Intelligent collaborative generation method and system for mineral resource coverage evaluation report

Abstract

The invention provides an intelligent collaborative generation method and system for a mineral resource coverage evaluation report, and the method comprises the steps: carrying out the intelligent analysis and classification of original data input based on a combined summarization double-process model, constructing a multi-dimensional data feature vector, and obtaining the standardized evaluation data input; performing modeling and safety analysis on standardized evaluation data input, performing feature extraction and correlation analysis on parameters in an analysis process through a peak time sequence learning mechanism, establishing a dynamic judgment model, and generating a comprehensive technical analysis report; performing predictive control by applying a predictive controller principle and combining an evaluation specification library, a regulation policy library and an excellent report template, matching an optimal report template and term elements, and outputting an evaluation report draft; a heuristic evolutionary algorithm is adopted, report content is optimized in the collaborative editing process, an intelligent question and answer system achieves optimal path planning of data retrieval and verification, and a complete evaluation report conforming to the format is generated. According to the invention, the generation efficiency and quality of the evaluation report are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mineral resources management, in particular to a mineral resource overburden assessment report intelligent collaborative generation method and system, which is applied to the fields of mineral resource overburden relationship assessment, report preparation and intelligent auxiliary decision-making. BACKGROUND

[0002] Mineral resource overburden assessment is a key link in mineral development, involving scientific assessment of the spatial relationship between underground mineral resources and surface construction projects. With the acceleration of urbanization and the growth of mineral development demand, the workload of overburden assessment report preparation has increased dramatically, and higher requirements have been put forward for assessment efficiency and quality.

[0003] The current main overburden assessment techniques include traditional manual assessment methods and semi-automatic assessment systems. The traditional manual assessment method relies on expert experience and manual calculation, collects geological exploration data and engineering planning data, uses classical geomechanics theory to analyze safe mining depth, and finally forms an assessment report. The other is a semi-automatic assessment system based on GIS technology, which can realize the visualization display and basic calculation of geological data, but still needs a lot of manual intervention for report generation.

[0004] The most relevant prior art is the assessment report auxiliary generation technology based on expert system, whose working principle is to encode the assessment specifications and expert knowledge into a rule base, process the input geological data through rule matching and reasoning mechanism, and generate a standardized assessment conclusion and report framework. This technology can reduce the amount of manual work to a certain extent and improve the format standardization of the report.

[0005] However, there are two key problems with the existing technology: first, the lack of intelligent data processing and analysis capabilities makes it difficult to effectively handle complex and variable geological conditions and engineering scenarios, resulting in low assessment accuracy; second, the report generation process lacks flexibility and collaboration, which cannot support the individual needs and real-time interaction of assessment personnel, seriously affecting work efficiency and report quality. SUMMARY

[0006] The purpose of the present application is to solve the technical problems of low intelligence and poor collaboration efficiency in the existing mineral resource overburden assessment report generation process, and to provide a mineral resource overburden assessment report intelligent collaborative generation method and system. Based on the combined summary double-process model, agent gradient support spiking neural network, spiking neuron predictive control and heuristic evolutionary algorithm for mixed integer linear programming solution, the method realizes intelligent analysis of overburden assessment data, accurate analysis of technical demonstration, intelligent matching of report templates and optimization of human-computer collaborative editing, and improves the generation efficiency and quality of the assessment report.

[0007] To achieve the above object, the present application provides a mineral resource pressure evaluation report intelligent collaborative generation method, comprising the following steps: The project basic information, planning data and core geological data input by the evaluation personnel are obtained as original data input, the original data input is intelligently analyzed and classified based on a combined summarization double process model, a structured multi-dimensional data feature vector is constructed, and standardized evaluation data input is obtained; The standardized evaluation data input is subjected to geological modeling and safety depth analysis, the spike timing learning mechanism of the proxy gradient support spike neural network is used to extract the timing characteristics and correlation analysis of the geological parameter changes, safety factor calculation and risk assessment results in the analysis process, a dynamic judgment model of the pressure relationship is established, and a comprehensive technical analysis report is generated; Based on the comprehensive technical analysis report, the principle of using spike neurons as a predictive controller of a linear system is used, combined with the built-in evaluation specification library, regulation policy library and excellent report template for predictive control, the optimal report template and clause elements are automatically matched, and a complete evaluation report draft is output; The evaluation report draft is used as the initial solution, a heuristic evolutionary algorithm based on generalized MILP mixed integer linear programming solution and data-driven is used to optimize the integrity and accuracy of the report content in the human-computer collaborative editing process, and the optimal path planning of material retrieval and verification is realized through an intelligent question and answer system, and a complete evaluation report meeting the official format requirements is generated.

[0008] As a preferred, the combined summarization double process model is used to intelligently analyze and classify the project basic information, planning data and core geological data, and construct a structured multi-dimensional data feature vector, which comprises: Based on the original data input, a fast processing process is started to process text type project information and extract keywords, and a slow processing process is started to mine deep features and understand semantics of numerical type geological parameters, and double processing results are obtained; The double processing results are combined and summarized, the consistency of text information and numerical data is ensured through attention mechanism and cross-validation algorithm, and a multi-dimensional feature vector containing geological conditions, engineering parameters, environmental factors and regulatory requirements is constructed; According to the multi-dimensional feature vector, quality evaluation is performed through data verification and completion mechanism, and the standardized evaluation data input is obtained.

[0009] As a preferred, the fast processing process is started to process text type project information and extract keywords, and the slow processing process is started to mine deep features and understand semantics of numerical type geological parameters, and double processing results are obtained, which comprises: Based on the text type project information, through pre-training of a geological professional vocabulary library and a named entity recognition model, natural language processing is performed, key information of mineral type, geographical location, geological age, and lithological characteristics is automatically extracted and semantically labeled and classified, and a semantic understanding result is obtained; Based on the numerical type geological parameters, a deep learning network is used to perform multi-level abstraction on rock mass strength parameters, groundwater level data, and tectonic stress field distribution, identify potential correlation between parameters, and infer rock mass stability level and aquifer dynamic characteristics, and a numerical analysis result is obtained. The semantic understanding result and the numerical analysis result are organically combined to obtain the dual processing result.

[0010] As preferred, the spike temporal learning mechanism of the proxy gradient support spiking neural network performs temporal feature extraction and correlation analysis on geological parameter changes, safety factor calculation, and risk assessment results during the analysis process, establishes a dynamic judgment model of the pressure-overburden relationship, including: Based on the standardized evaluation data input, geological modeling and safe mining depth analysis are performed, and geological parameter changes, safety factor calculation, and risk assessment results during the geological modeling and safe mining depth analysis process are converted into time series data streams. The numerical changes are converted into discrete spike sequences by the spiking neural network to obtain time series feature data. A smooth proxy function is constructed using the proxy gradient method to approximate the gradient of the spike function, the network weight is updated for the time series feature data, the characteristic spike temporal pattern corresponding to different risk levels is identified, and the dynamic judgment model of the pressure-overburden relationship is established. According to the dynamic judgment model of the pressure-overburden relationship, short-term fluctuations and long-term trends are analyzed through multi-scale time windows to obtain the comprehensive technical analysis report.

[0011] As preferred, the principle of using spiking neurons as a predictive controller of a linear system is used in combination with an embedded evaluation specification library, a regulation policy library, and an excellent report template for predictive control, automatically matching the optimal report template and clause elements, including: Based on the comprehensive technical analysis report, the report generation process is modeled as a multivariable linear control system, in which spiking neurons act as intelligent controllers. Each spiking neuron corresponds to a decision node for chapter structure selection, key content highlighting, and professional term usage, and a control system model is obtained. The control system model is used to predict the report generation effect of the control system model at each decision time, and when the membrane potential reaches a threshold, a spike triggers a corresponding report generation action, and a control strategy that optimizes the overall quality index is selected. According to the control strategy, the optimal report template and clause element combination are automatically matched from the evaluation specification library, the regulation policy library and the excellent report template to obtain the evaluation report draft.

[0012] As preferred, the data-driven heuristic evolutionary algorithm based on generalized MILP mixed integer linear programming solution optimizes the integrity and accuracy of the report content in the human-computer collaborative editing process, including: The quality evaluation of the evaluation report draft is converted into a multi-objective optimization problem, a target function including content integrity index, technical accuracy index, expression standardization index and logical coherence index and a constraint condition including mandatory requirements, format standard restrictions and page control requirements are established to form a multi-objective optimization model; An intelligent search strategy including crossover operator, mutation operator and selection operator is designed, the structural fusion of technical argumentation logic and risk analysis method between different excellent reports is simulated through the crossover operator, innovative chart display method and accurate professional term expression are introduced through the mutation operator, the diversity of solutions is maintained through the selection operator by adopting the multi-objective Pareto optimal principle, the solution space of the multi-objective optimization model is explored, and the optimized report content is obtained; Based on the optimized report content, the user query is modeled as a path planning problem of information retrieval by an intelligent question and answer system, the optimal retrieval strategy is found by comprehensively considering the relevance, reliability and timeliness of data, and the complete evaluation report meeting the official format requirements is obtained.

[0013] As preferred, the comprehensive technical analysis report includes detailed technical argumentation process, risk level evaluation, uncertainty quantification analysis and confidence evaluation.

[0014] As preferred, the evaluation specification library includes national, industry and local technical standards and specification requirements, the regulation policy library includes relevant legal provisions, policy explanations and implementation details, and the excellent report template includes typical evaluation report cases of different types and different regions.

[0015] As preferred, the intelligent question and answer system realizes the optimal path planning of data retrieval and verification, including: Based on the user query request, the user query request is modeled as a path planning problem of information retrieval, the data relevance, reliability and timeliness are set as evaluation indexes of path cost function, and the search space of retrieval path is constructed; A heuristic search algorithm is used to traverse the search space of the retrieval path, the optimal retrieval strategy with the minimum total cost is found by evaluating the cost function values of different retrieval paths, and the corresponding answer content is organized; Based on the optimal retrieval strategy and the answer content, data sources, confidence assessment and citation suggestions are provided, and through multiple rounds of human-computer interaction and algorithm optimization, the complete evaluation report meeting the official format requirements is obtained.

