A large model dynamic optimization method for survey report generation task

Abstract

This invention belongs to the field of exploration report generation technology and provides a dynamic optimization method for large-scale models in exploration report generation tasks, addressing the problem that expert feedback models cannot accurately interpret conclusion sections. This invention is the first to propose a deep optimization system of "conclusion section specialization" to achieve precise risk avoidance. For the "Conclusions and Recommendations" section, which is the highest-risk and most logically complex section in exploration reports, this invention does not adopt a general full-text generation model. Instead, it uses a specialized conclusion quality reward model and rule constraints to focus on solving the "illusion" problem in core aspects such as foundation evaluation and parameter recommendations. This significantly reduces the risk of geological disasters caused by random model output.

Description

Technical Field

[0001] This invention belongs to the field of exploration report generation technology, specifically relating to a dynamic optimization method for large models for exploration report generation tasks. Background Technology

[0002] With the rapid development of artificial intelligence and natural language processing technologies, Large Language Models (LLM) have begun to be applied in the field of engineering surveying. As a legally binding technical document, the accuracy of the generated engineering survey report directly affects the safety of the project. However, current intelligent generation systems for survey reports face the following technical challenges:

[0003] 1. "Black box" decision-making and risk coupling in the conclusion section: The conclusions and recommendations in the survey report (such as foundation stability assessment and foundation type recommendations) rely heavily on expert experience. Existing intelligent report generation systems lack a transparent logical link in their end-to-end generation method. The model may produce "illusions" at key nodes, randomly providing core parameters (such as bearing capacity), creating significant safety hazards.

[0004] 2. "Long-term dependence" and error accumulation in the processing flow: Survey reports are generally long and highly specialized, and the overall generation is prone to missing instructions or inconsistencies in contextual logic; while simple paragraph splicing leads to data isolation between chapters, lacking process monitoring and real-time quality intervention.

[0005] 3. The "gap" in the accumulation of expert knowledge and the shortage of training samples: Geological engineers' modification annotations are mostly scattered and unstructured text (such as "suggest a 20% reduction"). Existing intelligent report generation systems have difficulty converting their implicit judgment logic into adjustment instructions that the model can understand, resulting in the loss of high-value experience.

[0006] 4. The illusion of concealment caused by the lack of domain rule constraints: Although the reports generated by the model are consistent with the language logic, they often violate common sense of physics or current mandatory industry standards (such as GB 50021). Such errors are highly concealed and difficult to identify quickly through automated means. Summary of the Invention

[0007] To address the problem that expert feedback models cannot accurately understand conclusion sections, this invention provides a dynamic optimization method for large-scale models in the task of survey report generation.

[0008] To solve the technical problem, the technical solution adopted by this invention is as follows:

[0009] A dynamic optimization method for a large model for survey report generation tasks includes:

[0010] (1) The survey report is generated by the conclusion generation module;

[0011] (2) Experts provide feedback on the conclusions of the survey report, and the feedback information is collected by the feedback collection module.

[0012] (3) The conclusion feedback structuring and parsing module parses the collected expert feedback information and applies rules to constrain it;

[0013] (4) The conclusion quality reward model applies rewards and reinforcement learning to the conclusion section after the analysis and rule modules are constrained to obtain the final survey report.

[0014] (5) The conclusion reinforcement learning fine-tuning module fine-tunes the model of the conclusion generation module under the guidance of the conclusion quality reward model, so as to improve the ability of geological risk identification, stratigraphic logic consistency and standardized wording, and thus obtain the final exploration report.

[0015] In some embodiments, a dynamic iteration and conclusion flywheel module is also included. After obtaining the final survey report, the dynamic iteration and conclusion flywheel module calculates the uncertainty, logical consistency and normative risk of the conclusion section. When a high-risk conclusion is detected, a second round of expert review is automatically triggered, and the process proceeds to steps (2) to (5) in sequence.

[0016] In some embodiments, the conclusion feedback structuring and parsing module automatically converts expert feedback information into a structural pattern of problem type, correction suggestions, and supporting data / standards based on knowledge triple parsing.

