the name is changed, the original data is still located through the check value.

[0026] The meaning of data fields is identified by the field name, the data type of the field content, the unit of measurement, and the position of the field in the table. First, spaces, connectors, and unit brackets are removed from both ends of the field name. Then, the normalized field name is compared with the name and synonyms in the unified field table one by one. If the name matches, the field content is checked to see if it can be converted to the data type and unit of measurement required by the unified field. The meaning of the field is determined when both the name and the content type match.

[0027] The unified field table records the unified field name, allowed data types, target unit of measurement, and corresponding synonyms. When multiple unified fields are matched simultaneously, the matching results with the same data type are retained first, followed by the matching results with units of measurement that can be directly converted. If multiple matching results still exist, field merging is not performed, and the original field name, field content, and source information are saved as unmatched fields, thereby avoiding the incorrect writing of ambiguous data content into the unified fields.

[0028] The unit of measurement conversion reads the source unit, target unit, and conversion factor between them from the unified field table, multiplies the field value by the conversion factor, and retains the original data's significant digits; the time content is split into year, month, day, hour, minute, and second and written as a fixed-length number sequence; the text content is converted to a unified character encoding after identifying the source encoding through byte order markers and character distribution; when the conversion factor, time composition, or source encoding is unavailable, the original value, original unit, original encoding, and source information are retained.

[0029] Duplicate records are compared using the engineering object identifier, data category, spatial location or depth range, and inspection items. If the comparison items are completely identical, the converted content is compared field by field. If all field contents are identical, a standard record is retained, and the source information of the remaining records is appended to the same standard record. If the comparison items are identical but any field content is different, each record is retained separately, and the same conflict association number is written between the records. Conflicting content is not overwritten or averaged.

[0030] During field integration, matching fields are saved with a unified field name, while unmatched fields are saved with their original field names. Each field simultaneously records the converted field content, the original field content, the original field name, the source file checksum, and the original location. Multiple sources of information that are recorded repeatedly are appended sequentially, and conflict records are saved separately and located to each other through conflict association numbers. After integration, standardized survey data with unified field names, convertible content, and traceable sources are formed.

[0031] In this embodiment, S2 specifically includes: Spatial coordinates are read from the borehole, trench, or test point location records in the standardized exploration data; elevation is read from the borehole elevation or test point elevation records; and depth range consists of the starting depth and ending depth in the same record. When the record only contains absolute elevation, the difference between the borehole elevation and the recorded elevation is converted into depth range, and the positions of the two are swapped when the starting value is greater than the ending value.

[0032] The engineering space adopts the plane coordinate benchmark, elevation benchmark and depth benchmark that are uniformly used in the engineering project and increase downward from the orifice; if the benchmark name is missing, the engineering information page of the same source file is read. If it still cannot be obtained, the original location data and source information are retained and written into the spatial verification set. The data in the spatial verification set does not participate in the continuous spatial range division.

[0033] The planar coordinate transformation first reads the transformation parameters provided by the engineering control network, and then performs scale conversion, planar rotation, and coordinate translation on the source coordinates in sequence. If the engineering control network does not record the transformation parameters, it reads the control points that appear in both the source coordinate datum and the engineering space, uses no less than 3 non-collinear control points to calculate the scale, rotation angle, and translation, and determines the transformation result by minimizing the sum of the squared distances between the transformed coordinates of each control point and the engineering coordinates.

[0034] The elevation conversion reads the benchmarks shared by the two elevation datums, calculates the difference between the engineering elevation and the source elevation of each benchmark, and takes the median of all differences as the elevation correction. Then, it calculates the absolute value of the difference between each difference and the median, and takes the average of the absolute values. If a single absolute value exceeds three times the average, the corresponding benchmark is deleted and recalculated. If no valid benchmarks can be retained, the original elevation is written into the spatial set to be checked.

[0035] The depth benchmark conversion first identifies the depth recording direction. The depth value recorded from the borehole opening downwards remains unchanged. The absolute elevation record is obtained by subtracting the recorded elevation from the converted borehole elevation, and the depth unit is converted to the unified unit of the engineering project. When the recording direction cannot be identified or the borehole elevation is missing, the original depth value and source information are retained and written into the spatial set to be verified, and no spatial segment index is generated.

[0036] After completing the coordinate, elevation, and depth transformations, the standardized survey data are arranged sequentially according to their planar location and starting depth. The spatial coverage boundary of each record consists of the planar coordinate range, the starting depth, and the ending depth. The planar range of point records is taken from the same coordinate, while the planar range of line or area records is taken from the minimum and maximum values ​​of all spatial points in two coordinate directions.

[0037] Before comparing adjacent records, the coordinate and depth values ​​are uniformly retained to the lower decimal place of the two records; the boundaries are determined to be connected when the termination boundary of the previous record is equal to the starting boundary of the next record, the boundaries are determined to be coincident when the two ranges have a common interval, and the boundaries are determined to be separated when the termination boundary of the previous record is smaller than the starting boundary of the next record and the difference between the two is greater than the smallest unit of measurement corresponding to the uniform precision.

[0038] Records with adjacent or overlapping boundaries are continuously grouped into the same spatial range. The starting boundary of the continuous spatial range is the minimum value among the starting boundaries of the records being grouped into it, and the ending boundary is the maximum value among the ending boundaries. When boundary separation or non-intersecting planar positions are detected, the current continuous spatial range ends, and a new continuous spatial range is established based on the starting boundary of the next record.

[0039] The spatial segment index consists of engineering spatial identifier, planar position code, depth boundary code, and segment sequence code. The planar position code is generated by arranging the transformed coordinates according to the precision of the engineering data. The depth boundary code is generated by the starting and ending depths of the continuous spatial range. The segment sequence code is written in ascending order according to the arrangement of the continuous spatial range under the same planar position from shallow to deep.

[0040] Spatial segment indexes and standardized survey data are associated and stored through record identifiers. Each record is written into a spatial segment index that covers its spatial range. When a single record crosses a segment boundary, the spatial range is split at the boundary position. Each split part inherits the original field content, source information and record identifier, and is written into the corresponding spatial segment index to form a survey data fragment with continuous position, clear boundaries and the ability to locate the original record.

[0041] In this embodiment, S3 specifically includes: The survey results are obtained by extracting the engineering object identifier, plane coordinates, starting depth, and ending depth from the result tables, text paragraphs, and map annotations; when only the borehole number is recorded, the spatial coordinates and borehole elevation in the standardized survey data are retrieved by the borehole number, and the retrieved results are uniformly represented as the plane position range and depth range.

[0042] The planar location range of point-like results is represented by a single coordinate; the content of profiles or lines is represented by the starting coordinates and the ending coordinates; and the content of regional results is represented by a sequence of boundary points. When the starting depth or ending depth is missing, the corresponding boundary is read from the exploration data fragment associated with the result content. When both depth boundaries are missing, the result content and source information are retained, and spatial segment associations are not established.

[0043] Spatial range matching first compares the result content with the planar range of the spatial segment, and then compares the depth range. When the result position falls within the planar boundary of the spatial segment and the depth range of the result has a common interval with the depth range of the segment, an intersection is determined. For point-like result content, coordinate inclusion is used for judgment. For line-like and regional result content, the outer boundary is compared first, and then the actual intersection of the line segment or boundary region is calculated.

[0044] When the content of the exploration results intersects with only one spatial segment, the content identifier of the results is written into the associated record of the corresponding exploration data segment; when it spans multiple spatial segments, the intersecting spatial range within each segment is calculated separately, and an independent associated record is established for each intersecting segment. All associated records jointly store the original content identifier and source information of the results, and the main text of the results is neither split nor stored repeatedly.