[0016] The application also provides a mineral resource overburden evaluation report intelligent collaborative generation system, comprising: A data input module is configured to obtain project basic information, planning data and core geological data input by an evaluator as original data input, intelligently analyze and classify the original data input based on a combined summarization double-process model, construct a structured multi-dimensional data feature vector, and obtain standardized evaluation data input. A technical analysis module is configured to perform geological modeling and safe depth analysis on the standardized evaluation data input, perform time sequence feature extraction and correlation analysis on geological parameter changes, safety factor calculation and risk assessment results in the analysis process through a spike timing learning mechanism of a proxy gradient support spike neural network, establish a dynamic judgment model of overburden relationship, and generate a comprehensive technical analysis report. A report generation module is configured to use a spike neuron as a predictive controller of a linear system based on the comprehensive technical analysis report, combine an embedded evaluation specification library, a regulation and policy library and an excellent report template for predictive control, automatically match the optimal report template and clause elements, and output a complete evaluation report draft. A collaborative editing module is configured to use the evaluation report draft as an initial solution, adopt a data-driven heuristic evolutionary algorithm based on generalized MILP (mixed integer linear programming) solution, optimize the completeness and accuracy of the report content in the human-computer collaborative editing process, realize optimal path planning of material retrieval and verification through an intelligent question and answer system, and generate a complete evaluation report meeting the official format requirements.

[0017] Compared with the prior art, the application has the following beneficial effects: 1. The combined summarization double-process model is applied to mineral resource evaluation data processing, realizing intelligent analysis and standardized processing of multi-source heterogeneous geological data, and improving the intelligent level and accuracy of data processing. 2. The spike timing learning mechanism of the proxy gradient support spike neural network is used to establish a dynamic judgment model of overburden relationship, improve the accuracy and timeliness of technical demonstration, and make the evaluation results more objective and reliable. 3. The spike neuron predictive controller principle is introduced into the field of automatic report generation, realizing intelligent predictive control of the structure and content of the evaluation report, and greatly improving the intelligent level of report generation. 4. The heuristic evolutionary algorithm for generalized MILP solving is applied to human-computer collaborative editing optimization, which significantly improves the efficiency and quality of report generation, and realizes efficient collaboration between evaluation experts and intelligent systems. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0019] Figure 1 A flowchart of a mineral resource overburden assessment report intelligent collaborative generation method of the present application; Figure 2 A structural block diagram of a mineral resource overburden assessment report intelligent collaborative generation system of the present application. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application will be further described in detail below in combination with specific embodiments and drawings.

[0021] As shown in Figure 1 The present application provides a mineral resource overburden assessment report intelligent collaborative generation method, comprising the following steps: S1: Obtain the project basic information, planning data and core geological data input by the evaluation personnel as the original data input, intelligently analyze and classify the original data input based on the combined summary double process model, construct a structured multi-dimensional data feature vector, and obtain standardized evaluation data input; First, obtain the project basic information, planning data and core geological data input by the evaluation personnel as the original data input. The project basic information usually includes basic information such as project name, construction unit, project geographic location, construction scale and engineering type; the planning data includes data in related planning documents such as land use planning, urban master planning and mineral resource planning; and the core geological data includes professional technical data such as geological exploration report, drilling data, stratigraphic section, hydrogeological data and geotechnical engineering parameters. These original input data usually have various formats, complex structures and strong professional characteristics, and direct use will bring difficulties to subsequent analysis. Therefore, the combined summary double process model is used to intelligently analyze and classify these original data.

[0022] The combined summarization dual-process model is a data processing framework that simulates the cognitive mechanisms of the human brain. It is based on the dual-system theory in cognitive psychology, which posits that human thought consists of two complementary processing systems: System 1 is fast, automatic, and intuitive, while System 2 is slow, conscious, and analytical. In this method, the dual-process model simultaneously initiates a fast processing process and a slow processing process to handle different types of data. The fast processing process mainly handles text data and quickly extracts key information using natural language processing techniques, while the slow processing process conducts in-depth analysis of numerical geological parameters to uncover potential patterns and relationships. The results of these two processes are integrated through a combined summarization mechanism to form a unified cognitive representation. In this way, a structured multi-dimensional data feature vector can be constructed, converting the original heterogeneous data into standardized evaluation data input, which provides a high-quality data foundation for subsequent geological modeling and safe mining depth analysis.

[0023] S2: Perform geological modeling and safe mining depth analysis on the standardized evaluation data input. Use the spike temporal learning mechanism of the proxy gradient support spiking neural network to extract temporal features and analyze the relevance of geological parameter changes, safety factor calculations, and risk assessment results during the analysis process. Establish a dynamic judgment model for the overlying relationship and generate a comprehensive technical analysis report. Geological modeling refers to constructing a three-dimensional geological model of the study area based on geological exploration data, including factors such as stratum distribution, lithology changes, fault location, and aquifer distribution. Safe mining depth analysis determines the safe depth limit for mining mineral resources based on the geological model and engineering mechanics principles to ensure that surface buildings are not damaged by mining activities. In traditional methods, these analyses often rely on expert experience and simplified calculations, making it difficult to handle complex geological conditions. To improve analysis accuracy, this method introduces the spike temporal learning mechanism of the proxy gradient support spiking neural network to extract temporal features and analyze the relevance of geological parameter changes, safety factor calculations, and risk assessment results during the analysis process.

[0024] A spiking neural network is an artificial neural network that simulates the firing mechanism of biological neurons. Unlike traditional neural networks based on continuous activation functions, it uses discrete spikes (or pulses) as the means of information transmission between neurons. In biological neural systems, neurons communicate by generating potential spikes, and the firing time and frequency of spikes encode information. Spiking neural networks simulate this mechanism and can process time series data and extract time-dependent patterns. However, due to the discontinuity and non-differentiability of the spike function, traditional backpropagation algorithms cannot be directly applied. The proxy gradient method is a key technology to solve this problem, which approximates the non-differentiable spike function by constructing a differentiable proxy function, thereby realizing the effective training of the network. Through this mechanism, key features can be extracted from the time variation of geological parameters, characteristic spike timing patterns corresponding to different risk levels can be identified, and a dynamic judgment model of the pressure-overlap relationship considering time evolution characteristics can be established. Finally, a comprehensive technical analysis report containing detailed technical argumentation process, risk level assessment, uncertainty quantification analysis, and confidence assessment is generated.

[0025] S3: Based on the comprehensive technical analysis report, use spiking neurons as the predictive controller of linear systems, combine the built-in evaluation specification library, regulations and policies library, and excellent report templates for predictive control, automatically match the optimal report template and clause elements, and output a complete evaluation report draft; Based on the comprehensive technical analysis report, use spiking neurons as the predictive controller of linear systems, combine the built-in evaluation specification library, regulations and policies library, and excellent report templates to automatically generate an evaluation report draft. Spiking neurons as the predictive controller of linear systems is an innovative combination of neural computing and control theory. Spiking neurons simulate the membrane potential accumulation and firing process of biological neurons, triggering output spikes when input signals accumulate to the threshold. In control systems, this mechanism can be used to realize decision triggering and control signal generation. Predictive control is a model-based control method that optimizes current control decisions by predicting future system behavior, widely used in real-time control of complex systems.

[0026] In the report generation process, the report structure and content organization are regarded as a multivariable control system, where the control variables include chapter settings, content detail level, professional term usage, etc., and the multivariable control system states include report integrity, accuracy, standardization, etc. indicators. The spiking neural network acts as a controller, which predicts the effects of different report generation strategies according to the input of technical analysis results and built-in knowledge base, and triggers corresponding report generation actions at key decision points. The evaluation specification library includes mineral resources development and utilization scheme compilation content requirements, metal and non-metal open-pit mine slope engineering technical specifications, and other standards and industry specifications; the regulations and policies library includes mineral resources, safety production related laws and regulations and their implementation details; the excellent report template library collects high-quality evaluation report cases of different types of minerals and different regions. In this way, the most suitable report template and clause elements for the current evaluation task can be automatically matched, and a complete structure, standard format, and accurate content evaluation report draft can be output.

[0027] S4: Taking the evaluation report draft as the initial solution, using a data-driven heuristic evolutionary algorithm based on generalized MILP mixed integer linear programming solution, optimizing the integrity and accuracy of the report content in the human-machine collaborative editing process, and realizing the optimal path planning of data retrieval and verification through an intelligent question and answer system, a complete evaluation report meeting the official format requirements is generated.

[0028] Taking the evaluation report draft as the initial solution, using a data-driven heuristic evolutionary algorithm based on generalized MILP mixed integer linear programming solution, further optimizing the report content in the human-machine collaborative editing process. Mixed integer linear programming (MILP) is an optimization method in operations research, which contains integer variables and continuous variables in decision variables, and the objective function and constraint conditions are linear. Generalized MILP extends the traditional MILP, allowing more complex objective functions and constraints, which can more accurately model various requirements in practical problems. Heuristic evolutionary algorithm is a class of optimization methods that simulate the process of natural evolution, searching the solution space by simulating population evolution, natural selection and gene mutation, which is suitable for solving complex combinatorial optimization problems.

[0029] In the report optimization process, the report quality evaluation is first converted into a multi-objective optimization problem, and the objective function includes multiple dimensions such as content integrity, technical accuracy, expression standardization and logical coherence; the constraint conditions include mandatory requirements of regulations, format standard restrictions and page control requirements, etc. Then, an intelligent search strategy containing crossover operator, mutation operator and selection operator is designed, and through the simulation of report content structure fusion, innovation mutation and survival of the fittest, the report content is constantly explored and improved. At the same time, the intelligent question and answer function is integrated, the user query is modeled as a path planning problem of information retrieval, and the optimal retrieval strategy is found by considering the data relevance, reliability and timeliness to provide accurate technical support. Through human-computer collaborative editing and multiple rounds of optimization, a complete evaluation report that meets the official format requirements and has accurate and complete content is finally generated.

[0030] In a preferred embodiment, in step S1, the project basic information, planning data and core geological data are intelligently analyzed and classified based on the combined summary dual process model, and a structured multi-dimensional data feature vector is constructed, including: S1.1: Based on the original data input, start the fast processing process to perform natural language processing and keyword extraction on the text type project information, and start the slow processing process to perform deep feature mining and semantic understanding on the numerical type geological parameters, to obtain dual processing results; Step S1.1 is a dual processing process, based on the original data input, simultaneously starting the fast processing process and the slow processing process to process different types of data. The fast processing process is mainly aimed at text type project information, and adopts natural language processing technology for key information extraction and semantic understanding. Natural language processing (NLP) is an interdisciplinary field of computer science and linguistics, which studies the interaction between computers and human language, including text understanding, information extraction, semantic analysis and other technologies. The fast processing process uses NLP models specially trained for the geological field, including named entity recognition, keyword extraction and semantic role labeling technologies, to identify and extract key information from project description texts.

[0031] Named entity recognition (NER) is one of the basic tasks of NLP, aiming to identify and classify named entities such as names, places, organizations, etc. In the geological field, NER is extended to identify mineral names (such as "iron mine", "coal mine"), geological ages (such as "Jurassic", "Carboniferous"), rock types (such as "granite", "shale") and other professional entities. Keyword extraction is to identify the key terms that best represent the content of the text, usually based on word frequency statistics, TF-IDF or TextRank algorithms. Semantic role labeling further analyzes sentence structure to identify predicate-argument relationships, such as "located in", "covered by" and other relational predicates and related entities.