[0017] In some embodiments, the conclusion feedback structuring and analysis module introduces geological engineering-specific stratigraphic determination rules and site categories to determine rules and reasonable parameter ranges in order to achieve rule module constraints.

[0018] In some embodiments, the conclusion quality reward model includes a reward function that includes expert weights (experience values), error correction matching degree, specification consistency score, and data reference accuracy.

[0019] Compared with the prior art, the present invention has the following beneficial effects:

[0020] This invention presents a dynamic optimization method for large-scale models in exploration report generation, which for the first time proposes a deep optimization system of "conclusion section specialization" to achieve precise risk avoidance. For the "Conclusions and Recommendations" section, which is the highest-risk and most logically complex part of the exploration report, this invention does not adopt a general full-text generation model. Instead, it uses a specialized conclusion quality reward model (RM) and rule constraints to address the "illusion" problem in core aspects such as foundation evaluation and parameter recommendations. This significantly reduces the risk of geological disasters (such as incorrect subsidence and landslide judgments) caused by random model output.

[0021] This invention successfully transforms the scattered, unstructured annotations of geological engineers into standardized data that can be used for model fine-tuning through innovative knowledge triplet parsing technology. This mechanism solves the technical challenge of reusing and transferring high-value expert experience in algorithms, enabling large models to truly learn the deep interpretation logic of experts.

[0022] This invention constructs a unique "Conclusion Flywheel" dynamic evolution mechanism through dynamic iteration and a conclusion flywheel module, continuously improving model performance and establishing a closed-loop iterative system of "generation → annotation → structuring → reinforcement training → further improvement." As the system's scale of operation in enterprises expands and expert participation increases, the richer the cases absorbed by the model, the stronger its comprehensive judgment capability for complex geological scenarios becomes, forming a positive cycle where the model becomes more accurate and professional with use. By calculating uncertainties in real time and triggering a second round of expert review, it ensures that the output of high-risk projects is always under the dual monitoring of human experts and AI algorithms.

[0023] This invention, through a pre-defined hard constraint rule base and reward model, ensures that the generated conclusions strictly comply with current national and industry technical specifications (such as GB 50021). In practical applications, this significantly shortens the cycle of manual review and repeated revisions, reduces review costs, and enhances the professionalism and market competitiveness of surveying units in the field of intelligent report generation. Attached Figure Description

[0024] Figure 1 This is a system framework diagram for the large-scale dynamic optimization of the present invention;

[0025] Figure 2 This is a flowchart of the dynamic iteration and conclusion flywheel module of the present invention;

[0026] Figure 3 A flowchart illustrating the process of generating an exploration report for the conclusion generation module of this invention;

[0027] Figure 4 The flowchart illustrates the single-step (i.e., single atomic step) inner loop sampling, filtering, and rollback of the conclusion generation module of this invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The described embodiments are some, but not all, of the embodiments of the present invention. The following embodiments are only for specific illustration of the implementation methods of the present invention and do not limit the scope of protection of the present invention.

[0029] Combined with appendix Figure 1 and attached Figure 2The present invention provides a large-scale model dynamic optimization method for survey report generation tasks, comprising:

[0030] (1) The survey report is generated by the conclusion generation module;

[0031] (2) Experts provide feedback on the conclusions of the survey report, and the feedback information is collected by the feedback collection module.

[0032] (3) The conclusion feedback structuring and parsing module parses the collected expert feedback information and applies rules to constrain it;

[0033] (4) The conclusion quality reward model applies rewards and reinforcement learning to the conclusion segment after the parsing and rule modules are constrained;

[0034] (5) The conclusion reinforcement learning fine-tuning module fine-tunes the model of the conclusion generation module under the guidance of the conclusion quality reward model, so as to improve the ability of geological risk identification, stratigraphic logic consistency and standardized wording, and thus obtain the final exploration report.