[0045] Version identifiers are read from the deliverable file name, version field, and revision history. First, the major version number is compared, then the revision number is compared. If the numbers are the same, the deliverable creation time is compared. If the major version number, revision number, and creation time are all the same, they are considered to be the same deliverable version. The source file check value only determines the fixed arrangement order of deliverable content within the same version and is not used as a basis for determining the order of versions.

[0046] The survey data fragments and related results are divided into text content, category content, and numerical content. After word segmentation, the word vectors are read from the word vector table that is jointly updated during model training and the average is calculated. Out-of-vocabulary words are written into zero vectors and missing tags. The category content reads the category embedding vectors obtained from training through the category number. The numerical content is subtracted from the mean of the same field in the training set and then divided by the standard deviation. When the standard deviation is zero, it is written as zero.

[0047] Each field vector is concatenated according to a unified field order and then linearly projected to the archived input dimension. If there are not enough fields, they are padded with zero vectors. Fields that exceed the input range are stored in the source record. The planar position code, depth boundary code, and segment order code in the spatial segment index are converted into integer indices and then the spatial embedding vector obtained from training is read.

[0048] Version sequence encoding uses version order as position value. Even-numbered feature channels are written with the sine value of the position value divided by the channel ratio raised to the power of 10000. Adjacent odd-numbered feature channels are written with the same cosine value. The channel ratio is obtained by dividing the current channel number by the total encoding dimension.

[0049] Feature vectors, spatial segment index codes, and version sequence codes maintain the same dimension and are added item by item in the same feature channel. The superimposed vectors are first sorted by spatial segment index, then by version sequence. Within the same segment and version, the order of exploration data fragments is maintained before exploration results content, forming an archived input sequence.

[0050] In this embodiment, the improved Perceiver IO model includes a persistent latent array, a first encoding cross-attention layer, a latent processing module, a second encoding cross-attention layer, a residual write-back connection, an output query array, and a decoding cross-attention layer. The latent processing module includes a latent self-attention layer and a multilayer perceptron. The ordinary Perceiver IO model uses a learnable latent array as the encoding query, performs an encoding cross-attention once with the current input array as the key and value, and then processes the latent array by a latent self-attention layer and a multilayer perceptron. The output query array reads the processed latent array through decoding cross-attention and generates the output result.

[0051] The latent array of the ordinary Perceiver IO model serves a single forward propagation, without setting up a latent state that is continuously saved along the version order of the exploration results, and without reading the current version archived input sequence again after the latent processing is completed; when adjacent versions are input one after another, the newly added associations generated by the current version are easily mixed with existing latent features, and it is impossible to limit the changed content to the corresponding latent index and feature channel.

[0052] The improved Perceiver IO model includes a persistent latent array, a first encoding cross-attention layer, a latent processing module, a second encoding cross-attention layer, a residual write-back connection, an output query array, and a decoding cross-attention layer; the latent processing module consists of a latent self-attention layer and a multilayer perceptron connected sequentially.

[0053] The input of the first encoded cross-attention layer is connected to the copy result of the persistent latent array and the archived input sequence, respectively, and the output is connected to the latent self-attention layer; the output of the latent self-attention layer is connected to the multilayer perceptron, and the output of the multilayer perceptron is connected to the query input of the second encoded cross-attention layer.

[0054] The key and value inputs of the second encoding cross-attention layer are connected to the same version of the archived input sequence, and the output is connected to the persistent latent array via residual write-back; the query input of the decoding cross-attention layer is connected to the output query array, and the key and value inputs are connected to the persistent latent array after residual write-back.

[0055] The improved Perceiver IO model restricts the first read and potential processing to the transient state by adding a second encoding cross-attention layer and rewriting the residual connection position. Only the features added by the second read relative to the transient state are written into the persistent potential array saved across versions.

[0056] The archived input sequence is a three-dimensional tensor composed of batch, sequence position and feature channel. In one embodiment, each sequence position can take 256 feature channels; the persistent latent array can take 128×256. The first coding cross attention layer and the second coding cross attention layer are respectively set with independent projection parameters. The output query array is consistent with the number of effective sequence positions and maintains 256 feature channels.

[0057] The initial parameters of the persistent latent array are generated using a normal distribution with a mean of zero and a standard deviation equal to the reciprocal of the square root of the number of feature channels. The linear projection and multilayer perceptron weights are initialized using a uniform distribution whose upper and lower bounds are determined by the input and output dimensions, and the bias is initialized to zero. The parameters of input projection, spatial embedding, two-layer encoding cross attention, latent processing module, interpretation output mapping, evidence association mapping, input position projection, and decoding cross attention are all updated during model training.

[0058] The training data is taken from survey projects that have been archived and reviewed by geotechnical engineers. The training samples are composed of adjacent result versions within the same spatial segment, forming a version sequence. Each sample contains standardized survey data fragments, corresponding survey result content, spatial segment index, version order, and manually confirmed survey interpretation content.

[0059] Evidence association labels are generated by geotechnical engineers through item-by-item verification. Data fragments or results that directly support the content of the investigation and interpretation are marked as positively associated, while data that does not correspond to the current interpretation attribute or points to conflicting interpretation content are marked as negatively associated. When the same data involves multiple interpretation attributes, independent labels are created for each attribute and they are not merged into a single label.

[0060] The exploration interpretation labels are extracted from the soil and rock categories, layer interfaces, and engineering property evaluations in the results documents. The category content is converted into category numbers, and the numerical content retains the values ​​after being standardized by the unit. If the results of two labelers are inconsistent, a third geotechnical engineer will verify the original data and determine a unique label.

[0061] The training samples are divided into training set, validation set and test set based on engineering projects. Different versions of the same engineering project should not be included in different sets. In one implementation, the ratio of the number of samples in the training set, validation set and test set can be 8:1:1 to avoid the same project data from participating in parameter training and effect verification at the same time.

[0062] The total model loss is obtained by adding the evidence association loss, the explanatory content loss, the cross-version consistency loss and the residual sparsity loss. In one implementation, the weights of the four items can be 60%, 25%, 10% and 5% respectively. The loss is first averaged within the batch and then multiplied by the corresponding weight.

[0063] For evidence association loss, binary cross-entropy is calculated for each explanatory term and each valid input position. For positive association labels, the negative logarithm of the association probability is calculated, and for negative association labels, the negative logarithm of 1 minus the association probability is calculated. Positive and negative samples are multiplied by the reciprocal of their respective sample sizes. Cross-entropy is used for text and categories in the explanatory content, and smoothed absolute error between predicted and manually confirmed values ​​is used for numerical content.

[0064] Cross-version consistency loss is calculated as the average of the absolute values ​​of the decoding feature differences corresponding to positions where the interpretation content has not changed in adjacent versions. Positions that have been confirmed to have changed are not included in this calculation. Residual sparsity loss is calculated as the average of the absolute values ​​of all residual write-back parameters. Positions with missing labels are excluded by label masks.

[0065] The model uses the AdamW optimizer to update parameters. In one implementation, the initial learning rate can be 0.0001, a single batch can have 16 version sequences, and the weight decay coefficient can be 0.01. After each backpropagation, the overall norm of the gradients of all parameters is calculated. When the overall norm exceeds 1, the gradients of each parameter are reduced by the same proportion.

[0066] Training consists of one complete traversal of the training set as one round. After each round, the total loss is calculated on the validation set. If the total loss on the validation set does not decrease for three consecutive rounds, the learning rate is halved. If the decrease is less than 0.1% for five consecutive rounds, training is stopped. The maximum number of training rounds can be 100, and the model parameters with the lowest total loss on the validation set are saved.