[0032] The slow processing process targets numerical geological parameters and uses deep feature mining and semantic understanding methods. Deep feature mining uses deep learning networks to analyze numerical data and extract high-level abstract features. Deep learning is a branch of machine learning that uses multi-layer neural networks to automatically learn hierarchical representations of data, gradually abstracting high-level features from low-level features. In processing geological parameters, the main analysis is on rock mass strength parameters (such as uniaxial compressive strength, cohesion, internal friction angle, etc.), groundwater level data (such as water level elevation, seasonal variation, etc.), and tectonic stress field distribution (such as principal stress direction, stress concentration area, etc.). Through deep learning, potential relationships between these parameters can be identified, and high-level geological characteristics such as rock mass stability level and aquifer dynamic characteristics can be inferred. Through this parallel dual processing mechanism, the semantic richness of text information and the accuracy of numerical data can be utilized simultaneously to achieve a more comprehensive understanding of the data.

[0033] S1.2: Combine and summarize the dual processing results, ensure consistency between text information and numerical data through attention mechanism and cross-validation algorithm, and construct a multi-dimensional feature vector containing geological conditions, engineering parameters, environmental factors, and regulatory requirements; Step S1.2 is the combination and summarization process, which integrates the dual processing results and constructs a multi-dimensional feature vector. Combination and summarization is an information fusion method that analyzes the relevance and complementarity between different sources of information to form a unified and comprehensive knowledge representation. The core task of combination and summarization is to organically combine the semantic understanding results obtained from text processing and the analysis results obtained from numerical processing, ensuring the consistency and integrity of the information. Attention mechanism and cross-validation algorithm are used to achieve this goal.

[0034] Attention mechanism is a computational model derived from cognitive science, used to simulate the selective attention process of humans when processing information. In deep learning, attention mechanism allows the model to dynamically focus on different parts of the input, assigning weights based on their relevance and importance. In the combination and summarization process, attention mechanism is used to identify key information in text descriptions and numerical data and establish their relationships. For example, when the text describes "there are multiple NE-trending faults in this area," the attention mechanism will automatically focus on the parameters related to the faults in the numerical data, such as fault width, dip angle, and fault throw.

[0035] Cross-validation is a statistical method used to assess the generalization ability of a model by dividing the data into training and validation sets and repeating the training and validation process to evaluate model performance. In the combination summary, cross-validation is used to check the consistency between text information and numerical data. For example, it will check whether the "hard rock" described in the text matches the high compressive strength shown in the numerical data, and whether the "strong fault activity" mentioned in the text is consistent with the fault displacement rate data. In this way, inconsistencies between different data sources can be identified and resolved, ensuring the consistency of information. Finally, a structured feature vector is constructed that includes geological conditions (such as lithology distribution, tectonic characteristics), engineering parameters (such as strength indicators, stability ratings), environmental factors (such as hydrological conditions, topographic features), and regulatory requirements (such as safety distances, protection levels), etc. multiple dimensions to fully characterize the characteristics of the evaluation object.

[0036] S1.3: According to the multi-dimensional feature vector, quality assessment is carried out through data verification and completion mechanism, and the standardized evaluation data input is obtained.

[0037] Step S1.3 is the data verification and completion phase, according to the multi-dimensional feature vector, quality assessment is carried out through data verification and completion mechanism, and the standardized evaluation data input is obtained. Data verification refers to the process of checking the accuracy, completeness and consistency of data to ensure that the data meets the predetermined quality standards and business rules. In mineral resources overburden assessment, data verification is particularly important because inaccurate or incomplete data may lead to incorrect assessment results and affect engineering decisions. Data verification mechanism is based on a pre-defined rule set and statistical methods to comprehensively check the data in each dimension of the feature vector.

[0038] Data verification includes multiple levels: basic verification checks whether the data format and range are correct, such as whether the coordinate values are within a reasonable geographic range and whether the rock mass strength parameters are within the physically possible range; logical verification checks whether the relationship between data is reasonable, such as the consistency of overburden thickness and drilling depth, the orderliness of geological age division; professional verification is based on geological and engineering knowledge to check whether the data meets professional specifications, such as whether the fault parameters meet the principles of structural geology and whether the safety factor calculation meets the standards of engineering mechanics.

[0039] Data completion is a technique for handling missing data, filling in gaps through statistical inference, machine learning, or domain knowledge. In actual evaluation, due to limitations of exploration work or constraints of technical conditions, geological data often have missing values. Data completion mechanism uses various methods: for spatial data, use geostatistical methods such as Kriging interpolation, inverse distance weighting for spatial interpolation; for parameter data, analog inference based on similar lithology and geological conditions of historical cases; for complex relationships, use machine learning models to predict missing values. For example, when the elastic modulus data of a rock mass in a certain area is missing, the lithology, weathering degree and integrity index of the area can be used to infer the reasonable parameter value through a regression model based on historical data under similar geological conditions.

[0040] Through data verification and completion processing, the quality and integrity of the feature vector can be improved, and errors, contradictions and missing data in the data can be eliminated or reduced, ultimately obtaining standardized evaluation data input. These high-quality standardized data provide a reliable foundation for subsequent geological modeling, safe depth analysis and pressure relationship judgment, ensuring the accuracy and reliability of technical analysis. Standardized evaluation data input is usually stored in the form of a structured database, which is convenient for subsequent calling and analysis, and also facilitates data exchange and integrated analysis with other professional software.

[0041] In one specific embodiment, in step S1.1, a fast processing process is started to perform natural language processing and keyword extraction on the text-type project information, and a slow processing process is started to perform deep feature mining and semantic understanding on the numerical-type geological parameters, obtaining dual processing results, including: S1.1.1: Based on the text-type project information, perform natural language processing through a pre-trained geological professional vocabulary library and a named entity recognition model, automatically extract key information of mineral types, geographic locations, geological ages, and lithological characteristics, and perform semantic annotation and classification, to obtain semantic understanding results; In step S1.1.1, based on the text type project information, natural language processing is performed through a pre-trained geological professional vocabulary library and a named entity recognition model. The pre-trained geological professional vocabulary library is a knowledge base containing a large number of geological terms, concepts and relationships, providing professional background knowledge for text understanding. This vocabulary library usually contains professional terms such as mineral resource types (such as coal, metal mineral, non-metallic mineral, etc.), geological ages (such as Cambrian, Ordovician, Carboniferous, etc.), rock types (such as granite, limestone, shale, etc.), geological structures (such as faults, folds, joints, etc.) and their definitions and relationships. The named entity recognition model is a natural language processing model that specifically identifies entities of certain categories in text, usually using sequence labeling methods such as conditional random fields (CRF), bidirectional long short-term memory networks (BiLSTM) or pre-trained language models (such as BERT) and other technologies. Using these tools, key information such as mineral type, geographical location, geological age, lithological characteristics, etc. can be automatically extracted from project text and semantically labeled and classified. For example, when processing a text such as "a coal mine is located in the North China Plain, mainly exploiting the Carboniferous-Permian coal seam", the key information such as mineral type (coal mine), geographical location (North China Plain), geological age (Carboniferous-Permian) can be identified and structured into semantic understanding results.

[0042] S1.1.2: Based on the numerical geological parameters, a deep learning network is used to perform multi-level abstraction on rock mass strength parameters, groundwater level data, tectonic stress field distribution, identify potential correlation between parameters, infer rock mass stability level and aquifer dynamic characteristics, and obtain numerical analysis results; In step S1.1.2, based on numerical geological parameters, a deep learning network is used to perform multi-level abstraction on rock mass strength parameters, groundwater level data, tectonic stress field distribution, etc. Deep learning network is a kind of artificial neural network with multiple layers, which can automatically learn the hierarchical representation of data, and gradually abstract high-level features from low-level features. In geological parameter processing, commonly used deep learning architectures include multi-layer perceptron (MLP), convolutional neural network (CNN) and recurrent neural network (RNN) etc. Rock mass strength parameters are key indicators to describe rock mechanical properties, including uniaxial compressive strength, cohesion, internal friction angle, etc., which directly affect the safety and economy of mineral exploitation. Groundwater level data reflect regional hydrogeological conditions, affecting rock mass stability and mining drainage design. Tectonic stress field distribution describes the stress state within the region, which is crucial for predicting geological disasters and evaluating engineering stability. By analyzing the complex relationships between these parameters through deep learning network, potential correlation patterns such as the relationship between rock mass strength and geological structure, the influence of groundwater level change on rock mass stability, etc. are identified, so as to infer high-level geological characteristics such as rock mass stability level, aquifer dynamic characteristics, etc., forming numerical analysis results.

[0043] S1.1.3: organically combine the semantic understanding result with the numerical analysis result to obtain the dual-processing result.

[0044] In step S1.1.3, the semantic understanding result is organically combined with the numerical analysis result to obtain the dual-processing result. The core of this step is to establish the mapping relationship between the text description information and the numerical parameters to form a unified knowledge representation. Multi-modal fusion technology is used to integrate the results of two different types of processing together. Multi-modal fusion refers to the technology of integrating information from different modalities (such as text, numerical values, images, etc.). Common methods include feature-level fusion, decision-level fusion, and model-level fusion. The fusion process first establishes the correspondence between the text description and the numerical parameters, such as associating the description "strong fault activity" in the text with the fault displacement rate parameter in the numerical data. Then, through a weight distribution mechanism, the fusion weights are assigned according to the reliability and importance of different information sources. Finally, a unified feature representation is formed through vector connection or attention weighted summation, etc. This fusion method can fully utilize the semantic richness of text information and the accuracy of numerical data, complement and verify each other, and obtain more comprehensive and accurate dual-processing results, providing a solid foundation for subsequent feature vector construction.

[0045] In this embodiment, the project basic information, planning data and core geological data input by the evaluation personnel are received as the original data input, which usually includes project name, geographic coordinates, construction scale, stratum lithology, structural characteristics, hydrological conditions and other types of information. Based on the combined summary dual-process model, the dual-path information processing mechanism of the human brain is simulated, and the fast processing process and the slow processing process are started simultaneously. The fast processing process is mainly responsible for natural language processing of text-type information such as project name, site description and engineering type. Through the pre-trained geological professional vocabulary library and named entity recognition model, key information such as "limestone", "fault zone" and "aquifer" is automatically extracted, and semantic annotation and classification are performed. For example, when the input text is "a certain iron mine is located in the eastern foot of the Taihang Mountains, and the main ore-bearing layer is the limestone of the Middle Ordovician Fengfeng Group", the fast processing process can identify key information such as ore type (iron mine), geographic location (eastern foot of the Taihang Mountains), geological age (Middle Ordovician) and lithological characteristics (limestone).