[0035] In some embodiments, a dynamic iteration and conclusion flywheel module is also included. After obtaining the final survey report, the dynamic iteration and conclusion flywheel module calculates the uncertainty, logical consistency and normative risk of the conclusion section. When a high-risk conclusion is detected, a second round of expert review is automatically triggered, and the process proceeds to steps (2) to (4) in sequence.

[0036] In some embodiments, the conclusion feedback structuring and parsing module automatically converts expert feedback information into a structural pattern of problem type, correction suggestions, and supporting data / standards based on knowledge triple parsing.

[0037] In some embodiments, the conclusion feedback structuring and analysis module introduces geological engineering-specific stratigraphic determination rules and site categories to determine rules and reasonable parameter ranges in order to achieve rule module constraints.

[0038] In some embodiments, the conclusion quality reward model includes a reward function that includes expert weights (experience values), error correction matching degree, specification consistency score, and data reference accuracy.

[0039] This invention presents a dynamic optimization method for large-scale models in exploration report generation, which for the first time proposes a deep optimization system of "conclusion section specialization" to achieve precise risk avoidance. For the "Conclusions and Recommendations" section, which is the highest-risk and most logically complex part of the exploration report, this invention does not adopt a general full-text generation model. Instead, it uses a specialized conclusion quality reward model (RM) and rule constraints to address the "illusion" problem in core aspects such as foundation evaluation and parameter recommendations. This significantly reduces the risk of geological disasters (such as incorrect subsidence and landslide judgments) caused by random model output.

[0040] This invention successfully transforms the scattered, unstructured annotations of geological engineers into standardized data that can be used for model fine-tuning through innovative knowledge triplet parsing technology. This mechanism solves the technical challenge of reusing and transferring high-value expert experience in algorithms, enabling large models to truly learn the deep interpretation logic of experts.

[0041] This invention constructs a unique "Conclusion Flywheel" dynamic evolution mechanism through dynamic iteration and a conclusion flywheel module, continuously improving model performance and establishing a closed-loop iterative system of "generation → annotation → structuring → reinforcement training → further improvement." As the system's scale of operation in enterprises expands and expert participation increases, the richer the cases absorbed by the model, the stronger its comprehensive judgment capability for complex geological scenarios becomes, forming a positive cycle where the model becomes more accurate and professional with use. By calculating uncertainties in real time and triggering a second round of expert review, it ensures that the output of high-risk projects is always under the dual monitoring of human experts and AI algorithms.

[0042] This invention, through a pre-defined hard constraint rule base and reward model, ensures that the generated conclusions strictly comply with current national and industry technical specifications (such as GB 50021). In practical applications, this significantly shortens the cycle of manual review and repeated revisions, reduces review costs, and enhances the professionalism and market competitiveness of surveying units in the field of intelligent report generation.

[0043] In some embodiments, the conclusion generation module generates an exploration report, specifically including:

[0044] (i) Accessing multi-source heterogeneous data and performing standardization processing, which involves converting multi-source heterogeneous data into a unified structured JSON format to obtain standardized data. This then completes data quality testing and generates a data quality report, laying the foundation for subsequent processing.

[0045] In the specific implementation process, a data access module is used to access multi-source heterogeneous data. This multi-source heterogeneous data, based on its different sources, includes field logging data and indoor experimental data. Field logging data refers to the original records directly obtained by engineering geological survey personnel at the construction site (field) through drilling, pitting, and other methods. It mainly includes borehole numbers, borehole elevations, drilling depths, descriptions of the soil and rock mass (color, state, humidity, inclusions, etc.), and groundwater level observation records.

[0046] Based on the types of multi-source heterogeneous data, including tabular data, WORD documents, drawings, and unstructured descriptions (such as geotechnical descriptions), the data access module needs to convert different types of data into a unified structured JSON format.

[0047] For tabular data (such as field cataloging and indoor experiments), the system calls a data processing library (such as Python's Pandas or OpenPyXL) to read Excel / CSV source files. Through a pre-defined "field mapping table," inconsistent headers with different units (such as "moisture content" and "moisture percentage (%)") are uniformly mapped to standard system key values ​​(such as water_content); that is, the data processing library converts the tabular data into a unified structured JSON format.