[0067] After training, the initial parameters of the persistent latent array, the two-layer encoding cross-attention parameters, the latent processing module parameters, and the decoding cross-attention parameters are fixed. During actual archiving, only forward computation is performed, and the model training parameters are no longer updated. Changes between versions are handled by the residual write-back of the persistent latent array and the RIME optimization algorithm.

[0068] In this embodiment, the formation of the archived evidence association status specifically includes: The generation of archived evidence association states uses a persistent potential array as a cross-version state carrier. The persistent potential array is stored in a two-dimensional tensor, with array rows corresponding to potential indices and array columns corresponding to potential feature channels. Each potential index stores the potential association features formed between different spatial segments and result versions of the archived input sequence.

[0069] In one implementation, the persistent latent array can be 128×256, where 128 corresponds to the number of latent indices and 256 corresponds to the number of feature channels carried by each latent index. The number of latent indices and the number of feature channels are determined before model training and are consistent with the input and output dimensions of the encoding cross-attention layer, the latent processing module, and the decoding cross-attention layer.

[0070] The initial contents of the persistent latent array are taken from the trainable latent vector parameters saved after the model training is completed. The archived input sequence is read from the first version in the version order. The first version loads the initial latent vector, and the remaining versions load the persistent latent array saved after the residual is written back in the previous version.

[0071] The persistent potential array is saved as an array snapshot with a version identifier. After the current version is processed, the result is written back and registered in the current version snapshot. The next version reads the array that is adjacent in the version order and has been processed. When there is a gap in the version order, the last valid version before the gap is read. When there is no valid input position in the current version, the read array is saved and no residual write-back parameters are generated.

[0072] The persistent latent array is fully copied before entering the first encoding cross-attention layer. The copied result has the same latent index and feature channels as the persistent latent array. The copied result enters the temporary computation path, and the persistent latent array remains unchanged during the first read and latent processing.

[0073] The first encoding cross-attention layer includes layer normalization, query projection, key projection, value projection, multi-head cross-attention, output projection, and residual connection; query projection receives the copy result of the persistent latent array, while key projection and value projection receive the current version archived input sequence.

[0074] Layer normalization calculates the mean and variance of all feature channels of the latent vector row by row, subtracts the mean of the corresponding row from each channel, and then divides by the sum of the square root of the variance and the stability constant. In one implementation, the stability constant can be taken as 0.00001 to avoid division by zero when the variance is zero.

[0075] The normalized latent vector is multiplied by the query projection parameters and then superimposed with the query bias to form a query vector with the same number of latent indexes. The current version archive input sequence is subjected to the same form of layer normalization, then multiplied by the key projection parameters and value projection parameters respectively and superimposed with the corresponding biases to form key vectors and value vectors with the same number of sequence positions.

[0076] Each valid position in the archived input sequence corresponds to a fragment of exploration data or a piece of exploration results. For sequences with insufficient length within the same batch, zero vectors are added to the end. A masking flag with a value of 1 is written synchronously to each valid position, and a masking flag with a value of 0 is written synchronously to each padding position.

[0077] In one implementation, the query vector, key vector, and value vector can all be 256-dimensional, and are evenly divided into 8 attention heads along the feature channels. Each attention head contains 32 feature channels. Each attention head independently completes the attention score calculation and value vector aggregation, and then the aggregation results of all attention heads are connected along the feature channels.

[0078] The query subvector of each potential index is multiplied and accumulated with the key subvector of each sequence position in the same feature channel, and then divided by the square root of the number of channels in a single attention head to form the attention score between the corresponding potential index and the sequence position.

[0079] Before normalization, the attention score for the fill position is replaced with a sufficiently small negative value in the calculation, which can be -1000000000 in one implementation. The attention score for the effective position is first subtracted from the maximum score in the same query row, and then the index value is calculated and divided by the sum of the index values ​​of all effective positions in the same row to form an attention weight with a total sum of 1.

[0080] When all sequence positions corresponding to a single potential index are masked, the current attention head outputs a zero vector, and the residual connection retains the copied potential vector; when there are valid sequence positions, each attention weight is multiplied by the corresponding sub-vector and accumulated along the sequence position to form the aggregation result of the current attention head.

[0081] The aggregation results of multiple attention heads are concatenated along the feature channels, multiplied by the output projection parameters, and the output bias is superimposed. The output projection results are then added to the copy results of the persistent latent array item by item at the same latent index and feature channel to form the first read state.

[0082] The latent processing module includes a latent self-attention layer and a multilayer perceptron. The latent self-attention layer includes layer normalization, query projection, key projection, value projection, multi-head self-attention, output projection, and residual connection. The query vector, key vector, and value vector are all generated from the first read state.

[0083] The latent self-attention layer multiplies the query subvector of each latent index with the key subvector of other latent indices channel by channel and accumulates them. After scaling and normalization, the aggregated value subvector is formed. The attention calculation scope is changed from the archived input sequence position to the latent index, so that the associated features scattered in different input positions within the same version can be exchanged between latent indices.

[0084] After multiple attention heads of the latent self-attention layer are aggregated, they are connected along the feature channels, and then the output projection is performed and added to the input of the latent self-attention layer item by item; the input and output have the same number of latent indices and feature channels.

[0085] The multilayer perceptron includes a first linear transformation, a nonlinear activation, a second linear transformation, and a residual connection. In one embodiment, the first linear transformation can expand the 256-dimensional latent vector to 1024-dimensional, and the expanded result is superimposed with a bias and then activated by Gaussian error linear units. The second linear transformation then restores the 1024-dimensional features to 256-dimensional.

[0086] The vector recovered by the multilayer perceptron is added to the output of the latent self-attention layer item by item at the same latent index and feature channel to form a temporary latent state. When the latent processing module is set with multiple latent processing units, the output of the previous processing unit is directly used as the input of the next processing unit. In one embodiment, the number of processing units can be 4.

[0087] The second coded cross-attention layer includes layer normalization, query projection, key projection, value projection, multi-head cross-attention, output projection, and residual connection. The second coded cross-attention layer has independent projection parameters and does not share query projection parameters, key projection parameters, value projection parameters, and output projection parameters with the first coded cross-attention layer.

[0088] The second encoding cross-attention layer generates a query vector from a temporary latent state and regenerates the key and value vectors from the current version archived input sequence; the regenerated key and value vectors use projection parameters obtained independently from the second encoding cross-attention layer.

[0089] The second read continues to use the masking markers generated in the first read, and fills in the positions without entering the attention weight normalization; the potential index of the position where no effective sequence was read outputs a zero aggregate vector, and the corresponding temporary potential vector is retained through residual connection.

[0090] The second encoding cross-attention layer recalculates the attention score between each latent index and each valid sequence position in the current version. After scaling, normalization, and value vector aggregation, it connects multiple attention heads. The connection result is then projected onto the output and added to the temporary latent state item by item to form the second read state.

[0091] The second encoding cross-attention layer is located after the latent processing module. Its computation object is still the current version archived input sequence, but the query object has been changed from the copy result of the persistent latent array to the temporary latent state after feature exchange between latent indices; the second read thereby extracts the current version input increment not included in the temporary latent state.

[0092] The co-position residual calculation is located between the second coding cross-attention layer and the residual write-back connection. The second read state and the temporary latent state are subtracted item by item using the same latent index and the same feature channel. The subtraction result eliminates the temporary latent state retained by the residual connection in the second read state, leaving only the newly added feature components generated by the second output projection.

[0093] The two-dimensional tensor obtained by subtracting each term is used as the residual write-back parameter. The number of rows, columns, and arrangement order of the residual write-back parameter are completely consistent with the persistent latent array; each row corresponds to the same latent index, and each column corresponds to the same feature channel.

[0094] The residual write-back connection bypasses the first read state and the temporary latent state, directly superimposing each residual component onto the same latent index and the same feature channel of the persistent latent array before the current version read; the residual components are not passed across latent indices, nor are they remixed between different feature channels.