[0046] The slow processing process focuses on deep feature mining of numerical geological parameters, including rock mass strength parameters, groundwater level data, and tectonic stress field distribution. This process uses deep learning networks to perform multi-level abstraction on these numerical data and identify potential relationships between parameters. For example, by analyzing the combination patterns of rock uniaxial compressive strength, Poisson's ratio, and elastic modulus, the stability level of the rock mass can be inferred; by correlating groundwater level changes with seasonal precipitation data, the dynamic characteristics of the aquifer can be predicted. The core of the dual-process model is to combine the semantic understanding results of the fast processing with the numerical analysis results of the slow processing to build a unified knowledge representation framework. Through attention mechanisms and cross-validation algorithms, the consistency of text information and numerical data is ensured, such as verifying whether the "hard rock" described in the text matches the high compressive strength shown by the numerical data. Finally, all processing results are integrated into a standardized feature vector containing geological conditions, engineering parameters, environmental factors, regulatory requirements, and other dimensions, each dimension containing corresponding confidence assessment and data quality indicators, providing a reliable data foundation for subsequent technical analysis.

[0047] In a preferred embodiment, in step S2, the spiking temporal learning mechanism of the proxy gradient support spiking neural network is used to extract temporal features and analyze correlations of geological parameter changes, safety factor calculations, and risk assessment results during the analysis process, and to establish a dynamic judgment model of the overlying relationship, including: S2.1: Based on the standardized evaluation data input, perform geological modeling and safety depth analysis, and convert the geological parameter changes, safety factor calculations, and risk assessment results during the geological modeling and safety depth analysis process into time series data streams. The numerical changes are converted into discrete spike sequences by the spiking neural network to obtain time series feature data. Step S2.1 is the time series feature data extraction stage. Based on the standardized evaluation data input, first perform geological modeling and safety depth analysis. Geological modeling is the process of creating a digital representation of the three-dimensional spatial distribution of underground geological bodies. It usually uses three-dimensional modeling software and geostatistical methods to convert discrete exploration data into continuous geological body models. Safety depth analysis is based on the geological model and applies rock mechanics principles to calculate the impact of mineral extraction on surface buildings to determine the safe extraction depth limit. During these analysis processes, a large number of intermediate calculation results will be generated, including geological parameter changes (such as stress distribution, displacement field), safety factor calculations (such as stability coefficients under different working conditions), and risk assessment results (such as instability probability, hazard level).

[0048] Innovatively, these intermediate calculation results are converted into time series data streams for subsequent intelligent analysis. Time series data streams are data sequences arranged in chronological order. In this method, "time" can be actual calculation time or an abstract representation of calculation steps or iteration numbers. For example, in the finite element analysis process, the node displacement, element stress, and other results generated at each calculation step can form a time series data set; in the slope stability analysis, the process of safety factor changes with calculation parameters can also be regarded as time series data. By processing these time series data streams through a spiking neural network, continuous numerical changes are converted into discrete spike sequences.

[0049] A spiking neural network is a special type of artificial neural network that simulates the mechanism of biological neurons transmitting information through firing (generating spikes or pulses). In a spiking neural network, neurons do not produce continuous activation values, but rather discrete spikes when their membrane potential reaches a threshold. The firing time and frequency of spikes encode information, which is closer to the working principle of the biological nervous system and is beneficial for processing time-dependent patterns. Different changes in geological parameters are encoded as different spike patterns, for example, rapid growth in stress concentration areas can lead to high-frequency spike sequences, while parameter changes in stable areas correspond to low-frequency or regular spike sequences. Through this biologically inspired encoding method, key time series features can be extracted from complex geological data changes, resulting in time series feature data with high information compression.

[0050] S2.2: Using the proxy gradient method to construct a smooth proxy function to approximate the gradient of the spike function, updating the network weight of the time series feature data, identifying the characteristic spike time series patterns corresponding to different risk levels, and establishing a dynamic judgment model of the overlying relationship; Step S2.2 is the dynamic judgment model establishment stage of the overlying relationship, which uses the proxy gradient method to construct a smooth proxy function to approximate the gradient of the spike function, updates the network weight of the time series feature data, identifies the characteristic spike time series patterns corresponding to different risk levels, and establishes a dynamic judgment model of the overlying relationship. Although the spiking neural network has advantages in processing time series information, its training faces a key challenge: the spike function is essentially discontinuous and non-differentiable, which makes traditional gradient descent-based optimization algorithms cannot be directly applied. The proxy gradient method is an innovative technology to solve this problem, which constructs a differentiable proxy function to approximate the non-differentiable spike function, thereby realizing the effective training of the network.

[0051] The core idea of the proxy gradient method is to replace the discrete spike generating function with a continuous and differentiable function. During the forward propagation, the original spike function is used, while during the backward propagation to calculate the gradient, the proxy function is used. Common proxy functions include smooth approximations based on sigmoid functions, triangular function approximations, or exponential function approximations, etc. For example, a function like f(x) = α·sigmoid(β·x) can be used to replace the step function, where α and β are parameters that control the shape of the function. This method allows gradient information to be effectively propagated in the network, enabling the spike neural network to be trained through optimization algorithms such as gradient descent.

[0052] In the application of overburden relationship judgment, by learning a large number of historical evaluation cases, the characteristic spike timing patterns corresponding to different risk levels are identified. For example, when a spike pattern related to a high-risk state is detected, the alert level is automatically raised; when a spike sequence similar to a historical safe case is identified, it can be confirmed as a low-risk state. This dynamic judgment model based on pattern recognition can better adapt to complex and variable geological conditions, improving the accuracy and reliability of the evaluation. The model training process uses a supervised learning method, using historical cases labeled with risk levels as training data, and optimizing network parameters by minimizing the difference between predicted risk levels and actual risk levels.

[0053] S2.3: According to the dynamic judgment model of the overburden relationship, analyze short-term fluctuations and long-term trends through multi-scale time windows to obtain the comprehensive technical analysis report.

[0054] Step S2.3 is the comprehensive technical analysis report generation stage. According to the dynamic judgment model of the overburden relationship, short-term fluctuations and long-term trends are analyzed through multi-scale time windows to generate a comprehensive technical analysis report. Multi-scale time window analysis is a signal processing technique that observes signal changes at different time scales, capturing both rapid local features and slow global trends. In overburden evaluation, this analysis method can consider both short-term fluctuations in geological conditions (such as rapid rise in groundwater level caused by heavy rain) and long-term trends (such as slow evolution of regional geological structures).

[0055] By setting different lengths of time windows, sliding window analysis is performed on spike timing data. Short windows (such as data for a few days or weeks) are used to identify sudden risk events, such as abnormal groundwater outflow and sudden deformation of surrounding rock; medium windows (such as data for several months) are used to analyze the impact of seasonal factors, such as the impact of the rainy season on slope stability; long windows (such as annual data) are used to evaluate long-term trends in regional geological environments, such as tectonic activity and erosion rate. By integrating the results of these different time scale analyses, a comprehensive and dynamic overburden relationship judgment can be formed.

[0056] The comprehensive technical analysis report is the final output of this stage, which contains four key parts: the detailed technical argumentation process records the main steps, methods and key parameters of the geological modeling and safe mining depth analysis; the risk level assessment gives a qualitative and quantitative evaluation of the overburden relationship, usually divided into no overburden, mild overburden, moderate overburden and severe overburden levels; the uncertainty quantification analysis assesses the reliability of the evaluation results, identifies the main sources of uncertainty and their impact; the confidence level assessment gives the overall confidence level of the evaluation conclusion based on data quality, model applicability and validation results. This comprehensive report is not only a summary of the technical analysis results, but also an important input for the generation of subsequent reports.

[0057] In this embodiment, standardized evaluation data is input into the geological modeling and safe mining depth analysis engine, which usually uses mature geomechanical calculation methods such as finite element analysis and boundary element method. Based on traditional technology, the invention introduces the spike temporal learning mechanism of the proxy gradient support spiking neural network, considering the entire technical argumentation process as a dynamic temporal information processing system. In specific implementation, the intermediate results generated in each calculation step of the geological modeling process, such as stress distribution calculation, stability coefficient variation, safety factor iteration process, etc., are converted into temporal data streams. For example, when performing slope stability analysis, the adjustment of the sliding surface position, the calculated value of the safety factor, and the satisfaction degree of the convergence criterion in each iteration are recorded as temporal information. The spiking neural network converts these continuous numerical changes into discrete spike sequences by simulating the pulse firing mechanism of biological neurons, where the firing time and frequency of spikes contain important feature information.

[0058] The learning process of the network uses the proxy gradient method to solve the problem of non-differentiability of the spike function, and constructs a smooth proxy function to approximate the gradient of the spike function, so as to realize the effective update of the network weights. In the specific application of overburden relationship judgment, different types of geological risk factors (such as fault activity, goaf stability, surface subsidence possibility, etc.) are encoded as different spike patterns. By learning from a large number of historical evaluation cases, the characteristic spike temporal patterns corresponding to different risk levels are identified. For example, when detecting rapid growth of stress concentration area and sudden drop of safety factor, the network will produce a specific high-frequency spike sequence, warning potential instability risk. The temporal learning mechanism can also capture the lagging influence of geological condition changes on evaluation results, such as the long-term influence of groundwater level changes on rock mass strength, the periodic influence of seasonal factors on slope stability, etc. Through multi-scale time window analysis, a comprehensive judgment model considering short-term fluctuations and long-term trends is established, and finally a comprehensive technical analysis report containing detailed technical argumentation process, risk level assessment, uncertainty quantification analysis and confidence level assessment is output.

[0059] In a preferred embodiment, in step S3, the principle of spiking neurons as predictive controllers of linear systems is applied, combined with built-in evaluation specification library, regulatory policy library and excellent report template for predictive control, automatically matching the optimal report template and clause elements, including: S3.1: Based on the comprehensive technical analysis report, model the report generation process as a multivariable linear control system, where spiking neurons act as intelligent controllers, each corresponding to decision nodes for chapter structure selection, key content highlighting, and professional term usage, resulting in a control system model; Step S3.1 is the control system model establishment stage, based on the comprehensive technical analysis report, model the report generation process as a multivariable linear control system, where spiking neurons act as intelligent controllers, each corresponding to decision nodes for chapter structure selection, key content highlighting, and professional term usage, forming a control system model. The multivariable linear control system is a basic model in control theory, describing the linear relationship between the state, input and output of the control system, commonly represented by state space equations: x(k+1) = Ax(k) + Bu(k), y(k) = Cx(k) + Du(k), where x represents the state of the control system, u represents the control input, y represents the output of the control system, and A, B, C, D are parameter matrices of the control system.