[0048] For unstructured descriptions (such as descriptions of soil and rock masses), regular expressions or NLP entity extraction techniques are used to automatically decompose them into JSON objects of {color: "gray", state: "soft plastic", soil_type: "silty clay"}; that is, using regular expressions or NLP entity extraction techniques to convert unstructured descriptions into a unified structured JSON format.

[0049] In practice, regular expressions convert unstructured descriptions (such as descriptions of soil and rock masses) into a unified structured JSON format, specifically including:

[0050] (1.1.1) Analyze the text structure: Identify the fixed patterns in the data to be extracted (such as key-value pair delimiters, repeating blocks, special markers).

[0051] (1.1.2) Design target JSON structure: Determine the output format (object, nested array, etc.) as needed.

[0052] (1.1.3) Write regular expressions: Use regular expressions to match and capture key data groups.

[0053] (1.1.4) Extraction and transformation: Use hot modules to perform matching and build a dictionary list.

[0054] (1.1.5) Handling complex situations, including: non-fixed order: improve readability by using named grouping;

[0055] Multi-line blocks: Enable the re.DOTALL flag to match newlines;

[0056] Special escaping: Use re.escape() to process text containing regular expression metacharacters.

[0057] (1.1.6) Verification and Output:

[0058] Validate JSON data: Use json.loads(json_output) to test if it can be parsed.

[0059] Improve output quality: use json.dumps(..., indent=2) or customize indentation.

[0060] Regular expressions are not suitable for parsing nested structures (such as HTML / XML); in such cases, a dedicated parser is more reliable. When handling Unicode characters such as Chinese characters, add the `re.UNICODE` flag.

[0061] By following the steps above, text data matched by regular expressions can be systematically converted into structured JSON.

[0062] For drawings (topographic maps / plan maps): The system integrates a DXF parsing library (such as ezdxf or Teigha), directly reading the underlying data entities of DWG / DXF files without relying on a graphical interface. It extracts the circular entity coordinates (X, Y) from key layers (such as the "drill hole location layer") as drill hole coordinates, extracts text annotation entities as drill hole numbers, and finally converts them into geometric data in GeoJSON format; that is, it uses the DXF parsing library to convert the drawings into a unified structured JSON format.

[0063] In the specific implementation process, the drawings (CAD / BIM) will be converted into specific components, including:

[0064] (1.2.1) Requirements analysis and mapping table design, including:

[0065] Identify the input source: Determine the source of the drawings (such as CAD, PDF, scanned drawings) and their data characteristics;

[0066] Define target fields: Based on business requirements, define the fields that need to be included in the final JSON (e.g., "id", "name", "type", "coordinates", "area", "layer").

[0067] Create a mapping table: Establish a field mapping rule for each input source (usually a CSV or JSON configuration file).

[0068] (1.2.2) Raw data extraction and parsing:

[0069] Geometric information: vertex coordinates, line segments, polygon boundaries, etc.;

[0070] Attribute information: layer name, color, line type, associated text annotations, etc.

[0071] (1.2.3) Field alignment and data cleaning:

[0072] Based on the mapping table, rename the extracted original field names to the target field names;

[0073] Perform data cleaning: unit unification (convert units to standard units), coordinate system transformation (convert drawing coordinates to actual geographic coordinates if needed), and handling missing values ​​(fill missing fields with default values ​​or perform inference).

[0074] (1.2.4) Structured assembly:

[0075] The cleaned and aligned data is then assembled according to a predefined JSON Schema;

[0076] Ensure that geometric data uses a common structural representation (such as the geometry format).

[0077] (1.2.5) Output and Validation: Write the assembled data into a JSON file or store it directly in the database;

[0078] Perform data validation: format validation (ensure JSON conforms to the pre-designed schema), logic validation (check geometric validity (e.g., whether polygons are closed, whether required fields are complete), and visual spot checks (render the generated JSON in a simple front-end (e.g., Leaflet map, Three.js scenario), compare it with the original image, and ensure that the extraction is correct).