[0095] The effective write-back position is obtained by comparing the absolute value of the residual component with the numerical determination precision. The numerical determination precision is the smallest change that the model data type can stably distinguish. In one embodiment, the numerical determination precision can be 0.000001.

[0096] When the absolute value of the residual is not greater than the numerical determination precision, the residual component is written to zero, and the original value of the corresponding position in the persistent latent array is retained; when the absolute value of the residual is greater than the numerical determination precision, the corresponding latent index and feature channel are recorded together as a valid write-back position, and the residual write-back parameter at that position is recorded synchronously.

[0097] When only some feature channels of the same latent index produce valid residuals, only the feature channels that produce valid residuals are updated, while the other feature channels retain their values ​​before being written back. When the residual write-back parameter is non-numerical or infinite, the abnormal component is set to zero and the original value of the corresponding position in the persistent latent array is retained.

[0098] The persistent potential array after the residual write-back is completed is saved as a snapshot of the current version and used as the source of the persistent potential array for the next version. The effective write-back location and residual write-back parameters are saved synchronously with the current version identifier and participate in the RIME optimization algorithm to determine the updatable dimensions.

[0099] The output query array is stored in a two-dimensional tensor. The array rows correspond one-to-one with the valid sequence positions of the current version archived input sequence, and the array columns have the same number of feature channels as the persistent latent array. Each query position is obtained by adding the corresponding spatial segment index code and version sequence code item by item in the same feature channel. No output query vector is generated when the position is filled.

[0100] The decoding cross-attention layer includes query projection, key projection, value projection, multi-head cross-attention, and output projection; the output query array generates query vectors, and the persistent latent array that completes residual write-back generates key and value vectors. Attention scores are calculated between each output query position and all latent indices.

[0101] The decoding cross-attention layer scales and normalizes the attention scores, and then aggregates the value vectors in the persistent latent array; the aggregation results of multiple attention heads are connected along the feature channels and the output projection is performed to form a decoding vector consistent with the number of effective query locations.

[0102] The decoding vector generates the explanatory attribute category probability, text or category content probability, and numerical content prediction value through the interpretation output mapping, and generates the evidence query vector through the evidence association mapping; the valid input positions in the archived input sequence generate the evidence key vector through the input position projection.

[0103] The evidence association mapping parameters and input position projection parameters are updated together through the evidence association loss. Positive association labels increase the similarity between the corresponding evidence query vector and the evidence key vector, while negative association labels decrease the similarity between the two. After training, the two sets of parameters are fixed.

[0104] The evidence query vector is multiplied and accumulated channel by channel with each evidence key vector. Each accumulated result is compressed to between 0 and 1 by a logic function to form independent evidence association probabilities. Normalization to sum to 1 is not performed between different input positions.

[0105] After training, the evidence association probabilities of positively associated positions in the validation set are arranged from smallest to largest. In one implementation, the probability of 25% of the positions is taken as the association judgment value. During actual decoding, a formal evidence association is established when the evidence association probability is not lower than the association judgment value and not lower than the average evidence association probability of all current interpretation items.

[0106] If no input location satisfies both criteria simultaneously, the input location with the highest evidence association probability is recorded as the association to be reviewed; if the maximum probability of text or category content is lower than 25% of the position value of the maximum probability of the correct category in the validation set, it is recorded as the explanation item to be reviewed, and the numerical content directly retains the numerical prediction value and the corresponding evidence association probability.

[0107] The effective write-back position and the decoding result are established through the latent index. First, the attention weight of each query position to each latent index in the decoding cross attention is read, and then the latent index with a weight not lower than the average attention weight of the current query position is selected.

[0108] Select the valid feature channels and residual write-back parameters of the records under the potential index and attach them to the corresponding exploration interpretation items and their evidence association results; if there is no valid write-back location corresponding to the current query location, retain the exploration interpretation items and evidence association results, and do not attach residual write-back parameters.

[0109] The exploration interpretation items, the association results between exploration interpretation items and exploration data fragments, the association results between exploration interpretation items and exploration results content, the associations to be reviewed, the effective write-back locations, and the corresponding residual write-back parameters together constitute the archived evidence association status, and the spatial segment index and version order are kept consistent with the archived input sequence.

[0110] Compared to existing Perceiver IO, this invention changes the state responsibilities of the latent array within the model and the update topology of the encoding end. The latent array is transformed from a temporary feature carrier in a single forward propagation into a persistent state that is continuously stored along the result version and can be addressed through latent indexes and feature channels. The original continuous path from encoding results directly to latent processing and decoding is split into a temporary computation path that does not cover the persistent state and an independent differential write-back path. The persistent latent array only receives features with the same index and channel added relative to the temporary latent state during the second read. Thus, cross-version processing is changed from regenerating the entire set of latent features to a location-based incremental memory update. Historical evidence associations are not completely covered by the current version, and changes in position and magnitude can directly enter the residual optimization, interpretation conflict handling, and version boundary tracing processes.

[0111] In this embodiment, S5 specifically includes: The RIME optimization algorithm sequentially performs candidate solution population initialization, soft frost search, hard frost puncture, and forward greedy selection. The optimization targets only the valid write-back positions of archived evidence associated state records and the corresponding residual write-back parameters. Potential features not recorded as valid write-back positions remain unchanged.

[0112] Candidate solutions are first arranged by potential index in ascending order, and then by feature channel in ascending order within the same potential index. Each dimension stores a residual write-back parameter corresponding to a valid write-back position. The length of a candidate solution is equal to the number of valid write-back positions in the current version.

[0113] The lower and upper bounds of the parameters for each candidate solution dimension are obtained from the model training samples. First, the residuals of the same latent index and feature channels in the correctly associated evidence samples are extracted and written back to the parameters. Then, the minimum and maximum values ​​are taken respectively, and the current version of the residuals is included in the boundary range.

[0114] When there are no training parameters for the same latent index and feature channel, the effective residuals of the same feature channel in other latent indices are read, and their median absolute value is used as the search half-width and superimposed on both sides of the current parameter; when the median absolute value is zero, the search half-width is taken as 10 times the accuracy of the effective write-back position determination.

[0115] The first candidate solution in the candidate solution group directly uses the current version residual write-back parameters, and the remaining candidate solutions are uniformly sampled between the lower bound and the upper bound of the parameters in each dimension; when the upper and lower bounds are the same, the boundary value is fixed. In one implementation, the number of candidate solutions can be 30.

[0116] Each candidate solution is written back to a copy of the persistent latent array. During the write-back, the corresponding latent index and feature channel are found in the order of the candidate solution dimensions and the residual write-back parameters are replaced. After the replacement is completed, the decoding cross-attention layer is called to regenerate the exploration interpretation terms and evidence association results.

[0117] Each valid output query location constitutes a fixed comparison item. The exploration interpretation item generated by the query location is associated with both the exploration data fragment and the exploration result content. When both carry the same spatial segment index, they are recorded as a consistent item; when the spatial segment indexes of the two are different, either association is missing, or either spatial segment index is missing, they are recorded as an inconsistent item.

[0118] The fitness of a candidate solution is calculated by subtracting the number of inconsistent items from the number of consistent items among all fixed comparison items. Valid output query positions that do not generate exploration and interpretation items are also recorded as inconsistent items, thus preventing candidate solutions from improving their fitness by reducing decoding results.

[0119] At the start of each search round, the candidate solution with the highest fitness is selected as the current optimal candidate solution. If the fitness is the same, the candidate solution ranked higher in the candidate solution group is retained. The parameters of each dimension of the current optimal candidate solution are used as the same dimension reference values ​​for soft frost search and hard frost puncture.