[0060] In the application of report generation, the control system state x can represent the current completion, accuracy, and specification of the report; the control input u represents the content organization decisions that the control system can take, such as adding a certain type of technical explanation, emphasizing a certain risk factor, selecting a specific chapter structure, etc.; and the output y is the final generated report paragraph or chapter. The control system parameter matrices A, B, C, D are learned through analysis of a large number of high-quality report samples, reflecting the influence relationship of different control decisions on report quality.

[0061] As the core component of the intelligent controller, the spiking neuron simulates the membrane potential accumulation and discharge process of biological neurons. Each spiking neuron receives multiple input signals, which may come from technical analysis results, regulatory requirements or previous chapter content. The neuron maintains a membrane potential state inside, when the accumulated input makes the membrane potential exceed the threshold, the neuron generates a spike and triggers the corresponding report generation action. In specific implementation, different types of decision nodes are responsible by different spiking neurons: chapter structure selection neurons determine the overall framework of the report according to project type and evaluation purpose; key content highlighting neurons identify important findings and high-risk factors in technical analysis, ensuring sufficient emphasis in the report; professional term usage neurons select appropriate professional terms and expressions according to audience type and technical complexity. This neural computation-based control model has stronger adaptability and learning ability compared to traditional rule-based template systems.

[0062] S3.2: Adopting the control system model prediction control idea, predicting the report generation effect of the control system model in the future multiple steps at each decision time, producing a spike when the membrane potential reaches the threshold value to trigger the corresponding report generation action, and selecting the control strategy that optimizes the overall quality index; Step S3.2 is the prediction control strategy formulation stage. Adopting the control system model prediction control idea, predicting the report generation effect of the control system model in the future multiple steps at each decision time, producing a spike when the membrane potential reaches the threshold value to trigger the corresponding report generation action, and selecting the control strategy that optimizes the overall quality index. Model predictive control (MPC) is an advanced control algorithm. Its core idea is to predict the behavior in the future for a certain number of time steps based on the current system state and model at each control time, then solve a finite-time optimization problem to determine the optimal control sequence, and only execute the first control action, then repeat the process at the next time step.

[0063] Applying MPC in report generation will predict the impact of different decision choices on future report quality at each decision point (such as starting a new chapter, selecting content detail, etc.). For example, when deciding whether to describe a certain geological structure feature in detail, the impact of the two choices (detailed description or brief mention) on the subsequent risk assessment chapter will be predicted, and then the option that can maximize the overall report quality will be selected. The optimization objective of predictive control is usually a weighted combination of multiple indicators, including content integrity, technical accuracy, expression standardization, and logical coherence.

[0064] Spiking neurons play the role of decision triggers in this process. When analyzing a factor (such as the importance of a specific risk point), relevant information is input to the corresponding spiking neuron, which accumulates its membrane potential. When the accumulated evidence is strong enough (the membrane potential exceeds the threshold value), the neuron produces a spike, triggering the decision to add corresponding content to the report. This mechanism is particularly suitable for situations where multiple pieces of evidence need to be integrated to make a decision, such as determining whether a geological feature needs to be highlighted. By calculating the optimal control sequence through the predictive control algorithm, including chapter arrangement, content detail, chart configuration, and a series of decisions, a complete report generation control strategy is formed.

[0065] S3.3: According to the control strategy, automatically matching the optimal report template and clause element combination from the evaluation specification library, regulations and policies library, and excellent report template to obtain the evaluation report draft.

[0066] Step S3.3 is a report template matching and draft generation stage, according to the formulated control strategy, the optimal report template and clause element combination are automatically matched from the evaluation specification library, regulatory policy library and excellent report template, and the evaluation report draft is generated. The evaluation specification library is a built-in set of technical standards and specification requirements, including national standards, industry standards (such as coal, metal mine industry standards) and local standards (such as local technical specifications of provinces and cities). These specifications define the basic requirements, necessary content and technical indicators of the evaluation report. The regulatory policy library contains relevant legal provisions, policy explanations and implementing rules, such as mineral resources, safety production related laws and their implementing regulations, and policy documents issued by relevant departments.

[0067] The excellent report template library is an important knowledge base, which includes typical evaluation report cases of different types and different regions. These templates are classified according to the type of mineral (such as coal mine, metal mine, non-metallic mine), geological condition (such as simple, medium, complex), evaluation purpose (such as construction project pressure evaluation, mine right setting evaluation) and other dimensions, which can quickly locate the reference template suitable for the current evaluation task. Each template not only contains the overall structure of the report, but also includes the standard expression of typical chapters, the use of professional terms and the configuration scheme of charts.

[0068] According to the control strategy formulated in the foregoing steps, the report template and clause element most suitable for the current evaluation task are automatically matched from these knowledge bases. The matching process uses weighted similarity calculation, considering factors such as project type similarity, geological condition similarity, evaluation purpose consistency, etc. For example, for a coal mine pressure evaluation project in a complex structure area, the coal mine pressure evaluation report template in similar geological conditions will be preferentially selected. At the same time, the relevant regulatory provisions of the current project will be extracted from the regulatory policy library to ensure that the report content meets the latest legal and regulatory requirements.

[0069] Based on the matched template and clause element, the content is organized according to the control strategy, the technical analysis results are filled in, and the complete evaluation report draft is generated. The draft contains standard report structure, such as preface, geological overview, engineering overview, pressure relationship analysis, safe mining condition evaluation, conclusion and suggestion, etc. Each chapter is filled with content based on technical analysis results and relevant specification requirements. The generated report draft not only meets the specification requirements in format, but also fully reflects the core conclusions of technical analysis in content, and reasonably arranges the key chapters and details according to the characteristics of the project.

[0070] In this embodiment, the comprehensive technical analysis report is the core input, and the built-in evaluation specification library, regulations and policies library, and massive excellent report template database are called. The evaluation specification library includes national, industry, and local technical standards and specification requirements, such as mineral resources development and utilization scheme compilation content requirements, metal and non-metal open-pit mine slope engineering technical specifications, etc.; the regulations and policies library covers relevant legal provisions, policy explanations, and implementation details; the excellent report template library collects high-quality evaluation report cases of different types and different regions. Using a spike neuron as a predictive controller of a linear system, the complex report generation task is modeled as a multivariable linear control system. In this model, the input variables include various indicators of technical analysis results (such as safety factor, risk level, geological complexity, etc.), the output variables are the structure arrangement and content organization scheme of the report, and the spike neuron acts as an intelligent controller.

[0071] In specific implementation, each spike neuron corresponds to a decision node in the report generation process, such as chapter structure selection, key content highlighting, professional term use, chart configuration, etc. The membrane potential change of the neuron reflects the importance and urgency of the current decision, and when the membrane potential reaches the threshold, a spike is generated, triggering the corresponding report generation action. The predictive controller uses the idea of model predictive control (MPC) to predict the report generation effect of multiple steps in the future at each decision time, and selects the control strategy that optimizes the overall quality index. For example, when the technical analysis shows that there are high-risk factors, the controller will predict that detailed risk explanation and disposal recommendation chapters need to be added in the report, and the corresponding spike neuron will generate sustained high-frequency pulses to drive the system to select a report framework containing detailed risk analysis modules from the template library. It also has adaptive learning ability, constantly adjusts the weight parameters and threshold settings of the neurons by tracking user modifications to the generated report and satisfaction feedback, and improves the individualization and accuracy of report generation. The final output evaluation report draft not only meets the relevant specification requirements in structure, but also fully reflects the core conclusions of the technical analysis in content, and reasonably arranges the key chapters and details according to the project characteristics.

[0072] In a preferred embodiment, in step S4, a data-driven heuristic evolutionary algorithm based on generalized MILP mixed integer linear programming solution is used to optimize the integrity and accuracy of the report content in the human-computer collaborative editing process, including: S4.1: Convert the quality evaluation of the evaluation report draft into a multi-objective optimization problem, establish a target function including content integrity index, technical accuracy index, expression specification index, and logical coherence index, and a constraint condition including mandatory requirements, format standard restrictions, and page control requirements, forming a multi-objective optimization model; Step S4.1 is the multi-objective optimization model establishment stage, which converts the quality evaluation of the draft evaluation report into a multi-objective optimization problem, establishes the objective function including the content integrity index, the technical accuracy index, the expression standardization index, and the logical coherence index, and the constraint condition including the regulatory mandatory requirement, the format standard limitation, and the length control requirement, to form a multi-objective optimization model. Multi-objective optimization is an important method in operations research and decision science, which deals with optimization problems with multiple conflicting objectives. Unlike single-objective optimization, multi-objective optimization usually does not have a single optimal solution, but a set of Pareto optimal solutions, each of which performs well on some objectives but may be slightly worse on others.

[0073] In the report optimization application, four key target indicators are defined: the content integrity index measures whether the report covers all necessary evaluation elements such as geological condition description, engineering impact analysis, safety measure suggestion, etc.; the technical accuracy index assesses the consistency of the report content with the technical analysis results, ensuring that the data, conclusions and suggestions in the report are fully supported by technical analysis; the expression standardization index checks whether the report meets the official format requirements and the specification of professional terms; the logical coherence index evaluates the logical relationship between the chapters of the report and the completeness of the argument chain. Each indicator is quantified by a specific scoring function, such as content integrity can be calculated by the coverage rate of necessary elements, and technical accuracy can be measured by the consistency with the original technical analysis results.

[0074] The constraint condition limits the range of feasible solutions, including three main constraints: the regulatory mandatory requirement is a hard condition that must be met, such as the evaluation content and conclusion form that must be included; the format standard limitation specifies the form requirements of the report, such as page layout, chart specification, font size, etc.; the length control requirement limits the total length of the report and the relative length proportion of each chapter. These constraint conditions can be expressed in mathematical form, such as inequality constraints (x ≤ b) or equality constraints (Ax = b). In this way, the report optimization problem is formalized into a standard multi-objective optimization problem, providing a clear mathematical model for subsequent heuristic algorithm solution.