[0079] (1.2.6) Process automation: The above steps (i.e., the contents of steps 1.1 to 1.5) are written as scripts or pipelines to support batch processing.

[0080] (ii) The task orchestration engine in the large language model receives standardized data and initiates the task parsing process. Based on the stratigraphic layer number in the standardized data, the report template selected by the user, and by querying the complexity matrix, the task orchestration engine plans an execution sequence containing 20-80 atomic steps. That is, the task orchestration engine dynamically decomposes the exploration report to be generated into an execution sequence of 20-80 executable atomic steps.

[0081] In practice, the difference in the number of atomic steps is based on the following dimensions:

[0082] (2.1) Number of stratigraphic layers: If the site has a simple stratigraphic structure (e.g., only 3 layers), the profile generation only requires about 3-5 atomic steps of "reading data - calculating coordinates - receiving the connection". If the stratigraphic structure is complex (e.g., 15 layers, with lenses, faults or pinch-outs), then steps such as "lens closure processing", "fault line interpolation", and "stratigraphic pinch-out logic judgment" need to be added, which leads to a significant increase in the number of atomic steps.

[0083] (2.2) Report template type: The preliminary exploration report only requires basic logging and statistics (about 20-40 atomic steps), while the detailed exploration report needs to include advanced calculation steps such as liquefaction judgment, corrosivity evaluation, and pile foundation parameter calculation (about 60-80 atomic steps).

[0084] (2.3) Complexity matrix: The complexity matrix is ​​closely related to the number of attached plots. For each additional cross-sectional plot or bar chart, the task parser will automatically add a corresponding sub-sequence of "data extraction - coordinate transformation - plotting instruction generation".

[0085] In some embodiments, during the execution of each atomic step, a state manager manages the dependencies of each atomic step and tracks the execution status in real time. If a step fails, a rollback controller is used to roll back to the nearest stable checkpoint. In specific implementations, a directed acyclic graph (DAG) is used to manage atomic step dependencies.

[0086] The rollback controller primarily addresses the unreliability and state pollution issues inherent in Large Language Model (LLM) applications. Unreliability refers to the possibility that the LLM might output incorrect instructions, invoke incorrect tools, or generate unresolvable content. State pollution occurs when an erroneous output from one step is carried over to subsequent steps, leading to error accumulation and eventual task failure. Therefore, the rollback controller is used to restore the system state to a previously known correct point and attempt alternative paths when task execution deviates from expectations or encounters errors.

[0087] In practical implementation, the rollback controller mainly includes an execution monitor, a fault diagnostician, a rollback decision maker, a state restorer, and an alternative strategy executor. The rollback controller is tightly integrated with the task orchestration engine and the state manager. Specifically, the execution monitor is used to monitor the execution results of each step in real time; the fault diagnostician analyzes "what error occurred" and "why the error occurred"; the rollback decision maker, based on the diagnostic structure and predefined policies, decides whether to roll back, to which step, and how to retry; the state restorer collaborates with the state manager to execute the specific rollback operations; and the alternative strategy executor is used to adopt a different execution path after the rollback.

[0088] (iii) Single-step generation loop: The large language model performs closed-loop processing for each atomic step.

[0089] In some embodiments, the large language model performs closed-loop processing on each atomic step, specifically including:

[0090] (3.1) The big oracle model constructs a context based on the execution structure of the preceding steps and uses the generation module in the big language model to perform parallel sampling to generate 3-8 candidate JSON instructions.

[0091] Ultimately, the user only needs a single, explicit action instruction. However, to ensure the correctness of this instruction, the large language model generates 3-8 candidate actions in parallel for the same prompt. In practice, the Temperature parameter generated by the large language model is set to 0.5-0.7 to introduce a slight degree of randomness and generate multiple possible JSON instructions.