[0120] The soft-frost search first divides the current round by the total rounds to obtain the iteration progress, and then takes the square root of the iteration progress to obtain the parameter update trigger value. For each valid dimension of each candidate solution, a random number between 0 and 1 is generated. When the random number is less than the parameter update trigger value, the soft-frost search update is executed.

[0121] The random direction value for the soft frost search is obtained from -1 to 1. The direction change value is obtained by multiplying the current round by pi and then dividing by 10 times the total number of rounds, and then calculating the cosine value of the direction change value. The environmental change value is obtained by subtracting the current round's proportion of the total number of rounds from 1, rounding it to the nearest integer, and then dividing it by 5.

[0122] For each dimension to be updated, a stickiness value between 0 and 1 is generated. First, the stickiness value is multiplied by the difference between the upper and lower bounds of the parameters and then superimposed with the lower bound of the parameters. Then, it is multiplied by the random direction value, cosine value, and environmental change value in turn. The resulting update amount is superimposed on the same dimension parameter of the current optimal candidate solution.

[0123] When the new parameters generated by the soft-frost search are lower than the lower bound of the parameters, the lower bound of the parameters is used; when they are higher than the upper bound of the parameters, the upper bound of the parameters is used. Dimensions that have not been updated retain the original parameters of the current candidate solution. After processing all valid dimensions, the soft-frost search results are obtained.

[0124] Before hard frost puncture, the fitness of all candidate solutions in this round is mapped to the interval between 0 and 1. Each fitness is first subtracted from the minimum fitness of this round, and then divided by the difference between the maximum fitness and the minimum fitness of this round; when the maximum fitness and the minimum fitness are the same, all mapped values ​​are 0.

[0125] Hard frost puncture processes soft frost search results dimension by dimension in a predetermined order of potential index and feature channel. A random number between 0 and 1 is generated for each dimension. When the random number is lower than the fitness mapping value of the current candidate solution, the soft frost search parameters are replaced with the same dimension parameters of the current best candidate solution.

[0126] When the random number is not lower than the fitness mapping value, the soft frost search results are retained. The hard frost puncture does not traverse potential indexes and feature channels outside the effective write-back position, thereby avoiding the candidate solution update from changing the persistent potential features that have not undergone residual write-back.

[0127] Candidate solutions for soft frost search and hard frost puncture are written back to the persistent potential array copy and re-decoded. The fitness is then recalculated using the same consistent and inconsistent term statistics method to obtain the fitness pair before and after each candidate solution update.

[0128] The forward greedy selection compares the fitness of the same candidate solution before and after the update one by one. If the fitness after the update is greater than the fitness before the update, the original candidate solution is replaced by the updated candidate solution. If the fitness after the update is not greater than the fitness before the update, the original candidate solution is retained.

[0129] After the candidate solution is replaced, the candidate solution with the highest fitness is re-determined, and the next round of soft frost search and hard frost puncture is entered; in one embodiment, the total number of rounds can be 60. The search ends when the maximum fitness value has not increased for 8 consecutive rounds. If this state is not reached, the process continues until the end of the total number of rounds.

[0130] After the search is completed, the candidate solution with the highest fitness is read, and the parameters of each dimension are written back to the corresponding valid write-back positions in sequence. The potential indexes and feature channels not included in the candidate solutions continue to retain their values ​​before optimization. When none of the candidate solutions exceed the fitness of the initial candidate solutions, the initial residual write-back parameters are retained.

[0131] The persistent potential array that has completed parameter write-back re-enters the decoding cross-attention layer, regenerates the exploration interpretation terms and their association results with exploration data fragments and exploration results content, and together with the optimized effective write-back positions and residual write-back parameters, forms the optimized evidence association results.

[0132] In this implementation, the RIME optimization algorithm no longer performs a unified search on all feature channels of the persistent potential array. Instead, it establishes candidate solution dimensions along the effective write-back positions, allowing the candidate solution length to adaptively shrink as the residual changes actually occur in the current version. This improvement limits the search range to the internal state positions corresponding to the newly added information in the version, which reduces the historical association disturbances caused by irrelevant dimensions participating in soft frost search and hard frost puncture, reduces the computational load of candidate solution evaluation and repeated decoding, and enables the optimized residual parameters to directly correspond to the potential indexes and feature channels that cause changes in evidence association.

[0133] In this embodiment, S6 specifically includes: The optimized evidence association results use the spatial segment index as the retrieval key to read the exploration interpretation items and their associated exploration data fragments and exploration results within the same spatial segment; each association retains the input location, version identifier and normalized association probability at the same time, and the associations to be reviewed are saved separately and do not enter the automatic merging process.

[0134] An exploration interpretation item includes an interpretation attribute and an interpretation content. The interpretation attribute is read from the corresponding unified field name and model output category. When the unified field name and the model output category are the same, they are considered to be the same interpretation attribute. When a single exploration interpretation contains multiple interpretation attributes, it is split into multiple exploration interpretation items. Each split item retains the original interpretation content location and evidence association results.

[0135] The textual explanations are first standardized in terms of character encoding and full-width / half-width characters. Leading and trailing spaces are removed, and consecutive spaces are merged into single spaces. Then, synonymous geotechnical terms are replaced with standardized names through a standardized glossary. If the processed texts are completely identical, the explanations are considered identical; if the processed texts are not completely identical, the explanations are considered conflicting.

[0136] The explanatory content in category form is first converted into the category number used by the training label. If the category numbers corresponding to the same explanatory attribute are the same, the explanatory content is considered to be consistent. If the category numbers are different, the explanatory content is considered to be conflicting. If the category name cannot be converted into the category number, the original category name and source information are retained, and a separate branch for the explanatory content to be reviewed is created.

[0137] The numerical interpretations are first converted to a unified unit of measurement, and then the number of decimal places retained in each record is read. The smallest unit of measurement corresponding to the record with lower precision is taken as the allowable difference. If the absolute difference between two values ​​is not greater than the allowable difference, the interpretations are considered to be consistent. If the absolute difference is greater than the allowable difference, the interpretations are considered to be conflicting.

[0138] When the interpretation includes a starting boundary and a ending boundary, the two boundaries are compared separately. If both boundaries meet the consistency judgment, the interval interpretation is considered consistent. If either boundary exceeds the allowable difference, the interval interpretation is considered conflicting. If either boundary is missing, the inferred value is not added, and the original exploration interpretation item, together with the existing evidence, is saved separately as an interpretation branch to be reviewed.

[0139] Exploration interpretation items with identical interpretation attributes and interpretation content are merged into the same interpretation branch. When merging, the interpretation content with the earlier version number is retained as the branch content, and the exploration data fragments and exploration results associated with each exploration interpretation item are collected. If the record identifiers of the associated objects are the same, they are retained only once, and their source information and version identifiers are completely appended to the same associated record.

[0140] When the same interpretation attribute corresponds to different interpretation contents, interpretation branches are established for each interpretation content. Each interpretation branch retains the original exploration interpretation item, spatial segment index, version identifier and its associated exploration data fragments and exploration results content. Conflict separation only changes the attribution of the interpretation item, without copying or rewriting the original exploration data and results text.

[0141] The evidence pointer points to the exploration data fragment or exploration result content that is consistent with the content of the interpretation branch, and saves the target record identifier, spatial segment index, version identifier, content location, source file check value and original location; when the text, category or numerical content meets the corresponding consistency judgment, the pointer is written into the supporting evidence set of the interpretation branch.

[0142] The exclusion evidence pointer uses the same location content and points to the exploration data fragment or exploration result content that conflicts with the content of the current interpretation branch under the same interpretation attribute; when the associated object supports other interpretation branches and conflicts with the content of the current interpretation branch, the pointer is written to the exclusion evidence set of the current interpretation branch.