[0075] S4.2: Design an intelligent search strategy containing crossover operator, mutation operator, and selection operator, simulate the structural fusion of technical argument logic and risk analysis method between different excellent reports through the crossover operator, introduce innovative chart display method and accurate professional term expression through the mutation operator, maintain the diversity of solutions through the selection operator by adopting the multi-objective Pareto optimal principle, explore the solution space of the multi-objective optimization model, and obtain the optimized report content; Step S4.2 is the intelligent search strategy design and implementation phase. The intelligent search strategy is designed, including crossover operator, mutation operator and selection operator. The structure fusion of technical argumentation logic and risk analysis method between different excellent reports is simulated through the crossover operator. The innovative chart display method and accurate professional term expression are introduced through the mutation operator. The multi-objective Pareto optimal principle is adopted through the selection operator to maintain the diversity of solutions, explore the solution space of the multi-objective optimization model, and obtain the optimized report content. This step adopts the core idea of heuristic evolutionary algorithm, simulates the biological evolution process to search for the optimal solution of complex problems.

[0076] The crossover operator is a key operation in evolutionary algorithms, simulating gene recombination in biological reproduction, generating new offspring solutions by exchanging part of the information of two parent solutions. In report optimization, the crossover operation is embodied in the structure fusion of different report contents. For example, the framework structure of technical argumentation may be adopted from one template, while the methodology of risk analysis is borrowed from another template, creating a new report structure that combines the advantages of both. In specific implementation, single-point crossover (exchanging the subsequent contents of two templates at a certain chapter), multi-point crossover (exchanging at multiple chapters) or uniform crossover (deciding which template's content to use at each chapter according to a certain probability) can be used.

[0077] The mutation operator simulates gene mutation in biological evolution, increasing population diversity and exploration ability by randomly changing some parts of the solution. In report optimization, mutation operation mainly embodies the innovation of content expression. For example, it may be randomly decided to display a certain technical analysis result in chart form, or to replace general expressions with more accurate professional terms. Mutation operation is crucial for avoiding local optimal solution and discovering innovative expression methods. Common mutation types include replacement mutation (replacing existing content with alternative expressions), insertion mutation (adding supplementary notes or charts) and deletion mutation (removing redundant or secondary content).

[0078] The selection operator simulates the natural selection process, determining the survival and reproduction opportunities of individuals according to their fitness (objective function value). In multi-objective optimization, selection is usually based on Pareto dominance or non-dominated sorting. The Pareto optimal principle states that if one solution is superior to another in at least one objective, while not inferior in other objectives, the first solution dominates the second. The non-dominated solution set (i.e. all solutions not dominated by any other solution) constitutes the Pareto front, representing the best trade-off between objectives. Through this selection mechanism, the diversity of the solution set is maintained, while the Pareto front is pushed forward, gradually improving the overall quality of the report. Through multi-generation iterative evolution, combined with user feedback during the interaction process, the complex report optimization solution space can be explored, and the optimized report content that performs well in multiple quality dimensions can be obtained.

[0079] S4.3: Based on the optimized report content, model the user query as a path planning problem in information retrieval through an intelligent question-answering system, considering factors such as data relevance, reliability, and timeliness to find the optimal retrieval strategy, and obtain the complete evaluation report that meets the official format requirements.

[0080] Step S4.3 is the intelligent question-answering and final report generation phase. Based on the optimized report content, the user query is modeled as a path planning problem in information retrieval through an intelligent question-answering system, considering factors such as data relevance, reliability, and timeliness to find the optimal retrieval strategy, and obtain the complete evaluation report that meets the official format requirements. The intelligent question-answering system is an important component of human-computer interaction, enabling evaluation personnel to obtain the required technical support and reference materials during report editing. Unlike traditional keyword search, the intelligent question-answering system can understand natural language questions and provide targeted, structured answers.

[0081] When the user presents a query request (such as "Please provide the historical record of fault activity in this area" or "What is the typical safe drilling depth under similar geological conditions?"), first, the query is semantically understood, and the core entity and query intent are identified. Then, the query is modeled as a path planning problem in information retrieval. Path planning refers to finding the optimal path from the starting point to the ending point in a graph or network, commonly used in navigation and robot motion planning. In information retrieval, the knowledge space can be regarded as a graph, with nodes representing different knowledge points or data sources, and edges representing their association. The query process is to find the optimal path from the query intent to the relevant information in this graph.

[0082] Set data relevance, reliability, and timeliness as evaluation indicators of the path cost function. Relevance measures the matching degree of information and the query topic; reliability assesses the authority of the information source and the accuracy of the data; timeliness considers the update time and applicability of the information. The path cost function integrates these factors to assign a total cost value to each possible retrieval path. A heuristic search algorithm (such as A* algorithm) is used to traverse the search space of the retrieval path, find the path with the minimum total cost, and collect and organize relevant information along this path.

[0083] A* algorithm is a heuristic search algorithm that combines Dijkstra's algorithm and greedy best-first search, using a priority queue to order nodes by estimated total cost (known cost plus heuristic estimate). In information retrieval applications, the heuristic function can be designed based on text similarity, knowledge graph distance, or semantic matching degree of pre-trained models. Organizing the answer content according to the optimal retrieval path not only provides direct answers, but also includes data sources, confidence evaluation, and citation suggestions to help users make professional judgments. After multiple rounds of human-computer interaction and algorithm optimization, a complete evaluation report that meets the official format requirements is finally generated, maintaining professional technical standards while fully reflecting users' personalized needs and professional insights.

[0084] In this embodiment, the draft evaluation report is used as the initial solution of the optimization problem, and is iteratively improved in the process of continuous human-computer collaborative editing. A heuristic evolutionary method based on data algorithm solving generalized MILP is adopted, which first converts the report quality evaluation into a multi-objective optimization problem. The objective function includes multiple dimensions such as content integrity index (covering all necessary evaluation elements), technical accuracy index (consistency with known geological data and calculation results), expression standardization index (compliance with official format and professional terminology requirements), and logical coherence index (logical relationship and argument chain between chapters). The constraint conditions include mandatory requirements (such as mandatory evaluation content and conclusion form), format standard restrictions (such as page layout, chart specifications, font size, etc.), and linear or nonlinear constraints such as page length control requirements (reasonable length allocation of each chapter).

[0085] The core of the heuristic evolutionary algorithm is to design multiple intelligent search strategies to explore the solution space. The crossover operator simulates the fusion of structures between different excellent reports, such as combining the technical argumentation logic of one report with the risk analysis method of another report; the mutation operator introduces innovative expression methods and content organization, such as using new chart display methods or more precise professional terms; the selection operator maintains the diversity of solutions based on the multi-objective Pareto optimal principle. In the human-computer collaborative editing process, when the user modifies a chapter, the optimization objectives and constraints are updated in real time, and the evolutionary search process is restarted. The intelligent question-answering system, as an important part of collaborative editing, uses the same optimization framework to handle user queries. When the user asks "query the fault activity data of the relevant area", the user's query request is modeled as an information retrieval path planning problem, considering factors such as data relevance, reliability, timeliness, etc., and quickly finding the optimal retrieval strategy and answer organization method through heuristic algorithms. Not only does it return accurate data content, but also provides data sources, confidence assessment, and citation suggestions to help users make accurate professional judgments. After multiple rounds of human-computer interaction and algorithm optimization, the final evaluation report not only maintains the level of professional technology, but also fully reflects the user's personalized needs and professional insights, fully meeting the official release format standards and content requirements.

[0086] In a preferred embodiment, the comprehensive technical analysis report includes: detailed technical argumentation process, risk level assessment, uncertainty quantification analysis, and confidence assessment.

[0087] The detailed technical argumentation process is the core part of the evaluation report, recording the complete analysis chain from raw data to final conclusions, ensuring the traceability and reproducibility of the evaluation results. The technical argumentation process usually includes geological condition analysis, engineering impact calculation, safety condition evaluation, etc., and each link records the adopted method, key parameters and intermediate results in detail. For example, in the geological condition analysis, the stratigraphic division basis, lithology identification standard, structure interpretation method, etc. are recorded; in the engineering impact calculation, the calculation model selection reason, boundary condition setting, parameter determination basis, etc. are specified. Such detailed records not only meet the requirements of technical specifications, but also provide complete technical support for subsequent expert review and possible questioning.

[0088] Risk level assessment is the process of classifying and quantifying the potential risks caused by the overlying relationship. A multi-level risk assessment system is usually adopted, taking into account factors such as the degree of overlying, geological condition complexity, and engineering importance. Risk levels are usually classified into four levels: no risk, low risk, medium risk, and high risk. No risk means that there is no resource overlying or overlying relationship that can cause any safety hazards in the evaluation area; low risk means that there is an overlying relationship, but it can be safely handled through conventional technical measures; medium risk means that the overlying relationship has a significant impact on engineering safety and requires special engineering measures; high risk means that the overlying relationship may cause serious safety accidents, requiring a re-planning of engineering layout or mining scheme. Based on the feature pattern identified by the spike neural network, combined with the risk rating records in the historical case library, the risk level of the current project can be accurately determined.

[0089] Uncertainty quantification analysis is an important embodiment of scientific evaluation, aiming to identify and quantify the uncertainty factors in the evaluation results and their impact. In geological evaluation, uncertainty mainly comes from three aspects: data uncertainty, including measurement error, insufficient sampling representativeness, and uneven data coverage; model uncertainty, referring to the simplification and assumptions of geological models or calculation models to actual situations; parameter uncertainty, referring to the variability of model parameter values. Statistical methods such as Monte Carlo simulation and parameter sensitivity analysis are used to quantify these uncertainties. Monte Carlo simulation is a random sampling method that quantifies uncertainty by performing repeated calculations on a large number of random samples within the possible range of parameter values and observing the distribution characteristics of the results. Parameter sensitivity analysis identifies the key parameters that have the greatest impact on the final results by changing the value of a single parameter and observing the changes in the results. These analysis results are presented in the form of probability distribution, confidence interval, or error range, allowing decision-makers to fully understand the reliability range of the evaluation results.

[0090] Confidence assessment is a comprehensive judgment of the credibility of the overall evaluation conclusion, and is an important indicator of evaluation quality. A confidence evaluation system is established based on multiple indicators, including data quality indicators (such as exploration engineering density, test sample quantity, and data coverage completeness), model applicability indicators (such as the degree of agreement between model assumption conditions and actual conditions, and the number of validation cases), and result verification indicators (such as cross-validation results and historical case comparison results). These indicators are comprehensively calculated, usually using weighted average or analytic hierarchy process, to obtain the final confidence level, which is usually classified into high confidence, medium confidence, and low confidence. This confidence assessment provides an important reference for the reliability of the evaluation conclusion for decision-makers, especially in cases where data is incomplete or geological conditions are complex, reasonable confidence assessment can help decision-makers determine whether additional exploration is needed or more conservative safety measures should be taken.

[0091] In a preferred embodiment, the evaluation specification library includes technical standards and specification requirements at national, industry and local levels, the regulation policy library includes relevant legal provisions, policy explanations and implementation rules, and the excellent report template includes typical evaluation report cases of different types and different regions.