[0092] (3.2) The generated instructions are immediately passed through the exception filter (red flag mechanism), which removes any candidates with format errors or values ​​that exceed the specification boundaries.

[0093] Each of the 3-8 candidate actions (i.e., 3-8 candidate JSON instructions) is checked independently. The red flag mechanism principle is as follows: once a candidate action triggers a rule red line (such as the JSON format cannot be parsed, the required field target_layer is missing, the generated coordinate value exceeds the site boundary, or it contains rejection keywords such as "I don't know"), the candidate action (i.e., the JSON instruction) is marked as "red flag" and directly eliminated. Only compliant candidates are retained to enter the next voting stage.

[0094] (3.3) For the remaining compliant candidate actions (i.e. JSON instructions), the robust voting device executes the strategy of "stopping when the first K votes are reached" by statistically analyzing semantic redundancy, and quickly selects the optimal action (i.e. the optimal JSON instruction) with the highest consensus.

[0095] In its implementation, the robust voting system is based on the principle of "semantic consistency." Under high probability, the same answers repeatedly generated by a large language model (i.e., a large model) are often the correct answers. A threshold K is set (e.g., K=3), and the similarity between the remaining qualified candidate answers is calculated (for JSON commands, key-value exact matching is used). During the generation or comparison process, once the number of times a certain answer content appears repeatedly reaches the threshold K (e.g., K=3, meaning it gets 3 votes first), the system immediately terminates subsequent sampling or comparison, directly determining that answer as the "optimal action." This ensures accuracy while avoiding the waste of computational power waiting for all 8 sampling cycles to complete.

[0096] (iv) The optimal action is verified by a small model, and the verification is completed from at least two dimensions: compliance with engineering specifications and consistency with geological logic. If the verification is successful, the instruction is issued and executed; if it fails, a rollback and retry are triggered. This ensures the absolute reliability of each step.

[0097] Among these, Encoder-only architecture models (such as BERT, RoBERTa, or DistilBERT) with relatively small parameter counts (e.g., 110M-300M parameters) and extremely fast inference speeds are used for discriminative validation. These models excel at classification tasks and logical implication judgments.

[0098] In the specific implementation process, the optimal action is verified by a small model, which specifically includes:

[0099] Structure validation: Use Schema Validator to check the integrity of JSON structure, and use small models to determine whether data fields conform to the semantic definition of the current step (e.g., check whether the "stratum description" action includes features of three dimensions: color, density, and inclusions).

[0100] Compliance verification: Based on the Retrieval Enhanced Generation (RAG) approach, the numerical value in the optimal action (e.g., "cohesion = 5 kPa") is compared and verified with the provisions in the built-in standard database (e.g., the "Code for Investigation of Geotechnical Engineering" GB50021). The small model acts as a binary classifier, taking [standard provision, action parameter] as input and [compliant / non-compliant] as output. For example, if the standard stipulates that the void ratio of soft soil must be greater than 1.0, while the optimal action is 0.8, it is judged as non-compliant.

[0101] Logical consistency verification: Based on Natural Language Reasoning (NU), it checks whether the contextual logic is contradictory. For example, given the input pair: Premise: "The previous layer is the 3rd layer of silty clay", Hypothesis: "This layer is the 2nd layer of silt". According to the geological sedimentary pattern (older at the bottom, younger at the top), the small model will identify a logical contradiction in the stratigraphic sequence (the 2nd layer should not appear directly below the 3rd layer), thus rejecting the action.

[0102] (v) Verified instructions drive atomic executor operations. If execution is successful, the global state pool is updated; if execution fails or verification fails, the corresponding rollback strategy is immediately initiated.

[0103] (vi) Output of results synthesis: After all atomic steps are completed, the output synthesis model collects all execution results, arranges the content and optimizes the format according to the exploration report annotation template, and generates the final exploration report.

[0104] The conclusion generation module of this invention uses a multi-step decomposition principle (that is, decomposing the generation of a complex survey report into multiple atomic steps). Each step has a clear definition of input, processing and output, and the steps are connected through data dependencies.