[0143] When the same associated object can support one interpretation branch and exclude another interpretation branch, two pointers are created and bound to the corresponding interpretation branches respectively; when the target record cannot be located or the version identifier is missing, the original association result is retained and written into the evidence set to be reviewed, and no formal supporting evidence pointer or excluding evidence pointer is created.

[0144] After completing the merging of consistent interpretations, the separation of conflicting interpretations, and the binding of evidence pointers, each interpretation branch saves interpretation attributes, interpretation content, spatial segment index, version identifier, supporting evidence pointers, and excluding evidence pointers, forming interpretation branches that can distinguish between consistent interpretations and conflicting interpretations within the same segment.

[0145] In this embodiment, S7 specifically includes: The official exploration results version directly adopts the version order already determined in the archived input sequence, and is processed version by version starting from the earliest version; the official exploration results version and the interpretation branch correspond through spatial segment index and interpretation attribute name. When both have category numbers, the category numbers must also be consistent. When no interpretation branch with the same interpretation attribute is found, a new interpretation branch is established based on the content of the official exploration results and the corresponding evidence pointer is bound.

[0146] After completing the correspondence between the version and the interpretation branch, the interpretation content is read version by version starting from the earliest official exploration result version. The text content is compared after the terminology is standardized, the category content is compared by category number, the numerical content is compared by the difference after being converted to a unified unit of measurement, and the interval content is compared by the starting boundary and the ending boundary respectively.

[0147] When the content of the formal exploration results is consistent with the content of the interpretation branch, the interpretation branch continues to cover the current version of the formal exploration results, and the current version identifier is appended to the valid version record of the interpretation branch; existing supporting evidence pointers and excluding evidence pointers remain unchanged.

[0148] When there is a conflict between the content of the formal survey results and the content of the interpretation branch, the spatial segment index, interpretation attributes, pre-conflict interpretation content and evidence pointer of the current interpretation branch are copied, but the valid version record of the original interpretation branch is not copied. The original interpretation branch continues to save the pre-conflict content and the existing version range.

[0149] The copied interpretation branches only replace the text, categories, values, or interval boundaries that are explicitly modified in the formal survey results. Then, the interpretation content consistency comparison method is used to re-examine all evidence pointers. Pointers consistent with the replaced interpretation content are written into the supporting evidence set, while pointers conflicting with the replaced interpretation content are written into the exclusion evidence set.

[0150] Establish an evidence pointer pointing to the current formal survey results. The pointer stores the result content identifier, spatial segment index, version identifier, content location, source file check value, and original location, and writes it into the supporting evidence set of the copy branch. When the content location or source information cannot be obtained, the copy branch is transferred to the pending review state without defining the formal version boundary.

[0151] Evidence pointers added after copying and pointing to the current formal survey results are identified as evidence pointers that cause changes in interpretation when the content they point to is consistent with the interpretation content after replacement and conflicts with the interpretation content before copying; if either judgment is not met, the newly added pointer is retained and the process is put into a pending review state.

[0152] Version boundary records are set at the version of the formal survey results that caused the interpretation change. The boundary record stores the spatial segment index, interpretation attribute, the version identifier before the boundary, the boundary version identifier, and the evidence pointer that caused the interpretation change. When the same version of the formal survey results causes multiple interpretation attribute changes, version boundaries for the corresponding interpretation attributes are established separately.

[0153] The effective version endpoint of the interpretation branch before the boundary is written as the version of the formal survey result before the boundary, and the effective version starting point of the copied interpretation branch is written as the boundary version; the two interpretation branches are connected to each other through forward pointers and reverse pointers. The forward pointer points from the branch before the boundary to the branch after the boundary, and the reverse pointer points from the branch after the boundary back to the branch before the boundary.

[0154] When the same interpretation branch changes continuously in multiple official exploration result versions, each change copies a new branch from the currently valid interpretation branch, and uses the branch after the boundary of the previous version as the starting point of the connection. After all official exploration result versions have been processed, the interpretation branch, the start and end positions of the valid version, the version boundary, and the evidence pointers that caused the interpretation change together form the interpretation branch version relationship.

[0155] In this embodiment, S8 specifically includes: The interpretation branch version relationship is sorted by spatial segment index as the first sorting item, interpretation attribute as the second sorting item, and effective version start point as the third sorting item. Interpretation branches under the same spatial segment and the same interpretation attribute are arranged from earliest to latest. When the version start points are the same, the interpretation branch with the earlier version end point is arranged first.

[0156] The connection between adjacent interpretation branches reads the forward and reverse pointers from the version relationship. The connection is considered valid when the version end point of the previous interpretation branch and the version start point of the next interpretation branch are connected at the same version boundary. If the two connections point to different versions, the original interpretation branch is retained and written to the connection record to be reviewed, and the version relationship is not automatically modified.

[0157] The structured archive records use a single interpretation branch as the basic record object. Each record stores the spatial segment index, interpretation attributes, interpretation content, effective version start point, effective version end point, forward pointer, reverse pointer, supporting evidence pointer, exclusion evidence pointer, and version boundary identifier. The content of the same interpretation branch is stored centrally, and the original exploration data fragments and exploration results text are not repeatedly written into the archive records.

[0158] Version boundary records connect the interpretation branches before and after the boundary and save the evidence pointer that caused the interpretation change; text content compares the word sequence after unified terminology processing and records the different words; category content records the category number before and after the boundary; numerical or interval content records the changed numerical value, original value and replacement value; the reason for the interpretation change is composed of the difference record and the content pointed to by the evidence pointer.

[0159] Both support evidence pointers and exclusion evidence pointers use the same parsing process: first read the target record identifier, then read the spatial segment index, version identifier, and content location; the target record identifier locates the survey data fragment or survey result content, and the content location locates the fields, paragraphs, table cells, or map annotation areas in the target record.

[0160] When the evidence pointer points to a fragment of exploration data, the corresponding fragment is first retrieved using the spatial segment index, then the standardized exploration data is located using the target record identifier, and finally the original field content and source information are read using the content location. When the evidence pointer points to the content of the exploration results, the result version is first located using the version identifier, and then the corresponding result content is read using the target record identifier and content location.

[0161] When the target record identifier is not matched, the source data is located using the source file checksum stored in the evidence pointer, and then the target content is read from the original location. When the original location changes, the same field name and field content are first searched in the same page, the same worksheet, or the same paragraph area. If the same field name and field content are still not matched, the entire source file is searched. If the target cannot be located in the end, the evidence pointer is retained and written to the invalid pointer record.

[0162] The retrieval index consists of spatial segment index entries, explanatory attribute index entries, version boundary index entries, and evidence pointer index entries. Spatial segment index entries point to all explanatory branches within the same segment, explanatory attribute index entries point to the version relationship of explanatory branches with the same explanatory attribute, version boundary index entries point to the explanatory branches before and after the boundary, and evidence pointer index entries point to the original evidence content.

[0163] When writing the index, the spatial segment index is registered first, followed by the interpretation attributes and the effective version range. Then, the record positions of the version boundary and the interpretation branches before and after the boundary are written. Finally, the supporting evidence pointer, the excluding evidence pointer, and the evidence pointer that causes the interpretation change are written. Each index entry also saves the record identifier of the structured archive record to ensure that the index result can return the corresponding interpretation branch.

[0164] When querying a location containing spatial information, the query location is first converted to the engineering space and the spatial segment index is retrieved. Then, the interpretation branches within the segment that match the interpretation attributes and version range are read. When querying a location that does not contain interpretation attributes, all interpretation attributes within the same segment are returned. When querying a location that does not contain version information, the currently valid interpretation branches and their version change chains are returned.