[0092] The evaluation specification library is an important part of the knowledge base and includes technical standards and specifications at three levels. National standards mainly include GB (national standard) series documents such as building, water body, railway, and main roadway coal pillar setting and coal mining under pressure specifications, and metal and non-metal mine safety regulations. These standards have the highest effectiveness and specify the basic principles and technical requirements that must be followed in the evaluation work.

[0093] Industry standards are professional technical specifications developed by industry departments, such as coal mine water prevention and control regulations and mine geological environment protection and restoration management plan compilation specifications issued by relevant departments. These industry standards are usually more professional and specific than national standards, and provide detailed technical requirements for specific minerals or specific engineering types. Local standards are developed by provincial or municipal technical supervision departments and are supplementary provisions for special geological conditions or mineral characteristics in the region, such as coal seam gas surface development and utilization engineering design specifications in a certain area and coal and gas outburst risk prediction methods in coal mine areas. These different levels of specification requirements are structured and semantically annotated, and the reference relationship and hierarchical relationship between specification clauses are established to facilitate quick retrieval and application of relevant specification clauses in the report generation process.

[0094] The regulation policy library includes legal regulations, policy documents and explanations related to mineral resources evaluation. Laws and regulations establish the basic principles and institutional framework for the management and development of mineral resources. Policy explanations are specific explanation documents issued by relevant departments on legal provisions. These explanation documents provide more specific guidance for the implementation of laws and regulations.

[0095] Implementation rules are specific implementation plans developed by administrative departments at all levels to implement laws and regulations. These regulations and policy documents are content analyzed and classified, a professional term dictionary and law clause reference network are established, and the report generation process can accurately reference relevant legal provisions to ensure the legality and compliance of the evaluation report. The regulation policy library also has an update mechanism to regularly check and update the content of regulations to ensure that the regulations used are always the latest and effective versions.

[0096] The excellent report template library collects typical evaluation report cases of different types and different regions, which is an important reference resource for report generation. The report templates are classified according to multiple dimensions: according to the type of mineral resources, they are classified into coal mine pressure evaluation, metal mine pressure evaluation, and non-metallic mine pressure evaluation; according to the complexity of geological conditions, they are classified into simple geological conditions, medium complex geological conditions and complex geological conditions; according to the purpose of evaluation, they are classified into construction project pressure evaluation, mine right setting evaluation, and resource reserve verification evaluation; according to the regional characteristics, they are classified into eastern plain area, central hilly area, western mountainous area and special geological area (such as karst area, permafrost area, etc.). Each template not only contains the overall structural framework of the report, but also contains detailed information such as chapter title setting, paragraph organization method, professional term use specification, data table format and map configuration scheme.

[0097] These templates are subjected to in-depth semantic analysis, and the structural features, argumentation logic and expression style of the report are extracted, and a template feature vector is established. In the report generation process, according to the characteristics of the current evaluation task, the similarity with each template is calculated, the most matched template is selected as the basis, and then customized adjustment is made according to the specific needs. The excellent report template library not only improves the efficiency of report generation, but also ensures the consistency and professionalism of report quality. The template library is continuously expanded and optimized through a continuous learning mechanism, and when a new report generated is recognized by experts, it can be added to the template library as a new excellent case, forming a virtuous cycle and continuously improving the report generation capability.

[0098] In a preferred embodiment, the intelligent question and answer system realizes the optimal path planning of data retrieval and verification, comprising: Step S5.1: Based on the user query request, model the user query request as a path planning problem of information retrieval, set data relevance, reliability and timeliness as evaluation indexes of path cost function, and construct search space of retrieval path; Query modeling is the process of converting user's natural language problem into formalized representation that can be processed. First, natural language processing is performed on user query, including word segmentation, part-of-speech tagging, named entity recognition and semantic dependency analysis, to identify core concepts (such as "fault activity") and query intentions (such as "find historical records" and "ask typical values") in the query.

[0099] Path planning problem is to find the optimal path from the starting point to the end point in a graph or network, which is a classic problem in artificial intelligence and operations research. In information retrieval, the idea of path planning is applied to construct the knowledge space as a directed graph, where nodes represent different knowledge points or data sources (such as technical reports, regulatory provisions, historical cases, etc.), and edges represent their association relationships (such as citation relationships, expansion relationships, similarity relationships, etc.). The query process is then transformed into finding the optimal path from the query intent to the relevant information in this graph. The path cost function is a mathematical expression that evaluates the quality of the path, taking into account multiple indicators: data relevance measures the matching degree of information and query topic, usually calculated using vector space model or semantic similarity algorithm; data reliability assesses the authority of information source and the accuracy of data, considering factors including data source (such as official reports, academic papers or informal records), data generation method (such as measured data, estimated data or simulated data) and data verification degree; data timeliness considers the update time and applicability of information, new data is usually more reliable than old data, but in some historical analysis, early records may be more valuable.

[0100] These indicators are combined into a weighted total cost function: Cost = w1·(1-Relevance) + w2·(1-Reliability) + w3·(1-Timeliness), where w1, w2, w3 are weight coefficients, adjusted according to query type and user preference, Relevance represents relevance, used to represent the semantic matching degree between candidate information and user query request, its value range is [0, 1], the larger the value, the higher the relevance; Reliability represents reliability, used to represent the comprehensive level of candidate information in authenticity, authority and credibility, its value range is [0, 1], the larger the value, the higher the reliability; Timeliness represents timeliness, that is, to represent the applicability and effectiveness of candidate information in time dimension, its value range is [0, 1], the larger the value, the higher the timeliness. Based on this cost function, the search space of retrieval path is constructed, which is a state space composed of all possible retrieval paths, each path represents a way of information acquisition and organization from initial query to final answer. The size of search space is proportional to the scale and complexity of knowledge base, in order to improve retrieval efficiency, heuristic method is usually used to prune search space, only keep the paths that are most likely to produce high quality answers.

[0101] Step S5.2: Adopt heuristic search algorithm to traverse the search space of retrieval path, find the optimal retrieval strategy that minimizes the total cost by evaluating the cost function value of different retrieval paths, and organize the corresponding answer content; Heuristic search algorithms are a class of algorithms that incorporate problem-specific knowledge to improve search efficiency, finding approximate optimal solutions in large-scale search spaces. A* algorithm is primarily used for path search, which is a combination of Dijkstra's algorithm (guaranteed to find the shortest path) and greedy best-first search (utilizes heuristic functions to improve efficiency).

[0102] A* algorithm uses a priority queue to sort nodes by estimated total cost f(n) = g(n) + h(n), where g(n) is the known cost from the starting point to the current node n, and h(n) is the estimated cost from node n to the target (heuristic function). If h(n) is acceptable (not higher than the actual cost), A* algorithm guarantees to find the optimal solution. In information retrieval applications, heuristic function h(n) can be designed based on text similarity, knowledge graph distance, or semantic matching degree of pre-trained language models. For example, the cosine distance between the content vector of a node and the query vector can be used as a heuristic estimate, and nodes with more relevant content will have lower heuristic values and be explored first.

[0103] The algorithm starts from the starting node representing the query intent, gradually explores connected knowledge nodes, calculates the cumulative cost, and updates the priority queue. During the exploration process, different types of retrieval operations are considered, such as keyword search, concept expansion, case matching, and temporal analysis, each corresponding to different types of edges in the graph and having different cost calculation methods. For example, direct keyword matching usually has lower cost than complex semantic reasoning, but the latter may provide more relevant information. In this way, the retrieval path with the minimum total cost can be found, balancing the efficiency, relevance, and reliability of retrieval.

[0104] After finding the optimal retrieval strategy, relevant information is collected and organized along the path to form structured answer content. The answer organization process considers the logical structure and importance of information, transforming the collected raw data into coherent and clear answers. Template filling and natural language generation techniques are used to select appropriate expression forms based on query types and answer content. For example, for numerical queries, a table may be used to display data; for explanatory queries, structured text paragraphs are generated; for comparison queries, comparison lists or charts may be used. This diverse answer organization method ensures clear presentation and easy understanding of information.

[0105] Step S5.3: Based on the optimal retrieval strategy and the answer content, provide data sources, confidence evaluation, and citation suggestions, and through multiple rounds of human-computer interaction and algorithm optimization, obtain the complete evaluation report that meets the official format requirements.

[0106] In this stage, not only direct answers are provided, but also rich meta-information is attached to help users evaluate the reliability and applicability of the information. Data source explanations include the original origin of the information (such as specific technical reports, regulatory documents, or academic papers), the publishing agency, and the publishing time, allowing users to trace the original source of the information and judge its authority.

[0107] Confidence assessment is a self-evaluation of the reliability of the provided information, usually expressed in confidence levels (such as high, medium, low) or confidence scores (such as 0-100 points). This assessment is based on multiple indicators, including data source reliability, data timeliness, query and answer matching degree, information consistency test results, etc. For example, the latest data from official technical specifications usually has high confidence, while data from informal channels or from long ago has lower confidence. Consistency between different sources of information is also detected, and if multiple independent sources provide similar information, the confidence will increase; if obvious contradictions are found, the confidence will decrease.

[0108] Citation suggestions are provided for key information in the form of citation format and content, making it easy for users to correctly cite relevant materials in reports. Multiple citation formats are supported. Citation content includes basic information of the literature (such as author, title, publication information, page number, etc.) and citation context suggestions (such as which parts of the report, how to cite this material). These citation suggestions not only improve the academic standardization of the report, but also enhance the credibility and persuasiveness of the evaluation conclusions.

[0109] Multiple rounds of human-computer interaction are the key mechanism for continuously improving the quality of answers. In each round of interaction, users can provide feedback on the provided answers, such as requesting more detailed explanations, questioning the accuracy of certain data, or requiring additional supporting evidence, etc. Based on these feedback, the retrieval strategy is adjusted, relevant information is further excavated, or inaccurate content is corrected. For example, if a user has doubts about the numerical value of a technical parameter, a more in-depth search will be automatically initiated to find more supporting evidence or alternative data sources, and the reasons for the differences between different data sources will be explained. This interactive optimization process not only improves the quality of answers, but also helps users better understand the relevant issues.

[0110] Algorithm optimization is the process of continuously improving retrieval and answering strategies in multiple rounds of interaction. Reinforcement learning methods are used, with user feedback as a reward signal, to optimize the selection strategy of retrieval paths and the method of answer organization. For example, if users frequently request a specific type of supplementary information, the retrieval mode will be automatically adjusted to include this type of information in the initial answer; if a certain answer organization method receives positive feedback, the probability of using this method will be increased. Through this continuous learning mechanism, it can adapt to the preferences and needs of different users and provide more appropriate technical support.