[0105] Single-step generation principle: Large models generate only one atomic step's action description at a time, avoiding the error propagation problem in traditional chained generation. Each action uses a structured description to ensure parsing and execution. This ensures that the single-step output is highly consistent with engineering specifications.

[0106] The principle of multi-candidate selection: Multiple candidate actions are obtained through multiple samplings. After anomaly filtering, a voting mechanism based on multiple indicators is used to select the optimal solution, which significantly improves the reliability of single-step decision-making.

[0107] Real-time verification principle: A lightweight discrimination model is introduced, and RAG technology is used to search the survey specification database in real time to conduct engineering logic and compliance reviews on candidate actions.

[0108] Error isolation principle: By using atomic executors and rollback mechanisms, local execution exceptions can be captured in time and targeted retries can be triggered, ensuring the robustness of the global task chain.

[0109] In summary, based on the multi-step decomposition principle, single-step generation principle, multi-candidate selection principle, real-time verification principle, and error isolation principle of this invention, the conclusion generation module of this invention has the following specific advantages:

[0110] 1. Significantly improved reliability: Error propagation is effectively suppressed, and the overall logical consistency of the survey report is greatly improved.

[0111] 2. Improved accuracy: Single-step output maintains a high degree of consistency with engineering specifications and experience-based structures.

[0112] 3. Significantly enhanced consistency: Graphs, tables, and queries are generated from the same source, and data with the same name is automatically aligned.

[0113] 4. Overall efficiency leap: The entire process time is significantly shortened compared to the traditional chain method, and manual operation is significantly reduced.

[0114] 5. Stability is fully guaranteed: the output fluctuation is small after multiple runs, and the system can automatically roll back and continue running in case of failure.

[0115] Example 1 of the conclusion generation module of the present invention - cloud deployment

[0116] A container cluster is set up on a public cloud or industry private cloud. The data access module, task orchestration engine, large model generation module, discrimination and verification module, and atomic executor all run as microservices. Users drag and drop field records and CAD drawings into a browser → the backend automatically breaks them down into 60 atomic steps → each step is written by the large model with only one JSON action → the discriminator performs standard verification within 0.1 seconds → drives the execution of the existing "Lizheng + CAD + Excel" cloud container → write back on success and roll back on failure → download the entire report in 30 minutes. No software installation is required on the user side, and the original professional algorithms and mapping kernel are completely reused.

[0117] Example 2 of the conclusion generation module of the present invention - construction site integrated machine

[0118] The system utilizes a 2U industrial chassis with an embedded domestically produced GPU. It comes pre-installed with large language models, discrimination and verification models, and executable programs such as LiZheng, CAD, and Office. The process involves plugging in the box, importing data from a USB drive, local disassembly, gradual generation of a 32B model within the domestic GPU, real-time discrimination and scoring of small models, calling the pre-installed LiZheng / CAD / statistics executables, and automatically printing a PDF and burning it to a read-only CD. No data is left on the machine throughout the entire process, and a signed survey report is output.

[0119] Example 3 of the conclusion generation module of the present invention

[0120] 1. System Deployment and Configuration: Deploy the large language model and the discriminant verification model (small model) on the server; build a historical task knowledge base to store atomic action templates and specification constraints; configure the API interfaces of various professional software.

[0121] 2. Task initialization: Upload raw exploration data (drilling data, test results, site information, etc.); select report template and output specification requirements; set quality acceptance standards and mandatory clause checklist.

[0122] 3. Intelligent orchestration and execution: The large model (i.e., large language model) parsing module breaks down the overall task into a sequence of atomic steps; the multiple sampling module generates multiple candidate actions for each step; the discrimination and verification model checks the standard compliance of the candidate actions; the voting selection module determines the optimal action based on multi-index evaluation; and the atomic executor calls the corresponding professional software to execute the verified action.