[0165] When querying the reasons for the explanation changes, the version boundary index item is retrieved and the explanation content before and after the boundary is read. Then, the evidence pointer that caused the explanation change is parsed to locate the corresponding exploration data fragment or exploration result content. When querying the original basis, the supporting evidence pointer and the excluding evidence pointer are parsed respectively, and the content pointed to by the pointer and the location of its source file are returned.

[0166] The backtracking path from the structured archive record to the original evidence sequentially passes through the interpretation branch record, version boundary record, evidence pointer index item, target record identifier, and original content location; if any intermediate object cannot be hit, the automatic backtracking stops and returns the last successfully located archive object and the missed location to avoid incorrect connections to other spatial segments or result versions.

[0167] After the structured archive records and retrieval indexes are written, forward retrieval and reverse backtracking are performed on each interpretation branch. If the forward retrieval fails to locate the interpretation branch or the reverse backtracking fails to locate the original content, the corresponding archive record is written to the set to be reviewed. The remaining structured archive records and retrieval indexes that pass the two checks together form the geotechnical engineering investigation data archive results.

[0168] Example 1: To verify the feasibility of this invention in practice, geotechnical engineering investigation data of a certain building foundation project were archived. This project includes field logging data from 56 boreholes, in-situ test data, laboratory test reports, stratigraphic profiles, and multiple versions of investigation results. Different data were generated by field recording software, test data systems, and results compilation software, respectively. Different recording formats existed for the same borehole number, stratigraphic name, and depth range. During the revision process, stratigraphic boundaries were adjusted, soil and rock categories were modified, and test conclusions were supplemented. The existing file classification and archiving could only preserve each version of the results, making it impossible to directly determine the specific locations where the interpretation content changed, and it was also difficult to quickly find the original investigation data supporting the interpretation conclusions before and after the modifications.

[0169] During archiving, various types of exploration data are first parsed, synonymous fields are matched, and coordinate datum, depth units, time formats, and text encoding are standardized. Then, the data is mapped to the engineering space based on borehole location and recorded depth. Data with continuous depth ranges are grouped into the same spatial segment, while data is segmented where borehole location changes or depth ranges are broken. Exploration data fragments and corresponding results are encoded according to spatial segment and result version order and input into the improved Perceiver IO model. The content obtained from the first cross-attention read is processed in a provisional latent state. The residuals generated from the second read relative to the provisional latent state are written back to the persistent latent array. The RIME optimization algorithm only adjusts the residuals corresponding to the write-back locations. After model decoding, exploration interpretation items with consistent interpretation content within the same spatial segment are merged. Interpretation items with conflicting stratigraphic categories, stratigraphic interfaces, or engineering property evaluations are established as interpretation branches, and the corresponding exploration records and results are bound as supporting evidence pointers or excluding evidence pointers.

[0170] To ensure the verification data has a clear source, 320 units related to the survey data and results were selected from the archived materials as verification samples. After manual review, it was confirmed that 84 of these units exhibited cross-version interpretation changes, while the interpretations of the remaining 236 units remained consistent. Two geotechnical engineers respectively verified the survey interpretations, version boundaries, and original evidence; when the verification results were inconsistent, a third engineer confirmed the discrepancies. Using the same verification sample, the manual document classification and archiving method, the relational database index archiving method, and the method of this invention were compared. The statistical results are shown in Table 1 below.

[0171] Table 1 Comparison of Archival Evidence Correlation and Version Tracing Results in Geotechnical Engineering Investigation

[0172] As can be seen from the data in Table 1 above, the method of the present invention outperforms both the manual document classification and archiving method and the relational database index archiving method in five indicators: evidence association accuracy, conflict interpretation identification rate, version boundary location accuracy, original evidence backtracking completeness rate, and average time per backtracking. The manual document classification and archiving method relies on searching each document and verifying the content of the results. Its evidence association accuracy is 86.3%, conflict interpretation identification rate is 72.6%, and version boundary location accuracy is 69.0%, indicating that it is prone to missing conflict interpretations and misjudging changed versions when faced with multi-source data and continuously revised results. The relational database index archiving method improves query efficiency through field retrieval, achieving an evidence association accuracy of 91.9% and an original evidence backtracking completeness rate of 86.9%. However, its handling of cross-version interpretation relationships is still limited by fixed fields and static indexes. The method of this invention achieves an evidence association accuracy of 97.8%, a conflict interpretation identification rate of 96.4%, a version boundary positioning accuracy of 94.0%, and an original evidence backtracking completeness rate of 97.6%. The average time for a single backtracking is reduced to 8.7 seconds, indicating that it can accurately establish the correspondence between survey data and results, preserve the interpretation change link, and reduce the time for manual review.

[0173] This embodiment first performs field unification, spatial benchmark conversion, and continuous spatial range segmentation on geotechnical engineering investigation data from different sources, organizing borehole records, in-situ test results, laboratory test data, and investigation results according to spatial segments and version order. The improved Perceiver IO model retains established evidence associations from the previous version using a persistent latent array, only writing back the residuals formed by the second encoding cross-attention relative to the temporary latent state to the corresponding positions, avoiding new data overwriting existing version information. The RIME optimization algorithm further limits the write-back positions of records to the updatable dimensions of the hard frost puncture mechanism, focusing parameter adjustments on the actual changing latent features. After model decoding, consistent interpretation terms are merged, conflicting interpretation terms are separated into different interpretation branches, and supporting evidence pointers and excluding evidence pointers are bound. When the formal investigation results are revised, version boundaries are defined along the evidence pointers that caused the interpretation change, and the interpretation content before and after the change, the original investigation record, and the result version are connected. Thus, this embodiment can complete spatial correspondence, evidence association, conflict identification, version recording, and original evidence backtracking, solving the problems of result coverage, evidence chain breaks, and difficulty in locating the causes of changes in traditional archiving.

[0174] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for archiving geotechnical engineering investigation data based on deep learning, characterized in that, Includes the following steps: S1. Collect geotechnical engineering investigation data, and perform normalization processing on data from different sources to form standardized investigation data; S2. Map the standardized survey data to the engineering space, divide it into continuous spatial ranges and write it into the spatial segment index to form survey data fragments; S3. Organize the exploration data segments and exploration results according to the spatial segment index, retain the version order and encode them as the archived input sequence; S4. The archived input sequence is connected to the improved Perceiver IO model. The persistent latent array, which is retained in version order, is read through cross-attention after two encodings. The residual formed by the relative temporary latent state after the second reading is written back to the persistent latent array and the write-back position is recorded. The archived evidence association state is then decoded. The S5 and RIME optimization algorithms use the recorded write-back positions as the updatable dimension of the hard frost puncture mechanism, optimize the corresponding residuals and re-decode them to form optimized evidence association results. S6. Based on the results of optimized evidence association, merge exploration interpretation items with consistent attributes within the same spatial segment and separate conflicting items to form interpretation branches that bind supporting or excluding evidence pointers. S7. Correct the interpretation branches with the official survey results, divide the version boundaries along the evidence pointers that caused the interpretation changes, and form the interpretation branch version relationship; S8. Establish structured archive records according to the interpretation of branch version relationships, connect evidence pointers to the retrieval index, and form geotechnical engineering investigation data archive results that can trace back the original evidence and the reasons for changes.

2. The method for archiving geotechnical engineering investigation data based on deep learning according to claim 1, characterized in that, S1 specifically includes: S11. Analyze the data format of geotechnical engineering investigation data, and extract data fields, field content, and source information; S12. Match synonymous fields from different sources based on the meaning of the data fields, map the matched fields to a unified field, and retain the original field name and source information for unmatched fields. S13. Perform conversion on the unit of measurement, time format and text encoding in the unified field and the unmatched field. For fields that cannot be converted, retain the original value and source information. S14. Integrate the converted field content according to the unified field name and the original field name, merge duplicate records with the same content, and retain records with conflicting content and source information in parallel to form standardized survey data.