[0111] After multiple rounds of interaction and optimization, a complete evaluation report that meets the official format requirements is finally generated. This report not only accurately reflects the technical analysis results in terms of content, but also strictly follows industry standards and official standards in terms of form. The report format complies with the requirements of official documents such as the Mineral Resources Development and Utilization Plan, including standard chapter structure, standard text expression, professional chart configuration, and complete annex materials. A report quality self-checking function is also provided to check whether the report meets all mandatory requirements, such as whether the necessary chapters are complete, whether the key conclusions are clear, whether the data references are standardized, etc., to ensure the completeness and standardization of the final report.

[0112] As shown in Figure 2 The present application also provides a mineral resource overburden evaluation report intelligent collaborative generation system, comprising: A data input module 601 is used to obtain project basic information, planning data and core geological data input by evaluation personnel as raw data input, intelligently analyze and classify the raw data input based on a combined summary double-process model, construct a structured multi-dimensional data feature vector, and obtain standardized evaluation data input; A technical analysis module 602 is used to perform geological modeling and safe depth analysis on the standardized evaluation data input, perform time series feature extraction and correlation analysis on the geological parameter changes, safety factor calculation and risk assessment results during the analysis process through the spike timing learning mechanism of the proxy gradient support spike neural network, establish a dynamic judgment model of the overburden relationship, and generate a comprehensive technical analysis report; A report generation module 603 is used to generate a structured evaluation report draft based on the comprehensive technical analysis report, use spike neurons as predictive controllers of linear systems, combine built-in evaluation specification library, regulations and policies library and excellent report templates for predictive control, automatically match the optimal report template and clause elements, and output the structured evaluation report draft; A collaborative editing module 604 is used to take the evaluation report draft as the initial solution, adopt a data-driven heuristic evolutionary algorithm based on generalized MILP mixed integer linear programming solution, optimize the completeness and accuracy of the report content during human-computer collaborative editing, realize optimal path planning of material retrieval and verification through an intelligent question and answer system, and generate a complete evaluation report that meets the official format requirements.

[0113] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can make equivalent substitutions or changes to the technical solution and inventive concept of the present application within the scope of the disclosed technology, which should be covered within the protection scope of the present application.

Claims

1. A method for intelligent collaborative generation of mineral resource overlay assessment reports, characterized in that, Includes the following steps: The project's basic information, planning data, and core geological data input by the evaluators are obtained as raw data input. The raw data input is intelligently analyzed and classified based on a combined generalization dual-process model to construct a structured multi-dimensional data feature vector and obtain standardized evaluation data input. Geological modeling and safe mining depth analysis are performed on the standardized assessment data input. The temporal learning mechanism of the surrogate gradient support spike neural network is used to extract temporal features and perform correlation analysis on the geological parameter changes, safety factor calculation and risk assessment results during the analysis process. A dynamic judgment model of overburden relationship is established and a comprehensive technical analysis report is generated. Based on the comprehensive technical analysis report, the principle of peak neurons as predictive controllers for linear systems is applied. Combined with the built-in evaluation standard library, regulatory policy library and excellent report templates, predictive control is performed to automatically match the optimal report template and clause elements and output a structurally complete draft evaluation report. Using the draft evaluation report as the initial solution, a data-driven heuristic evolutionary algorithm based on generalized MILP mixed integer linear programming is employed to optimize the completeness and accuracy of the report content during the human-computer collaborative editing process. The optimal path planning for data retrieval and verification is achieved through an intelligent question-answering system, resulting in a complete evaluation report that meets the official format requirements.

2. The method according to claim 1, characterized in that, The combined generalization dual-process model intelligently analyzes and classifies the project's basic information, planning data, and core geological data, constructing a structured, multi-dimensional data feature vector, including: Based on the original data input, a fast processing procedure is initiated to perform natural language processing and keyword extraction on the text-based project information, and a slow processing procedure is initiated to perform deep feature mining and semantic understanding on the numerical geological parameters, resulting in dual processing results; The results of the dual processing are combined and summarized. The consistency between textual information and numerical data is ensured through attention mechanism and cross-validation algorithm. A multi-dimensional feature vector containing geological conditions, engineering parameters, environmental factors and regulatory requirements is constructed. Based on the multi-dimensional feature vector, a quality assessment is performed through a data validation and completion mechanism to obtain the standardized assessment data input.

3. The method according to claim 2, characterized in that, The fast processing step involves natural language processing and keyword extraction for textual project information, while the slow processing step involves deep feature mining and semantic understanding for numerical geological parameters, resulting in dual processing outcomes, including: Based on the textual project information, natural language processing is performed using a pre-trained geological lexicon and named entity recognition model to automatically extract key information such as mineral type, geographical location, geological age, and lithological characteristics, and then perform semantic annotation and classification to obtain semantic understanding results. Based on the numerical geological parameters, a deep learning network is used to perform multi-level abstraction of rock mass strength parameters, groundwater level data, and tectonic stress field distribution, identify potential correlations between parameters, infer the rock mass stability level and aquifer dynamic characteristics, and obtain numerical analysis results. The semantic understanding results are organically combined with the numerical analysis results to obtain the dual-processing results.

4. The method according to claim 1, characterized in that, The peak time-series learning mechanism, which uses a surrogate gradient-supported peak neural network, extracts time-series features and performs correlation analysis on geological parameter changes, safety factor calculations, and risk assessment results during the analysis process, establishing a dynamic judgment model for overburden relationships, including: Based on the standardized assessment data input, geological modeling and safe mining depth analysis are performed. The geological parameter changes, safety factor calculations, and risk assessment results during the geological modeling and safe mining depth analysis process are converted into time-series data streams. The numerical changes are converted into discrete peak sequences through a spike neural network to obtain time-series feature data. The gradient of the peak function is approximated by a smooth surrogate function using the surrogate gradient method. The network weights are updated on the time series feature data to identify characteristic peak time series patterns corresponding to different risk levels and to establish a dynamic judgment model for the overlay relationship. Based on the dynamic judgment model of the overburden relationship, the comprehensive technical analysis report is obtained by analyzing short-term fluctuations and long-term trends through multi-scale time windows.

5. The method according to claim 1, characterized in that, The principle of using spike neurons as a predictive controller for linear systems, combined with a built-in evaluation standard library, regulatory policy library, and excellent report templates, enables predictive control and automatically matches the optimal report template and clause elements, including: Based on the comprehensive technical analysis report, the report generation process is modeled as a multivariable linear control system, in which the spike neurons act as intelligent controllers, and each spike neuron corresponds to the decision nodes of chapter structure selection, key content highlighting, and professional terminology usage, thus obtaining the control system model; Using the predictive control concept of the control system model, the system model predicts the report generation effect of multiple future steps at each decision time. When the membrane potential reaches the threshold, a spike is generated to trigger the corresponding report generation action, and the control strategy that optimizes the overall quality index is selected. Based on the control strategy, the optimal report template and clause element combination is automatically matched from the assessment standard library, the regulatory policy library, and the excellent report templates to obtain the draft assessment report.

6. The method according to claim 1, characterized in that, The aforementioned data-driven heuristic evolutionary algorithm, based on generalized MILP mixed-integer linear programming, optimizes the completeness and accuracy of report content during human-computer collaborative editing, including: The quality assessment of the draft assessment report is transformed into a multi-objective optimization problem. An objective function is established, including indicators of content completeness, technical accuracy, expression standardization, and logical coherence, and constraints including mandatory legal requirements, format standard restrictions, and length control requirements, thus forming a multi-objective optimization model. The design incorporates an intelligent search strategy that includes crossover, mutation, and selection operators. The crossover operator simulates the structural integration of technical argumentation logic and risk analysis methods among different excellent reports. The mutation operator introduces innovative chart display methods and precise professional terminology. The selection operator adopts the multi-objective Pareto optimality principle to maintain the diversity of solutions, explores the solution space of the multi-objective optimization model, and obtains optimized report content. Based on the optimized report content, the user query is modeled as an information retrieval path planning problem through an intelligent question-answering system. The optimal retrieval strategy is found by comprehensively considering the relevance, reliability, and timeliness of the data, resulting in a complete evaluation report that meets the official format requirements.

7. The method according to claim 1, characterized in that, The comprehensive technical analysis report includes: a detailed technical demonstration process, risk level assessment, uncertainty quantification analysis, and confidence level assessment.

8. The method according to claim 1, characterized in that, The assessment standards library includes technical standards and specifications at the national, industry, and local levels; the regulations and policies library includes relevant legal provisions, policy interpretations, and implementation details; and the excellent report templates include typical assessment report cases of different types and regions.

9. The method according to claim 1, characterized in that, The intelligent question-answering system implements optimal path planning for information retrieval and verification, including: Based on user query requests, the user query requests are modeled as a path planning problem for information retrieval. Data relevance, reliability, and timeliness are set as evaluation indicators for the path cost function, and a search space for the retrieval path is constructed. A heuristic search algorithm is used to traverse the search space of the retrieval path. By evaluating the cost function values ​​of different retrieval paths, the optimal retrieval strategy that minimizes the total cost is found, and the corresponding answer content is organized. Based on the optimal retrieval strategy and the answer content, data sources, confidence assessments, and citation suggestions are provided. After multiple rounds of human-computer interaction and algorithm optimization, a complete evaluation report conforming to the official format requirements is obtained.

10. A smart collaborative generation system for mineral resource overlay assessment reports, characterized in that, include: The data input module is used to acquire basic project information, planning data, and core geological data input by the evaluators as raw data input. Based on the combined generalization dual-process model, the raw data input is intelligently analyzed and classified to construct a structured multi-dimensional data feature vector and obtain standardized evaluation data input. The technical analysis module is used to perform geological modeling and safe mining depth analysis on the standardized assessment data input. Through the peak time-series learning mechanism of the proxy gradient support peak neural network, it extracts time-series features and performs correlation analysis on the changes in geological parameters, safety factor calculation, and risk assessment results during the analysis process, establishes a dynamic judgment model of overburden relationship, and generates a comprehensive technical analysis report. The report generation module is used to perform predictive control based on the comprehensive technical analysis report, using the principle of peak neurons as predictive controllers for linear systems, combined with the built-in evaluation standard library, regulatory policy library and excellent report templates, to automatically match the optimal report template and clause elements, and output a structurally complete draft evaluation report. The collaborative editing module uses the draft evaluation report as an initial solution and employs a data-driven heuristic evolutionary algorithm based on generalized MILP mixed integer linear programming to optimize the completeness and accuracy of the report content during the human-computer collaborative editing process. It also uses an intelligent question-and-answer system to achieve optimal path planning for data retrieval and verification, thereby generating a complete evaluation report that meets the official format requirements.

Images