[0123] 4. Quality monitoring and exception handling: The execution engine monitors the execution status of each step in real time; when a failed step is encountered, it automatically rolls back to the previous checkpoint; the generation-verification-execution loop of that step is retried; error logs are recorded and the knowledge base is updated to avoid repeated errors.

[0124] 5. Results Output and Delivery: The output module summarizes the execution results of all successful steps; automatically generates a Word report document that conforms to the specifications; packages and outputs CAD drawings, statistical tables and other supporting materials; and generates a quality control report indicating the pass status of key inspection items.

[0125] In summary, the conclusion generation module of this invention, through an innovative architecture of multi-step decomposition, discriminant verification, and robust voting, achieves intelligent arrangement and stable execution of the entire process of "data input - stratigraphic division - parameter statistics - drawing generation - text writing - report synthesis." It effectively solves industry pain points in existing technologies such as "unreliable generation of large models, cross-step propagation of errors, and difficulty in implementing standard provisions." While maintaining the existing professional software system, it significantly improves the automation level and engineering reliability of exploration report generation, and has significant technological breakthrough value and promising prospects for widespread application.

[0126] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A large model dynamic optimization method for a survey report generation task, characterized in that, include: (1) The survey report is generated by the conclusion generation module; (2) Experts provide feedback on the conclusions of the survey report, and the feedback information is collected by the feedback collection module. (3) The conclusion feedback structuring and parsing module parses the collected expert feedback information and applies rules to constrain it; (4) The conclusion quality reward model applies rewards and reinforcement learning to the conclusion section after the analysis and rule modules are constrained to obtain the final survey report; (5) The conclusion reinforcement learning fine-tuning module fine-tunes the model of the conclusion generation module under the guidance of the conclusion quality reward model, so as to improve the ability of geological risk identification, stratigraphic logic consistency and standardized wording, and thus obtain the final exploration report.

2. The method of claim 1, wherein the method is characterized by, It also includes a dynamic iteration and conclusion flywheel module. After obtaining the final survey report, the dynamic iteration and conclusion flywheel module calculates the uncertainty, logical consistency and normative risk of the conclusion section. When a high-risk conclusion is detected, a second round of expert review is automatically triggered, and the process proceeds to steps (2) to (5) in sequence.

3. The method of claim 1, wherein the method further comprises: The aforementioned conclusion feedback structuring and parsing module automatically converts expert feedback information into a structural pattern of question type, correction suggestions, and supporting data / standards based on knowledge triple parsing.

4. The method of claim 1, wherein the method further comprises: The aforementioned conclusion feedback structured and analytical module introduces geological engineering-specific stratigraphic determination rules and site categories to determine rules and reasonable parameter ranges in order to achieve rule module constraints.

5. The method of claim 1, wherein the method further comprises: The conclusion quality reward model includes a reward function, which comprises expert weights, error correction matching degree, standard consistency score, and data citation accuracy.

6. The method of claim 1, wherein the method further comprises: The survey report generated by the conclusion generation module includes: (i) Accessing multi-source heterogeneous data and performing standardization processing, which converts multi-source heterogeneous data into a unified structured JSON format to obtain standardized data; (ii) The task orchestration engine receives standardized data and starts the task parsing process. Based on the number of stratigraphic layers in the standardized data, the report template selected by the user, and the complexity matrix, the task orchestration engine plans an execution sequence containing 20-80 atomic steps. (iii) Single-step generation loop: The large language model performs closed-loop processing for each atomic step; (iv) The optimal action is verified by a small model, and the verification is completed from at least two dimensions: compliance with engineering specifications and consistency with geological logic. If the verification is successful, the instruction is issued and executed; if it fails, a rollback and retry are triggered. (v) Verified instructions drive atomic executor operations: if the execution is successful, the global state pool is updated; if the execution fails or the verification fails, the corresponding rollback strategy is immediately initiated. (vi) Output of results synthesis: After all atomic steps are completed, the output synthesis model collects all execution results, arranges the content and optimizes the format according to the exploration report annotation template, and generates the final exploration report.

Images