3. The method for archiving geotechnical engineering investigation data based on deep learning according to claim 1, characterized in that, S2 specifically includes: S21. Read the spatial coordinates and depth range of standardized survey data, and convert the location data using different spatial benchmarks to the same engineering space; S22. Arrange the standardized survey data according to the converted location data, compare the spatial coverage boundaries of adjacent data, classify data with adjacent or overlapping boundaries into the same continuous spatial range, and cut at the boundary separation point to form a continuous spatial range. S23. Write the start and end boundaries and arrangement order of the continuous spatial range into the spatial segment index, and associate the spatial segment index with the standardized exploration data within the range to form exploration data fragments.

4. The method for archiving geotechnical engineering investigation data based on deep learning according to claim 1, characterized in that, S3 specifically includes: S31. Read the spatial range recorded in the survey results, retrieve the spatial segment index, and associate the survey results that intersect with the spatial segment with the corresponding survey data segment. S32. After parsing the version identifiers of the associated exploration results, arrange the exploration results within the same spatial segment according to the order of the version identifier records to form a version sequence. S33. Encode the exploration data fragments and associated exploration results into a feature vector of a unified dimension, and superimpose the spatial segment index code and version sequence code to form an archived input sequence.

5. The method for archiving geotechnical engineering investigation data based on deep learning according to claim 1, characterized in that, The improved Perceiver IO model includes a persistent latent array, a first encoding cross-attention layer, a latent processing module, a second encoding cross-attention layer, a residual write-back connection, an output query array, and a decoding cross-attention layer. The latent processing module includes a latent self-attention layer and a multilayer perceptron.

6. The method for archiving geotechnical engineering investigation data based on deep learning according to claim 1, characterized in that, The formation of the archived evidence association status specifically includes: The archived input sequence is read in version order. The first version initializes the persistent latent array with the latent vectors obtained from model training, and the remaining versions read the persistent latent arrays from the previous version after the residuals have been written back. The persistent latent array is copied, and the first encoding cross-attention layer projects the copied latent vectors into query vectors, and projects the current version of the archived input sequence into key vectors and value vectors; Calculate the attention weight of each potential index for each sequence position and aggregate the value vector. After projecting the aggregated result onto the output, add it to the copied potential vector to form the first read state. The continuous potential array keeps the potential vector before the read. The latent self-attention layer projects the first read state into a query vector, a key vector, and a value vector, aggregates latent features along the latent index, and adds them to the first read state; The multilayer perceptron performs linear transformation, nonlinear activation, and linear recovery along the feature channels, and adds the recovery result to the output of the latent self-attention layer to form a temporary latent state; The second encoding cross-attention layer projects the temporary latent state into a query vector and reprojects the current version of the archived input sequence into a key vector and a value vector, while masking the padding positions caused by length padding of the archived input sequence. Aggregate the value vector along the unmasked sequence position, project the aggregation result onto the output, and add it to the temporary latent state to form the second read state; Based on the same potential index and feature channel, the temporary potential state is subtracted from the second read state to form the residual write-back parameters; The residual write-back connection bypasses the first read state and the temporary latent state, adds the residual write-back parameters to the persistent latent array before the read, keeps the feature channels that do not generate residual components unchanged, and determines the latent index and feature channel that generate residual components as the valid write-back position; The spatial segment index encoding and version sequence encoding are arranged in the order of the archived input sequence to form the output query array. The decoding cross-attention layer uses the output query array as the query and the persistent potential array that has completed the residual write-back as the key and value to perform attention reading. Record the exploration interpretation items obtained from decoding and their association results with exploration data fragments and exploration results, and attach the valid write-back location and corresponding residual write-back parameters to form the associated status of archived evidence.

7. The method for archiving geotechnical engineering investigation data based on deep learning according to claim 1, characterized in that, S5 specifically includes: S51. Arrange the residual write-back parameters corresponding to the effective write-back positions in the order of potential index and feature channel, and use the arranged parameters as the candidate solution dimension to form a candidate solution group; S52. Write each candidate solution back to the persistent potential array and decode it again. Check whether the interrelated exploration data fragments and exploration results in the decoding results correspond to the same spatial segment. Determine the fitness of the candidate solution by subtracting the number of different spatial segments from the number of identical spatial segments. S53. The soft-frost search strategy uses the same-dimensional parameters of the candidate solution with the highest fitness as a benchmark, and combines the iteration progress and random direction to update the dimension corresponding to the effective write-back position. S54. Map the fitness of candidate solutions to a uniform numerical range. The hard frost puncture mechanism generates random numbers dimension by dimension. When the random number is lower than the mapped value, it is replaced with the same dimension parameter of the candidate solution with the highest fitness. The search results of soft frost are retained for the missing dimension. S55. Re-decode the updated candidate solution and calculate the fitness. The positive greedy selection mechanism replaces the original candidate solution when the fitness after the update is greater than the fitness before the update. Write the candidate solution with the maximum fitness back to the valid write-back position to form the optimized evidence association result.

8. The method for archiving geotechnical engineering investigation data based on deep learning according to claim 1, characterized in that, S6 specifically includes: S61. Read the exploration interpretation items in the optimized evidence association results according to the spatial segment index, as well as the exploration data fragments and exploration results associated with each exploration interpretation item; S62. Compare the interpretation content of the same interpretation attribute in the same spatial segment item by item, and merge the exploration interpretation items with consistent interpretation content and their associated exploration data fragments and exploration results. S63. The merged exploration interpretation items are identified as the same interpretation branch. Exploration interpretation items with conflicting interpretation content are established as interpretation branches according to different interpretation contents, and their associated exploration data fragments and exploration results are retained. S64. Establish supporting evidence pointers for exploration data fragments and exploration results associated with the same interpretation content as the interpretation branch, establish exclusion evidence pointers for exploration data fragments and exploration results associated with conflicting interpretation content, and bind them to the interpretation branch.

9. The method for archiving geotechnical engineering investigation data based on deep learning according to claim 1, characterized in that, Specifically, S7 includes: S71. According to the spatial segment index and interpretation attributes, match the official exploration results version with the interpretation branch item by item, and arrange them in the order of version. S72. Compare the content of the formal survey results with the interpretation content of the interpretation branch version by version. If the content is consistent, continue the interpretation branch. If the content is conflicting, copy the interpretation branch, replace the conflicting interpretation content with the content of the formal survey results, and bind the evidence pointer pointing to the corresponding survey results content. S73. Evidence pointers that are newly added after copying and point to the content of the formal survey results shall be identified as evidence pointers that cause changes in interpretation. S74. Divide the version boundaries by the version of the survey results indicated by the evidence pointer that causes the change in interpretation, connect the interpretation branches before and after the boundary, and form the interpretation branch version relationship.

10. A method for archiving geotechnical engineering investigation data based on deep learning according to claim 1, characterized in that, S8 specifically includes: S81. Arrange the interpretation branches according to the spatial segment index and version order, and connect the interpretation branches of adjacent versions according to the version boundary; S82. Write the interpretation branches, version boundaries, and interpretation content before and after the boundaries into the structured archive record, and associate the evidence pointers that cause interpretation changes with the corresponding version boundaries; S83. Analyze supporting evidence pointers and excluding evidence pointers, and locate the exploration data fragments and exploration results content pointed to by the pointers; S84. Incorporate spatial segment indexes, version boundaries, and evidence pointers into the retrieval index, and connect the structured archive records with the content pointed to by the pointers to form the archived results of geotechnical engineering investigation data.